statistique, théorie et gestion de portefeuille - Docs at ISFA
statistique, théorie et gestion de portefeuille - Docs at ISFA
statistique, théorie et gestion de portefeuille - Docs at ISFA
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
UNIVERSITE DE NICE-SOPHIA ANTIPOLIS - UFR SCIENCES<br />
Ecole Doctorale <strong>de</strong> Sciences Fondamentales <strong>et</strong> Appliquées<br />
THESE<br />
Présentée pour obtenir le titre <strong>de</strong><br />
Docteur en SCIENCES<br />
<strong>de</strong> l’Université <strong>de</strong> Nice - Sophia Antipolis<br />
Spécialité : Physique<br />
par<br />
Yannick MALEVERGNE<br />
Risques extrêmes en finance :<br />
<strong>st<strong>at</strong>istique</strong>, <strong>théorie</strong> <strong>et</strong> <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong><br />
Soutenue publiquement le 20 décembre 2002 <strong>de</strong>vant le jury composé <strong>de</strong> :<br />
Jean-Clau<strong>de</strong> AUGROS Professeur, Université Lyon I (Prési<strong>de</strong>nt)<br />
Jean-Paul LAURENT Professeur, Université Lyon I (Co-directeur <strong>de</strong> Thèse)<br />
Jean-François MUZY Chargé <strong>de</strong> recherche au CNRS<br />
Michael ROCKINGER Professeur, HEC Lausanne<br />
Bertrand ROEHNER Professeur, Université Paris VII (Rapporteur)<br />
Didier SORNETTE Directeur <strong>de</strong> recherche au CNRS (Directeur <strong>de</strong> Thèse)<br />
à l’Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Clau<strong>de</strong> Bernard - Lyon I
Table <strong>de</strong>s m<strong>at</strong>ières<br />
Avant-propos 11<br />
Introduction 13<br />
I Etu<strong>de</strong> <strong>et</strong> modélis<strong>at</strong>ion <strong>de</strong>s propriétés <strong>de</strong>s rentabilités <strong>de</strong>s actifs financiers 19<br />
1 Faits stylisés <strong>de</strong>s rentabilités boursières 21<br />
1.1 Rappel <strong>de</strong>s faits stylisés . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
1.1.1 La distribution <strong>de</strong>s ren<strong>de</strong>ments . . . . . . . . . . . . . . . . . . . . . . . . . . . 22<br />
1.1.2 Propriétés <strong>de</strong> dépendances temporelles . . . . . . . . . . . . . . . . . . . . . . 23<br />
1.1.3 Autres faits stylisés . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25<br />
1.2 De la difficulté <strong>de</strong> représenter la distribution <strong>de</strong>s ren<strong>de</strong>ments . . . . . . . . . . . . . . . 27<br />
1.3 Modélis<strong>at</strong>ion <strong>de</strong>s propriétés <strong>de</strong> dépendance <strong>de</strong>s ren<strong>de</strong>ments . . . . . . . . . . . . . . . . 28<br />
1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30<br />
2 Modèles phénoménologiques <strong>de</strong> cours 31<br />
2.1 Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses . . . . . . . . . . . . . . . 32<br />
2.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33<br />
2.1.2 Hypothèses du modèle <strong>de</strong> Blanchard <strong>et</strong> W<strong>at</strong>son . . . . . . . . . . . . . . . . . . 34<br />
2.1.3 Généralis<strong>at</strong>ion <strong>de</strong>s bulles r<strong>at</strong>ionnelles en dimensions quelconques . . . . . . . . 36<br />
2.1.4 Théorie du renouvellement <strong>de</strong>s produits <strong>de</strong> m<strong>at</strong>rices alé<strong>at</strong>oires . . . . . . . . . . 37<br />
2.1.5 Conséquences pour les bulles r<strong>at</strong>ionnelles . . . . . . . . . . . . . . . . . . . . . 39<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41<br />
2.2 Des bulles r<strong>at</strong>ionnelles aux krachs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42<br />
2.2.1 Les modèles <strong>de</strong> bulles r<strong>at</strong>ionnelles . . . . . . . . . . . . . . . . . . . . . . . . . 43<br />
3
4 Table <strong>de</strong>s m<strong>at</strong>ières<br />
2.2.2 Bulles r<strong>at</strong>ionnelles pour un actif isolé . . . . . . . . . . . . . . . . . . . . . . . 46<br />
2.2.3 Généralis<strong>at</strong>ion <strong>de</strong>s bulles r<strong>at</strong>ionnelles à un nombre arbitraire <strong>de</strong> dimensions . . . 49<br />
2.2.4 Modèle <strong>de</strong> krach avec taux <strong>de</strong> hasard . . . . . . . . . . . . . . . . . . . . . . . 54<br />
2.2.5 Modèle <strong>de</strong> croissance non st<strong>at</strong>ionnaire . . . . . . . . . . . . . . . . . . . . . . 56<br />
2.2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62<br />
3 Distributions exponentielles étirées contre distributions régulièrement variables 63<br />
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66<br />
3.2 St<strong>at</strong>istiques <strong>de</strong>scriptives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69<br />
3.2.1 Les données . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69<br />
3.2.2 Existence <strong>de</strong> dépendance temporelle . . . . . . . . . . . . . . . . . . . . . . . 70<br />
3.3 Propriétés extrêmes d’un processus à mémoire longue . . . . . . . . . . . . . . . . . . 70<br />
3.3.1 Résult<strong>at</strong>s théoriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71<br />
3.3.2 Exemples <strong>de</strong> convergence lente vers les distributions <strong>de</strong>s valeurs extrêmes <strong>et</strong> <strong>de</strong><br />
Par<strong>et</strong>o généralisées . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72<br />
3.3.3 Génér<strong>at</strong>ion d’un processus à mémoire longue avec <strong>de</strong>s marginales prescrites . . 73<br />
3.3.4 Simul<strong>at</strong>ions numériques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75<br />
3.3.5 Estim<strong>at</strong>eurs <strong>de</strong>s paramètres <strong>de</strong>s GEV <strong>et</strong> GPD pour les données réelles . . . . . . 77<br />
3.4 Estim<strong>at</strong>ion paramétrique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78<br />
3.4.1 Définition d’une famille <strong>de</strong> distributions à trois paramètres . . . . . . . . . . . . 78<br />
3.4.2 Méthodologie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80<br />
3.4.3 Résult<strong>at</strong>s empiriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81<br />
3.4.4 Résumé . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84<br />
3.5 Comparaison du pouvoir <strong>de</strong>scriptif <strong>de</strong>s différentes familles . . . . . . . . . . . . . . . . 84<br />
3.5.1 Comparaison entre les qu<strong>at</strong>re distributions paramétriques <strong>et</strong> la distribution globale 85<br />
3.5.2 Comparaison directe <strong>de</strong>s modèles <strong>de</strong> Par<strong>et</strong>o <strong>et</strong> Exponentiel-Etiré . . . . . . . . 86<br />
3.6 Discussion <strong>et</strong> conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103<br />
4 Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité 135<br />
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Table <strong>de</strong>s m<strong>at</strong>ières 5<br />
4.2 Mémoire longue <strong>et</strong> distinction entre chocs endogènes <strong>et</strong> exogènes . . . . . . . . . . . . 139<br />
4.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146<br />
5 Approche comportementale <strong>de</strong>s marchés financiers : l’apport <strong>de</strong>s modèles d’agents 149<br />
5.1 Prix d’un actif <strong>et</strong> excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150<br />
5.2 Modèles d’opinion contre modèles <strong>de</strong> marché . . . . . . . . . . . . . . . . . . . . . . . 152<br />
5.3 Modèles à agents adapt<strong>at</strong>ifs contre modèles à agents non adapt<strong>at</strong>ifs . . . . . . . . . . . . 154<br />
5.4 Conséquences <strong>de</strong>s phénomènes d’imit<strong>at</strong>ion <strong>et</strong> d’antagonisme... . . . . . . . . . . . . . . 155<br />
5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157<br />
6 Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos 159<br />
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162<br />
6.2 Le modèle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163<br />
6.3 Analyse qualit<strong>at</strong>ive <strong>de</strong>s propriétés dynamiques . . . . . . . . . . . . . . . . . . . . . . 165<br />
6.4 Analyse quantit<strong>at</strong>ive <strong>de</strong>s bulles spécul<strong>at</strong>ives en régime chaotique dans le cas symétrique 167<br />
6.5 Propriétés <strong>st<strong>at</strong>istique</strong>s <strong>de</strong>s ren<strong>de</strong>ments dans le cas symétrique . . . . . . . . . . . . . . 169<br />
6.6 Cas asymétriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170<br />
6.7 Eff<strong>et</strong>s <strong>de</strong> taille finie . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171<br />
6.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174<br />
Appendice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176<br />
II Etu<strong>de</strong> <strong>de</strong>s propriétés <strong>de</strong> dépendances entre actifs financiers 179<br />
7 Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s copules 181<br />
7.1 Les copules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182<br />
7.2 Quelques familles <strong>de</strong> copules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183<br />
7.2.1 Les copules gaussiennes <strong>et</strong> copules <strong>de</strong> Stu<strong>de</strong>nt . . . . . . . . . . . . . . . . . . 184<br />
7.2.2 Les copules archimédiennes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185<br />
7.3 Tests empiriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186<br />
7.3.1 Tests paramétriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
6 Table <strong>de</strong>s m<strong>at</strong>ières<br />
7.3.2 Tests non paramétriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188<br />
7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189<br />
8 Tests <strong>de</strong> copule gaussienne 191<br />
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194<br />
8.2 Généralités sur les copules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195<br />
8.2.1 Définitions <strong>et</strong> résult<strong>at</strong>s importants concernant les copules . . . . . . . . . . . . 195<br />
8.2.2 Dépendance entre <strong>de</strong>ux variables alé<strong>at</strong>oires . . . . . . . . . . . . . . . . . . . . 196<br />
8.2.3 La copule gaussienne . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198<br />
8.2.1 La copule <strong>de</strong> Stu<strong>de</strong>nt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199<br />
8.3 Tester l’hypothèse <strong>de</strong> copule gaussienne . . . . . . . . . . . . . . . . . . . . . . . . . . 200<br />
8.3.1 <strong>st<strong>at</strong>istique</strong> <strong>de</strong> test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200<br />
8.3.2 Procédure <strong>de</strong> test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201<br />
8.3.3 Sensibilité <strong>de</strong> la métho<strong>de</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203<br />
8.4 Résult<strong>at</strong>s empiriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206<br />
8.4.1 Les <strong>de</strong>vises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206<br />
8.4.2 Les m<strong>at</strong>ières premières : les métaux . . . . . . . . . . . . . . . . . . . . . . . . 209<br />
8.4.3 Les actions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209<br />
8.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212<br />
9 Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers 237<br />
9.1 Les différentes mesures <strong>de</strong> dépendances extrêmes . . . . . . . . . . . . . . . . . . . . . 238<br />
9.1.1 Coefficients <strong>de</strong> corrél<strong>at</strong>ion conditionnels . . . . . . . . . . . . . . . . . . . . . 243<br />
9.1.2 Mesures <strong>de</strong> dépendances conditionnelles . . . . . . . . . . . . . . . . . . . . . 249<br />
9.1.3 Coefficient <strong>de</strong> dépendance <strong>de</strong> queue . . . . . . . . . . . . . . . . . . . . . . . . 252<br />
9.1.4 Synthèse <strong>et</strong> discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279<br />
9.2 Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue . . . . . . . . . . . . . . . . . . . . . 299<br />
9.2.1 Mesure intrinsèque <strong>de</strong> dépendance extrême . . . . . . . . . . . . . . . . . . . . 303<br />
9.2.2 Coefficient <strong>de</strong> dépendance <strong>de</strong> queue pour un modèle à facteurs . . . . . . . . . 305<br />
9.2.3 Etu<strong>de</strong> empirique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Table <strong>de</strong>s m<strong>at</strong>ières 7<br />
9.2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326<br />
9.3 Synthèse <strong>de</strong> la <strong>de</strong>scription <strong>de</strong> la dépendance entre actifs financiers . . . . . . . . . . . . 340<br />
III Mesures <strong>de</strong>s risques extrêmes <strong>et</strong> applic<strong>at</strong>ion à la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> 343<br />
10 La mesure du risque 345<br />
10.1 La <strong>théorie</strong> <strong>de</strong> l’utilité . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346<br />
10.1.1 Théorie <strong>de</strong> l’utilité en environnement certain . . . . . . . . . . . . . . . . . . . 346<br />
10.1.2 Théorie <strong>de</strong> la décision face au risque . . . . . . . . . . . . . . . . . . . . . . . . 347<br />
10.1.3 Théorie <strong>de</strong> la décision face à l’incertain . . . . . . . . . . . . . . . . . . . . . . 350<br />
10.2 Les mesures <strong>de</strong> risque cohérentes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353<br />
10.2.1 Définition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353<br />
10.2.2 Quelques exemples <strong>de</strong> mesures <strong>de</strong> risque cohérentes . . . . . . . . . . . . . . . 354<br />
10.2.3 Représent<strong>at</strong>ion <strong>de</strong>s mesures <strong>de</strong> risque cohérentes . . . . . . . . . . . . . . . . . 355<br />
10.2.4 Conséquences sur l’alloc<strong>at</strong>ion <strong>de</strong> capital . . . . . . . . . . . . . . . . . . . . . . 356<br />
10.2.5 Critique <strong>de</strong>s mesures cohérentes <strong>de</strong> risque . . . . . . . . . . . . . . . . . . . . . 357<br />
10.3 Les mesures <strong>de</strong> fluctu<strong>at</strong>ions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359<br />
10.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361<br />
11 Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong> marché 363<br />
11.1 Les limites <strong>de</strong> l’approche moyenne - variance . . . . . . . . . . . . . . . . . . . . . . . 364<br />
11.2 Prise en compte <strong>de</strong>s grands risques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365<br />
11.2.1 Optimis<strong>at</strong>ion sous contrainte <strong>de</strong> capital économique . . . . . . . . . . . . . . . . 365<br />
11.2.2 Optimis<strong>at</strong>ion sous contrainte <strong>de</strong> fluctu<strong>at</strong>ions autour du ren<strong>de</strong>ment espéré . . . . . 366<br />
11.2.3 Optimis<strong>at</strong>ion sous d’autres contraintes . . . . . . . . . . . . . . . . . . . . . . . 366<br />
11.3 Equilibre <strong>de</strong> marché . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367<br />
11.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367<br />
11.5 Annexe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368<br />
12 Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes 373<br />
12.1 Comprendre <strong>et</strong> gérer les risques grands <strong>et</strong> extrêmes . . . . . . . . . . . . . . . . . . . . 374
8 Table <strong>de</strong>s m<strong>at</strong>ières<br />
12.1.1 Distributions <strong>de</strong>s ren<strong>de</strong>ments à queues épaisses . . . . . . . . . . . . . . . . . . 376<br />
12.1.2 Dépendance temporelle intermittente à l’origine <strong>de</strong>s gran<strong>de</strong>s pertes . . . . . . . 376<br />
12.1.3 Dépendance <strong>de</strong> queue <strong>et</strong> contagion . . . . . . . . . . . . . . . . . . . . . . . . 377<br />
12.1.4 N<strong>at</strong>ure multidimensionnelle <strong>de</strong>s risques . . . . . . . . . . . . . . . . . . . . . . 378<br />
12.1.5 P<strong>et</strong>its risques, grands risques <strong>et</strong> ren<strong>de</strong>ment . . . . . . . . . . . . . . . . . . . . 378<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379<br />
12.2 Minimiser l’impact <strong>de</strong>s grands co-mouvements . . . . . . . . . . . . . . . . . . . . . . 381<br />
12.2.1 Quantifier les grands co-mouvements . . . . . . . . . . . . . . . . . . . . . . . 382<br />
12.2.2 Dépendance <strong>de</strong> queue génerée par un modèle à facteur . . . . . . . . . . . . . . 383<br />
12.2.3 implément<strong>at</strong>ion pr<strong>at</strong>ique <strong>et</strong> conséquences . . . . . . . . . . . . . . . . . . . . . 384<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386<br />
13 Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique 387<br />
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389<br />
13.1 Définition <strong>et</strong> concepts importants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392<br />
13.1.1 Distributions <strong>de</strong> Weibull modifiées . . . . . . . . . . . . . . . . . . . . . . . . 392<br />
13.1.2 Equivalence <strong>de</strong> queue pour les fonctions <strong>de</strong> distribution . . . . . . . . . . . . . 393<br />
13.1.3 La copule gaussienne . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394<br />
13.2 Distribution <strong>de</strong> richesse d’un <strong>portefeuille</strong> pour différentes structures <strong>de</strong> dépendance . . . 395<br />
13.2.1 Distribution <strong>de</strong> richesse pour <strong>de</strong>s actifs indépendants . . . . . . . . . . . . . . . 395<br />
13.2.2 Distribution <strong>de</strong> richesse pour <strong>de</strong>s actifs comonotones . . . . . . . . . . . . . . . 396<br />
13.2.3 Distribution <strong>de</strong> richesse sous hypothèse <strong>de</strong> copule gaussienne . . . . . . . . . . 398<br />
13.3 Value-<strong>at</strong>-Risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400<br />
13.4 Portefeuilles Optimaux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401<br />
13.4.1 Portefeuilles à risque minimum . . . . . . . . . . . . . . . . . . . . . . . . . . 401<br />
13.4.2 Portefeuilles VaR-efficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404<br />
13.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417<br />
14 Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché 423<br />
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425<br />
14.2 Mesure <strong>de</strong>s grands risques d’un <strong>portefeuille</strong> . . . . . . . . . . . . . . . . . . . . . . . . 428
Table <strong>de</strong>s m<strong>at</strong>ières 9<br />
14.2.1 Pourquoi les moments d’ordres élevés perm<strong>et</strong>tent-ils <strong>de</strong> quantifier <strong>de</strong> plus grands<br />
risques ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428<br />
14.2.2 Quantific<strong>at</strong>ion <strong>de</strong>s fluctu<strong>at</strong>ions d’un actif . . . . . . . . . . . . . . . . . . . . . 429<br />
14.2.3 Exemples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430<br />
14.3 La frontière efficiente généralisée . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 432<br />
14.3.1 Frontière en l’absence d’actif sans risque . . . . . . . . . . . . . . . . . . . . . 432<br />
14.3.2 Frontière efficiente en présence d’actif sans risque . . . . . . . . . . . . . . . . 433<br />
14.3.3 Théorème <strong>de</strong> sépar<strong>at</strong>ion en <strong>de</strong>ux fonds . . . . . . . . . . . . . . . . . . . . . . 433<br />
14.3.4 Influence du taux d’intérêt sans risque . . . . . . . . . . . . . . . . . . . . . . . 434<br />
14.4 Classific<strong>at</strong>ions <strong>de</strong>s actifs <strong>et</strong> <strong>de</strong>s <strong>portefeuille</strong>s . . . . . . . . . . . . . . . . . . . . . . . . 434<br />
14.4.1 Métho<strong>de</strong> d’ajustement du risque . . . . . . . . . . . . . . . . . . . . . . . . . . 435<br />
14.4.2 Risque marginal d’un actif au sein d’un <strong>portefeuille</strong> . . . . . . . . . . . . . . . 436<br />
14.5 Un nouveau modèle d’équilibre <strong>de</strong> marché . . . . . . . . . . . . . . . . . . . . . . . . 437<br />
14.5.1 Equilibre en marché homogène . . . . . . . . . . . . . . . . . . . . . . . . . . 438<br />
14.5.2 Equilibre en marché hétérogène . . . . . . . . . . . . . . . . . . . . . . . . . . 438<br />
14.6 Estim<strong>at</strong>ion <strong>de</strong> la distribution <strong>de</strong> probabilité jointe du ren<strong>de</strong>ment <strong>de</strong> plusieurs actifs . . . 439<br />
14.6.1 Brève exposition <strong>et</strong> justific<strong>at</strong>ion <strong>de</strong> la métho<strong>de</strong> . . . . . . . . . . . . . . . . . . 440<br />
14.6.2 Transform<strong>at</strong>ion d’une variable alé<strong>at</strong>oire en une variable gaussienne . . . . . . . 441<br />
14.6.3 D<strong>et</strong>ermin<strong>at</strong>ion <strong>de</strong> la distribution jointe : maximum d’entropie <strong>et</strong> copule gaussienne442<br />
14.6.4 Test empirique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443<br />
14.7 Choix d’une famille exponentielle pour paramétrer les distributions marginales . . . . . 444<br />
14.7.1 Les distributions <strong>de</strong> Weibull modifiées . . . . . . . . . . . . . . . . . . . . . . 444<br />
14.7.2 Transform<strong>at</strong>ion <strong>de</strong>s lois <strong>de</strong> Weibull en gaussiennes . . . . . . . . . . . . . . . . 445<br />
14.7.3 Test empirique <strong>et</strong> estim<strong>at</strong>ion <strong>de</strong>s paramètres . . . . . . . . . . . . . . . . . . . 445<br />
14.8 Développement en cumulants <strong>de</strong> la distribution du <strong>portefeuille</strong> . . . . . . . . . . . . . . 446<br />
14.8.1 Lien entre moments <strong>et</strong> cumulants . . . . . . . . . . . . . . . . . . . . . . . . . 446<br />
14.8.2 Cas <strong>de</strong> distributions symétriques . . . . . . . . . . . . . . . . . . . . . . . . . 447<br />
14.8.3 Cas <strong>de</strong> distributions asymétriques . . . . . . . . . . . . . . . . . . . . . . . . . 449<br />
14.8.4 Tests empiriques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 449<br />
14.9 Peut-on avoir le beurre <strong>et</strong> l’argent du beurre ? . . . . . . . . . . . . . . . . . . . . . . . 450<br />
14.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451<br />
Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
10 Table <strong>de</strong>s m<strong>at</strong>ières<br />
Références . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464<br />
Conclusions <strong>et</strong> Perspectives 491<br />
A Evalu<strong>at</strong>ion <strong>de</strong> la conduite du proj<strong>et</strong> <strong>de</strong> thèse 495<br />
A.1 Bref résumé du suj<strong>et</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 495<br />
A.2 Eléments <strong>de</strong> contexte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496<br />
A.2.1 Choix du suj<strong>et</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496<br />
A.2.2 Choix <strong>de</strong> l’encadrement <strong>et</strong> du labor<strong>at</strong>oire d’accueil . . . . . . . . . . . . . . . . 496<br />
A.2.3 Financement du proj<strong>et</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497<br />
A.3 Evolution du proj<strong>et</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498<br />
A.3.1 Elabor<strong>at</strong>ion du proj<strong>et</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498<br />
A.3.2 Conduite du proj<strong>et</strong> . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498<br />
A.3.3 R<strong>et</strong>ombées scientifiques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499<br />
A.4 Compétences acquises <strong>et</strong> enseignements personnels . . . . . . . . . . . . . . . . . . . . 499<br />
A.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 500<br />
Bibliographie 503
Avant-propos<br />
La rédaction d’une thèse est un travail <strong>de</strong> longue haleine, exigeant, quelque déroutant voire décourageant,<br />
tout particulièrement lorsque les difficultés s’accumulent <strong>et</strong> semblent insurmontables. Mais au final, combien<br />
est captivante, passionnante <strong>et</strong> enthousiasmante c<strong>et</strong>te expérience, première réelle occasion <strong>de</strong> m<strong>et</strong>tre<br />
à profit la somme <strong>de</strong> connaissances p<strong>at</strong>iemment engrangée au cours <strong>de</strong> longues années d’étu<strong>de</strong>s.<br />
C’est lorsque vient le moment <strong>de</strong> la rédaction finale que l’on mesure l’ensemble du chemin parcouru.<br />
Les hésit<strong>at</strong>ions <strong>et</strong> difficultés rencontrées tout au long <strong>de</strong> la route s’évanouissent <strong>et</strong> seuls les succès <strong>et</strong><br />
s<strong>at</strong>isfactions reviennent en mémoire. Ressurgissent aussi les noms <strong>de</strong> ceux qui vous ont accompagné sur<br />
c<strong>et</strong>te route, <strong>et</strong> c’est à eux que je souhaiterai, avant toute chose, adresser ces quelques remerciements.<br />
Le travail que j’ai mené durant ces <strong>de</strong>ux années <strong>de</strong> thèse doit énormément à Didier Sorn<strong>et</strong>te <strong>et</strong> Jean-Paul<br />
Laurent qui ont accepté <strong>de</strong> le diriger. Je tiens avant tout à leur exprimer ma gr<strong>at</strong>itu<strong>de</strong> pour ces années<br />
passées à leur contact <strong>et</strong> voudrais dire le grand plaisir <strong>et</strong> l’immense s<strong>at</strong>isfaction que j’ai eu à travailler<br />
avec eux. Ils ont su à la fois me gui<strong>de</strong>r judicieusement <strong>et</strong> me laisser une gran<strong>de</strong> l<strong>at</strong>itu<strong>de</strong> quant aux choix<br />
<strong>de</strong>s thèmes <strong>de</strong> recherche que je souhaitais poursuivre, ce dont je leur suis infiniment reconnaissant.<br />
Je tiens aussi à remercier chaleureusement les professeurs Gouriéroux <strong>et</strong> Roehner qui m’ont fait l’honneur<br />
d’accepter d’être les rapporteurs <strong>de</strong> c<strong>et</strong>te thèse. Les questions <strong>et</strong> remarques qu’ils ont formulées,<br />
ainsi que les discussions que nous avons eues ont été pour moi du plus grand intérêt car elles ont<br />
été génér<strong>at</strong>rices d’idées nouvelles <strong>et</strong> m’ont ainsi permis d’entrevoir d’autres directions <strong>de</strong> recherche,<br />
complémentaires <strong>de</strong> celle suivies jusqu’ici. Je souhaite également exprimer ma reconnaissance aux professeurs<br />
Augros <strong>et</strong> Rockinger ainsi qu’à Jean-François Muzy – avec qui j’ai eu en outre le grand plaisir<br />
<strong>de</strong> collaborer – qui ont bien voulu consacrer une part <strong>de</strong> leur temps à juger mon travail <strong>et</strong> accepté <strong>de</strong><br />
participer à mon jury <strong>de</strong> thèse.<br />
J’aimerais également remercier les nombreuses personnes avec qui j’ai été amené à travailler sur l’un<br />
ou l’autre <strong>de</strong>s divers thèmes <strong>de</strong> recherche que j’ai abordé au cours <strong>de</strong> ma thèse. En premier lieu, je<br />
souhaiterais dire toute mon estime <strong>et</strong> ma reconnaissance à Vladilen Pisarenko, pour l’ai<strong>de</strong> précieuse<br />
qu’il m’a apporté tout au long <strong>de</strong> ces <strong>de</strong>ux années. Je voudrais aussi remercier Jorgen An<strong>de</strong>rsen avec<br />
qui j’ai passé les trois mois qui ont précédé le début <strong>de</strong> ma thèse durant mon stage <strong>de</strong> DEA ainsi qu’Ali<br />
Chabaane <strong>et</strong> Françoise Turpin du groupe BNP-Paribas avec qui j’ai entamé une collabor<strong>at</strong>ion fructueuse.<br />
Je souhaiterais aussi remercier Anne Sorn<strong>et</strong>te pour son soutien <strong>et</strong> ses encouragements constants mais<br />
également pour m’avoir incité à participer au proj<strong>et</strong> <strong>de</strong> “Nouveau chapitre <strong>de</strong> la thèse”, expérience qui<br />
s’est avérée être très intéressante <strong>et</strong> enrichissante. Inci<strong>de</strong>mment, je remercie Nadjia Hohweiller, consultante<br />
en bilan <strong>de</strong> compétence à la chambre <strong>de</strong> commerce <strong>et</strong> d’industrie <strong>de</strong> Nice, qui a participé à la<br />
rédaction <strong>de</strong> ce chapitre <strong>et</strong> Emmanuel Tric <strong>de</strong> l’Associ<strong>at</strong>ion Bernard Gregory qui soutient ce proj<strong>et</strong>.<br />
Je voudrais également remercier Bernard Gaypara <strong>et</strong> Yann Ageon pour l’ai<strong>de</strong> précieuse qu’ils m’ont<br />
apportée dans le domaine inform<strong>at</strong>ique, Bernard pour ce qui a été <strong>de</strong> résoudre les (nombreuses) pannes<br />
<strong>de</strong> mon ordin<strong>at</strong>eur <strong>et</strong> autres problèmes d’impression <strong>et</strong> Yann pour ses compétences <strong>de</strong> programmeur qui<br />
11
12<br />
m’ont souvent été d’un grand secours.<br />
Enfin, je tiens à remercier ma famille, mes ami(e)s, mes proches <strong>et</strong> tous ceux qui m’ont encouragé tout<br />
au long <strong>de</strong> ces <strong>de</strong>ux années. Qu’ils voient dans l’accomplissement <strong>de</strong> ce travail un témoignage <strong>de</strong> ma<br />
reconnaissance pour le soutien <strong>et</strong> la confiance qu’ils m’ont toujours manifestés.<br />
Nice, Janvier 2003
Introduction<br />
La multiplic<strong>at</strong>ion <strong>de</strong>s risques majeurs est l’une <strong>de</strong>s caractéristiques <strong>de</strong>s sociétés mo<strong>de</strong>rnes, à tel point que,<br />
reprenant le titre d’un <strong>de</strong>s plus célèbres ouvrages du sociologue allemand Ulrich Beck, on n’hésite plus<br />
à les qualifier <strong>de</strong> sociétés du risque. En outre, notre conception même du risque a évolué <strong>de</strong> sorte que les<br />
événements c<strong>at</strong>astrophiques ne semblent plus désormais perçus par nos concitoyens comme s’imposant<br />
à nos sociétés du fait d’un <strong>de</strong>stin injuste <strong>et</strong> implacable mais comme résultant essentiellement <strong>de</strong> notre<br />
propre développement technologique dont la complexité <strong>at</strong>teint un tel <strong>de</strong>gré qu’elle engendre “n<strong>at</strong>urellement”<br />
<strong>de</strong>s situ<strong>at</strong>ions <strong>de</strong> crises pour le présent (Tchernobyl, AZF...) comme pour l’avenir (réchauffement<br />
planétaire <strong>et</strong> changement clim<strong>at</strong>ique, par exemple). C<strong>et</strong>te prolifér<strong>at</strong>ion <strong>de</strong>s sources <strong>de</strong> risques impose<br />
alors aux organis<strong>at</strong>ions publiques, mais aussi aux institutions privées telles que les compagnies d’assurances<br />
ou les banques, <strong>de</strong> nouvelles oblig<strong>at</strong>ions quant à la détermin<strong>at</strong>ion, au contrôle <strong>et</strong> à la <strong>gestion</strong> <strong>de</strong> ces<br />
risques extrêmes afin d’en rendre les conséquences supportables pour la communauté sans pour autant<br />
m<strong>et</strong>tre en péril les organismes qui ont en charge <strong>de</strong> les assurer.<br />
Dans le secteur strictement financier, les krachs représentent peut-être les événements les plus frappants<br />
<strong>de</strong> toute une c<strong>at</strong>égorie d’événements extrêmes, <strong>et</strong> <strong>de</strong> plus en plus fréquemment l’activité financière doit<br />
en subir les eff<strong>et</strong>s néfastes. Que l’on songe, pour fixer les idées, que le krach d’octobre 1987 a englouti en<br />
quelques jours plus <strong>de</strong> mille milliards <strong>de</strong> dollars ou encore que le récent krach <strong>de</strong> la nouvelle économie<br />
a conduit à un effondrement <strong>de</strong> près d’un tiers <strong>de</strong> la capitalis<strong>at</strong>ion boursière mondiale par rapport à<br />
son niveau <strong>de</strong> 1999. Or, si le rôle <strong>de</strong> l’argent est <strong>de</strong> constituer une réserve <strong>de</strong> valeur, encore faut-il être<br />
capable d’en maîtriser les fluctu<strong>at</strong>ions afin notamment <strong>de</strong> ne pas voir, en un instant, englouti l’épargne<br />
<strong>de</strong> toute une vie, ruiné les proj<strong>et</strong>s d’expansion d’une entreprise ou anéanti l’économie d’une n<strong>at</strong>ion. Il<br />
est donc absolument nécessaire <strong>de</strong> se doter d’outils <strong>et</strong> <strong>de</strong> normes perm<strong>et</strong>tant <strong>de</strong> mieux appréhen<strong>de</strong>r les<br />
risques extrêmes sur les marchés financiers. Les instances bancaires mondiales, pleinement conscientes<br />
<strong>de</strong> c<strong>et</strong>te nécessité, ont émis <strong>de</strong>s instructions en se sens, au travers <strong>de</strong>s recommand<strong>at</strong>ions du comité <strong>de</strong><br />
Bâle (1996, 2001), en proposant <strong>de</strong>s modèles <strong>de</strong> <strong>gestion</strong> interne <strong>de</strong>s risques <strong>et</strong> en imposant <strong>de</strong>s montants<br />
minimaux <strong>de</strong> fonds propres en adéqu<strong>at</strong>ion avec les expositions aux risques. Cependant, quelques critiques<br />
se sont élevées contre ces recommand<strong>at</strong>ions (Szergö 1999, Danielsson, Embrechts, Goodhart, Ke<strong>at</strong>ing,<br />
Muennich, Renault <strong>et</strong> Shin 2001) jugées inadaptées <strong>et</strong> pouvant même conduire à une déstabilis<strong>at</strong>ion <strong>de</strong>s<br />
marchés. C<strong>et</strong>te controverse fait ressortir toute l’importance d’une meilleure compréhension <strong>de</strong>s risques<br />
extrêmes, <strong>de</strong> leurs conséquences <strong>et</strong> <strong>de</strong>s moyens <strong>de</strong> s’en prémunir.<br />
Ce challenge comporte à notre avis <strong>de</strong>ux vol<strong>et</strong>s. D’une part il est indispensable d’être capable <strong>de</strong> mieux<br />
quantifier les risques extrêmes, <strong>et</strong> cela passe par le développement d’outils <strong>st<strong>at</strong>istique</strong>s perm<strong>et</strong>tant <strong>de</strong><br />
dépasser le cadre gaussien dans lequel s’ancre la <strong>théorie</strong> financière classique héritée <strong>de</strong> Bachelier (1900),<br />
Markovitz (1959), <strong>et</strong> Black <strong>et</strong> Scholes (1973) notamment. D’autre part, il convient <strong>de</strong> s’interroger sur<br />
la façon dont peut s’intégrer à la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> la prise en compte <strong>de</strong>s risques extrêmes. En<br />
eff<strong>et</strong>, il est fondamental <strong>de</strong> savoir si, à l’instar <strong>de</strong> ce que nous apprend la <strong>théorie</strong> financière classique<br />
basée sur l’approche moyenne-variance, les risques extrêmes <strong>de</strong>meurent diversifiables ? Si tel n’est pas<br />
le cas, il conviendra alors d’envisager l’utilis<strong>at</strong>ion d’autres moyens que la constitution <strong>de</strong> <strong>portefeuille</strong>s<br />
13
14 Introduction<br />
pour espérer se couvrir contre ce type <strong>de</strong> risques en ayant recours notamment aux produits dérivés, pour<br />
autant qu’ils soient capables <strong>de</strong> fournir une réelle assurance contre les gran<strong>de</strong>s vari<strong>at</strong>ions <strong>de</strong> cours - ce<br />
qui fût loin d’être le cas lors du krach <strong>de</strong> 1987 - ou peut-être avoir recours à la mutualis<strong>at</strong>ion comme en<br />
assurance par exemple.<br />
Le problème du contrôle <strong>de</strong>s risques extrêmes en finance <strong>et</strong> plus particulièrement son applic<strong>at</strong>ion à la <strong>gestion</strong><br />
<strong>de</strong> <strong>portefeuille</strong> peuvent sembler totalement étrangers au domaine <strong>de</strong> la physique. Cela n’est cependant<br />
qu’une apparence puisque dès ses premiers pas, la finance a partagé avec la physique, la <strong>de</strong>vançant<br />
même parfois, <strong>de</strong> nombreuses métho<strong>de</strong>s <strong>et</strong> outils telle la marche alé<strong>at</strong>oire brownienne (Bachelier 1900,<br />
Einstein 1905), l’utilis<strong>at</strong>ion <strong>de</strong> la notion <strong>de</strong> probabilité subjective qui perm<strong>et</strong> <strong>de</strong> rendre compte du comportement<br />
<strong>de</strong>s agents économiques dans l’incertain (Savage 1954) mais aussi d’interpréter <strong>de</strong>s expériences<br />
<strong>de</strong> mécanique quantique (Caves, Fluchs <strong>et</strong> Schack 2002), ou encore la notion <strong>de</strong> compromis moyennevariance<br />
en <strong>théorie</strong> du <strong>portefeuille</strong> Markovitz (1959) comme en algorithmique quantique (Maurer 2001),<br />
ainsi que l’équ<strong>at</strong>ion <strong>de</strong> diffusion <strong>de</strong> la chaleur qui sert également à décrire l’évolution du prix d’une<br />
option dans l’univers <strong>de</strong> Black <strong>et</strong> Scholes (1973), <strong>et</strong> plus récemment l’utilis<strong>at</strong>ion <strong>de</strong> modèles d’agents en<br />
interactions (variante du modèle d’Ising par exemple) perm<strong>et</strong>tant <strong>de</strong> dépasser le cadre standard <strong>de</strong> l’agent<br />
économique représent<strong>at</strong>if <strong>et</strong> ainsi <strong>de</strong> mieux percevoir les mécanismes fondamentaux à l’oeuvre sur les<br />
marchés financiers.<br />
C’est donc dans les métho<strong>de</strong>s, mais aussi dans les concepts, qu’il faut chercher un lien entre <strong>de</strong>s m<strong>at</strong>ières<br />
en apparence si différentes que la finance <strong>et</strong> la physique. En eff<strong>et</strong>, la mesure du risque fait appel à <strong>de</strong>s<br />
notions <strong>de</strong> <strong>théorie</strong> <strong>de</strong> l’inform<strong>at</strong>ion, <strong>de</strong> la décision <strong>et</strong> du contrôle dynamique qui ont une longue tradition<br />
en physique, quant à la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>, <strong>de</strong> manière un peu schém<strong>at</strong>ique, on peut affirmer que<br />
ce n’est rien d’autre qu’un exemple particulier <strong>de</strong> problème d’optimis<strong>at</strong>ion sous contraintes, comme<br />
on en rencontre dans <strong>de</strong> nombreux domaines. En cela, d’un point <strong>de</strong> vue purement m<strong>at</strong>hém<strong>at</strong>ique, la<br />
recherche <strong>de</strong> <strong>portefeuille</strong>s optimaux n’est guère différente <strong>de</strong> la recherche <strong>de</strong>s ét<strong>at</strong>s d’équilibres d’un<br />
système thermodynamique. En eff<strong>et</strong>, dans le premier cas, on cherche à obtenir une valeur minimale du<br />
risque associé aux vari<strong>at</strong>ions <strong>de</strong> richesse du <strong>portefeuille</strong> pour une valeur fixée du ren<strong>de</strong>ment moyen <strong>et</strong><br />
un montant <strong>de</strong> la richesse initiale donnée. Dans le second cas, on cherche - pour ce qui est <strong>de</strong> l’étu<strong>de</strong> <strong>de</strong><br />
l’ensemble micro-canonique - à maximiser l’entropie (ou minimiser la néguentropie) du système sous<br />
la contrainte d’une énergie totale <strong>et</strong> d’un nombre <strong>de</strong> particules fixé. On peut donc, formellement, faire<br />
un parallèle (certes un peu rapi<strong>de</strong>) entre risque <strong>et</strong> néguentropie, ren<strong>de</strong>ment espéré <strong>et</strong> énergie totale <strong>et</strong><br />
richesse initiale <strong>et</strong> nombre <strong>de</strong> particules.<br />
Ceci étant, le parallèle s’arrête malheureusement là <strong>et</strong> l’on ne peut espérer d’une applic<strong>at</strong>ion simple<br />
<strong>et</strong> directe <strong>de</strong> certains résult<strong>at</strong>s généraux <strong>de</strong> la thermodynamique <strong>st<strong>at</strong>istique</strong> ou <strong>de</strong> mécanique quantique<br />
qu’elle nous perm<strong>et</strong>te d’obtenir d’intéressants enseignements pour la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>. La distinction<br />
essentielle entre ces <strong>de</strong>ux problèmes d’optimis<strong>at</strong>ion que sont la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> <strong>et</strong> la physique<br />
<strong>st<strong>at</strong>istique</strong> vient en fait <strong>de</strong> la différence entre les ordres <strong>de</strong> gran<strong>de</strong>ur en jeu dans ces <strong>de</strong>ux situ<strong>at</strong>ions : les<br />
plus gros <strong>portefeuille</strong>s contiennent au maximum quelques milliers d’actifs alors que les systèmes thermodynamiques<br />
auxquels nous sommes usuellement confrontés sont constitués <strong>de</strong> quelques cent milles<br />
milliards <strong>de</strong> milliards (NA 10 23 ) <strong>de</strong> particules. Pour un tel nombre <strong>de</strong> particules, le théorème <strong>de</strong> la<br />
limite centrale s’applique pleinement, y compris pour <strong>de</strong>s événements <strong>de</strong> très faibles probabilités, <strong>et</strong> fournit<br />
une simplific<strong>at</strong>ion fort utile. De plus, vu le nombre gigantesque d’entités considérées, les fluctu<strong>at</strong>ions<br />
rel<strong>at</strong>ives typiques <strong>de</strong>s gran<strong>de</strong>urs mesurées par rapport à leurs valeurs moyennes (espérées) sont <strong>de</strong> l’ordre<br />
<strong>de</strong> 1/ √ NA ∼ 10 −11 − 10 −12 , ce qui les rend tout à fait imperceptibles.<br />
Les choses sont toutes autres pour ce qui est <strong>de</strong> la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>, où il convient <strong>de</strong> caractériser<br />
avec rigueur la distribution <strong>de</strong> ren<strong>de</strong>ment dudit <strong>portefeuille</strong>. En eff<strong>et</strong>, du fait du nombre le plus souvent<br />
modéré d’actifs qui constituent un <strong>portefeuille</strong> <strong>et</strong> <strong>de</strong> par la structure sous-jacente <strong>de</strong>s distributions <strong>de</strong><br />
chacun <strong>de</strong>s actifs, la distribution <strong>de</strong> ren<strong>de</strong>ment du <strong>portefeuille</strong> s’avère très éloignée <strong>de</strong> la distribution
Introduction 15<br />
gaussienne, contrairement à ce que pourrait conduire à penser le théorème <strong>de</strong> la limite centrale <strong>et</strong> à ce<br />
que l’on observe le plus souvent en physique <strong>st<strong>at</strong>istique</strong>. Il est donc nécessaire d’estimer la distribution<br />
<strong>de</strong> ren<strong>de</strong>ment du <strong>portefeuille</strong>, ce qui peut être réalisé <strong>de</strong> manière directe pour une alloc<strong>at</strong>ion <strong>de</strong> capital<br />
donnée ou bien <strong>de</strong> manière plus générale en se penchant sur le problème <strong>de</strong> l’estim<strong>at</strong>ion <strong>de</strong> la distribution<br />
jointe <strong>de</strong> tous les actifs à inclure dans le <strong>portefeuille</strong>. La première approche est certes beaucoup plus<br />
simple <strong>et</strong> rapi<strong>de</strong> puisqu’elle ne consiste qu’en l’estim<strong>at</strong>ion d’une distribution monovariée, mais il est<br />
aussi très clair qu’elle n’est en fait guère s<strong>at</strong>isfaisante car elle néglige une gran<strong>de</strong> partie <strong>de</strong> l’inform<strong>at</strong>ion<br />
observable <strong>et</strong> extractible à partir <strong>de</strong>s données recueillies sur les marchés financiers <strong>et</strong> dont seule est<br />
capable <strong>de</strong> rendre compte la distribution multivariée <strong>de</strong>s actifs. Cependant, on peut gar<strong>de</strong>r à l’esprit que<br />
ces <strong>de</strong>ux approches se rejoignent dans la mesure où l’on sait que la connaissance <strong>de</strong>s distributions <strong>de</strong><br />
ren<strong>de</strong>ments <strong>de</strong> tous les <strong>portefeuille</strong>s (pour toutes les alloc<strong>at</strong>ions <strong>de</strong> capital possibles) est équivalente à<br />
la connaissance <strong>de</strong> la distribution multivariée. Au final, la secon<strong>de</strong> métho<strong>de</strong> semble préférable, <strong>et</strong> parait<br />
être celle qui mobilise aujourd’hui le plus d’énergie aussi bien dans la recherche académique que privée.<br />
L’<strong>at</strong>taque frontale du problème <strong>de</strong> la détermin<strong>at</strong>ion <strong>de</strong> la distribution multivariée <strong>de</strong>s actifs est ardue <strong>et</strong> à<br />
notre avis beaucoup moins instructive que l’étu<strong>de</strong> séparée du comportement <strong>de</strong>s distributions marginales<br />
<strong>de</strong> chaque actif d’une part <strong>et</strong> <strong>de</strong> la structure <strong>de</strong> dépendance entre ces actifs d’autre part. C’est pourquoi<br />
nous avons privilégié c<strong>et</strong>te démarche-ci, avec pour objectif principal <strong>de</strong> rendre compte <strong>de</strong> manière la plus<br />
précise possible <strong>de</strong>s diverses sources <strong>de</strong> risques : risques associés individuellement à chaque actif d’une<br />
part <strong>et</strong> risques collectivement liés à l’ensemble <strong>de</strong>s actifs d’autre part.<br />
Ceci nous a conduit dans un premier temps (partie I) à nous intéresser au comportement marginal <strong>de</strong>s<br />
actifs dans le but <strong>de</strong> faire ressortir les points essentiels qui nécessitaient une compréhension approfondie<br />
afin <strong>de</strong> mieux cerner l’origine <strong>de</strong>s risques inhérents aux actifs individuels. Dans notre optique, se focalisant<br />
sur l’impact <strong>de</strong>s risques extrêmes, les krachs - succédant aux bulles spécul<strong>at</strong>ives - présentent un<br />
<strong>at</strong>trait tout particulier, au même titre que les quelques plus grands mouvements observés sur les marchés.<br />
En eff<strong>et</strong>, conformément à la célèbre loi <strong>de</strong>s 80-20 <strong>de</strong> V. Par<strong>et</strong>o, une toute p<strong>et</strong>ite fraction d’événements (ici<br />
les krachs <strong>et</strong> autres grands mouvements) suffit à rendre compte <strong>de</strong> la plus gran<strong>de</strong> partie <strong>de</strong>s conséquences<br />
(ici le ren<strong>de</strong>ment à long terme <strong>de</strong>s actifs, par exemple) résultant <strong>de</strong> l’ensemble <strong>de</strong> tous les événements. En<br />
outre, <strong>de</strong> par leurs répétitions qui semblent <strong>de</strong> plus en plus fréquentes, on ne peut plus faire l’économie <strong>de</strong><br />
négliger ces événements extrêmes, en laissant au seul hasard / <strong>de</strong>stin le soin <strong>de</strong> déci<strong>de</strong>r <strong>de</strong> leurs eff<strong>et</strong>s tout<br />
autant macro-économiques que sur chacun d’entre nous. Il est donc souhaitable <strong>de</strong> mieux comprendre<br />
comment <strong>et</strong> pourquoi ces événements se produisent.<br />
Pour cela, nous avons mené notre étu<strong>de</strong> selon trois directions complémentaires. Nous avons tout d’abord<br />
commencé par une étu<strong>de</strong> empirique conduisant à une <strong>de</strong>scription <strong>st<strong>at</strong>istique</strong> <strong>et</strong> st<strong>at</strong>ique <strong>de</strong>s grands risques.<br />
Nous avons ensuite poursuivi par une étu<strong>de</strong> phénoménologique prenant en compte le caractère dynamique<br />
<strong>de</strong> l’occurrence <strong>de</strong>s extrêmes, <strong>et</strong> dont le but était <strong>de</strong> décrire <strong>de</strong> manière ad hoc les phénomènes<br />
observés. Nous avons enfin terminé par une étu<strong>de</strong> “microscopique” ou micro-structurelle nous perm<strong>et</strong>tant<br />
d’abor<strong>de</strong>r <strong>de</strong> manière fondamentale les mécanismes à l’œuvre sur les marchés <strong>et</strong> <strong>de</strong> relier les propriétés<br />
<strong>st<strong>at</strong>istique</strong>s st<strong>at</strong>iques <strong>et</strong> dynamiques <strong>de</strong>s cours à <strong>de</strong>s comportements individuels <strong>et</strong> <strong>de</strong>s organis<strong>at</strong>ions <strong>de</strong><br />
marchés.<br />
D’un point <strong>de</strong> vue <strong>st<strong>at</strong>istique</strong>, il est nécessaire d’estimer précisément la probabilité d’occurrence <strong>de</strong>s<br />
événements extrêmes, qui se traduisent par le fait désormais bien connu que les distributions <strong>de</strong> ren<strong>de</strong>ment<br />
<strong>de</strong>s actifs ont <strong>de</strong>s “queues épaisses”. Cependant, encore faut-il être capable <strong>de</strong> caractériser le plus<br />
parfaitement possible ces “queues épaisses”, c’est-à-dire sans négliger ni surestimer ces événements<br />
extrêmes. De plus, il faut gar<strong>de</strong>r à l’esprit que toute <strong>de</strong>scription paramétrique est suj<strong>et</strong>te à l’erreur <strong>de</strong><br />
modèle <strong>et</strong> qu’ainsi plusieurs représent<strong>at</strong>ions doivent être considérées afin <strong>de</strong> cerner correctement les eff<strong>et</strong>s<br />
d’une telle source d’erreur. C’est pour cela que nous nous sommes, tout au long <strong>de</strong> notre étu<strong>de</strong>,<br />
intéressés à <strong>de</strong>ux classes <strong>de</strong> distributions à queues épaisses : les distributions régulièrement variables <strong>et</strong>
16 Introduction<br />
les distributions exponentielles étirées dont nous avons discuté la pertinence <strong>et</strong> les conséquences vis-à-vis<br />
<strong>de</strong> la sous / surestim<strong>at</strong>ion <strong>de</strong>s risques (chapitres 1, 2 <strong>et</strong> 3).<br />
Nous nous sommes ensuite intéressés à la manière dont la vol<strong>at</strong>ilité r<strong>et</strong>ourne à un niveau d’“équilibre”<br />
après une longue pério<strong>de</strong> <strong>de</strong> forte variabilité <strong>de</strong>s cours (chapitre 4). On sait en eff<strong>et</strong> que la vol<strong>at</strong>ilité<br />
présente un phénomène <strong>de</strong> persistance très marqué, <strong>et</strong> qu’après une pério<strong>de</strong> anormalement haute (ou<br />
basse), elle relaxe vers un niveau moyen, que l’on peut en quelque sorte associer à un niveau d’équilibre.<br />
L’étu<strong>de</strong> <strong>de</strong> c<strong>et</strong>te relax<strong>at</strong>ion est très importante du point <strong>de</strong> vue <strong>de</strong> la <strong>gestion</strong> <strong>de</strong>s risques, car elle perm<strong>et</strong><br />
d’une part d’estimer la durée typique d’une pério<strong>de</strong> anormalement turbulente ou calme <strong>et</strong> d’autre part<br />
elle apporte un moyen <strong>de</strong> prédiction rel<strong>at</strong>ivement fiable <strong>de</strong> la valeur future <strong>de</strong> la vol<strong>at</strong>ilité, dont on connaît<br />
l’intérêt pour tout ce qui concerne le “pricing” <strong>de</strong> produits dérivés notamment. De plus, <strong>et</strong> c’est le point<br />
clé <strong>de</strong> notre développement, cela nous a permis <strong>de</strong> proposer un mécanisme expliquant l’impact du flux<br />
d’inform<strong>at</strong>ion sur la dynamique <strong>de</strong> la vol<strong>at</strong>ilité. Ceci ouvre la voie à une étu<strong>de</strong> systém<strong>at</strong>ique <strong>de</strong> l’eff<strong>et</strong><br />
<strong>de</strong> l’arrivée <strong>de</strong> telle ou telle inform<strong>at</strong>ion sur les marchés aussi bien ex post, pour tout ce qui concerne<br />
l’analyse <strong>de</strong>s causes <strong>de</strong>s grands mouvements <strong>de</strong> cours, qu’ex ante pour tout ce qui touche à l’étu<strong>de</strong> <strong>de</strong><br />
scénarii <strong>et</strong> l’anticip<strong>at</strong>ion <strong>de</strong>s eff<strong>et</strong>s <strong>de</strong>s grands chocs.<br />
Enfin, nous avons voulu explorer un peu plus avant les causes microscopiques responsables <strong>de</strong>s phénomènes<br />
observés sur les marchés (chapitres 5 <strong>et</strong> 6). Pour cela, nous avons construit un modèle d’agents en<br />
interaction, rel<strong>at</strong>ivement parcimonieux, visant à rendre compte <strong>de</strong> manière plus réaliste que ne le font<br />
les modèles actuels <strong>de</strong> la croissance super-exponentielle <strong>de</strong>s cours lors <strong>de</strong>s phases <strong>de</strong> bulles spécul<strong>at</strong>ives<br />
qui conduisent à d’importantes surestim<strong>at</strong>ions du prix <strong>de</strong>s actifs <strong>et</strong> finalement à <strong>de</strong> fortes corrections.<br />
Les résult<strong>at</strong>s obtenus sont intéressants <strong>et</strong> ont confirmé l’importance <strong>de</strong> certains types <strong>de</strong> comportements<br />
<strong>de</strong>s agents conduisant à <strong>de</strong>s “emballées” <strong>de</strong>s marchés <strong>et</strong> par suite aux pério<strong>de</strong>s <strong>de</strong> gran<strong>de</strong>s vol<strong>at</strong>ilités <strong>et</strong><br />
aux fluctu<strong>at</strong>ions extrêmes. Cependant, on ne peut, en l’ét<strong>at</strong>, espérer intégrer ces modèles, encore trop<br />
rudimentaires, à une chaîne <strong>de</strong> décision concernant la politique <strong>de</strong> <strong>gestion</strong> <strong>de</strong>s risques d’une institution.<br />
Néanmoins, nous pensons que ce genre d’outil présente un <strong>at</strong>trait particulier <strong>et</strong> <strong>de</strong>vrait, à terme, fournir<br />
d’utiles inform<strong>at</strong>ions perm<strong>et</strong>tant une meilleure estim<strong>at</strong>ion / prévision <strong>de</strong>s risques à venir. En particulier,<br />
ces modèles <strong>de</strong>vraient perm<strong>et</strong>tre d’obtenir, à l’ai<strong>de</strong> <strong>de</strong> la génér<strong>at</strong>ion <strong>de</strong> scénarii, <strong>de</strong>s estim<strong>at</strong>ions <strong>de</strong>s<br />
probabilités associées aux événements rares dans certaines phases <strong>de</strong> marchés, estim<strong>at</strong>ion beaucoup plus<br />
fiable que ne le sont les actuelles estim<strong>at</strong>ions subjectives (Johnson, Lamper, Jefferies, Hart <strong>et</strong> Howison<br />
2001).<br />
Le <strong>de</strong>uxième maillon <strong>de</strong> la chaîne <strong>de</strong> reconstruction <strong>de</strong> la distribution multivariée <strong>de</strong>s ren<strong>de</strong>ments d’actifs<br />
consiste en l’étu<strong>de</strong> <strong>de</strong> leur structure <strong>de</strong> dépendance, problème que nous abor<strong>de</strong>rons dans la partie II. En<br />
eff<strong>et</strong>, les risques ne sont pas uniquement dus au comportement marginal <strong>de</strong> chaque actif mais également<br />
à leur comportement collectif. Celui-ci peut être étudié à l’ai<strong>de</strong> d’obj<strong>et</strong>s m<strong>at</strong>hém<strong>at</strong>iques nommés copules<br />
qui perm<strong>et</strong>tent <strong>de</strong> capturer complètement la dépendance entre les actifs. En fait, nous verrons aux<br />
chapitres 7 <strong>et</strong> 8 que la détermin<strong>at</strong>ion <strong>de</strong> la copule est une affaire délic<strong>at</strong>e <strong>et</strong> que là encore le risque<br />
d’erreur <strong>de</strong> modèle est important, expressément pour ce qui est <strong>de</strong>s risques extrêmes. Ainsi, nous verrons<br />
qu’il s’avère nécessaire <strong>de</strong> mener une étu<strong>de</strong> spécifique <strong>de</strong> la dépendance entre les extrêmes, ce que<br />
nous ferrons au chapitre 9, en développant une métho<strong>de</strong> d’estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong><br />
queue, c’est-à-dire <strong>de</strong> la probabilité qu’un actif subisse une très gran<strong>de</strong> perte sachant qu’un autre actif<br />
par exemple ou que le marché dans son ensemble a lui aussi subi une très gran<strong>de</strong> perte. C<strong>et</strong>te quantité est<br />
à notre avis absolument cruciale car elle quantifie <strong>de</strong> manière très simple le fait que les risques extrêmes<br />
puissent être ou non diversifiés par agrég<strong>at</strong>ion au sein <strong>de</strong> <strong>portefeuille</strong>s. En eff<strong>et</strong>, soit le coefficient <strong>de</strong><br />
dépendance <strong>de</strong> queue est nul <strong>et</strong> les extrêmes se produisent <strong>de</strong> manière asymptotiquement indépendante<br />
ce qui perm<strong>et</strong> alors d’envisager <strong>de</strong> les diversifier, soit ils <strong>de</strong>meurent asymptotiquement dépendants <strong>et</strong> on<br />
ne peut qu’espérer minimiser la probabilité d’occurrence <strong>de</strong>s mouvements extrêmes concomitants.<br />
Synthétisant les résult<strong>at</strong>s <strong>de</strong> ces <strong>de</strong>ux premières parties, nous pourrons alors nous poser la question pro-
Introduction 17<br />
prement dite <strong>de</strong> la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>, ce qui sera l’obj<strong>et</strong> <strong>de</strong> la partie III. Avant d’en arriver au<br />
problème <strong>de</strong> la recherche <strong>de</strong>s <strong>portefeuille</strong>s optimaux, qui n’est en fait qu’un problème m<strong>at</strong>hém<strong>at</strong>ique,<br />
nous nous interrogerons sur la manière <strong>de</strong> choisir un <strong>portefeuille</strong> optimal, ce qui nous amènera à discuter,<br />
au chapitre 10, <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> la décision <strong>et</strong> <strong>de</strong> divers moyens <strong>de</strong> quantifier les risques associés à<br />
la distribution <strong>de</strong> ren<strong>de</strong>ment d’un <strong>portefeuille</strong>. La question n’est en fait pas triviale car elle revient à se<br />
<strong>de</strong>man<strong>de</strong>r comment synthétiser le mieux possible en un seul nombre l’inform<strong>at</strong>ion contenue dans une<br />
fonction (<strong>de</strong> distribution) toute entière <strong>et</strong> donc une infinité <strong>de</strong> nombres. Ce résumé <strong>de</strong> l’inform<strong>at</strong>ion ne<br />
pouvant qu’être partiel, il convient, afin d’en conserver la fraction la plus pertinente, <strong>de</strong> bien cerner nos<br />
objectifs <strong>et</strong> pour cela <strong>de</strong> bien comprendre les buts que l’on cherche à <strong>at</strong>teindre en définissant ce qu’est la<br />
notion <strong>de</strong> grand risque ou <strong>de</strong> risque extrême dans le contexte <strong>de</strong> la sélection <strong>de</strong> <strong>portefeuille</strong>s.<br />
A notre avis, la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> doit répondre à <strong>de</strong>ux objectifs. Premièrement, l’alloc<strong>at</strong>ion <strong>de</strong>s<br />
actifs au sein du <strong>portefeuille</strong> doit s<strong>at</strong>isfaire à <strong>de</strong>s contraintes portant sur le capital économique, c’està-dire<br />
sur la somme qui doit être investie dans un actif sans risque pour perm<strong>et</strong>tre au <strong>gestion</strong>naire <strong>de</strong><br />
faire face à ses oblig<strong>at</strong>ions en dépit <strong>de</strong>s fluctu<strong>at</strong>ions <strong>de</strong> la valeur <strong>de</strong> marché du <strong>portefeuille</strong>. L’objectif<br />
primordial étant, bien évi<strong>de</strong>mment d’éviter la ruine. Deuxièmement, le <strong>portefeuille</strong> doit répondre à un<br />
objectif <strong>de</strong> rentabilité fixé. Il convient donc d’être capable <strong>de</strong> quantifier la propension du <strong>portefeuille</strong> à<br />
réaliser l’objectif qui a été défini. Plus clairement, l’objectif <strong>de</strong> rentabilité est généralement déterminé<br />
par la valeur du ren<strong>de</strong>ment moyen (ou espéré) à <strong>at</strong>teindre <strong>et</strong> dont on souhaite ne pas voir trop s’écarter<br />
les ren<strong>de</strong>ments réalisés. Il est alors nécessaire d’être capable d’estimer les fluctu<strong>at</strong>ions que le <strong>portefeuille</strong><br />
subit typiquement. Donc, la notion <strong>de</strong> risque présente au minimum un double sens, dans la mesure où<br />
elle concerne d’une part le capital économique <strong>et</strong> d’autre part l’amplitu<strong>de</strong> <strong>de</strong>s fluctu<strong>at</strong>ions <strong>st<strong>at</strong>istique</strong>s du<br />
ren<strong>de</strong>ment autour du ren<strong>de</strong>ment à <strong>at</strong>teindre.<br />
Moyennant cela, nous verrons ensuite comment m<strong>et</strong>tre à profit ces mesures <strong>de</strong> risques grands <strong>et</strong> extrêmes<br />
afin <strong>de</strong> construire <strong>de</strong>s <strong>portefeuille</strong>s optimaux supportant le mieux possible les gran<strong>de</strong>s fluctu<strong>at</strong>ions <strong>de</strong>s<br />
marchés financiers (chapitre 11), <strong>et</strong> en particulier comment minimiser l’impact <strong>de</strong>s grands co-mouvements<br />
entre actifs qui ne peuvent être totalement diversifiés (chapitre 12), puis comment s<strong>at</strong>isfaire la contrainte<br />
portant sur le capital économique au travers <strong>de</strong> mesures telles que la Value-<strong>at</strong>-Risk ou l’Expected-shortall<br />
(chapitre 13) <strong>et</strong> enfin, comment construire <strong>de</strong>s <strong>portefeuille</strong>s dont les ren<strong>de</strong>ments réalisés s’écartent le<br />
moins possible <strong>de</strong> l’objectif <strong>de</strong> rentabilité initialement fixé sous l’eff<strong>et</strong> <strong>de</strong>s gran<strong>de</strong>s fluctu<strong>at</strong>ions (chapitre<br />
14), ce qui nous amènera aussi à discuter <strong>de</strong>s conséquences du choix <strong>de</strong> <strong>portefeuille</strong>s minimisant l’impact<br />
<strong>de</strong>s risques grands <strong>et</strong> extrêmes sur l’équilibre <strong>de</strong>s marchés.<br />
Enfin, nous synthétiserons l’ensemble <strong>de</strong>s résult<strong>at</strong>s obtenus, discuterons <strong>de</strong> leurs conséquences vis-à-vis<br />
<strong>de</strong> la <strong>gestion</strong> <strong>de</strong>s risques extrêmes <strong>et</strong> conclurons par quelques perspectives <strong>de</strong> recherches futures.
18 Introduction
Première partie<br />
Etu<strong>de</strong> <strong>et</strong> modélis<strong>at</strong>ion <strong>de</strong>s propriétés <strong>de</strong>s<br />
rentabilités <strong>de</strong>s actifs financiers<br />
19
Chapitre 1<br />
Faits stylisés <strong>de</strong>s rentabilités boursières<br />
L’obj<strong>et</strong> <strong>de</strong> ce chapitre est <strong>de</strong> présenter les principaux faits stylisés <strong>de</strong>s ren<strong>de</strong>ments boursiers, d’en discuter<br />
la pertinence vis-à-vis <strong>de</strong>s grands principes <strong>de</strong> la finance ou <strong>de</strong> certains modèles phénoménologiques,<br />
mais aussi <strong>de</strong> rem<strong>et</strong>tre en cause ou généraliser certains résult<strong>at</strong>s considérés comme acquis <strong>et</strong> bien établis<br />
alors même que d’autres caractéristiques toutes aussi réalistes semblent <strong>de</strong>voir être mises à jour.<br />
Par fait stylisé nous désignons toute caractéristique commune <strong>de</strong>s séries temporelles issues <strong>de</strong>s marchés<br />
financiers. Ainsi donc, nous ne nous intéressons pas à une étu<strong>de</strong> événementielle <strong>de</strong>s marchés, étu<strong>de</strong><br />
où l’on cherche à expliquer le mouvement <strong>de</strong>s cours par l’arrivée <strong>de</strong> telle ou telle inform<strong>at</strong>ion, mais<br />
plutôt à une approche <strong>st<strong>at</strong>istique</strong> dont le but est d’i<strong>de</strong>ntifier <strong>et</strong> d’isoler les traits communs <strong>de</strong>s séries<br />
temporelles issues <strong>de</strong>s cours <strong>de</strong>s actifs financiers. En écrivant cela, nous ne cherchons pas à opposer ces<br />
<strong>de</strong>ux approches, qui n’ont rien d’antinomiques, <strong>et</strong> peuvent même se révéler complémentaires comme<br />
nous le verrons à la fin <strong>de</strong> ce chapitre.<br />
L’étu<strong>de</strong> <strong>de</strong>s faits stylisés repose sur les données fournies par les séries temporelles <strong>de</strong>s prix d’actifs<br />
financiers <strong>et</strong> <strong>de</strong> leurs ren<strong>de</strong>ments. Dans la suite, nous définirons par Pt le prix d’un actif à l’instant t <strong>et</strong><br />
Pt+1 le prix <strong>de</strong> c<strong>et</strong> actif à l’instant t + ∆t, où ∆t désigne l’unité <strong>de</strong> temps considérée, qui peut aussi bien<br />
être la minute que la journée ou le mois. Le ren<strong>de</strong>ment <strong>de</strong> l’actif à l’instant t sera quant à lui noté rt <strong>et</strong><br />
défini par<br />
rt = ln Pt<br />
Pt−1<br />
Pt − Pt−1<br />
, (1.1)<br />
Pt−1<br />
qui est en toute rigueur le ren<strong>de</strong>ment logarithmique mais diffère peu du ren<strong>de</strong>ment traditionnel tant que<br />
celui-ci <strong>de</strong>meure très p<strong>et</strong>it <strong>de</strong>vant un. La différence essentielle vient <strong>de</strong> ce que le ren<strong>de</strong>ment traditionnel<br />
est borné inférieurement par moins un, tandis que le ren<strong>de</strong>ment logarithmique est définie sur la droite<br />
réelle toute entière, ce qui constitue une distinction importante pour la détermin<strong>at</strong>ion <strong>de</strong> la distribution<br />
marginale <strong>de</strong>s ren<strong>de</strong>ments. De plus, dans ces conditions le logarithme du prix ln Pt suit un procesus dont<br />
les incréments sont donnés par les ren<strong>de</strong>ments rt.<br />
Une question cruciale se pose alors, à savoir celle <strong>de</strong> la st<strong>at</strong>ionnarité <strong>de</strong>s données sur lesquelles se basent<br />
les étu<strong>de</strong>s <strong>st<strong>at</strong>istique</strong>s. L’hypothèse <strong>de</strong> la st<strong>at</strong>ionnarité <strong>de</strong>s prix <strong>de</strong>s actifs est évi<strong>de</strong>mment à rej<strong>et</strong>er, puisque<br />
ces <strong>de</strong>rniers dérivent d’un processus intégré, comme le traduit l’équ<strong>at</strong>ion (1.1). En revanche, hypothèse <strong>de</strong><br />
st<strong>at</strong>ionnarité semble beaucoup plus acceptable pour les séries <strong>de</strong> ren<strong>de</strong>ments <strong>et</strong> est très souvent faite, car<br />
nécessaire à l’analyse <strong>st<strong>at</strong>istique</strong> <strong>de</strong>s données 1 , même s’il nous faut reconnaitre qu’elle est parfois suj<strong>et</strong>te<br />
à caution. En eff<strong>et</strong>, certaines étu<strong>de</strong>s auxquelles nous feront référence dans la suite ont été menées sur <strong>de</strong>s<br />
1 Dans le cas <strong>de</strong> séries non st<strong>at</strong>ionnaires, l’étu<strong>de</strong> reste dans une certaine mesure possible, mais fait appel à <strong>de</strong>s métho<strong>de</strong>s plus<br />
complexes <strong>et</strong> dont l’emploi n’est valable que dans le cadre <strong>de</strong> modèles plus ou moins spécifiques (Gouriéroux <strong>et</strong> Jasiak 2001)<br />
21
22 1. Faits stylisés <strong>de</strong>s rentabilités boursières<br />
séries longues d’un siècle. Or, sur <strong>de</strong> telles durées, l’évolution du contexte économique, réglementaire<br />
mais aussi l’introduction <strong>de</strong> nouveaux produits financiers semblent laisser à penser que les mo<strong>de</strong>s <strong>de</strong><br />
fonctionnement <strong>de</strong>s marchés se sont considérablement modifiés au fil du temps ce qui <strong>de</strong>vrait se refléter<br />
dans les cours par <strong>de</strong>s non-st<strong>at</strong>ionnarités. De plus, même sur <strong>de</strong> brèves pério<strong>de</strong>s, <strong>de</strong>s eff<strong>et</strong>s saisonniers<br />
apparaissent <strong>et</strong> l’on note par exemple <strong>de</strong>s comportements anormaux en début <strong>et</strong> fin <strong>de</strong> journée, <strong>de</strong> semaine<br />
ou d’année. Cependant, ces eff<strong>et</strong>s étant i<strong>de</strong>ntifiés, il est possible <strong>de</strong> les corriger. En résumé, la<br />
st<strong>at</strong>ionnarité pose potentiellement un problème important, mais on peut espérer en limiter les eff<strong>et</strong>s par<br />
la <strong>de</strong>ssaisonalis<strong>at</strong>ion <strong>de</strong>s données <strong>et</strong> la considér<strong>at</strong>ion d’échantillons <strong>de</strong> taille raisonnable (<strong>de</strong> cinq à dix<br />
ans tout au plus) sélectionnant <strong>de</strong>s séquences <strong>de</strong> marchés rel<strong>at</strong>ivement homogènes.<br />
Dans ce qui suit, nous allons essentiellement présenter les propriétés <strong>de</strong>s ren<strong>de</strong>ments journaliers ou intradays,<br />
<strong>et</strong> sauf indic<strong>at</strong>ion contraire, le terme ren<strong>de</strong>ment fera implicitement référence aux ren<strong>de</strong>ments calculés<br />
à une échelle <strong>de</strong> temps inférieure ou égale à la journée. Lorsque nous considérerons <strong>de</strong>s ren<strong>de</strong>ments<br />
mensuels par exemple, il en sera explicitement fait mention.<br />
Enfin, nous voulons souligner que le rappel <strong>de</strong>s faits stylisés que nous donnons dans ce chapitre ne<br />
prétend pas à l’exhaustivité, mais se concentre sur la <strong>de</strong>scription <strong>de</strong> ceux qui nous seront les plus utiles<br />
pour la suite <strong>de</strong> notre exposé, <strong>et</strong> nous renvoyons le lecteur aux nombreux articles <strong>de</strong> revue sur le suj<strong>et</strong><br />
comme par exemple Pagan (1996), Cont (2001) ou encore Engle <strong>et</strong> P<strong>at</strong>ton (2001) ainsi qu’aux ouvrages<br />
<strong>de</strong> Campbell, Lo <strong>et</strong> MacKinlay (1997) ou Gouriéroux <strong>et</strong> Jasiak (2001) pour une approche économétrique<br />
<strong>et</strong> Mantegna <strong>et</strong> Stanley (1999), Bouchaud <strong>et</strong> Potters (2000) ou Roehner (2001, 2002) pour une vision <strong>de</strong><br />
physiciens.<br />
1.1 Rappel <strong>de</strong>s faits stylisés<br />
Nous allons maintenant exposer les principaux faits stylisés en commençant par la distribution <strong>de</strong>s ren<strong>de</strong>ments<br />
dont il semble aujourd’hui certain qu’elle possè<strong>de</strong> <strong>de</strong>s “queues épaisses”, puis nous présenterons<br />
les propriétés <strong>de</strong> dépendances temporelles <strong>de</strong>s séries financières essentiellement caractérisées par l’absence<br />
<strong>de</strong> corrél<strong>at</strong>ion entre les ren<strong>de</strong>ments mais une forte persistance <strong>de</strong> la vol<strong>at</strong>ilité, après quoi nous<br />
terminerons par quelques propriétés complémentaires <strong>de</strong> ces séries financières.<br />
1.1.1 La distribution <strong>de</strong>s ren<strong>de</strong>ments<br />
La toute première particularité <strong>de</strong>s ren<strong>de</strong>ments financiers est <strong>de</strong> suivre <strong>de</strong>s lois dites “à queues épaisses”.<br />
Par loi à queues épaisses nous désignons l’ensemble <strong>de</strong>s lois régulièrement variables à l’infini, c’està-dire<br />
pour simplifier, les lois équivalentes à l’infini à une loi <strong>de</strong> puissance (pour une définition exacte<br />
<strong>de</strong> la notion <strong>de</strong> vari<strong>at</strong>ion régulière voir Bingham, Goldie <strong>et</strong> Teugel (1987)). Ce comportement est radicalement<br />
différent <strong>de</strong> celui généralement admis durant la première moitié du XXème siècle. En eff<strong>et</strong>,<br />
<strong>de</strong>puis <strong>de</strong>s travaux pionniers <strong>de</strong> Bachelier (1900) repris <strong>et</strong> étendus par Samuelson (1965, 1973), tout<br />
le mon<strong>de</strong> s’accordait sur le fait que les distributions <strong>de</strong> ren<strong>de</strong>ments suivaient <strong>de</strong>s distributions gaussiennes.<br />
Il faudra <strong>at</strong>tendre Man<strong>de</strong>lbrot (1963) <strong>et</strong> son étu<strong>de</strong> <strong>de</strong>s prix pr<strong>at</strong>iqués sur le marché du coton, puis<br />
Fama (1963, 1965a) pour clairement rej<strong>et</strong>er c<strong>et</strong>te hypothèse <strong>et</strong> en venir à considérer <strong>de</strong>s distributions en<br />
lois <strong>de</strong> puissance, <strong>et</strong> plus particulièrement les lois stables <strong>de</strong> Lévy, dont l’une <strong>de</strong>s caractéristiques est<br />
<strong>de</strong> ne pas adm<strong>et</strong>tre <strong>de</strong> second moment <strong>et</strong> donc d’avoir une variance infinie. Or, <strong>de</strong>s tests directs sur la<br />
plupart <strong>de</strong>s séries financières perm<strong>et</strong>tent clairement <strong>de</strong> conclure à l’existence <strong>de</strong> ce second moment. En<br />
eff<strong>et</strong>, par agrég<strong>at</strong>ion temporelle <strong>de</strong>s ren<strong>de</strong>ments, c’est-à-dire lorsque l’on passe <strong>de</strong>s ren<strong>de</strong>ments journaliers<br />
aux ren<strong>de</strong>ments mensuels ou à <strong>de</strong>s ren<strong>de</strong>ments calculés à <strong>de</strong>s échelles encore plus importantes, on<br />
note une (lente) convergence <strong>de</strong> la distribution vers la gaussienne (Bouchaud <strong>et</strong> Potters 2000, Campbell
1.1. Rappel <strong>de</strong>s faits stylisés 23<br />
<strong>et</strong> al. 1997, Mantegna <strong>et</strong> Stanley 1999). Donc, les lois stables <strong>de</strong> Lévy ne sauraient elles non plus convenir<br />
à la <strong>de</strong>scription <strong>de</strong>s distributions <strong>de</strong> rentabilités boursières, du moins dans leur globalité.<br />
Durant les années 80-90, <strong>de</strong> gigantesques bases <strong>de</strong> données contenant les cours <strong>de</strong>s actifs financiers sur<br />
<strong>de</strong> longues pério<strong>de</strong>s <strong>et</strong> enregistrées à <strong>de</strong> très hautes fréquences voient le jours. Dans le même temps,<br />
l’économétrie <strong>et</strong> plus particulièrement l’économétrie financière développe <strong>de</strong> nouveaux outils, si bien<br />
qu’à la fois sur le plan méthodologique que sur le plan <strong>de</strong> la quantité <strong>de</strong> données accessibles, une p<strong>et</strong>ite<br />
révolution se produit. C’est alors que l’on en vient à cerner <strong>de</strong> manière beaucoup plus précise le comportement<br />
asymptotique <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments <strong>de</strong>s actifs financiers. De nombreuses étu<strong>de</strong>s<br />
montrent qu’effectivement les distributions <strong>de</strong> ren<strong>de</strong>ments se comportent comme <strong>de</strong>s lois régulièrement<br />
variables dont l’exposant <strong>de</strong> queue est compris entre <strong>de</strong>ux <strong>et</strong> <strong>de</strong>mi <strong>et</strong> qu<strong>at</strong>re <strong>de</strong> sorte que les <strong>de</strong>ux premiers<br />
moments <strong>de</strong> ces distributions existent <strong>et</strong> peut-être aussi les troisième <strong>et</strong> qu<strong>at</strong>rième. Ces étu<strong>de</strong>s ont<br />
été menées sur diverses sortes d’actifs mais conduisent toutes à <strong>de</strong>s conclusions similaires. On pourra<br />
notamment consulter Longin (1996), Lux (1996), Pagan (1996), Gopikrishnan, Meyer, Amaral <strong>et</strong> Stanley<br />
(1998) pour <strong>de</strong>s étu<strong>de</strong>s concernant les marchés d’actions ainsi que Dacorogna, Müller, Pict<strong>et</strong> <strong>et</strong> <strong>de</strong> Vries<br />
(1992), <strong>de</strong> Vries (1994) ou encore Guillaume, Dacorogna, Davé, Müller, Olsen <strong>et</strong> Pict<strong>et</strong> (1997) pour ce<br />
qui est <strong>de</strong>s taux <strong>de</strong> change.<br />
Cependant, quelques voix se sont récemment élevées pour suggérer qu’il serait peut-être sage <strong>de</strong> tempérer<br />
l’enthousiasme général à l’égard <strong>de</strong>s lois <strong>de</strong> puissances. En eff<strong>et</strong>, selon Mantegna <strong>et</strong> Stanley (1995) les<br />
résult<strong>at</strong>s <strong>de</strong> Man<strong>de</strong>lbrot (1963) décrivent certes <strong>de</strong> façon correcte une gran<strong>de</strong> partie <strong>de</strong> la distribution<br />
<strong>de</strong>s ren<strong>de</strong>ments mais, <strong>de</strong> manière ultime, le comportement en loi stable <strong>de</strong> Lévy est erroné <strong>et</strong> doit être<br />
tronqué par une décroissance exponentielle. On sait d’ailleurs que ce type <strong>de</strong> distribution converge - par<br />
convolution - <strong>de</strong> manière extraordinairement lente vers la gaussienne (Mantegna <strong>et</strong> Stanley 1994), ce qui<br />
est conforme aux observ<strong>at</strong>ions empiriques. Dans le même esprit, Gouriéroux <strong>et</strong> Jasiak (1998) concluent<br />
que la <strong>de</strong>nsité <strong>de</strong> probabilité du titre Alc<strong>at</strong>el décroît plus vite que toute loi <strong>de</strong> puissance. Enfin, Laherrère<br />
<strong>et</strong> Sorn<strong>et</strong>te (1999) affirment <strong>de</strong> manière plus précise que les distributions dites exponentielles étirées<br />
semblent fournir une meilleure <strong>de</strong>scription <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments que les lois <strong>de</strong> puissance,<br />
non seulement dans les queues extrêmes mais aussi sur une large gamme <strong>de</strong> ren<strong>de</strong>ments. Aussi à <strong>de</strong>s fins<br />
<strong>de</strong> généralité, il semble judicieux d’étendre la notion <strong>de</strong> distributions à queues épaisses aux distributions<br />
sous-exponentielles, c’est-à-dire qui ne décroissent pas plus vite qu’une exponentielle à l’infini. Pour une<br />
définition rigoureuse ainsi qu’une synthèse <strong>de</strong>s propriétés <strong>de</strong> ces distributions, nous renvoyons le lecteur<br />
à Embrechts, Klüppelberg <strong>et</strong> Mikosh (1997) <strong>et</strong> Goldie <strong>et</strong> Klüppelberg (1998).<br />
Au vu du flou qui <strong>de</strong>meure quant au comportement exact <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments, un réexamen<br />
<strong>de</strong> leur comportement asymptotique semble nécessaire. En particulier, il parait indispensable <strong>de</strong> comparer<br />
le pouvoir <strong>de</strong>scriptif <strong>de</strong>s <strong>de</strong>ux familles <strong>de</strong> distributions que nous venons d’évoquer, <strong>et</strong> donc nous<br />
reviendrons en détail sur ce problème dans un prochain paragraphe. Nous délaissons temporairement<br />
ce point pour nous tourner maintenant vers le <strong>de</strong>uxième trait caractéristique <strong>de</strong>s séries financières, à savoir<br />
l’existence <strong>de</strong> structures <strong>de</strong> dépendances temporelles non triviales. Ceci est d’ailleurs un élément<br />
fondamental <strong>de</strong> la compréhension <strong>de</strong> la difficulté à déterminer précisément la distribution marginale <strong>de</strong>s<br />
ren<strong>de</strong>ments.<br />
1.1.2 Propriétés <strong>de</strong> dépendances temporelles<br />
Les propriétés <strong>de</strong> dépendances temporelles que nous allons résumer dans ce paragraphe dérivent toutes<br />
<strong>de</strong> l’un <strong>de</strong>s grands principes fondamentaux <strong>de</strong> la <strong>théorie</strong> financière, à savoir le principe <strong>de</strong> non arbitrage.<br />
Sans rentrer ici dans le détail, nous nous bornerons à dire <strong>de</strong> manière sommaire que ce principe postule<br />
l’impossibilité d’obtenir un gain certain sur les marchés financiers. La logique d’un tel postul<strong>at</strong> tient dans<br />
le fait que si un agent détecte une telle opportunité, il va immédi<strong>at</strong>ement en tirer profit <strong>et</strong> celle-ci va donc
24 1. Faits stylisés <strong>de</strong>s rentabilités boursières<br />
Vol<strong>at</strong>ility<br />
0.01<br />
0.009<br />
0.008<br />
0.007<br />
0.006<br />
0.005<br />
0.004<br />
0.003<br />
0.002<br />
0.001<br />
0<br />
Jan 70 Dec 85<br />
FIG. 1.1 – Vol<strong>at</strong>ilité réalisée (carrés <strong>de</strong>s ren<strong>de</strong>ments) <strong>de</strong> l’indice Standard & Poor’s 500 entre janvier<br />
1970 <strong>et</strong> décembre 1985.<br />
disparaître. Ainsi, hormis <strong>de</strong> manière fugitive, une telle situ<strong>at</strong>ion ne peut se produire. Cela implique en<br />
particulier que les ren<strong>de</strong>ments futurs ne peuvent être prédits.<br />
Une conséquence immédi<strong>at</strong>e <strong>et</strong> très communément observée est l’absence d’auto-corrél<strong>at</strong>ion entre les<br />
ren<strong>de</strong>ments 2 (Fama 1971, Pagan 1996, Gouriéroux <strong>et</strong> Jasiak 2001). C<strong>et</strong>te absence <strong>de</strong> corrél<strong>at</strong>ions temporelles<br />
est aisément justifiable par l’absence d’opportunité d’arbitrage (<strong>st<strong>at</strong>istique</strong>). En eff<strong>et</strong>, s’il existe<br />
<strong>de</strong>s corrél<strong>at</strong>ions signific<strong>at</strong>ives, il <strong>de</strong>vient possible <strong>de</strong> prédire les prix futurs, ce qui offre - du moins <strong>st<strong>at</strong>istique</strong>ment<br />
- une possibilité <strong>de</strong> gagner <strong>de</strong> l’argent à coup sûr. Ainsi, selon Man<strong>de</strong>lbrot (1971) l’absence<br />
d’opportunité d’arbitrage conduit à blanchir le spectre <strong>de</strong>s changements <strong>de</strong> prix <strong>et</strong> donc à faire disparaître<br />
les corrél<strong>at</strong>ions temporelles. Ceci a longtemps était l’un <strong>de</strong>s supports <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> l’efficience<br />
<strong>de</strong>s marchés financiers. En eff<strong>et</strong>, l’absence <strong>de</strong> corrél<strong>at</strong>ion justifie notamment l’hypothèse remontant à<br />
Bachelier (1900) selon laquelle les prix suivent une marche alé<strong>at</strong>oire (mouvement Brownien). Cela dit, à<br />
partir du moment où l’on adm<strong>et</strong> que le processus stochastique décrivant l’évolution au cours du temps <strong>de</strong>s<br />
ren<strong>de</strong>ments n’est pas gaussien, sa seule fonction d’auto-corrél<strong>at</strong>ion ne peut suffire à le définir. En particulier,<br />
l’absence <strong>de</strong> corrél<strong>at</strong>ion ne peut perm<strong>et</strong>tre <strong>de</strong> conclure que le prix (plus précisément le logarithme<br />
du prix) suit un processus à incréments indépendants, ce qui serait faux.<br />
Ceci se vérifie d’ailleurs simplement lorsque l’on observe une série <strong>de</strong> ren<strong>de</strong>ments. En eff<strong>et</strong>, il apparaît<br />
que les pério<strong>de</strong>s <strong>de</strong> gran<strong>de</strong> vol<strong>at</strong>ilité alternent avec les pério<strong>de</strong>s <strong>de</strong> vol<strong>at</strong>ilité plus faible montrant ainsi<br />
que la vol<strong>at</strong>ilité 3 présente un phénomène <strong>de</strong> persistance (voir figure 1.1). Ce phénomène est très similaire<br />
2 En toute rigueur, il existe <strong>de</strong>ux limites à c<strong>et</strong>te assertion. Premièrement, en <strong>de</strong>ssous d’un intervalle <strong>de</strong> temps <strong>de</strong> l’ordre <strong>de</strong><br />
quelques secon<strong>de</strong>s à quelques minutes, les eff<strong>et</strong>s <strong>de</strong> microstructure <strong>de</strong>s marchés <strong>de</strong>viennent importants <strong>et</strong> justifient que l’on<br />
observe l’existence <strong>de</strong> corrél<strong>at</strong>ions signific<strong>at</strong>ives (Campbell <strong>et</strong> al. 1997). Deuxièmement, pour les ren<strong>de</strong>ments calculés à <strong>de</strong>s<br />
échelles <strong>de</strong> temps <strong>de</strong> l’ordre du mois ou plus, il semble là aussi que <strong>de</strong>s corrél<strong>at</strong>ions signific<strong>at</strong>ives apparaissent. Cela peut<br />
s’expliquer par le fait que sur <strong>de</strong>s échelles <strong>de</strong> temps <strong>de</strong> c<strong>et</strong> ordre, le ren<strong>de</strong>ment d’un actif possè<strong>de</strong> un contenu économique<br />
beaucoup plus marqué qu’un ren<strong>de</strong>ment journalier essentiellement dominé par <strong>de</strong>s activités <strong>de</strong> (noise) trading pouvant être<br />
totalement découplées <strong>de</strong> toutes réalités économiques.<br />
3 Nous parlons ici <strong>de</strong> vol<strong>at</strong>ilité dans un sens rel<strong>at</strong>ivement vague, qui peut recouvrir à la fois l’écart type <strong>de</strong>s ren<strong>de</strong>ments<br />
ou simplement leur amplitu<strong>de</strong>. On peut toutefois noter qu’une telle persistance s’observe aussi sur la sknewness <strong>et</strong> la kurtosis<br />
(Jon<strong>de</strong>au <strong>et</strong> Rockinger 2000, par exemple).
1.1. Rappel <strong>de</strong>s faits stylisés 25<br />
au phénomène d’intermittence observé en turbulence (Frisch 1995). Il caractérise clairement le fait qu’il<br />
existe une certaine dépendance dans les séries financières : la séquence <strong>de</strong>s ren<strong>de</strong>ments n’apparaît pas <strong>de</strong><br />
manière indépendante. En fait, on const<strong>at</strong>e que la valeur absolue <strong>de</strong>s ren<strong>de</strong>ments ou <strong>de</strong> leurs carrés sont<br />
fortement corrélés <strong>et</strong> que c<strong>et</strong>te corrél<strong>at</strong>ion persiste : la fonction d’auto-corrél<strong>at</strong>ion <strong>de</strong>s valeurs absolues<br />
<strong>de</strong>s ren<strong>de</strong>ments décroît typiquement comme une loi <strong>de</strong> puissance dont l’exposant est compris entre 0.2<br />
<strong>et</strong> 0.4 (Cont, Potters <strong>et</strong> Bouchaud 1997, Liu, Cizeau, Meyer, Peng <strong>et</strong> Stanley 1997). Une fois encore,<br />
cela est conforme au principe <strong>de</strong> non arbitrage, car la vol<strong>at</strong>ilité n’étant pas une gran<strong>de</strong>ur signée, elle<br />
ne perm<strong>et</strong> en rien <strong>de</strong> déduire les ren<strong>de</strong>ments futurs. De manière générale, toute fonction <strong>de</strong> la vol<strong>at</strong>ilité<br />
(ou <strong>de</strong> l’amplitu<strong>de</strong> du ren<strong>de</strong>ment) est autorisée à présenter <strong>de</strong>s corrél<strong>at</strong>ions temporelles signific<strong>at</strong>ives.<br />
Pour quelques exemples voir notamment Cont (2001) mais aussi Muzy, Delour <strong>et</strong> Bacry (2000), où il est<br />
montré que la corrél<strong>at</strong>ion du logarithme <strong>de</strong> la valeur absolue <strong>de</strong>s ren<strong>de</strong>ments décroît comme le logarithme<br />
du décalage temporel entre les ren<strong>de</strong>ments considérés <strong>et</strong> possè<strong>de</strong> donc une mémoire longue.<br />
Ce phénomène <strong>de</strong> persistance (ou mémoire longue) est caractéristique <strong>de</strong> la vol<strong>at</strong>ilité mais sa mise en<br />
évi<strong>de</strong>nce est quelques fois délic<strong>at</strong>e <strong>et</strong> suj<strong>et</strong>te à caution comme l’ont montré Granger <strong>et</strong> Teräsvirta (1999),<br />
qui ont construit un processus auto-régressif du premier ordre non-linéaire passant les tests standards <strong>de</strong><br />
mémoire longue ou An<strong>de</strong>rsson, Eklund <strong>et</strong> Lyhagen (1999) qui ont prouvé le contraire, c’est-à-dire qu’un<br />
processus à mémoire longue pouvait passer les tests <strong>de</strong> linéarité.<br />
Pour en terminer avec les exemples sur les corrél<strong>at</strong>ions temporelles, il convient <strong>de</strong> citer l’eff<strong>et</strong> <strong>de</strong> levier<br />
mis en évi<strong>de</strong>nce par Black (1976) puis Christie (1982) <strong>et</strong> Bouchaud, M<strong>at</strong>acz <strong>et</strong> Potters (2001) notamment,<br />
<strong>et</strong> que l’on rencontre pour les ren<strong>de</strong>ments d’actions <strong>et</strong> les taux d’intérêt mais pas les taux <strong>de</strong> change.<br />
C<strong>et</strong> eff<strong>et</strong> <strong>de</strong> levier se traduit par une corrél<strong>at</strong>ion nég<strong>at</strong>ive entre la vol<strong>at</strong>ilité à un instant donné <strong>et</strong> les<br />
ren<strong>de</strong>ments antérieurs, ce qui signifie que les mouvements <strong>de</strong> prix à la baisse ont tendance à entraîner<br />
une augment<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité. Il semble d’ailleurs que c<strong>et</strong> eff<strong>et</strong> <strong>de</strong> levier soit à l’origine <strong>de</strong> la casca<strong>de</strong><br />
causale observée par Arnéodo, Muzy <strong>et</strong> Sorn<strong>et</strong>te (1998) sur les marchés d’actions 4 .<br />
En résumé, <strong>de</strong> par la contrainte <strong>de</strong> non arbitrage, il est impossible d’observer <strong>de</strong>s corrél<strong>at</strong>ions signific<strong>at</strong>ives<br />
entre <strong>de</strong>s fonctions <strong>de</strong>s variables passées <strong>et</strong> les ren<strong>de</strong>ments futurs qui perm<strong>et</strong>traient <strong>de</strong> prédire ces<br />
<strong>de</strong>rniers. En revanche, il est tout à fait possible d’observer - <strong>et</strong> c’est effectivement le cas - d’importantes<br />
corrél<strong>at</strong>ions entre <strong>de</strong>s fonctions <strong>de</strong>s variables passées <strong>et</strong> la vol<strong>at</strong>ilité future, ceci ne perm<strong>et</strong>tant pas <strong>de</strong><br />
réaliser <strong>de</strong> prédiction sur les ren<strong>de</strong>ments.<br />
Enfin, outre l’existence <strong>de</strong> fortes dépendances entre vol<strong>at</strong>ilité future d’une part, <strong>et</strong> vol<strong>at</strong>ilité <strong>et</strong> ren<strong>de</strong>ments<br />
passés d’autre part, il convient <strong>de</strong> remarquer que globalement, celle-ci a tendance à r<strong>et</strong>ourner à un niveau<br />
moyen, que l’on peut interpréter comme son niveau “normal”. Ce phénomène <strong>de</strong> r<strong>et</strong>our à la moyenne<br />
est bien observé lorsque l’on s’intéresse à la vol<strong>at</strong>ilité prédite par la plupart <strong>de</strong>s modèles usuels : lorsque<br />
l’horizon <strong>de</strong> prévision augmente, on const<strong>at</strong>e que la vol<strong>at</strong>ilité tend vers une même valeur moyenne (voir<br />
Engle <strong>et</strong> P<strong>at</strong>ton (2001) par exemple). Il est alors très intéressant d’étudier le mo<strong>de</strong> <strong>de</strong> relax<strong>at</strong>ion <strong>de</strong> la<br />
vol<strong>at</strong>ilité vers son niveau moyen, ce qui a été réalisé par Sorn<strong>et</strong>te, Malevergne <strong>et</strong> Muzy (2002) <strong>et</strong> dont<br />
nous parlerons un peu plus en détail dans un paragraphe ultérieur.<br />
1.1.3 Autres faits stylisés<br />
Les faits stylisés précé<strong>de</strong>mment évoqués sont les faits stylisés principaux universellement admis. Ce sont<br />
ceux dont doit au moins rendre compte tout modèle <strong>de</strong> cours réaliste. Cela peut sembler peu, mais en fait,<br />
il est déjà assez difficile <strong>de</strong> trouver <strong>de</strong> manière ad hoc <strong>de</strong>s processus s<strong>at</strong>isfaisant ces quelques restrictions.<br />
Nous en donnerons quelques exemples dans la suite <strong>de</strong> ce chapitre. A coté <strong>de</strong> ces grands faits stylisés,<br />
existent plusieurs autres caractéristiques importantes dont notamment la dissymétrie gains/pertes pour<br />
4 C<strong>et</strong>te interprét<strong>at</strong>ion nous a été suggérée par J.F. Muzy lors d’une convers<strong>at</strong>ion privée.
26 1. Faits stylisés <strong>de</strong>s rentabilités boursières<br />
Hang-Seng<br />
10000<br />
1000<br />
↓<br />
↓<br />
↓<br />
↓<br />
100<br />
70 75 80 85 90 95 100<br />
FIG. 1.2 – Indice Hang Seng (bourse <strong>de</strong> Hong Kong) entre janvier 1970 <strong>et</strong> Decembre 2000.<br />
les actions, les accélér<strong>at</strong>ions super-exponentielles <strong>de</strong>s cours lors <strong>de</strong>s phases <strong>de</strong> bulles spécul<strong>at</strong>ives ou<br />
encore la multifractalité <strong>et</strong> la corrél<strong>at</strong>ion entre vol<strong>at</strong>ilité <strong>et</strong> volume <strong>de</strong> transaction, ce <strong>de</strong>rnier point étant<br />
intuitivement évi<strong>de</strong>nt (Gopikrishnan, Plerou, Gabiax <strong>et</strong> Stanley 2000).<br />
Concernant la dissymétrie gains/pertes, il a été clairement établi par Johansen <strong>et</strong> Sorn<strong>et</strong>te (1998) <strong>et</strong> Johansen<br />
<strong>et</strong> Sorn<strong>et</strong>te (2002) que les marchés d’actions sont suj<strong>et</strong>s à <strong>de</strong>s pertes cumulées <strong>de</strong> très gran<strong>de</strong>s<br />
amplitu<strong>de</strong>s, alors que l’on n’observe pas systém<strong>at</strong>iquement <strong>de</strong> phénomènes <strong>de</strong> telle ampleur pour les<br />
gains. De plus, ce phénomène ne semble pas toucher les marchés <strong>de</strong> change, où une plus gran<strong>de</strong> symétrie<br />
parait <strong>de</strong> mise. Ceci est en fait n<strong>at</strong>urel, car la hausse ou la baisse d’un taux <strong>de</strong> change n’est que rel<strong>at</strong>ive à<br />
la position que l’on adopte : une hausse du taux euro/dollar par exemple correspond évi<strong>de</strong>mment à une<br />
baisse du taux dollar/euro.<br />
Pour ce qui est <strong>de</strong> la croissance super-exponentielle du prix <strong>de</strong>s actifs, celle-ci est illustrée sur la figure<br />
1.2. On y observe que le prix (représenté en échelle logarithmique) croit en moyenne <strong>de</strong> manière linéaire<br />
en fonction du temps, ce qui caractérise une croissance exponentielle <strong>et</strong> donc un taux <strong>de</strong> croissance<br />
constant. Mais, un examen plus approfondi montre qu’en fait se succè<strong>de</strong>nt une série <strong>de</strong> phases <strong>de</strong> croissance<br />
accélérée - <strong>et</strong> donc plus rapi<strong>de</strong>s que la moyenne - interrompues par <strong>de</strong> fortes corrections du prix.<br />
Ce phénomène a été qualifié par Roehner <strong>et</strong> Sorn<strong>et</strong>te (1998) <strong>de</strong> “sharp peak, fl<strong>at</strong> trough” <strong>et</strong> semble tirer<br />
son origine <strong>de</strong>s phénomènes d’imit<strong>at</strong>ion <strong>et</strong> eff<strong>et</strong>s <strong>de</strong> foule observés sur les marchés. Nous reviendrons<br />
beaucoup plus en détail sur ce point précis aux chapitres 5 <strong>et</strong> 6.<br />
Enfin le <strong>de</strong>rnier point que nous abor<strong>de</strong>rons concernera la quantific<strong>at</strong>ion <strong>de</strong> la régularité <strong>de</strong> la trajectoire<br />
<strong>de</strong>s prix d’actifs financiers. Arnéodo <strong>et</strong> al. (1998) <strong>et</strong> Fisher, Calv<strong>et</strong> <strong>et</strong> Man<strong>de</strong>lbrot (1998) ont conclu au<br />
caractère multifractal <strong>de</strong> ces trajectoires. En eff<strong>et</strong>, ils ont montré que quel que soit l’actif considéré, son<br />
spectre multifractal est compris strictement entre zéro <strong>et</strong> un, ce qui assure que sa trajectoire est presque<br />
partout continue mais non dérivable, <strong>et</strong> a la forme d’une parabole dont le maximum est voisin <strong>de</strong> 0.6 5 .<br />
Par comparaison, le mouvement brownien étant auto similaire - <strong>et</strong> donc (mono) fractal - son spectre vaut<br />
simplement un <strong>de</strong>mi. Ceci conduit à <strong>de</strong> très fortes contraintes sur les processus perm<strong>et</strong>tant d’obtenir ce<br />
type <strong>de</strong> trajectoires, mais un bémol s’impose quant aux résult<strong>at</strong>s <strong>de</strong> ces étu<strong>de</strong>s car <strong>de</strong> nombreux auteurs,<br />
5 Il est intéressant <strong>de</strong> remarquer que ces observ<strong>at</strong>ions conduisent à rej<strong>et</strong>er un grand nombre <strong>de</strong> modèles en temps continu tels<br />
que les processus <strong>de</strong> diffusion, les processus <strong>de</strong> Lévy ou les processus à saut.<br />
D<strong>at</strong>e<br />
↓<br />
↓<br />
↓<br />
↓
1.2. De la difficulté <strong>de</strong> représenter la distribution <strong>de</strong>s ren<strong>de</strong>ments 27<br />
dont Veneziano, Moglen <strong>et</strong> Bras (1995), Avnir, Biham, Lidar <strong>et</strong> Malcai (1998), Bouchaud, Potters <strong>et</strong><br />
Meyer (2000), LeBaron (2001) ou encore Sorn<strong>et</strong>te <strong>et</strong> An<strong>de</strong>rsen (2002), ont montré qu’il était très facile<br />
<strong>de</strong> m<strong>et</strong>tre en évi<strong>de</strong>nce un caractère multifractal pour <strong>de</strong>s processus dont il est connu qu’ils ne possè<strong>de</strong>nt<br />
aucune propriété <strong>de</strong> fractalité, simplement du fait <strong>de</strong> la taille finie <strong>de</strong>s séries temporelles considérées ou<br />
d’autres mécanismes faisant intervenir <strong>de</strong>s non-linéarités.<br />
1.2 De la difficulté <strong>de</strong> représenter la distribution <strong>de</strong>s ren<strong>de</strong>ments<br />
Nous venons <strong>de</strong> voir dans la première partie du paragraphe précé<strong>de</strong>nt que la distribution <strong>de</strong>s ren<strong>de</strong>ments<br />
semblait pouvoir être décrite par <strong>de</strong>s distributions régulièrement variables d’indice <strong>de</strong> queue <strong>de</strong> l’ordre<br />
<strong>de</strong> trois ou qu<strong>at</strong>re. Cependant, une autre hypothèse a récemment vu le jour, hypothèse selon laquelle les<br />
distributions <strong>de</strong> ren<strong>de</strong>ments suivraient <strong>de</strong>s lois exponentielles étirées. Or, si ces <strong>de</strong>ux types <strong>de</strong> distributions<br />
possè<strong>de</strong>nt beaucoup <strong>de</strong> points communs, elles présentent néanmoins une différence majeure dont<br />
l’impact théorique est très important.<br />
Pour ce qui est <strong>de</strong>s points communs, ces <strong>de</strong>ux familles <strong>de</strong> distributions appartiennent à la classe <strong>de</strong>s<br />
distributions sous-exponentielles <strong>et</strong> sont donc à même <strong>de</strong> produire <strong>de</strong>s événements extrêmes avec une<br />
probabilité élevée. En particulier, pour c<strong>et</strong>te classe <strong>de</strong> variables alé<strong>at</strong>oires, les gran<strong>de</strong>s dévi<strong>at</strong>ions <strong>de</strong> la<br />
somme d’un grand nombre <strong>de</strong> ces variables sont dominées par les gran<strong>de</strong>s dévi<strong>at</strong>ions d’une seule d’entre<br />
elles : la somme est gran<strong>de</strong> si l’une <strong>de</strong>s variables est gran<strong>de</strong> (Embrechts <strong>et</strong> al. 1997, Sorn<strong>et</strong>te 2000), ce<br />
qui contraste énormément avec les distributions dites super-exponentielles où chaque variable contribue<br />
<strong>de</strong> manière signific<strong>at</strong>ive (<strong>et</strong> globalement équivalente) à la somme (Frisch <strong>et</strong> Sorn<strong>et</strong>te 1997).<br />
Cela dit, il existe une différence majeure : les distributions sous-exponentielles adm<strong>et</strong>tent <strong>de</strong>s moments<br />
<strong>de</strong> tous ordres alors que les distributions régulièrement variables n’adm<strong>et</strong>tent <strong>de</strong> moments que d’ordre<br />
inférieur à leur indice <strong>de</strong> queue. C<strong>et</strong>te remarque montre combien il est crucial <strong>de</strong> pouvoir trancher entre<br />
ces <strong>de</strong>ux types <strong>de</strong> distributions, car nous verrons en partie III que ces moments vont jouer un rôle essentiel<br />
dans la représent<strong>at</strong>ion du comportement qu’adoptent les agents économiques face au risque ainsi que<br />
dans la modélis<strong>at</strong>ion <strong>de</strong>s décisions qu’ils prennent. Il est donc absolument fondamental d’être sûr <strong>de</strong> leur<br />
existence.<br />
L’incertitu<strong>de</strong> qui semble naître quant à la n<strong>at</strong>ure exacte <strong>de</strong> la distribution <strong>de</strong>s ren<strong>de</strong>ments dans les<br />
extrêmes <strong>et</strong> les difficultés à la déterminer précisément est à notre avis en gran<strong>de</strong> partie liée aux dépendances<br />
temporelles présentent dans les séries financières. En eff<strong>et</strong>, lorsque les observ<strong>at</strong>ions sont supposées<br />
indépendantes <strong>et</strong> i<strong>de</strong>ntiquement distribuées, la <strong>théorie</strong> montre que les estim<strong>at</strong>eurs usuels - tel l’estim<strong>at</strong>eur<br />
<strong>de</strong> Hill (1975), qui perm<strong>et</strong> <strong>de</strong> déterminer l’exposant <strong>de</strong> queue d’une distribution régulièrement variable,<br />
mais aussi l’estim<strong>at</strong>eur <strong>de</strong> Pickands (1975) - sont asymptotiquement consistants <strong>et</strong> normalement distribués<br />
(Embrechts <strong>et</strong> al. 1997), ce qui perm<strong>et</strong>, par exemple, d’estimer <strong>de</strong> façon précise l’incertitu<strong>de</strong><br />
sur la mesure d’un indice <strong>de</strong> queue <strong>et</strong> éventuellement d’effectuer <strong>de</strong>s tests d’égalité entre les indices <strong>de</strong><br />
queues <strong>de</strong> différents actifs ou entre la queue positive <strong>et</strong> nég<strong>at</strong>ive d’un même actif afin <strong>de</strong> m<strong>et</strong>tre, ou non,<br />
à jour d’éventuelles dissymétries (Jon<strong>de</strong>au <strong>et</strong> Rockinger 2001).<br />
Mais, lorsque les données présentent une certaine dépendance temporelle, on ne peut plus guère espérer<br />
que la consistance asymptotique <strong>de</strong> ces estim<strong>at</strong>eurs (Rootzèn, Leadb<strong>et</strong>ter <strong>et</strong> <strong>de</strong> Haan 1998), l’incertitu<strong>de</strong><br />
<strong>de</strong> la mesure étant quant à elle beaucoup plus importante que celle fournit sous hypothèse d’asymptotique<br />
normalité. En eff<strong>et</strong>, Kearns <strong>et</strong> Pagan (1997) ont montré que pour <strong>de</strong>s processus <strong>de</strong> type (G)ARCH,<br />
qui perm<strong>et</strong>tent <strong>de</strong> rendre raisonnablement compte <strong>de</strong>s dépendances <strong>de</strong>s séries financières (voir paragraphe<br />
suivant), la dévi<strong>at</strong>ion standard <strong>de</strong> l’estim<strong>at</strong>eur <strong>de</strong> Hill peut être sept à huit fois plus gran<strong>de</strong> que<br />
celle estimée pour <strong>de</strong>s séries dont les données sont in<strong>de</strong>ntiquement <strong>et</strong> indépendamment distribuées. Ces
28 1. Faits stylisés <strong>de</strong>s rentabilités boursières<br />
résult<strong>at</strong>s sont encore pires pour l’estim<strong>at</strong>eur <strong>de</strong> Pickand, dont on sait qu’il joue un rôle important dans<br />
la détermin<strong>at</strong>ion empirique du domaine d’<strong>at</strong>traction <strong>de</strong>s lois extrêmes auquel appartient une distribution<br />
donnée.<br />
C<strong>et</strong>te <strong>de</strong>rnière remarque est cruciale <strong>et</strong> conduit à rem<strong>et</strong>tre potentiellement en cause bon nombre d’étu<strong>de</strong>s<br />
sur le comportement asymptotique <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments. En eff<strong>et</strong>, négliger les dépendances<br />
temporelles revient à minorer l’incertitu<strong>de</strong> <strong>de</strong>s estim<strong>at</strong>eurs comme dans le cas <strong>de</strong>s étu<strong>de</strong>s se fondant sur la<br />
<strong>théorie</strong> <strong>de</strong>s valeurs extrêmes menées par Longin (1996) ou Lux (1997) notamment, où l’on peut dire que<br />
l’hypothèse selon laquelle les distributions <strong>de</strong> ren<strong>de</strong>ments sont régulièrement variables à l’infini semble<br />
avoir été un peu rapi<strong>de</strong>ment <strong>et</strong> peut-être inconsidérément acceptée.<br />
De plus, sur le plan théorique, il a été démontré que certaines classes <strong>de</strong> modèles supportant l’hypothèse<br />
<strong>de</strong> distributions régulièrement variables <strong>de</strong>vaient être abandonnées. En eff<strong>et</strong> Lux <strong>et</strong> Sorn<strong>et</strong>te (2002) <strong>et</strong><br />
Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2001a) ont prouvé que les modèles phénoménologiques à bruit multiplic<strong>at</strong>if du<br />
type Blanchard (1979) <strong>et</strong> Blanchard <strong>et</strong> W<strong>at</strong>son (1982), <strong>et</strong> qui perm<strong>et</strong>tent <strong>de</strong> justifier la notion <strong>de</strong> bulles<br />
r<strong>at</strong>ionnelles (Coll<strong>et</strong>az <strong>et</strong> Gourlaouen 1989, Broze, Gouriéroux <strong>et</strong> Szafarz 1990), conduisent, <strong>de</strong> par la<br />
condition <strong>de</strong> non arbitrage, à <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ment régulièrement variables dont l’exposant <strong>de</strong><br />
queue est nécessairement inférieur à un (voir chapitre 2), ce que toutes les étu<strong>de</strong>s empiriques menées<br />
jusqu’à ce jour ont réfuté. Enfin, parmi les divers modèles <strong>de</strong> prix d’actifs passés en revue dans Sorn<strong>et</strong>te<br />
<strong>et</strong> Malevergne (2001), il ressort notamment que les modèles <strong>de</strong> Johansen, Sorn<strong>et</strong>te <strong>et</strong> Ledoit (1999) <strong>et</strong><br />
Johansen, Ledoit <strong>et</strong> Sorn<strong>et</strong>te (2000) - dans lesquels est introduit un taux <strong>de</strong> krach qui en assure la st<strong>at</strong>ionnarité<br />
- adm<strong>et</strong>tent <strong>de</strong>s dynamiques <strong>de</strong> prix comp<strong>at</strong>ibles aussi bien avec <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments<br />
régulièrement variables qu’exponentielles étirées.<br />
Au vu <strong>de</strong> ces incertitu<strong>de</strong>s aussi bien empiriques que théoriques, nous avons jugé nécessaire <strong>de</strong> mener<br />
une étu<strong>de</strong> compar<strong>at</strong>ive du pouvoir <strong>de</strong>scriptif <strong>de</strong>s <strong>de</strong>ux représent<strong>at</strong>ions possibles <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments<br />
(voir Malevergne, Pisarenko <strong>et</strong> Sorn<strong>et</strong>te (2002) présenté au chapitre 3). Il ressort <strong>de</strong> c<strong>et</strong>te étu<strong>de</strong><br />
que, d’un point <strong>de</strong> vue <strong>st<strong>at</strong>istique</strong>, les exponentielles étirées ren<strong>de</strong>nt au moins aussi bien compte <strong>de</strong>s distributions<br />
<strong>de</strong> ren<strong>de</strong>ments que les lois <strong>de</strong> puissances <strong>et</strong> autres lois à vari<strong>at</strong>ions régulières, mais que du fait<br />
<strong>de</strong> la dépendance temporelle, il ne nous est pas possible <strong>de</strong> dire que l’une est réellement meilleure que<br />
l’autre. Nous <strong>de</strong>vons donc nous borner à dire que les exponentielles étirées fournissent une altern<strong>at</strong>ive<br />
parfaitement crédible aux distributions régulièrement variables, mais nous sommes malheureusement<br />
incapable <strong>de</strong> trancher la question <strong>de</strong> l’existence ou non <strong>de</strong>s moments au-<strong>de</strong>là d’un certain ordre.<br />
1.3 Modélis<strong>at</strong>ion <strong>de</strong>s propriétés <strong>de</strong> dépendance <strong>de</strong>s ren<strong>de</strong>ments<br />
Le paragraphe précé<strong>de</strong>nt vient <strong>de</strong> nous perm<strong>et</strong>tre <strong>de</strong> comprendre que d’un point <strong>de</strong> vue <strong>st<strong>at</strong>istique</strong>, il n’est<br />
pas vraiment justifié <strong>de</strong> privilégier les distributions régulièrement variables au profit <strong>de</strong>s distributions exponentielles<br />
étirées ou vice-versa. En fait, comme nous allons le voir maintenant, l’étu<strong>de</strong> <strong>de</strong> l’évolution<br />
dynamique <strong>de</strong>s cours fournit un moyen <strong>de</strong> spécifier un peu mieux ce choix. En eff<strong>et</strong>, la prise en compte<br />
<strong>de</strong>s faits stylisés concernant la dépendance temporelle <strong>de</strong> la vol<strong>at</strong>ilité impose <strong>de</strong>s contraintes suffisamment<br />
fortes pour perm<strong>et</strong>tre <strong>de</strong> définir <strong>de</strong> façon assez s<strong>at</strong>isfaisante le processus stochastique décrivant<br />
l’évolution <strong>de</strong>s cours, ce qui, en r<strong>et</strong>our, fixe la distribution marginale <strong>de</strong>s ren<strong>de</strong>ments comp<strong>at</strong>ible avec le<br />
processus spécifié. L’intérêt <strong>de</strong> trouver un processus stochastique est donc double : décrire la dynamique<br />
<strong>de</strong>s cours en tenant compte <strong>de</strong> la dépendance exhibée par la vol<strong>at</strong>ilité <strong>et</strong> fournir <strong>de</strong>s jalons quant à la<br />
distribution marginale <strong>de</strong>s ren<strong>de</strong>ments.<br />
Les premières <strong>de</strong>scriptions <strong>de</strong>s processus <strong>de</strong> prix en terme <strong>de</strong> marches alé<strong>at</strong>oires à incréments indépendants<br />
ou la modélis<strong>at</strong>ion <strong>de</strong> la dynamique <strong>de</strong>s ren<strong>de</strong>ments à l’ai<strong>de</strong> <strong>de</strong> processus ARMA ont rapi<strong>de</strong>ment
1.3. Modélis<strong>at</strong>ion <strong>de</strong>s propriétés <strong>de</strong> dépendance <strong>de</strong>s ren<strong>de</strong>ments 29<br />
été abandonnées en raison <strong>de</strong> leur incapacité à rendre compte <strong>de</strong> la mémoire longue <strong>de</strong> la vol<strong>at</strong>ilité <strong>et</strong><br />
notamment <strong>de</strong>s bouffées <strong>de</strong> vol<strong>at</strong>ilité. En eff<strong>et</strong>, par construction, les marches alé<strong>at</strong>oires à incréments<br />
indépendants ne présentent aucune dépendance <strong>de</strong> vol<strong>at</strong>ilité tandis que la fonction d’auto-corrél<strong>at</strong>ion <strong>de</strong>s<br />
processus ARMA décroît <strong>de</strong> manière exponentielle (voir Gouriéroux <strong>et</strong> Jasiak (2001) par exemple) <strong>et</strong><br />
non <strong>de</strong> manière algébrique. Une altern<strong>at</strong>ive aux processus ARMA est fournie par les processus ARIMA<br />
fractionnaires (Granger <strong>et</strong> Joyeux 1980, Hosking 1981) qui présentent <strong>de</strong>s corrél<strong>at</strong>ions à longues portées<br />
mais dont l’usage n’est pas bien adapté à la modélis<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité.<br />
La première réelle ébauche <strong>de</strong> solution a en fait été proposée par Engle (1982) <strong>et</strong> les modèles autorégressifs<br />
conditionnellement hétéroskédastiques (ARCH) qui furent ensuite généralisés par Bollerslev<br />
(1986) avec les modèles GARCH. Dans c<strong>et</strong>te famille <strong>de</strong> modèles, les ren<strong>de</strong>ments sont décomposés<br />
comme un produit :<br />
rt = σt · εt, (1.2)<br />
où la vol<strong>at</strong>ilité σt suit un processus auto-régressif <strong>et</strong> εt est un bruit blanc indépendant <strong>de</strong> σt, ce qui assure<br />
l’indépendance <strong>de</strong>s ren<strong>de</strong>ments à <strong>de</strong>s d<strong>at</strong>es successives. Dans un modèle ARCH, la vol<strong>at</strong>ilité présente σt<br />
ne dépend que <strong>de</strong>s réalis<strong>at</strong>ions passées <strong>de</strong>s ren<strong>de</strong>ments rt−1, rt−2, · · · alors que pour un modèle GARCH,<br />
elle est aussi fonction <strong>de</strong>s vol<strong>at</strong>ilités passées σt−1, σt−2, · · ·. En pr<strong>at</strong>ique, le modèle GARCH possè<strong>de</strong> un<br />
indéniable avantage sur le modèle ARCH car il est beaucoup plus parcimonieux : prendre en compte la<br />
<strong>de</strong>rnière vol<strong>at</strong>ilité <strong>et</strong> les <strong>de</strong>ux <strong>de</strong>rniers ren<strong>de</strong>ments (<strong>et</strong> donc trois paramètres) est en général suffisant alors<br />
que pour un modèle ARCH, il n’est pas rare <strong>de</strong> <strong>de</strong>voir considérer les dix ou quinze <strong>de</strong>rniers ren<strong>de</strong>ments<br />
réalisés.<br />
Du point <strong>de</strong> vue <strong>de</strong> l’adéqu<strong>at</strong>ion vis-à-vis <strong>de</strong>s faits stylisés, ces processus perm<strong>et</strong>tent <strong>de</strong> rendre compte<br />
<strong>de</strong> la lente décroissance <strong>de</strong> l’auto-corrél<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité, ce qui était leur premier objectif <strong>et</strong> du r<strong>et</strong>our<br />
à la moyenne. De plus, on peut monter facilement (voir Embrechts <strong>et</strong> al. (1997) par exemple) à l’ai<strong>de</strong> <strong>de</strong><br />
la <strong>théorie</strong> <strong>de</strong>s équ<strong>at</strong>ions <strong>de</strong> renouvellement stochastique (Kesten 1973, Goldie 1991) que la distribution<br />
marginale <strong>de</strong>s ren<strong>de</strong>ments est à vari<strong>at</strong>ion régulière.<br />
Une <strong>de</strong>s limites <strong>de</strong> c<strong>et</strong>te modélis<strong>at</strong>ion est <strong>de</strong> ne pas rendre compte convenablement <strong>de</strong> la lente décroissance<br />
<strong>de</strong> la corrél<strong>at</strong>ion du logarithme <strong>de</strong> la vol<strong>at</strong>ilité, <strong>de</strong> la multifractalité 6 ni <strong>de</strong> l’eff<strong>et</strong> <strong>de</strong> levier. Cependant,<br />
ce <strong>de</strong>rnier point peut être corrigé en considérant le modèle EGARCH <strong>de</strong> Nelson (1991) qui n’est<br />
rien d’autre qu’un processus GARCH sur le logarithme <strong>de</strong> la vol<strong>at</strong>ilité <strong>et</strong> non la vol<strong>at</strong>ilité elle-même ou<br />
encore les modèles TARCH <strong>de</strong> Glosten, Jagannanthan <strong>et</strong> Runkle (1993) <strong>et</strong> Zakoian (1994) qui sont <strong>de</strong><br />
modèles GARCH à seuil. Nous n’irons pas plus avant dans la <strong>de</strong>scription <strong>de</strong> la vaste famille <strong>de</strong>s modèles<br />
ARCH <strong>et</strong> <strong>de</strong> leur généralis<strong>at</strong>ion <strong>et</strong> nous renvoyons le lecteur aux nombreux articles <strong>de</strong> revues <strong>et</strong> livres<br />
sur le suj<strong>et</strong> dont notamment Bollerslev, Chou <strong>et</strong> Kroner (1992), Bollerslev, Engle <strong>et</strong> Nelson (1994) <strong>et</strong><br />
Gouriéroux (1997).<br />
La secon<strong>de</strong> altern<strong>at</strong>ive a été présentée plus récemment par Bacry, Delour <strong>et</strong> Muzy (2001) qui ont développé<br />
un processus <strong>de</strong> marche alé<strong>at</strong>oire multifractale (MRW), processus en temps continu infiniment logdivisible<br />
(Bacry <strong>et</strong> Muzy 2002, Muzy <strong>et</strong> Bacry 2002). Comme dans un processus EGARCH, c’est ici le<br />
logarithme <strong>de</strong> la vol<strong>at</strong>ilité qui est modélisé. Ce processus est en fait un processus gaussien st<strong>at</strong>ionnaire<br />
dont l’auto-corrél<strong>at</strong>ion du logarithme <strong>de</strong> la vol<strong>at</strong>ilité est spécifiée <strong>de</strong> sorte qu’elle décroisse proportionnellement<br />
au logarithme du décalage entre les vol<strong>at</strong>ilités, ce qui est un <strong>de</strong>s faits stylisés présentés plus<br />
haut <strong>et</strong> dont ne rendait pas compte les modèles (G)ARCH. C’est en fait l’avancée majeure proposée<br />
par ce type <strong>de</strong> modèle par rapport aux autres processus déjà existant, d’autant qu’il rend compte, par<br />
construction même, <strong>de</strong>s corrél<strong>at</strong>ions à longue portée <strong>de</strong> la vol<strong>at</strong>ilité <strong>et</strong> est l’un <strong>de</strong>s rares processus connus<br />
à s<strong>at</strong>isfaire <strong>de</strong> manière théorique, <strong>et</strong> non artificiellement, les contraintes liées à la multifractalité.<br />
6 Sur <strong>de</strong>s échantillons <strong>de</strong> taille finie, Baviera, Biferale, Mantegna <strong>et</strong> Vulpiani (1998) ont cependant noté la présence - pure-<br />
ment artificielle - d’une apparente multifractalité.
30 1. Faits stylisés <strong>de</strong>s rentabilités boursières<br />
Pour ce qui est <strong>de</strong> la distribution marginale <strong>de</strong>s ren<strong>de</strong>ments, celle-ci est aussi à queues épaisses <strong>et</strong><br />
régulièrement variables mais avec <strong>de</strong>s exposants <strong>de</strong> queue beaucoup plus élevés que ceux précé<strong>de</strong>mment<br />
rencontrés. En eff<strong>et</strong>, lorsque le modèle est calibré sur les données réelles, il indique n<strong>et</strong>tement que les<br />
moments existent jusqu’à un ordre supérieur à vingt, ce qui est bien plus important que ce qui est habituellement<br />
admis. Cela montre une fois <strong>de</strong> plus combien il est difficile <strong>de</strong> déci<strong>de</strong>r du comportement<br />
asymptotique <strong>de</strong> la distribution marginale <strong>de</strong>s ren<strong>de</strong>ments.<br />
Enfin, ce modèle a permis à Sorn<strong>et</strong>te <strong>et</strong> al. (2002) <strong>de</strong> prédire d’un point <strong>de</strong> vue théorique le mo<strong>de</strong> <strong>de</strong><br />
relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité, c’est-à-dire comment décroît (ou croit) la vol<strong>at</strong>ilité lorsqu’elle r<strong>et</strong>ourne vers sa<br />
valeur moyenne. Lillo <strong>et</strong> Mantegna (2002) ont récemment observé qu’après un grand choc, la vol<strong>at</strong>ilité<br />
relaxait selon une loi <strong>de</strong> puissance dont l’exposant était étrangement proche <strong>de</strong> celui <strong>de</strong> la fonction d’autocorrél<strong>at</strong>ion<br />
<strong>de</strong> la vol<strong>at</strong>ilité, ce qui pouvait conduire à penser à une rel<strong>at</strong>ion <strong>de</strong> type fluctu<strong>at</strong>ion-dissip<strong>at</strong>ion<br />
entre les corrél<strong>at</strong>ions <strong>de</strong>s fluctu<strong>at</strong>ions <strong>de</strong> la vol<strong>at</strong>ilité “ à l’équilibre” <strong>et</strong> la dissip<strong>at</strong>ion entraînant le r<strong>et</strong>our à<br />
l’équilibre après un choc. En fait il n’en n’est rien, <strong>et</strong> nous avons montré que le mo<strong>de</strong> <strong>de</strong> relax<strong>at</strong>ion (plus<br />
précisément l’exposant <strong>de</strong> la loi <strong>de</strong> puissance) dépendait <strong>de</strong> la n<strong>at</strong>ure du choc. Si une importante nouvelle<br />
est suffisante par elle-même pour faire bouger le marché (comme le coup d’ét<strong>at</strong> contre Gorb<strong>at</strong>chev en<br />
1991, par exemple) la vol<strong>at</strong>ilité relaxe avec un exposant -1/2, indépendamment <strong>de</strong> l’amplitu<strong>de</strong> du choc.<br />
Au contraire, si le choc est dû à une accumul<strong>at</strong>ion <strong>de</strong> “p<strong>et</strong>ites” mauvaises nouvelles, la relax<strong>at</strong>ion se fait<br />
avec un exposant en général plus faible dont la valeur dépend <strong>de</strong> l’amplitu<strong>de</strong> du choc.<br />
C<strong>et</strong>te approche, vérifiée expérimentalement, a apporté une nouvelle confirm<strong>at</strong>ion <strong>de</strong> l’intérêt du modèle<br />
MRW, puisqu’il semble le seul modèle à même <strong>de</strong> décrire ce changement <strong>de</strong> mo<strong>de</strong> <strong>de</strong> relax<strong>at</strong>ion selon<br />
la n<strong>at</strong>ure du choc. En particulier, la relax<strong>at</strong>ion observée sur un processus GARCH n’en dépend pas.<br />
Enfin, c<strong>et</strong>te approche perm<strong>et</strong> d’entrevoir une métho<strong>de</strong> <strong>de</strong> classement systém<strong>at</strong>ique <strong>de</strong>s chocs <strong>de</strong> vol<strong>at</strong>ilité<br />
selon leur n<strong>at</strong>ure “endogène” ou “exogène”. Cependant, une limite actuelle <strong>de</strong> ce modèle est <strong>de</strong> traiter<br />
<strong>de</strong> manière totalement équivalente les gains <strong>et</strong> les pertes, négligeant <strong>de</strong> ce fait l’eff<strong>et</strong> <strong>de</strong> levier. Une<br />
solution envisageable serait <strong>de</strong> considérer un processus MRW asymétrique tel celui proposé par Pochart<br />
<strong>et</strong> Bouchaud (2002).<br />
1.4 Conclusion<br />
L’obj<strong>et</strong> <strong>de</strong> ce chapitre était d’exposer les principaux faits stylisés <strong>de</strong>s rentabilités boursières, que l’on<br />
peut résumer comme suit :<br />
– distribution <strong>de</strong>s ren<strong>de</strong>ments à queues épaisses,<br />
– dépendance temporelle complexe : absence <strong>de</strong> corrél<strong>at</strong>ion temporelle entre les ren<strong>de</strong>ments, mais corrél<strong>at</strong>ion<br />
persistante <strong>de</strong> la vol<strong>at</strong>ilité,<br />
– croissance super-exponentielle transitoire m<strong>et</strong>tant en évi<strong>de</strong>nce l’existence <strong>de</strong> différentes phases <strong>de</strong><br />
marché,<br />
– <strong>et</strong> caractère multifractal <strong>de</strong>s trajectoires <strong>de</strong>s prix <strong>de</strong>s actifs financiers.<br />
Certaines justific<strong>at</strong>ions théoriques peuvent être apportées à ces faits stylisés en faisant appel aussi bien<br />
à <strong>de</strong>s principes fondamentaux <strong>de</strong> la <strong>théorie</strong> financière qu’à <strong>de</strong> simples modèles phénoménologiques. En<br />
r<strong>et</strong>our, cela perm<strong>et</strong> <strong>de</strong> limiter le champ <strong>de</strong>s processus acceptables pour la modélis<strong>at</strong>ion <strong>de</strong>s prix <strong>de</strong>s actifs<br />
financiers. Pour autant, c<strong>et</strong>te démarche reste superficielle en ce qu’elle ne perm<strong>et</strong> pas une meilleure<br />
compréhension <strong>de</strong>s mécanismes <strong>de</strong> fonctionnement <strong>de</strong>s marchés financiers <strong>et</strong> <strong>de</strong>s comportements <strong>de</strong>s<br />
agents prenant position sur ces marchés. C’est pourquoi il sera nécessaire <strong>de</strong> nous intéresser à l’aspect<br />
“microscopique” ou microstructurel <strong>de</strong>s marchés financiers, ce qui sera l’obj<strong>et</strong> <strong>de</strong>s chapitres 5 <strong>et</strong> 6. Mais<br />
avant cela, nous allons, dans les chapitres qui suivent, présenter certains <strong>de</strong>s résult<strong>at</strong>s que nous avons<br />
obtenus <strong>et</strong> qui perm<strong>et</strong>tent <strong>de</strong> justifier les assertions que nous avons formulées tout au long <strong>de</strong> ce chapitre.
Chapitre 2<br />
Modèles phénoménologiques <strong>de</strong> cours<br />
Nous présentons dans ce chapitre quelques modèles phénoménologiques <strong>de</strong> cours d’actifs afin d’illustrer<br />
notre affirm<strong>at</strong>ion du chapitre 1, section 1.2 selon laquelle il était tout aussi possible <strong>de</strong> construire <strong>de</strong>s<br />
modèles perm<strong>et</strong>tant <strong>de</strong> justifier que la distribution st<strong>at</strong>ionnaire <strong>de</strong>s ren<strong>de</strong>ments est régulièrement variable<br />
ou exponentielle étirée.<br />
Pour cela, nous commençons par montrer que les modèles simples <strong>de</strong> bulles r<strong>at</strong>ionnelles 1 à la Blanchard<br />
<strong>et</strong> W<strong>at</strong>son (1982) ne sont en fait pas comp<strong>at</strong>ibles avec les données empiriques, car s’ils conduisent effectivement<br />
à justifier l’existence <strong>de</strong> distributions <strong>de</strong> ren<strong>de</strong>ments régulièrement variables, l’indice <strong>de</strong> queue<br />
auquel conduisent ces modèles est beaucoup plus faible que celui réellement estimé sous c<strong>et</strong>te hypothèse.<br />
Nous passons ensuite en revue <strong>de</strong>ux modèles qui s<strong>at</strong>isfont aux contraintes empiriques en vérifiant la plupart<br />
<strong>de</strong>s faits stylisés exposés au chapitre 1, <strong>et</strong> qui perm<strong>et</strong>tent <strong>de</strong> justifier tout autant l’hypothèse <strong>de</strong><br />
distribution hyperbolique que l’hypothèse <strong>de</strong> distribution exponentielle étirée.<br />
1 Le lecteur est invité à se référer à Blanchard <strong>et</strong> W<strong>at</strong>son (1982), Coll<strong>et</strong>az <strong>et</strong> Gourlaouen (1989), Broze <strong>et</strong> al. (1990) ou<br />
Adams <strong>et</strong> Szarfarz (1992) notamment pour une revue sur le suj<strong>et</strong>.<br />
31
32 2. Modèles phénoménologiques <strong>de</strong> cours<br />
2.1 Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses<br />
Lux <strong>et</strong> Sorn<strong>et</strong>te (2002) ont démontré que les queues <strong>de</strong>s distributions inconditionnelles <strong>de</strong>s incréments <strong>de</strong><br />
prix <strong>et</strong> <strong>de</strong> ren<strong>de</strong>ments associées au modèle <strong>de</strong> bulles r<strong>at</strong>ionnelles <strong>de</strong> Blanchard <strong>et</strong> W<strong>at</strong>son (1982) suivent<br />
<strong>de</strong>s lois <strong>de</strong> puissance (décroissent <strong>de</strong> manière hyperbolique), dont l’exposant <strong>de</strong> queue µ est inférieur à<br />
un sur un large domaine. Bien que les queues en lois <strong>de</strong> puissance soient une caractéristique marquante<br />
relevée sur les données empiriques, la valeur numérique µ < 1 est en désaccord avec les estim<strong>at</strong>ions habituelles<br />
qui, selon ce type <strong>de</strong> modèle, donnent µ 3. Parmi les qu<strong>at</strong>re hypothèses soutenant le modèle<br />
<strong>de</strong> bulles r<strong>at</strong>ionnelles <strong>de</strong> Blanchard <strong>et</strong> W<strong>at</strong>son (r<strong>at</strong>ionalité <strong>de</strong>s agents, condition <strong>de</strong> non arbitrage, dynamique<br />
multiplic<strong>at</strong>ive, bulles indépendantes pour chaque actif), nous démontrons que le même résult<strong>at</strong><br />
µ < 1 reste valable lorsque que l’on relaxe la <strong>de</strong>rnière <strong>de</strong> ces hypothèses, i.e en perm<strong>et</strong>tant le couplage<br />
entre les différentes bulles <strong>de</strong>s divers actifs. En conséquence, <strong>de</strong>s extensions non linéaires <strong>de</strong> la dynamique<br />
<strong>de</strong>s bulles ou une relax<strong>at</strong>ion partielle du principe d’évalu<strong>at</strong>ion r<strong>at</strong>ionnel <strong>de</strong>s prix sont nécessaires<br />
si l’on souhaite rendre compte <strong>de</strong>s observ<strong>at</strong>ions empiriques.<br />
Reprint from : Y. Malevergne and D. Sorn<strong>et</strong>te (2001), “Multi-dimensional bubbles and f<strong>at</strong> tails”, Quantit<strong>at</strong>ive<br />
Finance 1, 533-541.
2.1. Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses 33<br />
Q UANTITATIVE F INANCE V OLUME 1 (2001) 533–541 RESEARCH PAPER<br />
I NSTITUTE OF P HYSICS P UBLISHING quant.iop.org<br />
Multi-dimensional r<strong>at</strong>ional bubbles<br />
and f<strong>at</strong> tails<br />
Y Malevergne 1,2 and D Sorn<strong>et</strong>te 1,3<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622,<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances, Université Lyon I, 43 Bd du<br />
11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x, France<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and<br />
Space Science, University of California, Los Angeles, CA 90095, USA<br />
E-mail: Yannick.Malevergne@unice.fr and sorn<strong>et</strong>te@unice.fr<br />
Received 20 July 2001, in final form 20 August 2001<br />
Published 14 September 2001<br />
Online <strong>at</strong> stacks.iop.org/Quant/1/533<br />
Abstract<br />
Lux and Sorn<strong>et</strong>te have <strong>de</strong>monstr<strong>at</strong>ed th<strong>at</strong> the tails of the unconditional<br />
distributions of price differences and of r<strong>et</strong>urns associ<strong>at</strong>ed with the mo<strong>de</strong>l of<br />
r<strong>at</strong>ional bubbles of Blanchard and W<strong>at</strong>son follow power laws (i.e. exhibit<br />
hyperbolic <strong>de</strong>cline), with an asymptotic tail exponent µ
34 2. Modèles phénoménologiques <strong>de</strong> cours<br />
Y Malevergne and D Sorn<strong>et</strong>te Q UANTITATIVE F INANCE<br />
on the process Xt. There is a huge liter<strong>at</strong>ure on theor<strong>et</strong>ical<br />
refinements of this mo<strong>de</strong>l and on the empirical d<strong>et</strong>ectability of<br />
RE bubbles in financial d<strong>at</strong>a (see Camerer (1989) and Adam<br />
and Szafarz (1992) for surveys of this liter<strong>at</strong>ure).<br />
Recently, Lux and Sorn<strong>et</strong>te (1999) studied the implic<strong>at</strong>ions<br />
of the RE bubble mo<strong>de</strong>ls for the unconditional distribution<br />
of prices, price changes and r<strong>et</strong>urns resulting from a<br />
more general discr<strong>et</strong>e-time formul<strong>at</strong>ion extending (2) by allowing<br />
the multiplic<strong>at</strong>ive factor <strong>at</strong> to take arbitrary values and<br />
be i.i.d. random variables drawn from some non-<strong>de</strong>gener<strong>at</strong>e<br />
probability <strong>de</strong>nsity function (pdf) Pa(a). The mo<strong>de</strong>l can also<br />
be generalized by consi<strong>de</strong>ring non-normal realiz<strong>at</strong>ions of bt<br />
with distribution Pb(b) with E[bt] = 0, where E[·] is the expect<strong>at</strong>ion<br />
oper<strong>at</strong>or. Since in (2) the bubble Xt <strong>de</strong>notes the difference<br />
b<strong>et</strong>ween the observed price and the fundamental price,<br />
the ‘bubble’ regimes refer to the cases when Xt explo<strong>de</strong>s exponentially<br />
un<strong>de</strong>r the action of successive multiplic<strong>at</strong>ions by<br />
factor <strong>at</strong>,<strong>at</strong>+1,...with a majority of them larger than 1 in absolute<br />
value but different, thus adding a stochastic component<br />
to the standard mo<strong>de</strong>l of Blanchard and W<strong>at</strong>son (1982).<br />
For this large class of stochastic processes, Lux and<br />
Sorn<strong>et</strong>te (1999) have shown th<strong>at</strong> the distribution of r<strong>et</strong>urns<br />
is a power law whose exponent µ is enforced by the nofree-lunch<br />
condition to remain lower than one. Although<br />
power-law tails are a pervasive fe<strong>at</strong>ure of empirical d<strong>at</strong>a,<br />
these characteriz<strong>at</strong>ions are in strong disagreement with the<br />
usual empirical estim<strong>at</strong>es which find µ ≈ 3 (<strong>de</strong> Vries 1994,<br />
Lux 1996, Pagan 1996, Guillaume <strong>et</strong> al 1997, Gopikrishnan<br />
<strong>et</strong> al 1998). Thus, Lux and Sorn<strong>et</strong>te (1999) conclu<strong>de</strong>d th<strong>at</strong><br />
exogenous r<strong>at</strong>ional bubbles are hardly comp<strong>at</strong>ible with the<br />
empirical distribution d<strong>at</strong>a.<br />
At this stage, one could argue th<strong>at</strong> there is a logical trap<br />
in the finding µ
2.1. Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses 35<br />
Q UANTITATIVE F INANCE Multi-dimensional r<strong>at</strong>ional bubbles<br />
for instance Hommes (2001) and Chall<strong>et</strong> <strong>et</strong> al (2001) and<br />
references therein). Following many others before us, we<br />
conclu<strong>de</strong> th<strong>at</strong> the hypotheses of r<strong>at</strong>ional expect<strong>at</strong>ions and of noarbitrage<br />
condition are useful approxim<strong>at</strong>ion or starting points<br />
for mo<strong>de</strong>l constructions.<br />
Within this framework, the price Pt of an ass<strong>et</strong> <strong>at</strong> time t<br />
should obey the equ<strong>at</strong>ion<br />
Pt = δ · EQ[Pt+1|Ft]+dt ∀{Pt}t0, (4)<br />
where dt is an exogeneous ‘divi<strong>de</strong>nd’, δ is the discount factor<br />
and EQ[Pt+1|Ft] is the expect<strong>at</strong>ion of Pt+1 conditioned upon the<br />
knowledge of the filtr<strong>at</strong>ion up to time t un<strong>de</strong>r the risk-neutral<br />
probability Q.<br />
It is important to stress th<strong>at</strong> the r<strong>at</strong>ionality of both<br />
expect<strong>at</strong>ions and behaviour does not necessarily imply th<strong>at</strong><br />
the price of an ass<strong>et</strong> be equal to its fundamental value. In<br />
other words, there can be r<strong>at</strong>ional <strong>de</strong>vi<strong>at</strong>ions of the price from<br />
this value, called r<strong>at</strong>ional bubbles. A r<strong>at</strong>ional bubble can arise<br />
when the actual mark<strong>et</strong> price <strong>de</strong>pends positively on its own<br />
expected r<strong>at</strong>e of change. This is thought to som<strong>et</strong>imes occur in<br />
ass<strong>et</strong> mark<strong>et</strong>s and constitutes the very mechanism un<strong>de</strong>rlying<br />
the mo<strong>de</strong>ls of Blanchard (1979) and Blanchard and W<strong>at</strong>son<br />
(1982). In<strong>de</strong>ed, the ‘forward’ solution of (4) is well-known to<br />
be<br />
+∞<br />
Ft =<br />
i=0<br />
δ i · EQ[dt+i|Ft]. (5)<br />
It is straightforward to check by replacement th<strong>at</strong> the general<br />
solution of (4) is the sum of the forward solution (5) and of an<br />
arbitrary component Xt<br />
where Xt has to obey the single condition:<br />
Pt = Ft + Xt, (6)<br />
Xt = δ · EQ[Xt+1|Ft]. (7)<br />
With only two fundamental and quite reasonable assumptions,<br />
it is thus possible to <strong>de</strong>rive an equ<strong>at</strong>ion (6) (with (7)) justifying<br />
the existence of price fluctu<strong>at</strong>ions and <strong>de</strong>vi<strong>at</strong>ions from the<br />
fundamental value Ft, for equilibrium mark<strong>et</strong>s with r<strong>at</strong>ional<br />
agents. However, the dynamics of the bubbles remains<br />
unknown and is a priori compl<strong>et</strong>ely arbitrary apart from the<br />
no-arbitrage constraint (7).<br />
2.2. Bubble dynamics<br />
One of the main contributions of the mo<strong>de</strong>l of Blanchard<br />
and W<strong>at</strong>son (1982), and its possible generaliz<strong>at</strong>ions (8), is<br />
to propose a bubble dynamics which is both comp<strong>at</strong>ible with<br />
most of the empirical stylized facts of price time series and<br />
sufficiently simple to allow for a tractable analytical tre<strong>at</strong>ment.<br />
In<strong>de</strong>ed, the stochastic autoregressive process<br />
Xt+1 = <strong>at</strong>Xt + bt, (8)<br />
where {<strong>at</strong>} and {bt} are i.i.d. random variables, is well-known<br />
to lead to f<strong>at</strong>-tailed distributions and vol<strong>at</strong>ility clustering.<br />
Kesten (1973) (see Goldie (1991) for a mo<strong>de</strong>rn extension)<br />
has shown th<strong>at</strong>, if E[ln |a|] < 0, the stochastic process {Xt}<br />
admits a st<strong>at</strong>ionary solution (i.e. with a st<strong>at</strong>ionary distribution<br />
function), whose distribution <strong>de</strong>nsity P(X) is a power law<br />
P(X) ∼ X −1−µ with exponent µ s<strong>at</strong>isfying<br />
E[|a| µ ] = 1, (9)<br />
provi<strong>de</strong>d th<strong>at</strong> E[|b| µ ] < ∞. It is easy to show th<strong>at</strong><br />
without other constraints every exponent µ > 0 can be<br />
reached. Consi<strong>de</strong>r, for example, a sequence of variables {<strong>at</strong>}<br />
in<strong>de</strong>pen<strong>de</strong>ntly and i<strong>de</strong>ntically distributed according to a lognormal<br />
law with localiz<strong>at</strong>ion param<strong>et</strong>er a0 < 1 and scale<br />
ln a0<br />
param<strong>et</strong>er σ . Equ<strong>at</strong>ion (9) simply yields µ =−2 σ 2 , which<br />
shows th<strong>at</strong>, varying σ , µ can range over the entire positive real<br />
line.<br />
The second interesting point is th<strong>at</strong> the process (8) allows<br />
vol<strong>at</strong>ility clustering. It is easy to see th<strong>at</strong> a large value of Xt will<br />
be followed by a large value of Xt+1 with a large probability.<br />
The change of variables Xt = Y 2<br />
t with <strong>at</strong> = vZ2 t and bt = uZ2 t<br />
maps exactly the process (8) onto an ARCH(1) process<br />
<br />
u + vYt 2 , (10)<br />
Yt+1 = Zt<br />
where Zt is a Gaussian random variable. The ARCH processes<br />
are known to account for vol<strong>at</strong>ility clustering. Therefore, the<br />
process (8) also exhibits vol<strong>at</strong>ility clustering.<br />
This class (8) of stochastic processes thus provi<strong>de</strong>s<br />
several interesting fe<strong>at</strong>ures of real price series (Roman <strong>et</strong> al<br />
2001). There is however an objection rel<strong>at</strong>ed to the fact th<strong>at</strong>,<br />
without additional assumptions, the bubble price can become<br />
arbitrarily neg<strong>at</strong>ive and can then lead to a neg<strong>at</strong>ive price:<br />
Pt < 0. In fact, a neg<strong>at</strong>ive price is not as meaningless as<br />
often taken for granted, as shown in Sorn<strong>et</strong>te (2000). But,<br />
even without allowing for neg<strong>at</strong>ive prices, it is reasonable to<br />
argue th<strong>at</strong> near Pt = 0, other mechanisms come into play and<br />
modify equ<strong>at</strong>ion (8) in the neighbourhood of a vanishing price.<br />
For instance, when a mark<strong>et</strong> un<strong>de</strong>rgoes a too strong and abrupt<br />
loss, quot<strong>at</strong>ions are interrupted.<br />
2.3. The in<strong>de</strong>pen<strong>de</strong>nt bubbles assumption<br />
The Blanchard and W<strong>at</strong>son mo<strong>de</strong>l assumes th<strong>at</strong> there is only<br />
one ass<strong>et</strong> and one bubble. In other words, the evolution of<br />
each ass<strong>et</strong> does not <strong>de</strong>pend on the dynamics of the others.<br />
But, in reality, there is no such thing as an isol<strong>at</strong>ed ass<strong>et</strong>.<br />
Stock mark<strong>et</strong>s exhibit a vari<strong>et</strong>y of inter-<strong>de</strong>pen<strong>de</strong>nces, based<br />
in part on the mutual influences b<strong>et</strong>ween the USA, European<br />
and Japanese mark<strong>et</strong>s. In addition, individual stocks may be<br />
sensitive to the behaviour of the specific industry as a whole<br />
to which they belong and to a few other indic<strong>at</strong>ors, such as<br />
the main indices, interest r<strong>at</strong>es and so on. Mantegna (1999)<br />
and Bonanno <strong>et</strong> al (2001) have in<strong>de</strong>ed shown the existence<br />
of a hierarchical organiz<strong>at</strong>ion of stock inter<strong>de</strong>pen<strong>de</strong>nces.<br />
Furthermore, bubbles often appear to be not isol<strong>at</strong>ed fe<strong>at</strong>ures<br />
of a s<strong>et</strong> of mark<strong>et</strong>s. For instance, Flood <strong>et</strong> al (1984) tested<br />
wh<strong>et</strong>her a bubble simultaneously existed across the n<strong>at</strong>ions,<br />
such as Germany, Poland, and Hungary, th<strong>at</strong> experienced<br />
hyperinfl<strong>at</strong>ion in the early 1920s. Coordin<strong>at</strong>ed corrections<br />
to wh<strong>at</strong> may be consi<strong>de</strong>red to be correl<strong>at</strong>ed bubbles can<br />
som<strong>et</strong>imes be d<strong>et</strong>ected. One of the most prominent examples<br />
535
36 2. Modèles phénoménologiques <strong>de</strong> cours<br />
Y Malevergne and D Sorn<strong>et</strong>te Q UANTITATIVE F INANCE<br />
is found in the mark<strong>et</strong> appreci<strong>at</strong>ions observed in many of the<br />
world mark<strong>et</strong>s prior to the world mark<strong>et</strong> crash in October 1987<br />
(Barro <strong>et</strong> al 1989). The collective growth of most of the<br />
mark<strong>et</strong>s worldwi<strong>de</strong> was interrupted by a worldwi<strong>de</strong> mark<strong>et</strong><br />
crash: from the opening on 14 October 1987 through the<br />
mark<strong>et</strong> close on 19 October major indices of mark<strong>et</strong> valu<strong>at</strong>ion<br />
in the United St<strong>at</strong>es <strong>de</strong>clined by 30% or more. Furthermore,<br />
all major mark<strong>et</strong>s in the world <strong>de</strong>clined substantially in the<br />
month: out of 23 major industrial countries, 19 had a <strong>de</strong>cline<br />
gre<strong>at</strong>er than 20%. This is one of the most striking evi<strong>de</strong>nces<br />
of the existence of correl<strong>at</strong>ions b<strong>et</strong>ween corrections to bubbles<br />
across the world mark<strong>et</strong>s. Similar intermittent coordin<strong>at</strong>ion<br />
of bubbles have been d<strong>et</strong>ected among the significant bubbles<br />
followed by large crashes or severe corrections in L<strong>at</strong>in-<br />
American and Asian stock mark<strong>et</strong>s (Johansen and Sorn<strong>et</strong>te<br />
2000).<br />
These empirical facts suggest an improvement on the onedimensional<br />
bubble mo<strong>de</strong>l by introducing a multi-dimensional<br />
generaliz<strong>at</strong>ion.<br />
2.4. Position of the problem<br />
As shown in Lux and Sorn<strong>et</strong>te (1999), the one-ass<strong>et</strong> mo<strong>de</strong>l<br />
(8) suffers from a <strong>de</strong>ficiency: the power law tail exponents<br />
predicted by the Blanchard and W<strong>at</strong>son mo<strong>de</strong>l are not<br />
comp<strong>at</strong>ible with the empirical facts. The proof relies on<br />
the following ingredients: the no-free-lunch condition and<br />
the one-dimensional dynamics of the bubble. The onedimensional<br />
dynamics of the bubble implies th<strong>at</strong> the equ<strong>at</strong>ion<br />
(9) holds, while the no-free-lunch condition imposes<br />
E[a] > 1. (11)<br />
Putting these two equ<strong>at</strong>ions tog<strong>et</strong>her allows Lux and Sorn<strong>et</strong>te<br />
(1999) to conclu<strong>de</strong> th<strong>at</strong> necessarily µ
2.1. Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses 37<br />
Q UANTITATIVE F INANCE Multi-dimensional r<strong>at</strong>ional bubbles<br />
A generaliz<strong>at</strong>ion of the two-dimensional case to arbitrary<br />
dimensions leads to the following stochastic random equ<strong>at</strong>ion<br />
(SRE)<br />
Xt = AtXt−1 + Bt<br />
(16)<br />
where (Xt, Bt) are d-dimensional vectors. Each component<br />
of Xt can be thought of as the difference b<strong>et</strong>ween the price<br />
of the ass<strong>et</strong> and its fundamental price. The m<strong>at</strong>rices (At)<br />
are i<strong>de</strong>ntically in<strong>de</strong>pen<strong>de</strong>nt distributed d × d-dimensional<br />
stochastic m<strong>at</strong>rices. We assume th<strong>at</strong> Bt are i<strong>de</strong>ntically<br />
in<strong>de</strong>pen<strong>de</strong>nt distributed random vectors and th<strong>at</strong> (Xt) is a<br />
causal st<strong>at</strong>ionary solution of (16). Generaliz<strong>at</strong>ions introducing<br />
additional arbitrary linear terms <strong>at</strong> larger time lags such as<br />
Xt−2,... can be tre<strong>at</strong>ed with slight modific<strong>at</strong>ions of our<br />
approach and yield the same conclusions. We shall thus<br />
confine our <strong>de</strong>monstr<strong>at</strong>ion on the SRE of or<strong>de</strong>r 1, keeping in<br />
mind th<strong>at</strong> our results apply analogously to arbitrary or<strong>de</strong>rs of<br />
regressions.<br />
To formalize the SRE in a rigorous manner, we introduce<br />
in a standard way the probability space (, F, P) and a<br />
filtr<strong>at</strong>ion (Ft). Here P represents the product measure P =<br />
PX ⊗ PA ⊗ PB, where PX, PA and PB are the probability<br />
measures associ<strong>at</strong>ed with {Xt}, {At} and {Bt}. We further<br />
assume as is customory th<strong>at</strong> the stochastic process (Xt) is<br />
adapted to the filtr<strong>at</strong>ion (Ft).<br />
In the following, we <strong>de</strong>note by |·|the Eucli<strong>de</strong>an norm and<br />
by ||·||the corresponding norm for any d × d-m<strong>at</strong>rix A<br />
||A|| = sup |Ax|. (17)<br />
|x|=1<br />
Now, we will formalize the ‘no-free-lunch’ condition for<br />
the SRE (16) and show th<strong>at</strong> it entails in particular th<strong>at</strong> the<br />
spectral radius (largest eigenvalue) of E[At] must be equal to<br />
the inverse of the discount factor, hence it must be larger than<br />
1.<br />
3.3. The no-free-lunch condition<br />
3.3.1. No-free-lunch condition un<strong>de</strong>r the risk-neutral<br />
probability measure The ‘no-free-lunch’ condition is<br />
equivalent to the existence of a probability measure Q<br />
equivalent to P such th<strong>at</strong>, for all self-financing portfolios t,<br />
t<br />
S0,t is a Q-martingale, where S0,t = t−1 i=0 δi −1 , δi = (1+ri) −1<br />
is the discount factor for period i and ri is the corresponding<br />
risk-free interest r<strong>at</strong>e.<br />
It is n<strong>at</strong>ural to assume th<strong>at</strong>, for a given period i, the<br />
discount r<strong>at</strong>es ri are the same for all ass<strong>et</strong>s. In frictionless<br />
mark<strong>et</strong>s, a <strong>de</strong>vi<strong>at</strong>ion for this hypothesis would lead to arbitrage<br />
opportunities. Furthermore, since the sequence of m<strong>at</strong>rices<br />
{At} is i.i.d. and therefore st<strong>at</strong>ionary, this implies th<strong>at</strong> δt or rt<br />
must be constant and equal respectively to δ and r.<br />
Un<strong>de</strong>r those conditions, we have the following<br />
proposition:<br />
Proposition 1. The stochastic process<br />
Xt = AtXt−1 + Bt<br />
s<strong>at</strong>isfies the no-arbitrage condition if and only if<br />
(18)<br />
EQ[A] = 1<br />
δ Id. (19)<br />
The proof is given in appendix A.<br />
The condition (19) imposes some stringent constraints on<br />
admissible m<strong>at</strong>rices At. In<strong>de</strong>ed, while At are not diagonal in<br />
general, their expect<strong>at</strong>ion must be diagonal. This implies th<strong>at</strong><br />
the off-diagonal terms of the m<strong>at</strong>rices At must take neg<strong>at</strong>ive<br />
values, sufficiently often so th<strong>at</strong> their averages vanish. The offdiagonal<br />
coefficients quantify the influence of other bubbles on<br />
a given one. The condition (19) thus means th<strong>at</strong> the average<br />
effect of other bubbles on any given one must vanish. It is<br />
straightforward to check th<strong>at</strong>, in this linear framework, this<br />
implies an absence of correl<strong>at</strong>ion (but not of <strong>de</strong>pen<strong>de</strong>nce)<br />
b<strong>et</strong>ween the different bubble components E[X (k) X (ℓ) ] = 0<br />
for any k = ℓ, where X (j) <strong>de</strong>notes the jth component of the<br />
bubble X.<br />
In contrast, the diagonal elements of At must be mostly<br />
positive in or<strong>de</strong>r for EP[Aii] = δ−1 , for all i, to hold true. In<br />
fact, on economic grounds, we can exclu<strong>de</strong> the cases where<br />
the diagonal elements take neg<strong>at</strong>ive values. In<strong>de</strong>ed, a neg<strong>at</strong>ive<br />
value of Aii <strong>at</strong> a given time t would imply th<strong>at</strong> X (i)<br />
t might<br />
abruptly change sign b<strong>et</strong>ween t − 1 and t, wh<strong>at</strong> does not seem<br />
to be a reasonable financial process.<br />
3.3.2. Consequence for the no-free-lunch condition un<strong>de</strong>r<br />
historic probability measure The historical P and riskneutral<br />
Q probability measures are equivalent. This means<br />
th<strong>at</strong> there exists a non-neg<strong>at</strong>ive m<strong>at</strong>rix h(θ) = hij (θij ) such<br />
th<strong>at</strong>, for each element in<strong>de</strong>xed by i, j, wehave<br />
EP[Aij ] = EQ[hij · Aij ] (20)<br />
= hij (θ 0 ij ) · EQ[Aij ] for some θ 0 ij ∈ R. (21)<br />
The second equ<strong>at</strong>ion comes from the well-known result:<br />
<br />
<br />
f(θ)· g(θ)dµ(θ) = g(θ0) · f(θ)dµ(θ) for some θ0 ∈ R.<br />
(22)<br />
We thus g<strong>et</strong><br />
EP[Aij ] = 0 if i = j (23)<br />
EP[Aij ] = 1<br />
δ (i)<br />
if i = j, (24)<br />
where the δ (i) can be different. We can thus write<br />
EP[A] = δ −1 where δ −1 = diag[δ (1)−1 ,...,δ (d)−1 ]. (25)<br />
Appendix B gives a proof showing th<strong>at</strong> δ (i) is in<strong>de</strong>ed the<br />
genuine discount factor for the ith bubble component.<br />
4. Renewal theory for products of<br />
random m<strong>at</strong>rices<br />
In the following, we will consi<strong>de</strong>r th<strong>at</strong> the random d × d<br />
m<strong>at</strong>rices At are invertible m<strong>at</strong>rices with real entries. We will<br />
<strong>de</strong>note by GLd(R) the group of these m<strong>at</strong>rices.<br />
537
38 2. Modèles phénoménologiques <strong>de</strong> cours<br />
Y Malevergne and D Sorn<strong>et</strong>te Q UANTITATIVE F INANCE<br />
4.1. Definitions<br />
Definition 1 (Feasible m<strong>at</strong>rix). A m<strong>at</strong>rix M ∈ GLd(R) is<br />
P-feasible if there exists an n ∈ N and M1,...,Mn ∈<br />
supp(P) such th<strong>at</strong> M = M1 ...Mn and if M has a simple real<br />
eigenvalue q(M) which, in modulus, exceeds all other<br />
eigenvalues of M.<br />
Definition 2. For any m<strong>at</strong>rix M ∈ GLd(R) and M ′ its<br />
transpose, MM ′ is a symm<strong>et</strong>ric positive <strong>de</strong>finite m<strong>at</strong>rix. We<br />
<strong>de</strong>fine λ(M) as the square root of the smallest eigenvalue of<br />
MM ′ .<br />
4.2. Theorem<br />
We extend the theorem 2.7 of Davis <strong>et</strong> al (1999), which<br />
synthesized Kesten’s theorems 3 and 4 in Kesten (1973), to the<br />
case of real valued m<strong>at</strong>rices. The proof of this theorem is given<br />
in Le Page (1983). We stress th<strong>at</strong> the conditions listed below do<br />
not require the m<strong>at</strong>rices (An) to be non-neg<strong>at</strong>ive. Actually, we<br />
have seen th<strong>at</strong>, in or<strong>de</strong>r for the r<strong>at</strong>ional expect<strong>at</strong>ion condition<br />
not to lead to trivial results, the off-diagonal coefficients of<br />
(An) have to be neg<strong>at</strong>ive with sufficiently large probability<br />
such th<strong>at</strong> their means vanish.<br />
Theorem 1. L<strong>et</strong> (An) be an i.i.d. sequence of m<strong>at</strong>rices in<br />
GLd(R) s<strong>at</strong>isfying the following s<strong>et</strong> of conditions:<br />
H1: for some ɛ>0, EPA [||A||ɛ ] < 1,<br />
H2: for every open U ⊂ Sd−1 (the unit sphere in Rd ) and for<br />
all x ∈ Sd−1 there exists an n such th<strong>at</strong><br />
<br />
xA1 ...An<br />
Pr<br />
∈ U > 0. (26)<br />
||xA1 ...An||<br />
H3: the group {ln |q(M)|,M is PA-feasible} is <strong>de</strong>nse in R .<br />
H4: for all r ∈ R d , Pr{A1r + B1 = r} < 1.<br />
H5: there exists a κ0 > 0 such th<strong>at</strong><br />
EPA ([λ(A1)] κ0 ) 1. (27)<br />
H6: with the same κ0 > 0 as for the previous condition, there<br />
exists a real number u>0 such th<strong>at</strong><br />
sup{||A1||, EPA<br />
||B1||} κ0+u <br />
< ∞ ,<br />
<br />
||A1|| −u (28)<br />
< ∞.<br />
EPA<br />
Provi<strong>de</strong>d th<strong>at</strong> these conditions hold,<br />
• there exists a unique solution κ1 ∈ (0,κ0] to the equ<strong>at</strong>ion<br />
1 <br />
lim ln EPA ||A1 ...An||<br />
n→∞ n κ1 = 0, (29)<br />
• if (Xn) is the st<strong>at</strong>ionary solution to the SRE in (16) then<br />
X is regularly varying with in<strong>de</strong>x κ1. In other words, the<br />
tail of the marginal distribution of each of the<br />
components of the vector X is asymptotically a power<br />
law with exponent κ1.<br />
538<br />
4.3. Comments on the theorem<br />
4.3.1. Intuitive meaning of the hypotheses A suitable<br />
property for an economic mo<strong>de</strong>l is the existence of a st<strong>at</strong>ionary<br />
solution, i.e. the solution Xt of the SRE (16) should not<br />
blow up. This condition is ensured by the hypothesis (H1).<br />
In<strong>de</strong>ed, EPA [ln ||A||] < 0 implies th<strong>at</strong> the Lyapunov exponent<br />
of the sequence {An} of i.i.d. m<strong>at</strong>rices is neg<strong>at</strong>ive (Davis<br />
<strong>et</strong> al 1999). And it is well known th<strong>at</strong> the neg<strong>at</strong>ivity of the<br />
Lyapunov exponent is a sufficient condition for the existence<br />
of a st<strong>at</strong>ionary solution Xt, provi<strong>de</strong>d th<strong>at</strong> E[ln + ||B||] < ∞.<br />
However, this condition can lead to too fast a <strong>de</strong>cay of the<br />
tail of the distibution of {X}. This phenomenon is prevented<br />
by (H5) which means intuitively th<strong>at</strong> the multiplic<strong>at</strong>ive<br />
factors given by the elements of At som<strong>et</strong>imes produce an<br />
amplific<strong>at</strong>ion of Xt. In the one-dimensional bubble case, this<br />
condition reduces to the simple rule th<strong>at</strong> <strong>at</strong> must som<strong>et</strong>imes<br />
be larger than 1 so th<strong>at</strong> a bubble can <strong>de</strong>velop. Otherwise, the<br />
power law tail will be replaced by an exponential tail.<br />
So, (H1) and (H5) keep the balance b<strong>et</strong>ween two opposite<br />
objectives: to obtain a st<strong>at</strong>ionary solution and to observe a<br />
f<strong>at</strong>-tailed distribution for the process (Xt).<br />
Another <strong>de</strong>sirable property for the mo<strong>de</strong>l is ergodicity: we<br />
expect the price process (Xt) to explore the entire space Rd .<br />
This is ensured by hypotheses (H2) and (H4): hypothesis (H2)<br />
allows Xt to visit the neighbourhood of any point in Rd , and<br />
(H4) forbids the trajectory to be trapped <strong>at</strong> some points.<br />
Hypotheses (H3) and (H6) are more technical ones. The<br />
hypothesis (H6) simply ensures th<strong>at</strong> the tails of the distributions<br />
of At and Bt are thinner than the tail cre<strong>at</strong>ed by the SRE (16),<br />
so th<strong>at</strong> the observed tail in<strong>de</strong>x is really due to the dynamics of<br />
the system and not to the heavy-tail n<strong>at</strong>ure of the distributions<br />
of At or Bt. The hypothesis (H3) expresses some kind of<br />
aperiodicity condition.<br />
4.3.2. Intuitive meaning of (29) The equ<strong>at</strong>ion (29)<br />
d<strong>et</strong>ermining the tail exponent κ1 reduces to (9) in the onedimensional<br />
case, which is simple to handle. In the multidimensional<br />
case, the novel fe<strong>at</strong>ure is the non-diagonal n<strong>at</strong>ure<br />
of the multiplic<strong>at</strong>ion of m<strong>at</strong>rices which does not allow in<br />
general for an explicit equ<strong>at</strong>ion similar to (9).<br />
4.4. Constraint on the tail in<strong>de</strong>x<br />
The first conclusion of the theorem above shows th<strong>at</strong> the tail<br />
in<strong>de</strong>x κ1 of the process (Xt) is driven by the behaviour of the<br />
m<strong>at</strong>rices (At). We will then st<strong>at</strong>e a proposition in which we<br />
give a major<strong>at</strong>ion of the tail in<strong>de</strong>x.<br />
Proposition 2. A necessary condition for the process (16) to<br />
have a tail exponent κ1 > 1 is th<strong>at</strong> the spectral radius of<br />
[A] be smaller than 1:<br />
EPA<br />
κ1 > 1 ⇒ ρ(EPA [A])
2.1. Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses 39<br />
Q UANTITATIVE F INANCE Multi-dimensional r<strong>at</strong>ional bubbles<br />
5. Consequences for r<strong>at</strong>ional expect<strong>at</strong>ion<br />
bubbles<br />
We have seen in section 3.3 from proposition 1 th<strong>at</strong>, as a<br />
result of the no-arbitrage condition, the spectral radius of the<br />
m<strong>at</strong>rix EP[A] = 1<br />
δ Id is gre<strong>at</strong>er than 1. As a consequence, by<br />
applic<strong>at</strong>ion of the converse of proposition 2, this proves th<strong>at</strong><br />
the tail in<strong>de</strong>x κ1 of the distribution of (X) is smaller than 1.<br />
Using the same arguments as in Lux and Sorn<strong>et</strong>te (1999),<br />
th<strong>at</strong> we do not recall here, it can be shown th<strong>at</strong> the distribution<br />
of price differences and price r<strong>et</strong>urns follows, <strong>at</strong> least over<br />
an exten<strong>de</strong>d range of large r<strong>et</strong>urns, a power law distribution<br />
whose exponent remains lower than 1. This result generalizes<br />
to arbitrary d-dimensional processes the result of Lux and<br />
Sorn<strong>et</strong>te (1999). As a consequence, d-dimensional r<strong>at</strong>ional<br />
expect<strong>at</strong>ion bubbles linking several ass<strong>et</strong>s suffer from the<br />
same discrepancy compared to empirical d<strong>at</strong>a as the onedimensional<br />
bubbles. It therefore appears th<strong>at</strong> accounting for<br />
possible <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween bubbles is not sufficient to cure<br />
the Blanchard and W<strong>at</strong>son mo<strong>de</strong>l: a linear multi-dimensional<br />
bubble dynamics such as (16) is hardly reconcilable with some<br />
of the most fundamental stylized facts of financial d<strong>at</strong>a <strong>at</strong> a very<br />
elementary level.<br />
This result does not rely on the diagonal property of the<br />
m<strong>at</strong>rices E[At] but only on the value of its spectral radius<br />
imposed by the no-arbitrage condition. In other words, the fact<br />
th<strong>at</strong> the introduction of <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween ass<strong>et</strong> bubbles is<br />
not sufficient to cure the mo<strong>de</strong>l can be traced to the constraint<br />
introduced by the no-arbitrage condition, which imposes th<strong>at</strong><br />
the averages of the off-diagonal terms of the m<strong>at</strong>rices At<br />
vanish. As we indic<strong>at</strong>ed before, this implies zero correl<strong>at</strong>ions<br />
(but not the absence of <strong>de</strong>pen<strong>de</strong>nce) b<strong>et</strong>ween ass<strong>et</strong> bubbles.<br />
It thus seems th<strong>at</strong> the multi-dimensional generaliz<strong>at</strong>ion is<br />
constrained so much by the no-arbitrage condition th<strong>at</strong> the<br />
multi-dimensional bubble mo<strong>de</strong>l almost reduces to an average<br />
one-dimensional mo<strong>de</strong>l. With this insight, our present result<br />
generalizing th<strong>at</strong> of Lux and Sorn<strong>et</strong>te (1999) is n<strong>at</strong>ural.<br />
To our knowledge, there are two possible remedies. The<br />
first one is based on the r<strong>at</strong>ional bubble and crash mo<strong>de</strong>l<br />
of Johansen <strong>et</strong> al (1999, 2000) which abandons the linear<br />
stochastic dynamics (8) in favour of an essentially arbitrary<br />
and nonlinear dynamics controlled by a crash hazard r<strong>at</strong>e.<br />
A jump process for crashes is ad<strong>de</strong>d to the process, with a<br />
crash hazard r<strong>at</strong>e evolving with time such th<strong>at</strong> the r<strong>at</strong>ional<br />
expect<strong>at</strong>ion condition is ensured. This mo<strong>de</strong>l is squarely<br />
based on the r<strong>at</strong>ional expect<strong>at</strong>ion framework and shows th<strong>at</strong><br />
changing the dynamics of the Blanchard and W<strong>at</strong>son mo<strong>de</strong>l<br />
allows s<strong>at</strong>isfying results to be obtained, as the corresponding<br />
r<strong>et</strong>urn distributions can be ma<strong>de</strong> to exhibit reasonable f<strong>at</strong> tails.<br />
The second solution (Sorn<strong>et</strong>te 2001) requires the existence<br />
of an average exponential growth of the fundamental price <strong>at</strong><br />
some r<strong>et</strong>urn r<strong>at</strong>e rf > 0 larger than the discount r<strong>at</strong>e. With<br />
the condition th<strong>at</strong> the price fluctu<strong>at</strong>ions associ<strong>at</strong>ed with bubbles<br />
must on average grow with the mean mark<strong>et</strong> r<strong>et</strong>urn rf , it can be<br />
shown th<strong>at</strong> the exponent of the power law tail of the r<strong>et</strong>urns is<br />
no more boun<strong>de</strong>d by 1 as soon as rf is larger than the discount<br />
r<strong>at</strong>e r and can take essentially arbitrary values. This second<br />
approach amounts to abandoning the r<strong>at</strong>ional pricing theory<br />
(6) with (4) and keeping only the no-arbitrage constraint (7)<br />
on the bubble component. This hypothesis may be true in the<br />
case of a firm, som<strong>et</strong>imes in the case of an industry (railways in<br />
the 19th century, oil and computer in the 20th for instance), but<br />
is hard to <strong>de</strong>fend in the case of an economy as a whole. The<br />
real long run interest r<strong>at</strong>e in the US is approxim<strong>at</strong>ely 3.5%,<br />
and the real r<strong>at</strong>e of growth of profits since World War II is<br />
about 2.1%. Thus, for the economy as a whole, the discounted<br />
sum is always finite. It would be interesting to investig<strong>at</strong>e the<br />
interplay b<strong>et</strong>ween inter-<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween several bubbles<br />
and this exponential growth mo<strong>de</strong>l.<br />
As a final positive remark balancing our neg<strong>at</strong>ive result,<br />
we stress th<strong>at</strong> the stochastic multiplic<strong>at</strong>ive multi-ass<strong>et</strong> r<strong>at</strong>ional<br />
bubble mo<strong>de</strong>l presented here provi<strong>de</strong>s a n<strong>at</strong>ural mechanism for<br />
the existence of a ‘universal’ tail exponent µ across different<br />
mark<strong>et</strong>s.<br />
Acknowledgments<br />
We acknowledge helpful discussions and exchanges with<br />
J P Laurent, T Lux, V Pisarenko and M Taqqu. We also<br />
thank T Mikosch for providing access to Le Page (1983) and<br />
V Pisarenko for a critical reading of the manuscript. Any<br />
remaining error is ours.<br />
Appendix A. Proof of proposition 1 on<br />
the no-arbitrage condition<br />
L<strong>et</strong> t be the value <strong>at</strong> time t of any self-financing portfolio:<br />
t = W ′<br />
t Xt, (A.1)<br />
where W ′<br />
t = (W1,...,Wd) is the vector whose components<br />
are the weights of the different ass<strong>et</strong>s and the prime <strong>de</strong>notes<br />
the transpose. The no-free-lunch condition reads:<br />
t = δ · EQ[t+1|Ft] ∀{t}t0. (A.2)<br />
Therefore, for all self-financing str<strong>at</strong>egies (Wt), wehave<br />
W ′<br />
<br />
t+1 EQ[A] − 1<br />
δ Id<br />
<br />
Xt = 0 ∀ Xt ∈ R d , (A.3)<br />
where we have used the fact th<strong>at</strong> (Wt+1) is (Ft)-measurable<br />
and th<strong>at</strong> the sequence of m<strong>at</strong>rices {At} is i.i.d.<br />
The str<strong>at</strong>egy W ′<br />
t = (0,...,0, 1, 0,...,0) (1 in ith<br />
position), for all t, is self-financing and implies<br />
(ai1,ai2,...,aii − 1<br />
δ ,...,aid)<br />
·(X (1)<br />
t ,X (2)<br />
t ,...,X (i)<br />
t ,...,X (d)<br />
t ) ′ = 0, (A.4)<br />
for all Xt ∈ Rd . We have called aij the (i, j)th coefficient of<br />
the m<strong>at</strong>rix EQ[A]. As a consequence,<br />
(ai1,ai2,...,aii − 1<br />
δ ,...,aid) = 0 ∀i, (A.5)<br />
and<br />
EQ[A] = 1<br />
δ Id. (A.6)<br />
539
40 2. Modèles phénoménologiques <strong>de</strong> cours<br />
Y Malevergne and D Sorn<strong>et</strong>te Q UANTITATIVE F INANCE<br />
The no-arbitrage condition thus implies: EQ[A] = 1<br />
δ Id.<br />
We now show th<strong>at</strong> the converse is true, namely th<strong>at</strong> if<br />
EQ[A] = 1<br />
δ Id is true, then the no-arbitrage condition is<br />
verified. L<strong>et</strong> us thus assume th<strong>at</strong> EQ[A] = 1<br />
δ Id holds. Then<br />
W ′<br />
EQ[Pt+1|Ft] = EQ[W ′<br />
t+1 · Xt+1|Ft] (A.7)<br />
= W ′<br />
t+1 · EQ[Xt+1|Ft] (A.8)<br />
= W ′<br />
t+1 · EQ[At+1Xt + Bt+1|Ft] (A.9)<br />
= W ′<br />
t+1 · EQ[At+1|Ft] · Xt (A.10)<br />
= W ′<br />
t+1 · EQ[A] · Xt<br />
(A.11)<br />
= 1 ′<br />
W t+1<br />
δ · Xt. (A.12)<br />
The condition th<strong>at</strong> the portfolio is self-financing is<br />
t+1 Xt = W ′<br />
t Xt, which means th<strong>at</strong> the weights can be<br />
rebalanced a priori arbitrarily b<strong>et</strong>ween the ass<strong>et</strong>s with the<br />
constraint th<strong>at</strong> the total wealth <strong>at</strong> the same time remains<br />
constant. We can thus write<br />
EQ[t+1|Ft] = 1 ′<br />
W t+1 · Xt (A.13)<br />
δ<br />
= 1<br />
δ t. (A.14)<br />
Therefore, the discounted process {t} is a Q-martingale.<br />
Appendix B. Proof th<strong>at</strong> δ (i) is the<br />
discount factor for the ith bubble<br />
component in the historical space<br />
Here, we express the no-free-lunch condition in the historical<br />
space (or real space). The condition we will obtain is the socalled<br />
‘r<strong>at</strong>ional expect<strong>at</strong>ion condition’, which is a little bit less<br />
general than the condition d<strong>et</strong>ailed in the previous appendix A.<br />
Given the prices {X (i)<br />
k }kt of an ass<strong>et</strong>, labelled by i, until<br />
the d<strong>at</strong>e t, the best estim<strong>at</strong>ion of its price <strong>at</strong> t +1isEP[X (i)<br />
t+1 |Ft].<br />
So, the RE condition leads to<br />
EP[X (i)<br />
t+1 |Ft] − X (i)<br />
t<br />
X (i) = r<br />
t<br />
(i)<br />
t , (A.15)<br />
where r (i)<br />
t is the r<strong>et</strong>urn of the ass<strong>et</strong> i b<strong>et</strong>ween t and t +1. As<br />
previously, we will assume in wh<strong>at</strong> follows th<strong>at</strong> r (i)<br />
t = r (i) is<br />
time in<strong>de</strong>pen<strong>de</strong>nt. Thus, the r<strong>at</strong>ional expect<strong>at</strong>ion condition for<br />
the ass<strong>et</strong>s i reads<br />
<br />
X (i)<br />
t = δ (i) · EPX<br />
= δ (i) · EP<br />
X (i)<br />
t+1 |Ft<br />
<br />
X (i)<br />
t+1 |Ft<br />
, (A.16)<br />
<br />
, (A.17)<br />
where δ (i) is the discount factor.<br />
A priori, each ass<strong>et</strong> has a different r<strong>et</strong>urn. Thus,<br />
introducing the vector ˜Xt whose ith component is X(i) t<br />
δ (i) ,we<br />
can summarize the r<strong>at</strong>ional expect<strong>at</strong>ion condition as<br />
540<br />
˜Xt = EP [Xt+1|Ft] . (A.18)<br />
Again, we evalu<strong>at</strong>e the conditional expect<strong>at</strong>ion of (16),<br />
and using the fact th<strong>at</strong> {At} are i.i.d., we have<br />
<br />
= EP[A]Xt−1. (A.19)<br />
<br />
EP Xt|Ft−1<br />
This equ<strong>at</strong>ion tog<strong>et</strong>her with (A.18), leads to<br />
˜Xt−1 = EP[A]Xt−1, (A.20)<br />
which can be rewritten<br />
<br />
EP[A] − δ−1 <br />
Xt−1 = 0, (A.21)<br />
where δ−1 = diag[δ (1)−1 ...δ (d)−1 ] is the m<strong>at</strong>rix whose ith<br />
diagonal component is δ (i)−1 and 0 elsewhere.<br />
The equ<strong>at</strong>ion (A.21) must be true for every Xt−1 ∈ Rd ,<br />
thus<br />
EP[A] = δ−1 , (A.22)<br />
which is the result announced in section (3.3.2).<br />
Appendix C. Proof of proposition 2 on<br />
the condition κ1 < 1<br />
First step. Behaviour of the function<br />
1 <br />
f(κ)= lim ln EPA ||An ...A1||<br />
n→∞ n κ1<br />
in the interval [0,κ0].<br />
(A.23)<br />
In the fourth step of the proof of theorem 3, Kesten (1973)<br />
shows th<strong>at</strong> the function f has the following properties:<br />
• f is continuous on [0,κ0],<br />
• f(0) = 0 and f(κ0) >0,<br />
• f ′ (0) 1, it is necessary th<strong>at</strong><br />
f(1)
2.1. Bulles r<strong>at</strong>ionnelles multi-dimensionnelles <strong>et</strong> queues épaisses 41<br />
Q UANTITATIVE F INANCE Multi-dimensional r<strong>at</strong>ional bubbles<br />
f(k)<br />
f(1)<br />
0<br />
Figure A1. Schem<strong>at</strong>ic shape of the function f(κ)<strong>de</strong>fined in (A.23).<br />
where ρ(EPA [A]) is the spectral radius of A. By <strong>de</strong>finition,<br />
|| EPA [A] n <br />
|| | EPA [A] n<br />
xmax| =λ n max . (A.29)<br />
Then<br />
1<br />
lim<br />
n→∞ n ln EPA ||An<br />
1<br />
...A1|| lim ln ρ(EPA [A])n<br />
n→∞ n<br />
= ln ρ(EPA [A]) . (A.30)<br />
Now, suppose th<strong>at</strong> ρ(EPA [A]) 1. We obtain<br />
1<br />
f(1) = lim<br />
n→∞ n ln EPA ||An ...A1|| 0, (A.31)<br />
which is in contradiction with the necessary condition (A.24).<br />
Thus,<br />
References<br />
f(1)
42 2. Modèles phénoménologiques <strong>de</strong> cours<br />
2.2 Des bulles r<strong>at</strong>ionnelles aux krachs<br />
Nous étudions <strong>et</strong> généralisons <strong>de</strong> diverses manières le modèle <strong>de</strong> bulles r<strong>at</strong>ionnelles introduit dans la<br />
littér<strong>at</strong>ure économique par Blanchard <strong>et</strong> W<strong>at</strong>son (1982). Les bulles sont présentées comme équivalentes<br />
aux mo<strong>de</strong>s <strong>de</strong> Goldstone du prix fondamental <strong>de</strong> l’équ<strong>at</strong>ion <strong>de</strong> “pricing” r<strong>at</strong>ionnel, associés à la brisure<br />
<strong>de</strong> symétrie liée à l’existence d’un divi<strong>de</strong>n<strong>de</strong> non nul. Généralisant les bulles en terme <strong>de</strong> processus stochastique<br />
multiplic<strong>at</strong>if, nous résumons le résult<strong>at</strong> <strong>de</strong> Lux <strong>et</strong> Sorn<strong>et</strong>te (2002) selon lequel la condition<br />
<strong>de</strong> non arbitrage impose que la queue <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ment est régulièrement variable, d’indice<br />
<strong>de</strong> queue µ < 1. Nous rappelons ensuite le principal résult<strong>at</strong> <strong>de</strong> Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2001a),<br />
qui étend le modèle <strong>de</strong> bulles r<strong>at</strong>ionnelles à un nombre quelconque <strong>de</strong> dimensions d : un nombre d<br />
<strong>de</strong> séries financières sont rendues linéairement interdépendantes via une m<strong>at</strong>rice <strong>de</strong> couplage alé<strong>at</strong>oire.<br />
Nous dérivons la condition <strong>de</strong> non arbitrage dans ce contexte <strong>et</strong>, à l’ai<strong>de</strong> <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> renouvellement<br />
<strong>de</strong>s m<strong>at</strong>rices alé<strong>at</strong>oires, nous étendons le théorème <strong>de</strong> Lux <strong>et</strong> Sorn<strong>et</strong>te (2002) <strong>et</strong> démontrons que les<br />
queues <strong>de</strong>s distributions inconditionnelles <strong>de</strong>s ren<strong>de</strong>ments <strong>de</strong> chaque actif sont régulièrement variables<br />
<strong>et</strong> <strong>de</strong> même indice <strong>de</strong> queue inférieur à un. Bien que le comportement asymptotique en loi <strong>de</strong> puissance<br />
<strong>de</strong>s queues <strong>de</strong> distributions soit une caractéristique marquante ressortant <strong>de</strong> l’étu<strong>de</strong> <strong>de</strong>s données empiriques,<br />
l’exposant <strong>de</strong> queue inférieur à un est en contradiction flagrante avec l’estim<strong>at</strong>ion empirique<br />
généralement admise qui est <strong>de</strong> l’ordre <strong>de</strong> trois. Nous discutons alors <strong>de</strong>ux extensions du modèle <strong>de</strong><br />
bulles r<strong>at</strong>ionnelles en accord avec les faits stylisés.<br />
Reprint from : D. Sorn<strong>et</strong>te <strong>et</strong> Y. Malevergne (2001),“From r<strong>at</strong>ional bubbles to crashes”, Physica A 299,<br />
40-59.
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 43<br />
Physica A 299 (2001) 40–59<br />
From r<strong>at</strong>ional bubbles to crashes<br />
www.elsevier.com/loc<strong>at</strong>e/physa<br />
D. Sorn<strong>et</strong>te a;b; ∗ , Y. Malevergne a<br />
a Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>iere Con<strong>de</strong>nsee, CNRS UMR 6622 and Universite <strong>de</strong> Nice-Sophia<br />
Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
b Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science,<br />
University of California, Los Angeles, CA 90095, USA<br />
Abstract<br />
We study and generalize in various ways the mo<strong>de</strong>l of r<strong>at</strong>ional expect<strong>at</strong>ion (RE) bubbles<br />
introduced by Blanchard and W<strong>at</strong>son in the economic liter<strong>at</strong>ure. Bubbles are argued to be the<br />
equivalent of Goldstone mo<strong>de</strong>s of the fundamental r<strong>at</strong>ional pricing equ<strong>at</strong>ion, associ<strong>at</strong>ed with<br />
the symm<strong>et</strong>ry-breaking introduced by non-vanishing divi<strong>de</strong>nds. Generalizing bubbles in terms of<br />
multiplic<strong>at</strong>ive stochastic maps, we summarize the result of Lux and Sorn<strong>et</strong>te th<strong>at</strong> the no-arbitrage<br />
condition imposes th<strong>at</strong> the tail of the r<strong>et</strong>urn distribution is hyperbolic with an exponent ¡1.<br />
We then outline the main results of Malevergne and Sorn<strong>et</strong>te, who extend the RE bubble mo<strong>de</strong>l<br />
to arbitrary dimensions d: a number d of mark<strong>et</strong> time series are ma<strong>de</strong> linearly inter<strong>de</strong>pen<strong>de</strong>nt<br />
via d × d stochastic coupling coe cients. We <strong>de</strong>rive the no-arbitrage condition in this context<br />
and, with the renewal theory for products of random m<strong>at</strong>rices applied to stochastic recurrence<br />
equ<strong>at</strong>ions, we extend the theorem of Lux and Sorn<strong>et</strong>te to <strong>de</strong>monstr<strong>at</strong>e th<strong>at</strong> the tails of the<br />
unconditional distributions associ<strong>at</strong>ed with such d-dimensional bubble processes follow power<br />
laws, with the same asymptotic tail exponent ¡1 for all ass<strong>et</strong>s. The distribution of price<br />
di erences and of r<strong>et</strong>urns is domin<strong>at</strong>ed by the same power-law over an exten<strong>de</strong>d range of<br />
large r<strong>et</strong>urns. Although power-law tails are a pervasive fe<strong>at</strong>ure of empirical d<strong>at</strong>a, the numerical<br />
value ¡1 is in disagreement with the usual empirical estim<strong>at</strong>es ≈ 3. We then discuss two<br />
extensions (the crash hazard r<strong>at</strong>e mo<strong>de</strong>l and the non-st<strong>at</strong>ionary growth r<strong>at</strong>e mo<strong>de</strong>l) of the RE<br />
bubble mo<strong>de</strong>l th<strong>at</strong> provi<strong>de</strong> two ways of reconcili<strong>at</strong>ion with the stylized facts of nancial d<strong>at</strong>a.<br />
c○ 2001 Elsevier Science B.V. All rights reserved.<br />
1. The mo<strong>de</strong>l of r<strong>at</strong>ional bubbles<br />
Blanchard [1] and Blanchard and W<strong>at</strong>son [2] originally introduced the mo<strong>de</strong>l of<br />
r<strong>at</strong>ional expect<strong>at</strong>ions (RE) bubbles to account for the possibility, often discussed in<br />
∗ Corresponding author. Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>iere Con<strong>de</strong>nsee, CNRS UMR 6622 and Universite<br />
<strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France. Fax: +33-4-92-07-67-54.<br />
E-mail addresses: sorn<strong>et</strong>te@unice.fr (D. Sorn<strong>et</strong>te), yannick.malevergne@unice.fr (Y. Malevergne).<br />
0378-4371/01/$ - see front m<strong>at</strong>ter c○ 2001 Elsevier Science B.V. All rights reserved.<br />
PII: S 0378-4371(01)00281-3
44 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 41<br />
the empirical liter<strong>at</strong>ure and by practitioners, th<strong>at</strong> observed prices may <strong>de</strong>vi<strong>at</strong>e signi -<br />
cantly and over exten<strong>de</strong>d time intervals from fundamental prices. While allowing for<br />
<strong>de</strong>vi<strong>at</strong>ions from fundamental prices, r<strong>at</strong>ional bubbles keep a fundamental anchor point<br />
of economic mo<strong>de</strong>lling, namely th<strong>at</strong> bubbles must obey the condition of r<strong>at</strong>ional expect<strong>at</strong>ions.<br />
In contrast, recent works stress th<strong>at</strong> investors are not fully r<strong>at</strong>ional, or have<br />
<strong>at</strong> most boun<strong>de</strong>d r<strong>at</strong>ionality, and th<strong>at</strong> behavioral and psychological mechanisms, such<br />
as herding, may be important in the shaping of mark<strong>et</strong> prices [3–5]. However, for<br />
uid ass<strong>et</strong>s, dynamic investment str<strong>at</strong>egies rarely perform over simple buy-and-hold<br />
str<strong>at</strong>egies [6], in other words, the mark<strong>et</strong> is not far from being e cient and little arbitrage<br />
opportunities exist as a result of the constant search for gains by sophistic<strong>at</strong>ed<br />
investors. Here, we shall work within the conditions of r<strong>at</strong>ional expect<strong>at</strong>ions and of<br />
no-arbitrage condition, taken as useful approxim<strong>at</strong>ions. In<strong>de</strong>ed, the r<strong>at</strong>ionality of both<br />
expect<strong>at</strong>ions and behavior often does not imply th<strong>at</strong> the price of an ass<strong>et</strong> be equal to its<br />
fundamental value. In other words, there can be r<strong>at</strong>ional <strong>de</strong>vi<strong>at</strong>ions of the price from<br />
this value, called r<strong>at</strong>ional bubbles. A r<strong>at</strong>ional bubble can arise when the actual mark<strong>et</strong><br />
price <strong>de</strong>pends positively on its own expected r<strong>at</strong>e of change, as som<strong>et</strong>imes occurs in<br />
ass<strong>et</strong> mark<strong>et</strong>s, which is the mechanism un<strong>de</strong>rlying the mo<strong>de</strong>ls of Refs. [1,2].<br />
In or<strong>de</strong>r to avoid the unrealistic picture of ever-increasing <strong>de</strong>vi<strong>at</strong>ions from fundamental<br />
values, Blanchard [2] proposed a mo<strong>de</strong>l with periodically collapsing bubbles in<br />
which the bubble component of the price follows an exponential explosive p<strong>at</strong>h (the<br />
price being multiplied by <strong>at</strong> =a¿1) with probability and collapses to zero (the price<br />
being multiplied by <strong>at</strong> = 0) with probability 1− . It is clear th<strong>at</strong>, in this mo<strong>de</strong>l, a bubble<br />
has an exponential distribution of lif<strong>et</strong>imes with a nite average lif<strong>et</strong>ime =(1 − ).<br />
Bubbles are thus transient phenomena. The condition of r<strong>at</strong>ional expect<strong>at</strong>ions imposes<br />
th<strong>at</strong> a =1=( ), where is the discount factor. In or<strong>de</strong>r to allow for the start of new<br />
bubbles after the collapse, a stochastic zero mean normally distributed component bt<br />
is ad<strong>de</strong>d to the system<strong>at</strong>ic part of Xt. This leads to the following dynamical equ<strong>at</strong>ion<br />
Xt+1 = <strong>at</strong>Xt + bt ; (1)<br />
where, as we said, <strong>at</strong> =a with probability and <strong>at</strong> = 0 with probability 1 − . Both<br />
variables <strong>at</strong> and bt do not <strong>de</strong>pend on the process Xt. There is a huge liter<strong>at</strong>ure on theor<strong>et</strong>ical<br />
re nements of this mo<strong>de</strong>l and on the empirical d<strong>et</strong>ectability of RE bubbles in<br />
nancial d<strong>at</strong>a (see Refs. [7,8], for surveys of this liter<strong>at</strong>ure). Mo<strong>de</strong>l (1) has also been<br />
explored in a large vari<strong>et</strong>y of contexts, for instance in ARCH processes in econom<strong>et</strong>ry<br />
[9], 1D random- eld Ising mo<strong>de</strong>ls [10] using Mellin transforms, and more recently<br />
using extremal properties of the G-harmonic functions on non-compact groups [11]<br />
and the Wiener–Hopf technique [12]. See also Ref. [13] for a short review of other<br />
domains of applic<strong>at</strong>ions including popul<strong>at</strong>ion dynamics with external sources, epi<strong>de</strong>mics,<br />
immigr<strong>at</strong>ion and investment portfolios, the intern<strong>et</strong>, directed polymers in<br />
random media ::: :<br />
Large |Xk| are gener<strong>at</strong>ed by intermittent ampli c<strong>at</strong>ions resulting from the multiplic<strong>at</strong>ion<br />
by several successive values of |a| larger than one. We now o er a simple<br />
“mean- eld” type argument th<strong>at</strong> clari es the origin of the power law f<strong>at</strong> tail. L<strong>et</strong> us
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 45<br />
42 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
call p¿ the probability th<strong>at</strong> the absolute value of the multiplic<strong>at</strong>ive factor a is found<br />
larger than 1. The probability to observe n successive multiplic<strong>at</strong>ive factors |a| larger<br />
than 1 is thus pn ¿. L<strong>et</strong> us call |a¿| the average of |a| conditioned on being larger than<br />
1:|a¿| is thus the typical absolute value of the ampli c<strong>at</strong>ion factor. When n successive<br />
multiplic<strong>at</strong>ive factors occur with absolute values larger than 1, they typically lead to<br />
an ampli c<strong>at</strong>ion of the amplitu<strong>de</strong> of X by |a¿| n . Using the fact th<strong>at</strong> the additive term<br />
bk ensures th<strong>at</strong> the amplitu<strong>de</strong> of Xk remains of the or<strong>de</strong>r of the standard <strong>de</strong>vi<strong>at</strong>ion or<br />
of other measures of typical scales b of the distribution Pb(b) when the multiplic<strong>at</strong>ive<br />
factors |a| are less than 1, this shows th<strong>at</strong> a value of Xk of the or<strong>de</strong>r of |X |≈ b|a¿| n<br />
occurs with probability<br />
p n <br />
<br />
ln |X |= b 1<br />
¿ = exp (n ln p¿) ≈ exp ln p¿ =<br />
(2)<br />
ln |a¿| (|X |= b)<br />
with =lnp¿=ln |a¿|, which can be rewritten as p¿|a¿| = 1. Note the similarity<br />
b<strong>et</strong>ween this last “mean- eld” equ<strong>at</strong>ion and the exact solution (8) given below. The<br />
power law distribution is thus the result of an exponentially small probability of cre<strong>at</strong>ing<br />
an exponentially large value [14]. Expression (2) does not provi<strong>de</strong> a precise<br />
d<strong>et</strong>ermin<strong>at</strong>ion of the exponent , only an approxim<strong>at</strong>e one since we have used a kind<br />
of mean- eld argument in the <strong>de</strong> nition of |a¿|.<br />
In the next section, we recall how bubbles appear as possible solutions of the fundamental<br />
pricing equ<strong>at</strong>ion and play the role of Goldstone mo<strong>de</strong>s of a price-symm<strong>et</strong>ry<br />
broken by the divi<strong>de</strong>nd ow. We then <strong>de</strong>scribe the Kesten generaliz<strong>at</strong>ion of r<strong>at</strong>ional<br />
bubbles in terms of random multiplic<strong>at</strong>ive maps and present the fundamental result [15]<br />
th<strong>at</strong> the no-arbitrage condition leads to the constraint th<strong>at</strong> the exponent of the power<br />
law tail is less than 1. We then present an extension to arbitrary multidimensional<br />
random multiplic<strong>at</strong>ive maps: a number d of mark<strong>et</strong> time series are ma<strong>de</strong> linearly inter<strong>de</strong>pen<strong>de</strong>nt<br />
via d × d stochastic coupling coe cients. We show th<strong>at</strong> the no-arbitrage<br />
condition imposes th<strong>at</strong> the non-diagonal impacts of any ass<strong>et</strong> i on any other ass<strong>et</strong> j = i<br />
has to vanish on average, i.e., must exhibit random altern<strong>at</strong>ive regimes of reinforcement<br />
and contrarian feedbacks. In contrast, the diagonal terms must be positive and<br />
equal on average to the inverse of the discount factor. Applying the results of renewal<br />
theory for products of random m<strong>at</strong>rices to stochastic recurrence equ<strong>at</strong>ions (SRE), we<br />
extend the theorem of Ref. [15] and <strong>de</strong>monstr<strong>at</strong>e th<strong>at</strong> the tails of the unconditional<br />
distributions associ<strong>at</strong>ed with such d-dimensional bubble processes follow power laws<br />
(i.e., exhibit hyperbolic <strong>de</strong>cline), with the same asymptotic tail exponent ¡1 for all<br />
ass<strong>et</strong>s. The distribution of price di erences and of r<strong>et</strong>urns is domin<strong>at</strong>ed by the same<br />
power-law over an exten<strong>de</strong>d range of large r<strong>et</strong>urns. In or<strong>de</strong>r to unlock the paradox,<br />
we brie y discuss the crash hazard r<strong>at</strong>e mo<strong>de</strong>l [16,17] and the non-st<strong>at</strong>ionary growth<br />
mo<strong>de</strong>l [18]. We conclu<strong>de</strong> by proposing a link with the theory of specul<strong>at</strong>ive pricing<br />
through a spontaneous symm<strong>et</strong>ry-breaking [19].<br />
We should stress th<strong>at</strong>, due to the no-arbitrage condition th<strong>at</strong> forms the backbone of<br />
our theor<strong>et</strong>ical approach, correl<strong>at</strong>ions of r<strong>et</strong>urns are vanishing. In addition, the multiplic<strong>at</strong>ive<br />
stochastic structure of the mo<strong>de</strong>ls ensures the phenomenon of vol<strong>at</strong>ility
46 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 43<br />
clustering. These two stylized facts, taken for granted in our present approach, will<br />
not be discussed further.<br />
2. R<strong>at</strong>ional bubbles of an isol<strong>at</strong>ed ass<strong>et</strong> [15]<br />
2.1. R<strong>at</strong>ional expect<strong>at</strong>ion bubble mo<strong>de</strong>l and Goldstone mo<strong>de</strong>s<br />
We rst brie y recall th<strong>at</strong> pricing of an ass<strong>et</strong> un<strong>de</strong>r r<strong>at</strong>ional expect<strong>at</strong>ions theory is<br />
based on the two following hypothesis: the r<strong>at</strong>ionality of the agents and the “no-free<br />
lunch” condition.<br />
Un<strong>de</strong>r the r<strong>at</strong>ional expect<strong>at</strong>ion condition, the best estim<strong>at</strong>ion of the price pt+1 of<br />
an ass<strong>et</strong> <strong>at</strong> time t + 1 viewed from time t is given by the expect<strong>at</strong>ion of pt+1 conditioned<br />
upon the knowledge of the ltr<strong>at</strong>ion {Ft} (i.e., sum of all available inform<strong>at</strong>ion<br />
accumul<strong>at</strong>ed) up to time t: E[pt+1|Ft].<br />
The “no-free lunch” condition imposes th<strong>at</strong> the expected r<strong>et</strong>urns of every ass<strong>et</strong>s are<br />
all equal un<strong>de</strong>r a given probability measure Q equivalent to the historical probability<br />
measure P. In particular, the expected r<strong>et</strong>urn of each ass<strong>et</strong> is equal to the r<strong>et</strong>urn r of<br />
the risk-free ass<strong>et</strong> (which is assumed to exist), and thus the probability measure Q is<br />
named the risk neutral probability measure.<br />
Putting tog<strong>et</strong>her these two conditions, we are led to the following valu<strong>at</strong>ion formula<br />
for the price pt:<br />
pt = EQ[pt+1|Ft]+dt ∀{pt}t¿0 ; (3)<br />
where dt is an exogeneous “divi<strong>de</strong>nd”, and =(1+r) −1 is the discount factor. The rst<br />
term in the r.h.s. quanti es the usual fact th<strong>at</strong> som<strong>et</strong>hing tomorrow is less valuable than<br />
today by a factor called the discount factor. Intuitively, the second term, the divi<strong>de</strong>nd,<br />
is ad<strong>de</strong>d to express the fact th<strong>at</strong> the expected price tomorrow has to be <strong>de</strong>creased by<br />
the divi<strong>de</strong>nd since the value before giving the divi<strong>de</strong>nd incorpor<strong>at</strong>es it in the pricing.<br />
The “forward” solution of (3) is well-known to be the fundamental price<br />
p f<br />
t =<br />
+∞<br />
i=0<br />
i EQ[dt+i|Ft] : (4)<br />
It is straightforward to check by replacement th<strong>at</strong> the sum of the forward solution (4)<br />
and of an arbitrary component Xt<br />
pt = p f<br />
t + Xt ; (5)<br />
where Xt has to obey the single condition of being an arbitrary martingale<br />
Xt = EQ[Xt+1|Ft] (6)<br />
is also the solution of (3). In fact, it can be shown [20] th<strong>at</strong> (5) is the general solution<br />
of (3).<br />
Here, it is important to note th<strong>at</strong>, in the framework of the Blanchard and W<strong>at</strong>son<br />
mo<strong>de</strong>l, the specul<strong>at</strong>ive bubbles appear as a n<strong>at</strong>ural consequence of the valu<strong>at</strong>ion formula
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 47<br />
44 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
(3), i.e., of the no free-lunch condition and the r<strong>at</strong>ionality of the agents. Thus, the<br />
concept of bubbles is not an addition to the theory, as som<strong>et</strong>imes believed, but is<br />
entirely embed<strong>de</strong>d in it.<br />
Notice also th<strong>at</strong> the component Xt in (5) plays a role analogous to the Goldstone<br />
mo<strong>de</strong> in nuclear, particle and con<strong>de</strong>nsed-m<strong>at</strong>ter physics [21,22]. Goldstone mo<strong>de</strong>s are<br />
the zero-wavenumber zero-energy modal uctu<strong>at</strong>ions th<strong>at</strong> <strong>at</strong>tempt to restore a broken<br />
symm<strong>et</strong>ry. For instance, consi<strong>de</strong>r a Bloch wall b<strong>et</strong>ween two semi-in nite magn<strong>et</strong>ic<br />
domains of opposite spin directions selected by opposite magn<strong>et</strong>ic eld <strong>at</strong> boundaries<br />
far away. At non-zero temper<strong>at</strong>ure, capillary waves are excited by thermal uctu<strong>at</strong>ions.<br />
The limit of very long-wavelength capillary mo<strong>de</strong>s correspond to Goldstone mo<strong>de</strong>s th<strong>at</strong><br />
tend to restore the transl<strong>at</strong>ional symm<strong>et</strong>ry broken by the presence of the Bloch wall<br />
[23].<br />
In the present context, as shown in Ref. [19], the<br />
p →−p parity symm<strong>et</strong>ry (7)<br />
is broken by the “external” eld embodied in the divi<strong>de</strong>nd ow dt. In<strong>de</strong>ed, as can be<br />
seen from (3) and its forward solution (4), the fundamental price is i<strong>de</strong>ntically zero in<br />
absence of divi<strong>de</strong>nds. Ref. [19] has stressed the fact th<strong>at</strong> it makes perfect sense to think<br />
of neg<strong>at</strong>ive prices. For instance, we are ready to pay a (positive) price for a commodity<br />
th<strong>at</strong> we need or like. However, we will not pay a positive price to g<strong>et</strong> som<strong>et</strong>hing we<br />
dislike or which disturb us, such as garbage, waste, broken and useless car, chemical<br />
and industrial hazards, <strong>et</strong>c. Consi<strong>de</strong>r a chunk of waste. We will be ready to buy it<br />
for a neg<strong>at</strong>ive price, in other words, we are ready to take the unwanted commodity if<br />
it comes with cash. Positive divi<strong>de</strong>nds imply positive prices, neg<strong>at</strong>ive divi<strong>de</strong>nds lead<br />
to neg<strong>at</strong>ive prices. Neg<strong>at</strong>ive divi<strong>de</strong>nds correspond to the premium to pay to keep an<br />
ass<strong>et</strong> for instance. From an economic view point, wh<strong>at</strong> makes a share of a company<br />
<strong>de</strong>sirable is its earnings, th<strong>at</strong> provi<strong>de</strong> divi<strong>de</strong>nds, and its potential appreci<strong>at</strong>ion th<strong>at</strong> give<br />
rise to capital gains. As a consequence, in absence of divi<strong>de</strong>nds and of specul<strong>at</strong>ion,<br />
the price of share must be nil and the symm<strong>et</strong>ry (7) holds. The earnings leading to<br />
divi<strong>de</strong>nds d thus act as a symm<strong>et</strong>ry-breaking “ eld”, since a positive d makes the share<br />
<strong>de</strong>sirable and thus <strong>de</strong>velop a positive price.<br />
It is now clear th<strong>at</strong> the addition of the bubble Xt, which can be anything but for<br />
the martingale condition (6), is playing the role of the Goldstone mo<strong>de</strong>s restoring the<br />
broken symm<strong>et</strong>ry: the bubble price can wan<strong>de</strong>r up or down and, in the limit where it<br />
becomes very large in absolute value, domin<strong>at</strong>e over the fundamental price, restoring<br />
the in<strong>de</strong>pen<strong>de</strong>nce with respect to divi<strong>de</strong>nd. Moreover, as in con<strong>de</strong>nsed-m<strong>at</strong>ter physics<br />
where the Goldstone mo<strong>de</strong> appears spontaneously since it has no energy cost, the<br />
r<strong>at</strong>ional bubble itself can appear spontaneously with no divi<strong>de</strong>nd.<br />
The “bubble” Goldstone mo<strong>de</strong> turns out to be intim<strong>at</strong>ely rel<strong>at</strong>ed to the “money” Goldstone<br />
mo<strong>de</strong> introduced by Bak <strong>et</strong> al. [24]. Ref. [24] introduces a dynamical many-body<br />
theory of money, in which the value of money in equilibrium is not xed by the<br />
equ<strong>at</strong>ions, and thus obeys a continuous symm<strong>et</strong>ry. The dynamics breaks this continuous<br />
symm<strong>et</strong>ry by x<strong>at</strong>ing the value of money <strong>at</strong> a level which <strong>de</strong>pends on initial
48 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 45<br />
conditions. The uctu<strong>at</strong>ions around the equilibrium, for instance in the presence of<br />
noise, are governed by the Goldstone mo<strong>de</strong>s associ<strong>at</strong>ed with the broken symm<strong>et</strong>ry. In<br />
apparent contrast, a bubble represents the arbitrary <strong>de</strong>vi<strong>at</strong>ion from fundamental valu<strong>at</strong>ion.<br />
Introducing money, a given valu<strong>at</strong>ion or price is equivalent to a certain amount<br />
of money. A growing bubble thus corresponds to the same ass<strong>et</strong> becoming equivalent<br />
to more and more cash. Equivalently, from the point of view of the ass<strong>et</strong>, this can be<br />
seen as cash <strong>de</strong>valu<strong>at</strong>ion, i.e., in <strong>at</strong>ion. The “bubble” Goldstone mo<strong>de</strong> and the “money”<br />
Goldstone mo<strong>de</strong> are thus two fac<strong>et</strong>s of the same fundamental phenomenon: they both<br />
are left unconstrained by the valu<strong>at</strong>ion equ<strong>at</strong>ions.<br />
2.2. The no-arbitrage condition and f<strong>at</strong> tails<br />
Following Ref. [15], we study the implic<strong>at</strong>ions of the RE bubble mo<strong>de</strong>ls for the<br />
unconditional distribution of prices, price changes and r<strong>et</strong>urns resulting from a general<br />
discr<strong>et</strong>e-time formul<strong>at</strong>ion extending (1) by allowing the multiplic<strong>at</strong>ive factor <strong>at</strong> to take<br />
arbitrary values and be i.i.d. random variables drawn from some non-<strong>de</strong>gener<strong>at</strong>e probability<br />
<strong>de</strong>nsity function (pdf) Pa(a). The mo<strong>de</strong>l can also be generalized by consi<strong>de</strong>ring<br />
non-normal realiz<strong>at</strong>ions of bt with distribution Pb(b) with EP[bt] = 0, where EP[ · ]is<br />
the unconditional expect<strong>at</strong>ion with respect to the probability measure P.<br />
Provi<strong>de</strong>d EP[ln a] ¡ 0 (st<strong>at</strong>ionarity condition) and if there is a number such th<strong>at</strong><br />
0 ¡EP[|b| ] ¡ + ∞, such th<strong>at</strong><br />
EP[|a| ]=1 (8)<br />
and such th<strong>at</strong> EP[|a| ln |a|] ¡ + ∞, then the tail of the distribution of X is asymptotically<br />
(for large X ’s) a power law [25,26]<br />
PX (X )dX ≈ C<br />
dX ; (9)<br />
|X | 1+<br />
with an exponent given by the real positive solution of (8).<br />
R<strong>at</strong>ional expect<strong>at</strong>ions require in addition th<strong>at</strong> the bubble component of ass<strong>et</strong> prices<br />
obeys the “no free-lunch” condition<br />
EQ[Xt+1|Ft]=Xt<br />
(10)<br />
where ¡1 is the discount factor and the expect<strong>at</strong>ion is taken conditional on the<br />
knowledge of the<br />
rst<br />
ltr<strong>at</strong>ion (inform<strong>at</strong>ion) until time t. Condition (10) with (1) imposes<br />
and then<br />
EQ[a]=1= ¿ 1 ; (11)<br />
EP[a] ¿ 1 ; (12)<br />
on the distribution of the multiplic<strong>at</strong>ive factors <strong>at</strong>.<br />
Consi<strong>de</strong>r the function<br />
M( )=EP[a ] : (13)
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 49<br />
46 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
Fig. 1. Convexity of M( ). This enforces th<strong>at</strong> the exponent solution of (8) is ¡1.<br />
It has the following properties:<br />
(1) M(0) = 1 by <strong>de</strong> nition,<br />
(2) M ′ (0) = EP[ln a] ¡ 0 from the st<strong>at</strong>ionarity condition,<br />
(3) M ′′ ( )=EP[(ln a) 2 |a| ] ¿ 0, by the positivity of the square,<br />
(4) M(1)=1= ¿ 1 by the no-arbitrage result (12).<br />
M( ) is thus convex and is shown in Fig. 1. This <strong>de</strong>monstr<strong>at</strong>e th<strong>at</strong> ¡1 autom<strong>at</strong>ically<br />
(see Ref. [15] for a d<strong>et</strong>ailed <strong>de</strong>monstr<strong>at</strong>ion). It is easy to show [15] th<strong>at</strong> the<br />
distribution of price di erences has the same power law tail with the exponent ¡1<br />
and the distribution of r<strong>et</strong>urns is domin<strong>at</strong>ed by the same power-law over an exten<strong>de</strong>d<br />
range of large r<strong>et</strong>urns [15]. Although power-law tails are a pervasive fe<strong>at</strong>ure of empirical<br />
d<strong>at</strong>a, these characteriz<strong>at</strong>ions are in strong disagreement with the usual empirical<br />
estim<strong>at</strong>es which nd ≈ 3 [27–31]. Lux and Sorn<strong>et</strong>te [15] conclu<strong>de</strong>d th<strong>at</strong> exogenous<br />
r<strong>at</strong>ional bubbles are thus hardly reconcilable with some of the stylized facts of nancial<br />
d<strong>at</strong>a <strong>at</strong> a very elementary level.<br />
3. Generaliz<strong>at</strong>ion of r<strong>at</strong>ional bubbles to arbitrary dimensions [32]<br />
3.1. Generaliz<strong>at</strong>ion to several coupled ass<strong>et</strong>s<br />
In reality, there is no such thing as an isol<strong>at</strong>ed ass<strong>et</strong>. Stock mark<strong>et</strong>s exhibit a vari<strong>et</strong>y<br />
of inter-<strong>de</strong>pen<strong>de</strong>nces, based in part on the mutual in uences b<strong>et</strong>ween the USA,<br />
European and Japanese mark<strong>et</strong>s. In addition, individual stocks may be sensitive to<br />
the behavior of the speci c industry as a whole to which they belong and to a few<br />
other indic<strong>at</strong>ors, such as the main indices, interest r<strong>at</strong>es and so on. Mantegna <strong>et</strong> al.<br />
[33,34] have in<strong>de</strong>ed shown the existence of a hierarchical organiz<strong>at</strong>ion of stock inter<strong>de</strong>pen<strong>de</strong>nces.<br />
Furthermore, bubbles often appear to be not isol<strong>at</strong>ed fe<strong>at</strong>ures of a s<strong>et</strong> of<br />
mark<strong>et</strong>s. For instance, Ref. [35] tested wh<strong>et</strong>her a bubble simultaneously existed across
50 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 47<br />
the n<strong>at</strong>ions, such as Germany, Poland, and Hungary, th<strong>at</strong> experienced hyperin <strong>at</strong>ion<br />
in the early 1920s. Coordin<strong>at</strong>ed bubbles can som<strong>et</strong>imes be d<strong>et</strong>ected. One of the most<br />
prominent example is found in the mark<strong>et</strong> appreci<strong>at</strong>ions observed in many of the world<br />
mark<strong>et</strong>s prior to the world mark<strong>et</strong> crash in October 1987 [36]. Similar intermittent<br />
coordin<strong>at</strong>ion of bubbles have been d<strong>et</strong>ected among the signi cant bubbles followed<br />
by large crashes or severe corrections in L<strong>at</strong>in American and Asian stock mark<strong>et</strong>s<br />
[37]. It is therefore <strong>de</strong>sirable to generalize the one-dimensional RE bubble mo<strong>de</strong>l (1)<br />
to the multi-dimensional case. One could also hope a priori th<strong>at</strong> this generaliz<strong>at</strong>ion<br />
would modify the result ¡1 obtained in the one-dimensional case and allow for a<br />
b<strong>et</strong>ter a<strong>de</strong>qu<strong>at</strong>ion with empirical results. In<strong>de</strong>ed, 1d-systems are well-known to exhibit<br />
anamalous properties often not shared by higher dimensional systems. Here however,<br />
it turns out th<strong>at</strong> the same result ¡1 holds, as we shall see.<br />
In the case of several ass<strong>et</strong>s, r<strong>at</strong>ional pricing theory again dict<strong>at</strong>es th<strong>at</strong> the fundamental<br />
price of each individual ass<strong>et</strong> is given by a formula like (3), where the speci c<br />
divi<strong>de</strong>nd ow of each ass<strong>et</strong> is used, with the same discount factor. The corresponding<br />
forward solution (4) is again valid for each ass<strong>et</strong>. The general solution for each ass<strong>et</strong><br />
is (5) with a bubble component Xt di erent from an ass<strong>et</strong> to the next. The di erent<br />
bubble components can be coupled, as we shall see, but they must each obey the martingale<br />
condition (6), component by component. This imposes speci c conditions on<br />
the coupling terms, as we shall see.<br />
Following this reasoning, we can therefore propose the simplest generaliz<strong>at</strong>ion of a<br />
bubble into a “two-dimensional” bubble for two ass<strong>et</strong>s X and Y with bubble prices, respectively,<br />
equal to Xt and Yt <strong>at</strong> time t. We express the generaliz<strong>at</strong>ion of the Blanchard–<br />
W<strong>at</strong>son mo<strong>de</strong>l as follows:<br />
Xt+1 = <strong>at</strong>Xt + btYt + t ; (14)<br />
Yt+1 = ctXt + dtYk + t ; (15)<br />
where <strong>at</strong>, bt, ct and dt are drawn from some multivari<strong>at</strong>e probability <strong>de</strong>nsity function.<br />
The two additive noises t and t are also drawn from some distribution function with<br />
zero mean. The diagonal case bt = ct = 0 for all t recovers the previous one-dimensional<br />
case with two uncoupled bubbles, provi<strong>de</strong>d t and t are in<strong>de</strong>pen<strong>de</strong>nt.<br />
R<strong>at</strong>ional expect<strong>at</strong>ions require th<strong>at</strong> Xt and Yt obey both the “no-free lunch”<br />
condition (10), i.e., · EQ[Xt+1|Ft]=Xt and · EQ[Yt+1|Ft]=Yt. With (14; 15); this<br />
gives<br />
(EQ[<strong>at</strong>] − −1 )Xt + EQ[bt]Yt =0; (16)<br />
EQ[ct]Xt +(EQ[dt] − −1 )Yt =0; (17)<br />
where we have used th<strong>at</strong> t and t are centered. The two equ<strong>at</strong>ions (16; 17) must<br />
be true for all times, i.e., for all values of Xt and Yt visited by the dynamics. This
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 51<br />
48 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
imposes EQ[bt]=EQ[ct]=0 and EQ[<strong>at</strong>]=EQ[dt]= −1 . We are going to r<strong>et</strong>rieve this<br />
result more formally in the general case.<br />
3.2. General formul<strong>at</strong>ion<br />
A generaliz<strong>at</strong>ion to arbitrary dimensions leads to the following stochastic random<br />
equ<strong>at</strong>ion (SRE):<br />
Xt = AtXt−1 + Bt ; (18)<br />
where (Xt; Bt) are d-dimensional vectors. Each component of Xt can be thought of<br />
as the price of an ass<strong>et</strong> above its fundamental price. The m<strong>at</strong>rices (At) are i<strong>de</strong>ntically<br />
in<strong>de</strong>pen<strong>de</strong>nt distributed d × d-dimensional stochastic m<strong>at</strong>rices. We assume th<strong>at</strong> Bt are<br />
i<strong>de</strong>ntically in<strong>de</strong>pen<strong>de</strong>nt distributed random vectors and th<strong>at</strong> (Xt) is a causal st<strong>at</strong>ionary<br />
solution of (18). Generaliz<strong>at</strong>ions introducing additional arbitrary linear terms <strong>at</strong> larger<br />
time lags such as Xt−2;::: can be tre<strong>at</strong>ed with slight modi c<strong>at</strong>ions of our approach<br />
and yield the same conclusions. We shall thus con ne our <strong>de</strong>monstr<strong>at</strong>ion on the SRE<br />
of or<strong>de</strong>r 1, keeping in mind th<strong>at</strong> our results apply analogously to arbitrary or<strong>de</strong>rs of<br />
regressions.<br />
In the following, we <strong>de</strong>note by |·| the Eucli<strong>de</strong>an norm and by · the corresponding<br />
norm for any d × d-m<strong>at</strong>rix A<br />
A =sup|Ax|<br />
: (19)<br />
|x|=1<br />
Technical d<strong>et</strong>ails are given in [32].<br />
3.3. The no-free lunch condition<br />
The valu<strong>at</strong>ion formula (3) and the martingale condition (6) given for a single ass<strong>et</strong><br />
easily extends to a bask<strong>et</strong> of ass<strong>et</strong>s. It is n<strong>at</strong>ural to assume th<strong>at</strong>, for a given period t, the<br />
discount r<strong>at</strong>e rt(i), associ<strong>at</strong>ed with ass<strong>et</strong> i, are all the same. In frictionless mark<strong>et</strong>s, a<br />
<strong>de</strong>vi<strong>at</strong>ion for this hypothesis would lead to arbitrage opportunities. Furthermore, since<br />
the sequence of m<strong>at</strong>rices {At} is assumed to be i.i.d. and therefore st<strong>at</strong>ionary, this<br />
implies th<strong>at</strong> t or rt must be constant and equal, respectively, to and r.<br />
Un<strong>de</strong>r those conditions, we have the following proposition.<br />
Proposition 1. The stochastic process<br />
Xt = AtXt−1 + Bt<br />
s<strong>at</strong>is es the no-arbitrage condition if and only if<br />
(20)<br />
EQ[A]= 1 Id : (21)<br />
The proof is given in Ref. [32] in which this condition (21) is also shown to hold true<br />
un<strong>de</strong>r the historical probability measure P.
52 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 49<br />
The condition (21) imposes some stringent constraints on admissible m<strong>at</strong>rices At.<br />
In<strong>de</strong>ed, while At are not diagonal in general, their average must be diagonal. This<br />
implies th<strong>at</strong> the o -diagonal terms of the m<strong>at</strong>rices At must take neg<strong>at</strong>ive values, sufciently<br />
often so th<strong>at</strong> their averages vanish. The o -diagonal coe cients quantify the<br />
in uence of other bubbles on a given one. The condition (21) thus means th<strong>at</strong> the<br />
average e ect of other bubbles on any given one must vanish. It is straightforward to<br />
check th<strong>at</strong>, in this linear framework, this implies an absence of correl<strong>at</strong>ion (but not an<br />
absence of <strong>de</strong>pen<strong>de</strong>nce) b<strong>et</strong>ween the di erent bubble components E[X (k) X (‘) ]=0 for<br />
any k = ‘.<br />
In contrast, the diagonal elements of At must be positive in majority in or<strong>de</strong>r for<br />
EP[Aii]= (i)−1,<br />
for all i’s, to hold true. In fact, on economic grounds, we can exclu<strong>de</strong><br />
the cases where the diagonal elements take neg<strong>at</strong>ive values. In<strong>de</strong>ed, a neg<strong>at</strong>ive value<br />
of Aii <strong>at</strong> a given time t would imply th<strong>at</strong> X (i)<br />
t abruptly change sign b<strong>et</strong>ween t − 1 and<br />
t, wh<strong>at</strong> does not seem to be a reasonable nancial process.<br />
3.4. Renewal theory for products of random m<strong>at</strong>rices<br />
In the following, we will consi<strong>de</strong>r th<strong>at</strong> the random d × d m<strong>at</strong>rices At are invertible<br />
m<strong>at</strong>rices with real entries. We use Theorem 2:7 of Davis <strong>et</strong> al. [38], which synthesized<br />
Kesten’s Theorems 3 and 4 in Ref. [25], to the case of real valued m<strong>at</strong>rices. The proof<br />
of this theorem is given in Ref. [39]. We stress th<strong>at</strong> the conditions listed below do<br />
not require the m<strong>at</strong>rices (An) to be non-neg<strong>at</strong>ive. Actually, we have seen th<strong>at</strong>, in or<strong>de</strong>r<br />
for the r<strong>at</strong>ional expect<strong>at</strong>ion condition not to lead to trivial results, the o -diagonal<br />
coe cients of (An) have to be neg<strong>at</strong>ive with su ciently large probability such th<strong>at</strong><br />
their means vanish.<br />
Theorem 1. L<strong>et</strong> (An) be an i.i.d. sequence of m<strong>at</strong>rices in GLd(R) s<strong>at</strong>isfying the following<br />
s<strong>et</strong> of conditions th<strong>at</strong> we st<strong>at</strong>e in a heuristic manner (see Ref. [32] for technical<br />
d<strong>et</strong>ails). Provi<strong>de</strong>d th<strong>at</strong> the following conditions hold;<br />
H1: st<strong>at</strong>ionarity condition;<br />
H2: ergodicity;<br />
H3: intermittent ampli c<strong>at</strong>ion of the random m<strong>at</strong>rices;<br />
H4: the f<strong>at</strong>tailness of the distribution is not controlled by th<strong>at</strong> of the additive<br />
part Bt;<br />
then;<br />
• there exists a unique solution ∈ (0; 0] to the equ<strong>at</strong>ion<br />
1<br />
lim<br />
n→∞ n ln EP[A1 ···An ]=0; (22)<br />
• If (Xn) is the st<strong>at</strong>ionary solution to the stochastic recurrence equ<strong>at</strong>ion in (18)<br />
then X is regularly varying with in<strong>de</strong>x . In other words; the tail of the marginal
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 53<br />
50 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
distribution of each of the components of the vector X is asymptotically a power<br />
law with exponent .<br />
Eq. (22) d<strong>et</strong>ermining the tail exponent reduces to (8) in the one-dimensional case,<br />
which is simple to handle. In the multi-dimensional case, the novel fe<strong>at</strong>ure is the<br />
non-diagonal n<strong>at</strong>ure of the multiplic<strong>at</strong>ion of m<strong>at</strong>rices which does not allow in general<br />
for an explicit equ<strong>at</strong>ion similar to (8).<br />
It is important to stress th<strong>at</strong> the tails of the distribution of r<strong>et</strong>urns for all the components<br />
of the bubble <strong>de</strong>crease with the same tail in<strong>de</strong>x . This mo<strong>de</strong>l thus provi<strong>de</strong>s<br />
a n<strong>at</strong>ural s<strong>et</strong>ting for r<strong>at</strong>ionalizing the well-documented empirical observ<strong>at</strong>ion th<strong>at</strong> the<br />
exponent is found to be approxim<strong>at</strong>ely the same for most ass<strong>et</strong>s. The constraint on<br />
its value discussed in the next paragraph does not diminish the value of this remark,<br />
as explained in Section 5.<br />
3.5. Constraint on the tail in<strong>de</strong>x<br />
The rst conclusion of the theorem above shows th<strong>at</strong> the tail in<strong>de</strong>x of the process<br />
(Xt) is driven by the behavior of the m<strong>at</strong>rices (At). We will then st<strong>at</strong>e a proposition<br />
in which we give a major<strong>at</strong>ion of the tail in<strong>de</strong>x.<br />
Proposition 2. A necessary condition to have ¿1 is th<strong>at</strong> the spectral radius (largest<br />
eigenvalue) of EP[A] be smaller than 1:<br />
¿1 ⇒ (EP[A]) ¡ 1 : (23)<br />
The proof is given in Ref. [32]. This proposition, put tog<strong>et</strong>her with Proposition 1 above,<br />
allows us to <strong>de</strong>rive our main result. We have seen in Section 3.3 from Proposition 1<br />
th<strong>at</strong>, as a result of the no-arbitrage condition, the spectral radius of the m<strong>at</strong>rix EP[A]<br />
is gre<strong>at</strong>er than 1. As a consequence, by applic<strong>at</strong>ion of the converse of Proposition 2,<br />
we nd th<strong>at</strong> the tail in<strong>de</strong>x of the distribution of (X) is smaller than 1. This result<br />
does not rely on the diagonal property of the m<strong>at</strong>rices EP[At] but only on the value<br />
of the spectral radius imposed by the no-arbitrage condition.<br />
This result generalizes to arbitrary d-dimensional processes the result of Ref. [15]. As<br />
a consequence, d-dimensional r<strong>at</strong>ional expect<strong>at</strong>ion bubbles linking several ass<strong>et</strong>s su er<br />
from the same discrepancy compared to empirical d<strong>at</strong>a as the one-dimensional bubbles.<br />
It would therefore appear th<strong>at</strong> exogenous r<strong>at</strong>ional bubbles are hardly reconcilable with<br />
some of the most fundamental stylized facts of nancial d<strong>at</strong>a <strong>at</strong> a very elementary level.<br />
At this stage, we have to ask the question: wh<strong>at</strong> is wrong with the mo<strong>de</strong>l of r<strong>at</strong>ional<br />
bubbles? Two altern<strong>at</strong>ive answers are explored below: either we believe in the standard<br />
valu<strong>at</strong>ion formula and we are led to extend the restricted framework <strong>de</strong>scribed by the<br />
Blanchard and W<strong>at</strong>son’s mo<strong>de</strong>l; or we believe in the existence of the bubbles within<br />
their framework and we have to generalize the valu<strong>at</strong>ion formula. In the next two<br />
sections, we will discuss these two points of view.
54 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 51<br />
4. The crash hazard r<strong>at</strong>e mo<strong>de</strong>l [16,17]<br />
In the stylised framework of a purely specul<strong>at</strong>ive ass<strong>et</strong> th<strong>at</strong> pays no divi<strong>de</strong>nds—i.e.,<br />
with zero fundamental price—and in which we ignore inform<strong>at</strong>ion asymm<strong>et</strong>ry and the<br />
mark<strong>et</strong>-clearing condition, the price of the ass<strong>et</strong> equals the price of the bubble and the<br />
valu<strong>at</strong>ion formula (3) leads to the familiar martingale hypothesis for the bubble price:<br />
for all t ′ ¿t t→t ′EQ[X (t ′ )|Ft]=X (t) : (24)<br />
This equ<strong>at</strong>ion is nothing but a generalis<strong>at</strong>ion of Eq. (6) to a continuous time formul<strong>at</strong>ion,<br />
in which t→t ′ <strong>de</strong>notes the discount factor from time t to time t′ .<br />
We consi<strong>de</strong>r a general bubble dynamics given by<br />
dX = m(t) X (t)dt − X (t)dj; (25)<br />
where m(t) can be any non-linear causal function of X itself. We add a jump process<br />
j to capture the possibility th<strong>at</strong> the bubble exhibits a crash. j is thus zero before the<br />
crash and one afterwards. The random n<strong>at</strong>ure of the crash occurrence is mo<strong>de</strong>led by the<br />
cumul<strong>at</strong>ive distribution function Q(t) of the time of the crash, the probability <strong>de</strong>nsity<br />
function q(t)=dQ=dt and the hazard r<strong>at</strong>e h(t)=q(t)=[1 − Q(t)]. The hazard r<strong>at</strong>e is the<br />
probability per unit of time th<strong>at</strong> the crash will happen in the next instant provi<strong>de</strong>d it<br />
has not happened y<strong>et</strong>, i.e.:<br />
EQ[dj|Ft]=h(t)dt: (26)<br />
Expression (25) assumes th<strong>at</strong>, during a crash, the bubble drops by a xed percentage<br />
∈ (0; 1), say b<strong>et</strong>ween 20% and 30% of the bubble price.<br />
Using EQ[X (t +dt)|Ft]=(1+r dt)X (t), where r is the riskless discount r<strong>at</strong>e, taking<br />
the expect<strong>at</strong>ion of (25) conditioned on the ltr<strong>at</strong>ion up to time t and using Eq. (26),<br />
we g<strong>et</strong><br />
EQ[dX |Ft]=m(t)X (t)dt − X (t)h(t)dt = rX (t)dt; (27)<br />
which yields<br />
m(t) − r = h(t) : (28)<br />
If the crash hazard r<strong>at</strong>e h(t) increases, the r<strong>et</strong>urn m(t)−r above the riskless interest r<strong>at</strong>e<br />
increases to compens<strong>at</strong>e the tra<strong>de</strong>rs for the increasing risk. Reciprocally, if the dynamics<br />
of the bubble shoots up, the r<strong>at</strong>ional expect<strong>at</strong>ion condition imposes an increasing crash<br />
risk in or<strong>de</strong>r to ensure the absence of arbitrage opportunities: the risk-adjusted r<strong>et</strong>urn<br />
remains constant equal to the risk-free r<strong>at</strong>e. The corresponding equ<strong>at</strong>ion for the bubble<br />
price, conditioned on the crash not to have occurred, is<br />
t<br />
X (t)<br />
log = rt + h(t<br />
X (t0)<br />
t0<br />
′ )dt ′<br />
before the crash : (29)<br />
The integral t<br />
t0 h(t′ )dt ′ is the cumul<strong>at</strong>ive probability of a crash until time t. This gives<br />
the logarithm of the bubble price as the relevant observable. It has successfully been
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 55<br />
52 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
applied to the 1929 and 1987 Wall Stre<strong>et</strong> crashes up to about 7:5 years prior to the<br />
crash [40,17].<br />
The higher the probability of a crash, the faster the bubble must increase (conditional<br />
on having no crash) in or<strong>de</strong>r to s<strong>at</strong>isfy the martingale condition. Reciprocally,<br />
the higher the bubble, the more dangerous is the probability of a looming crash. Intuitively,<br />
investors must be compens<strong>at</strong>ed by a higher r<strong>et</strong>urn in or<strong>de</strong>r to be induced to<br />
hold an ass<strong>et</strong> th<strong>at</strong> might crash. This is the only e ect th<strong>at</strong> this mo<strong>de</strong>l captures. Note<br />
th<strong>at</strong> the bubble dynamics can be anything and the bubble can in particular be such<br />
th<strong>at</strong> the distribution of r<strong>et</strong>urns are f<strong>at</strong> tails with an exponent ≈ 3 without loss of<br />
generality [41].<br />
Ilinski [42] raised the concern th<strong>at</strong> the martingale condition (24) leads to a mo<strong>de</strong>l<br />
which “assumes a zero r<strong>et</strong>urn as the best prediction for the mark<strong>et</strong>”. He continues:<br />
“No need to say th<strong>at</strong> this is not wh<strong>at</strong> one expects from a perfect mo<strong>de</strong>l of mark<strong>et</strong><br />
bubble! Buying shares, tra<strong>de</strong>rs expect the price to rise and it is re ected (or caused)<br />
by their prediction mo<strong>de</strong>l. They support the bubble and the bubble support them!”.<br />
In other words, Ilinski [42] criticises a key economic element of the mo<strong>de</strong>l [16,17]:<br />
mark<strong>et</strong> r<strong>at</strong>ionality. This point is captured by assuming th<strong>at</strong> the mark<strong>et</strong> level is expected<br />
to stay constant (up to the riskless discount r<strong>at</strong>e) as written in equ<strong>at</strong>ion (24). Ilinski<br />
claims th<strong>at</strong> this equ<strong>at</strong>ion (24) is wrong because the mark<strong>et</strong> level does not stay constant<br />
in a bubble: it rises, almost by <strong>de</strong> nition.<br />
This misun<strong>de</strong>rstanding addresses a r<strong>at</strong>her subtle point of the mo<strong>de</strong>l and stems from<br />
the di erence b<strong>et</strong>ween two di erent types of r<strong>et</strong>urns:<br />
(1) The unconditional r<strong>et</strong>urn is in<strong>de</strong>ed zero as seen from (24) and re ects the fair<br />
game condition.<br />
(2) The conditional r<strong>et</strong>urn, conditioned upon no crash occurring b<strong>et</strong>ween time t and<br />
time t ′ , is non-zero and is given by Eq. (28). If the crash hazard r<strong>at</strong>e is increasing with<br />
time, the conditional r<strong>et</strong>urn will be acceler<strong>at</strong>ing precisely because the crash becomes<br />
more probable and the investors need to be remuner<strong>at</strong>ed for their higher risk.<br />
Thus, the expect<strong>at</strong>ion which remains constant in Eq. (24) takes into account the<br />
probability th<strong>at</strong> the mark<strong>et</strong> may crash. Therefore, conditionally on staying in the bubble<br />
(no crash y<strong>et</strong>), the mark<strong>et</strong> must r<strong>at</strong>ionally rise to compens<strong>at</strong>e buyers for having taken<br />
the risk th<strong>at</strong> the mark<strong>et</strong> could have crashed.<br />
The mark<strong>et</strong> price re ects the equilibrium b<strong>et</strong>ween the greed of buyers who hope the<br />
bubble will in <strong>at</strong>e and the fear of sellers th<strong>at</strong> it may crash. A bubble th<strong>at</strong> goes up is<br />
just one th<strong>at</strong> could have crashed but did not. The mo<strong>de</strong>l [16,17] is well summarised<br />
by borrowing the words of another economist: “(...) the higher probability of a crash<br />
leads to an acceler<strong>at</strong>ion of [the mark<strong>et</strong> price] while the bubble lasts”. Interestingly, this<br />
cit<strong>at</strong>ion is culled from the very same article by Blanchard [1] th<strong>at</strong> Ilinski [42] cites as<br />
an altern<strong>at</strong>ive mo<strong>de</strong>l more realistic than the mo<strong>de</strong>l [16,17]. We see th<strong>at</strong> this is in fact<br />
more of an endorsement than an altern<strong>at</strong>ive.<br />
A simple way to incorpor<strong>at</strong>e a di erent level of risk aversion into the mo<strong>de</strong>l [16,17]<br />
is to say th<strong>at</strong> the probability of a crash in the next instant is perceived by tra<strong>de</strong>rs as<br />
being K times bigger than it objectively is. This amounts to multiplying our hazard
56 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 53<br />
r<strong>at</strong>e h(t) byK, and once again this makes no substantive di erence as long as K is<br />
boun<strong>de</strong>d away from zero and in nity. Risk aversion is a central fe<strong>at</strong>ure of economic<br />
theory, and it is generally thought to be stable within a reasonable range, associ<strong>at</strong>ed<br />
with slow-moving secular trends such as changes in educ<strong>at</strong>ion, social structures and<br />
technology. Ilinski [42] rightfully points out th<strong>at</strong> risk perceptions are constantly changing<br />
in the course of real-life bubbles, but wrongfully claims th<strong>at</strong> the mo<strong>de</strong>l [16,17]<br />
viol<strong>at</strong>es this intuition. In this mo<strong>de</strong>l, risk perceptions do oscill<strong>at</strong>e dram<strong>at</strong>ically throughout<br />
the bubble, even though subjective aversion to risk remains stable, simply because<br />
it is the objective <strong>de</strong>gree of risk th<strong>at</strong> the bubble may burst th<strong>at</strong> goes through wild<br />
swings. For these reasons, the criticisms put forth by Ilinski, far from making a <strong>de</strong>nt<br />
in the economic mo<strong>de</strong>l [16,17], serve instead to show th<strong>at</strong> it is robust, exible and<br />
intuitive.<br />
To summarize, the crash hazard r<strong>at</strong>e mo<strong>de</strong>l is such th<strong>at</strong> the price dynamics can be<br />
essentially arbitrary, and in particular such th<strong>at</strong> the corresponding r<strong>et</strong>urns exhibit a<br />
reasonable f<strong>at</strong> tail. A jump process for crashes is ad<strong>de</strong>d, with a crash hazard r<strong>at</strong>e such<br />
th<strong>at</strong> the r<strong>at</strong>ional expect<strong>at</strong>ion condition is ensured.<br />
5. The non-st<strong>at</strong>ionary growth mo<strong>de</strong>l [18]<br />
In the previous section, we have presented a mo<strong>de</strong>l which assumes th<strong>at</strong> the fundamental<br />
valu<strong>at</strong>ion formula remains valid and have generalized Blanchard and W<strong>at</strong>son’s<br />
framework by reformul<strong>at</strong>ing the r<strong>at</strong>ional expect<strong>at</strong>ion condition with a jump crash process.<br />
We now consi<strong>de</strong>r the second view point which consists in rejecting the validity<br />
of the valu<strong>at</strong>ion formula while keeping the <strong>de</strong>composition of the price of an ass<strong>et</strong> into<br />
the sum of a fundamental price and a bubble term. In this aim, we present a possible<br />
modi c<strong>at</strong>ion of the r<strong>at</strong>ional bubble mo<strong>de</strong>l of Blanchard and W<strong>at</strong>son, recently proposed<br />
in Ref. [18], which involves an average exponential growth of the fundamental price<br />
<strong>at</strong> some r<strong>et</strong>urn r<strong>at</strong>e rf ¿ 0 larger than the discount r<strong>at</strong>e.<br />
5.1. Exponentially growing economy<br />
Recall th<strong>at</strong> (5) shows th<strong>at</strong> the observable mark<strong>et</strong> price is the sum of the bubble<br />
component Xt and of a “fundamental” price p f<br />
t<br />
pt = p f<br />
t + Xt : (30)<br />
Thus, waiving o the valu<strong>at</strong>ion formula (3), l<strong>et</strong> us assume th<strong>at</strong> the fundamental price<br />
p f<br />
t is growing exponentially as<br />
p f<br />
t = p0e rft<br />
(31)<br />
<strong>at</strong> the r<strong>at</strong>e rf and the bubble price is following (1).<br />
Note th<strong>at</strong> this formul<strong>at</strong>ion is comp<strong>at</strong>ible with the standard valu<strong>at</strong>ion formula as long<br />
as rf ¡r, provi<strong>de</strong>d the cash- ow dt <strong>at</strong> time t also grows with the same exponential
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 57<br />
54 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
r<strong>at</strong>e rf, i.e.: dt = d0erft . In<strong>de</strong>ed, the standard valu<strong>at</strong>ion formula then applies and leads<br />
to<br />
p f<br />
∞ dte<br />
t =<br />
k=0<br />
krf<br />
ekr <br />
d0<br />
e<br />
r − rf<br />
rft<br />
(32)<br />
up to the rst or<strong>de</strong>r in rf − r. The discussion of the case rf ¿r, for which (32) loses<br />
its meaning, is the subject of the sequel in which we follow [19].<br />
Putting (1) and (31) tog<strong>et</strong>her with (30), we obtain<br />
pt+1 = p f<br />
t+1 + <strong>at</strong>Xt + bt = <strong>at</strong>pt +(e rf − <strong>at</strong>)p f<br />
t + bt :<br />
Replacing p<br />
(33)<br />
f<br />
t in (30) by p0erft given in (31) leads to<br />
pt =e rft (p0 +ât) ; (34)<br />
where we have <strong>de</strong> ned the “reduced” bubble price following:<br />
ât+1 = <strong>at</strong>e −rf ât +e −rf e −rft bt : (35)<br />
Thus, if we allow the additive term bt in (1) to also grow exponentially as<br />
bt =e rft ˆbt ; (36)<br />
where ˆbt is a st<strong>at</strong>ionary stochastic white noise process, we obtain<br />
ât+1 = <strong>at</strong>e −rf ât +e −rf ˆbt ; (37)<br />
which is of the usual form. Intuitively, the additive term represents the background of<br />
“normal” uctu<strong>at</strong>ions around the fundamental price (31). Their “normal” uctu<strong>at</strong>ions<br />
have thus to grow with the same growth r<strong>at</strong>e in or<strong>de</strong>r to remain st<strong>at</strong>ionary in rel<strong>at</strong>ive<br />
value.<br />
In addition, replacing p f<br />
t in (33) again by p0e rft leads to<br />
pt+1 = <strong>at</strong>pt +e rft [p0(e rf − <strong>at</strong>)+ ˆbt] : (38)<br />
The expression (38) has the same form as (1) with a di erent additive term [p0(e r −<br />
<strong>at</strong>)+ ˆbt] replacing bt. The structure of this new additive term makes clear the origin of<br />
the factor e rft : as we said, it re ects nothing but the average exponential growth of the<br />
un<strong>de</strong>rlying economy. The contributions bt =e rft ˆbt are then nothing but the uctu<strong>at</strong>ions<br />
around this average growth.<br />
5.2. The value of the tail exponent<br />
The condition E[ln a] ¡rf ensures th<strong>at</strong> E[ln(ae −rf )] ¡ 0 which is now the st<strong>at</strong>ionarity<br />
condition for the process ât <strong>de</strong> ned by (37). The conditions 0 ¡E[|e rf ˆbt| ] ¡ + ∞<br />
(which is the same condition 0 ¡E[|bt| ] ¡ + ∞ as before) and the solution of<br />
E[|ae −rf | ] = 1 (39)<br />
tog<strong>et</strong>her with the constraint E[|ae −rf | ln |ae −rf |] ¡ + ∞ (which is the same as<br />
E[|a| ln |a|] ¡ + ∞) leads to an asymptotic power law distribution for the reduced
58 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 55<br />
price variable ât of the form Pâ(â) ≈ Câ=|â| 1+ , where is the real positive solution<br />
of (39). Note th<strong>at</strong> the condition E[ln(ae −rf )] ¡ 0 which is E[ln(a)] ¡rf now allows<br />
for positive average growth r<strong>at</strong>e of the product <strong>at</strong><strong>at</strong>−1<strong>at</strong>−2 ···a2a1a0.<br />
Consi<strong>de</strong>r the illustr<strong>at</strong>ive case where the multiplic<strong>at</strong>ive factors <strong>at</strong> are distributed according<br />
to a log-normal distribution such th<strong>at</strong> E[ln a]=lna0 (where a0 is thus the most<br />
probable value taken by <strong>at</strong>) and of variance 2 . Then<br />
E[|ae −rf<br />
<br />
| ] = exp −rf + ln a0 + 2<br />
2<br />
2<br />
: (40)<br />
Equ<strong>at</strong>ing (40) to 1 to g<strong>et</strong> according to Eq. (39) gives<br />
=2 rf − ln a0<br />
2<br />
= rf − ln a0<br />
r − ln a0<br />
=1+ rf − r<br />
r − ln a0<br />
: (41)<br />
We have used the not<strong>at</strong>ion 1= =1+r for the discount r<strong>at</strong>e r <strong>de</strong> ned in terms of the<br />
discount factor . The second equality in (41) uses E[a]=a0e 2 =2 .<br />
First, we r<strong>et</strong>rieve the result [15] th<strong>at</strong> ¡1 for the initial RE mo<strong>de</strong>l (1) for which<br />
rf = 0 and ln a0 ¡ 0. However, as soon as rf ¿r −ln ; we g<strong>et</strong><br />
¿1 ; (42)<br />
and can take arbitrary values. Technically, this results fundamentally from the structure<br />
of the process in which the additive noise grows exponentially to mimick the<br />
growth of the bubble which allevi<strong>at</strong>es the bound ¡1. Note th<strong>at</strong> rf does not need<br />
to be large for the result (42) to hold. Take for instance an annualized discount r<strong>at</strong>e<br />
ry =2%; an annualized r<strong>et</strong>urn r y<br />
f = 4% and a0 =1:0004. Expression (41) predicts<br />
which is comp<strong>at</strong>ible with empirical d<strong>at</strong>a.<br />
=3;<br />
5.3. Price r<strong>et</strong>urns<br />
The observable r<strong>et</strong>urn is<br />
where<br />
Rt = pt+1 − pt<br />
pt<br />
= t<br />
<br />
f<br />
pt+1 − pf t<br />
p f<br />
t<br />
= pf<br />
t+1 − pf t + Xt+1 − Xt<br />
p f<br />
t + Xt<br />
+ Xt+1 − Xt<br />
p f<br />
t<br />
<br />
= t rf + ât+1 − ât<br />
p0<br />
<br />
; (43)<br />
t = pf t<br />
p f<br />
1<br />
=<br />
: (44)<br />
t + Xt 1+(ât=p0)<br />
In or<strong>de</strong>r to <strong>de</strong>rive the last equality in the right-hand-si<strong>de</strong> of (43), we have used the<br />
<strong>de</strong> nition of the r<strong>et</strong>urn of the fundamental price (neglecting the small second or<strong>de</strong>r<br />
di erence b<strong>et</strong>ween e rf − 1 and rf). Expression (43) shows th<strong>at</strong> the distribution of
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 59<br />
56 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
r<strong>et</strong>urns Rt of the observable prices is the same as th<strong>at</strong> of the product of the random<br />
variable t by rf+(ât+1−ât)=p0. Now, the tail of the distribution of rf+(ât+1−ât)=p0 is<br />
the same as the tail of the distribution of ât+1 −ât; which is a power law with exponent<br />
solution of (39), as shown rigorously in Ref. [15].<br />
It remains to show th<strong>at</strong> the product of this variable rf +(ât+1 − ât)=p0 by t has<br />
the same tail behavior as rf +(ât+1 − ât)=p0 itself. If rf +(ât+1 − ât)=p0 and t were<br />
in<strong>de</strong>pen<strong>de</strong>nt, this would follow from results in Ref. [43] who <strong>de</strong>monstr<strong>at</strong>es th<strong>at</strong> for two<br />
in<strong>de</strong>pen<strong>de</strong>nt random variables and with Proba(| | ¿x) ≈ cx − and E[ + ] ¡ ∞<br />
for some ¿0; the random product obeys Proba(| | ¿x) ≈ E[ ]x − .<br />
However, rf+(ât+1−ât)=p0 and t are not in<strong>de</strong>pen<strong>de</strong>nt as both contain a contribution<br />
from the same term ât. However, when âtp0; t is close to 1 and the previous result<br />
should hold. The impact of ât in becomes important when ât becomes comparable<br />
to p0.<br />
It is then convenient to rewrite (43) using (44) as<br />
Rt =<br />
rf<br />
1+(ât=p0) + ât+1 − At<br />
=<br />
p0 +ât<br />
We can thus distinguish two regimes:<br />
rf<br />
1+(ât=p0) + (<strong>at</strong>e −rf − 1)ât +e −rf bt<br />
p0 +ât<br />
: (45)<br />
• for not too large values of the reduced bubble term ât; speci cally for ât ¡p0, the<br />
<strong>de</strong>nomin<strong>at</strong>or p0 +ât changes more slowly than the numer<strong>at</strong>or of the second term, so<br />
th<strong>at</strong> the distribution of r<strong>et</strong>urns will be domin<strong>at</strong>ed by the vari<strong>at</strong>ions of this numer<strong>at</strong>or<br />
(<strong>at</strong>e −rf − 1)ât +e −rf bt and, hence, will follow approxim<strong>at</strong>ely the same power-law<br />
as for ât; according to the results of Ref. [43].<br />
• For large bubbles, ât of the or<strong>de</strong>r of or gre<strong>at</strong>er than p0; the situ<strong>at</strong>ion changes,<br />
however: from (45), we see th<strong>at</strong> when the reduced bubble term ât increases without<br />
bound, the rst term rf=(1+(ât=p0)) goes to 0 while the second term becomes<br />
asymptotically <strong>at</strong>e −r − 1. This leads to the existence of an absolute upper bound for<br />
the absolute value of the r<strong>et</strong>urns.<br />
To summarize, we expect th<strong>at</strong> the distribution of r<strong>et</strong>urns will therefore follow a<br />
power-law with the same exponent as for ât, but with a nite cut-o (see [18]<br />
for d<strong>et</strong>ails). This is valid<strong>at</strong>ed by numerical simul<strong>at</strong>ions shown in Fig. 2 taken from<br />
Ref. [18].<br />
Thus, when the price uctu<strong>at</strong>ions associ<strong>at</strong>ed with bubbles on average grow with the<br />
mean mark<strong>et</strong> r<strong>et</strong>urn rf; we nd th<strong>at</strong> the exponent of the power law tail of the r<strong>et</strong>urns<br />
is no more boun<strong>de</strong>d by 1 as soon as rf is larger than the discount r<strong>at</strong>e r and can take<br />
essentially arbitrary values. It is remarkable th<strong>at</strong> this condition rf ¿r corresponds to<br />
the paradoxical and unsolved regime in fundamental valu<strong>at</strong>ion theory where the forward<br />
valu<strong>at</strong>ion solution (4) loses its meaning, as discussed in Ref. [19]. In analogy<br />
with the theory of bifurc<strong>at</strong>ions and their normal forms, Ref. [19] proposed th<strong>at</strong> this<br />
regime might be associ<strong>at</strong>ed with a spontaneous symm<strong>et</strong>ry breaking phase corresponding
60 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 57<br />
Fig. 2. Double logarithmic scale represent<strong>at</strong>ion of the complementary cumul<strong>at</strong>ive distribution of the “monthly”<br />
r<strong>et</strong>urns Rt <strong>de</strong> ned in (43) of the synth<strong>et</strong>ic total price sum of the exponential growing fundamental price and<br />
the bubble price. The continuous (resp. dashed) line corresponds to the positive (resp. neg<strong>at</strong>ive) r<strong>et</strong>urns. The<br />
distribution is well-<strong>de</strong>scribed by an asymptotic power law with an exponent in agreement with the prediction<br />
≈ 3:3 given by the equ<strong>at</strong>ions (39) and (41) and shown as the straight line. The small di erences b<strong>et</strong>ween<br />
the predicted slope and the numerically gener<strong>at</strong>ed ones are within the error bar of ±0:3 obtained from a<br />
standard maximum likelihood Hill estim<strong>at</strong>ion. From Ref. [18].<br />
to a spontaneous valu<strong>at</strong>ion in absence of divi<strong>de</strong>nds by pure specul<strong>at</strong>ive imit<strong>at</strong>ive<br />
processes.<br />
6. Conclusion<br />
Despite its elegant formul<strong>at</strong>ion of the bubble phenomenon, the Blanchard and W<strong>at</strong>son’s<br />
mo<strong>de</strong>l su ers from a l<strong>et</strong>hal discrepancy: it does not seem to comply with the<br />
empirical d<strong>at</strong>a, i.e., it cannot gener<strong>at</strong>e power law tails whose exponent is gre<strong>at</strong>er than<br />
1, in disagreement with the empirical tail in<strong>de</strong>x found around 3. We have summarized<br />
the <strong>de</strong>monstr<strong>at</strong>ion th<strong>at</strong> this result holds true both for a bubble <strong>de</strong> ned for a single ass<strong>et</strong><br />
as well as for bubbles on any s<strong>et</strong> of coupled ass<strong>et</strong>s, as long as the r<strong>at</strong>ional expect<strong>at</strong>ion<br />
condition holds.<br />
In or<strong>de</strong>r to reconcile the theory with the empirical facts on the tails of the distributions<br />
of r<strong>et</strong>urns, two altern<strong>at</strong>ive mo<strong>de</strong>ls have been presented. The “crash hazard r<strong>at</strong>e”<br />
mo<strong>de</strong>l extends the formul<strong>at</strong>ion of Blanchard and W<strong>at</strong>son by replacing the linear stochastic<br />
bubble price equ<strong>at</strong>ion by an arbitrary dynamics solely constrained by the no-arbitrage<br />
condition ma<strong>de</strong> to hold with the introduction of a jump process. The “growth r<strong>at</strong>e<br />
mo<strong>de</strong>l” <strong>de</strong>parts more audaciously from standard economic mo<strong>de</strong>ls since it discards<br />
one of the pillars of the standard valu<strong>at</strong>ion theory, but putting itself rmly in the<br />
regime rf ¿r for which the fundamental valu<strong>at</strong>ion formula breaks down. In addition to<br />
allowing for correct values of the tail exponent, it provi<strong>de</strong>s a generaliz<strong>at</strong>ion of the
2.2. Des bulles r<strong>at</strong>ionnelles aux krachs 61<br />
58 D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59<br />
fundamental valu<strong>at</strong>ion formula by providing an un<strong>de</strong>rstanding of its breakdown as<br />
<strong>de</strong>eply associ<strong>at</strong>ed with a spontaneous breaking of the price symm<strong>et</strong>ry. Its implement<strong>at</strong>ion<br />
for multi-dimensional bubbles is straightforward and the results obtained in<br />
Section 5 carry over n<strong>at</strong>urally in this case. This provi<strong>de</strong>s an explan<strong>at</strong>ion for why the<br />
tail in<strong>de</strong>x seems to be the same for any group of ass<strong>et</strong>s as observed empirically. This<br />
work begs for the introduction of a generalized eld theory which would be able to<br />
capture the spontaneous breaking of symm<strong>et</strong>ry, recover the fundamental valu<strong>at</strong>ion formula<br />
in the normal economic case rf ¡r and extend it to the still unexplored regime<br />
rf ¿r.<br />
Acknowledgements<br />
We acknowledge helpful discussions and exchanges with J.V. An<strong>de</strong>rsen, A. Johansen,<br />
J.P. Laurent, O. Ledoit, T. Lux, V. Pisarenko and M. Taqqu and thank T. Mikosch for<br />
providing access to Ref. [39].<br />
References<br />
[1] O.J. Blanchard, Econ. L<strong>et</strong>t. 3 (1979) 387–389.<br />
[2] O.J. Blanchard, M.W. W<strong>at</strong>son, Bubbles, r<strong>at</strong>ional expect<strong>at</strong>ions and specul<strong>at</strong>ive mark<strong>et</strong>s, in: P. Wachtel<br />
(Ed.), Crisis in Economic and Financial Structure: Bubbles, Bursts, and Shocks, Lexington Books,<br />
Lexington, 1982.<br />
[3] R.H. Thaler (Ed.), Advances in Behavioral Finance, Russell Sage Found<strong>at</strong>ion, New York, 1993.<br />
[4] H. Shefrin, Beyond Greed and Fear: Un<strong>de</strong>rstanding Behavioral Finance and the Psychology of Investing,<br />
Harvard Business School Press, Boston, MA, 2000.<br />
[5] A. Shleifer, Ine cient Mark<strong>et</strong>s: An Introduction to Behavioral Finance, Oxford University Press, Oxford,<br />
2000.<br />
[6] B.G. Malkiel, A random walk down Wall Stre<strong>et</strong>, WW Norton & Company, New York, 1999.<br />
[7] C. Camerer, J. Econ. Surveys 3 (1989) 3–41.<br />
[8] M.C. Adam, A. Szafarz, Oxford Economic Papers 44 (1992) 626–640.<br />
[9] L. <strong>de</strong> Haan, S.I. Resnick, H. Rootzen, C.G. <strong>de</strong> Vries, Stochastic Processes Appl. 32 (1989) 213.<br />
[10] C. <strong>de</strong> Calan, J.-M. Luck, T.M. Nieuwenhuizen, D. P<strong>et</strong>ritis, J. Phys. A 18 (1985) 501.<br />
[11] M. Levy, S. Solomon, G. Ram, Int. J. Mod. Phys. C 7 (1996) 65.<br />
[12] D. Sorn<strong>et</strong>te, R. Cont, J. Phys. I France 7 (1997) 431–444.<br />
[13] D. Sorn<strong>et</strong>te, Physica A 250 (1998) 295–314.<br />
[14] D. Sorn<strong>et</strong>te, Critical Phenomena in N<strong>at</strong>ural Sciences (Chaos, Fractals, Self-organiz<strong>at</strong>ion and Disor<strong>de</strong>r:<br />
Concepts and Tools), Springer Series in Synerg<strong>et</strong>ics, Springer, Hei<strong>de</strong>lberg, 2000.<br />
[15] T. Lux, D. Sorn<strong>et</strong>te, On r<strong>at</strong>ional bubbles and f<strong>at</strong> tails, J. Money, Credit Banking, 1999, in press, (e-print<br />
<strong>at</strong> http:==xxx.lanl.gov=abs=cond-m<strong>at</strong>=9910141).<br />
[16] A. Johansen, D. Sorn<strong>et</strong>te, O. Ledoit, J. Risk 1 (4) (1999) 5–32.<br />
[17] A. Johansen, O. Ledoit, D. Sorn<strong>et</strong>te, Int. J. Theor<strong>et</strong>. Appl. Finance 3 (2) (2000) 219–255.<br />
[18] D. Sorn<strong>et</strong>te, “Slimming” of power law tails by increasing mark<strong>et</strong> r<strong>et</strong>urns, preprint, Univ. Nice (e-print<br />
<strong>at</strong> http:==arXiv.org=abs=cond-m<strong>at</strong>=0010112).<br />
[19] D. Sorn<strong>et</strong>te, Physica A 284 (2000) 355–375.<br />
[20] C. Gourieroux, J.J. La ont, A. Monfort, Econom<strong>et</strong>rica 50 (1982) 409–425.<br />
[21] R.A. Broglia, J. Terasaki, N. Giovanardi, Phys. Rep. 335 (2000) 2–18.<br />
[22] C.P. Burgess, Phys. Rep. 330 (2000) 194–261.<br />
[23] J.W. Vanholten, Phys. L<strong>et</strong>t. 352 (1995) 220–223.<br />
[24] P. Bak, S.F. Norrelykke, M. Shubik, Phys. Rev. E 60 (1999) 2528–2532.
62 2. Modèles phénoménologiques <strong>de</strong> cours<br />
D. Sorn<strong>et</strong>te, Y. Malevergne / Physica A 299 (2001) 40–59 59<br />
[25] H. Kesten, Acta M<strong>at</strong>h. 131 (1973) 207–248.<br />
[26] C.M. Goldie, Ann. Appl. Probab. 1 (1991) 126–166.<br />
[27] C.G. <strong>de</strong> Vries, Stylized facts of nominal exchange r<strong>at</strong>e r<strong>et</strong>urns, in: F. van <strong>de</strong>r Ploeg (Ed.), The Handbook<br />
of Intern<strong>at</strong>ional Macroeconomics, Blackwell, Oxford, 1994, S. 348–389.<br />
[28] L. Lux, Appl. Financial Econ. 6 (1996) 463–475.<br />
[29] A. Pagan, J. Empirical Finance 3 (1996) 15–102.<br />
[30] D.M. Guillaume, M.M. Dacorogna, R.R. Dave, J.A. Muller, R.B. Olsen, O.V. Pict<strong>et</strong>, Finance Stochastics<br />
1 (1997) 95–129.<br />
[31] P. Gopikrishnan, M. Meyer, L.A.N. Amaral, H.E. Stanley, Eur. Phys. J. B 3 (1998) 139–140.<br />
[32] Y. Malevergne, D. Sorn<strong>et</strong>te, Multi-dimensional r<strong>at</strong>ional bubbles and f<strong>at</strong> tails, preprint <strong>at</strong><br />
http:==arXiv.org=abs=cond-m<strong>at</strong>=0101371, 2001.<br />
[33] R.N. Mantegna, Eur. Phys. J. B 11 (1999) 193–197.<br />
[34] G. Bonanno, F. Lillo, R.N. Mantegna, Quantit<strong>at</strong>ive Finance 1 (1) (2001) 96–104.<br />
[35] R.P. Flood, P.M. Garber, L.O. Scott, J. Econ. Dyn. Control 8 (1984) 329–340.<br />
[36] R.J. Barro, E.F. Fama, D.R. Fischel, A.H. Meltzer, R. Roll, L.G. Telser, in: R.W. Kamphuis Jr.,<br />
R.C. Kormendi, J.W.H. W<strong>at</strong>son (Eds.), Black Monday and the Future of Financial Mark<strong>et</strong>s, Mid<br />
American Institute for Public Policy Research, Inc. and Dow Jones-Irwin, Inc., 1989.<br />
[37] A. Johansen, D. Sorn<strong>et</strong>te, Bubbles and anti-bubbles in L<strong>at</strong>in-American, Asian and Western stock mark<strong>et</strong>s:<br />
an empirical study, Int. J. Theor<strong>et</strong>. Appl. Finance, 2000, in press (e-print <strong>at</strong> http:==xxx.lanl.gov=<br />
abs=cond-m<strong>at</strong>=9907270).<br />
[38] R.A. Davis, T. Mikosch, B. Basrak, Sample ACF of multivari<strong>at</strong>e stochastic recurrence equ<strong>at</strong>ions<br />
with applic<strong>at</strong>ions to GARCH. Technical Report, University of Groeningen, available <strong>at</strong> www.m<strong>at</strong>h.<br />
rug.nl=mikosch, 1999.<br />
[39] E. Le Page, Theoremes <strong>de</strong> renouvellement pour les produits <strong>de</strong> m<strong>at</strong>rices ale<strong>at</strong>oires. Equ<strong>at</strong>ions aux<br />
di erences ale<strong>at</strong>oires. Seminaires <strong>de</strong> probabilites Rennes 1938, Publ. Sem. M<strong>at</strong>h. Univ. Rennes I, 1983,<br />
116pp.<br />
[40] D. Sorn<strong>et</strong>te, A. Johansen, Physica A 245 (1997) 411–422.<br />
[41] D. Sorn<strong>et</strong>te, J.V. An<strong>de</strong>rsen, Quantifying herding during specul<strong>at</strong>ive nancial bubbles, preprint, Univ.<br />
Nice (e-print <strong>at</strong> http:==arXiv.org=abs=cond-m<strong>at</strong>=0104341).<br />
[42] K. Ilinski, Int. J. Mod. Phys. C 10 (1999) 741–746.<br />
[43] L. Breiman, Theory Probab. Appl. 10 (1965) 323–329.
Chapitre 3<br />
Distributions exponentielles étirées contre<br />
distributions régulièrement variables<br />
Un large consensus semble s’être développé en faveur <strong>de</strong> l’hypothèse selon laquelle les distributions<br />
empiriques <strong>de</strong> ren<strong>de</strong>ment <strong>de</strong>s séries financières sont régulièrement variables, avec un indice <strong>de</strong> queue <strong>de</strong><br />
l’ordre <strong>de</strong> trois. Nous montrons tout d’abord, à l’ai<strong>de</strong> <strong>de</strong> tests synthétiques réalisés sur <strong>de</strong>s séries temporelles<br />
dont la vol<strong>at</strong>ilité présente <strong>de</strong> longues dépendances <strong>et</strong> dont les distributions marginales sont soit<br />
<strong>de</strong>s lois <strong>de</strong> puissance soit <strong>de</strong>s exponentielles étirées, que les estim<strong>at</strong>eurs basés sur les distributions <strong>de</strong>s<br />
extrêmes généralisés ou les distributions <strong>de</strong> Par<strong>et</strong>o généralisées ne sont pas à même <strong>de</strong> distinguer ces <strong>de</strong>ux<br />
classes <strong>de</strong> distributions, en contradiction avec <strong>de</strong> nombreuses étu<strong>de</strong>s antérieures. Ensuite, nous utilisons<br />
une représent<strong>at</strong>ion paramétrique <strong>de</strong> la distribution <strong>de</strong> ren<strong>de</strong>ments qui englobe à la fois <strong>de</strong>s distributions<br />
régulièrement variables ainsi que rapi<strong>de</strong>ment variables comme notamment les exponentielles-étirées.<br />
Utilisant le théorème <strong>de</strong> Wilk, nous comparons la validité <strong>de</strong> ces différentes hypothèses <strong>et</strong> concluons que<br />
les distributions exponentielles-étirées <strong>et</strong> les distributions <strong>de</strong> Par<strong>et</strong>o fournissent une <strong>de</strong>scription convenable<br />
<strong>de</strong>s données <strong>et</strong> ne peuvent être distinguées lorsque l’on s’intéresse à <strong>de</strong>s quantiles suffisamment<br />
élevés. Nous introduisons ensuite un nouveau test basé sur le fait que dans une certaine limite, les distributions<br />
exponentielles-étirées ten<strong>de</strong>nt vers les distributions <strong>de</strong> Par<strong>et</strong>o. La conclusion <strong>de</strong> notre ensemble<br />
<strong>de</strong> tests est que probablement les distributions <strong>de</strong> ren<strong>de</strong>ments d’actifs financiers décroissent moins vite<br />
que toute exponentielle-étirée mais plus vite que toute loi <strong>de</strong> puissance d’exposant raisonnable. Pour terminer,<br />
nous discutons les implic<strong>at</strong>ions <strong>de</strong> nos résult<strong>at</strong>s sur le problème d’existence <strong>de</strong>s moments <strong>et</strong> pour<br />
l’estim<strong>at</strong>ion du risque.<br />
63
64 3. Distributions exponentielles étirées contre distributions régulièrement variables
Empirical Distributions of Log-R<strong>et</strong>urns:<br />
Exponential or Power-like?<br />
Y. Malevergne 1,2 , V. Pisarenko 3 , and D. Sorn<strong>et</strong>te 1,4<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée<br />
CNRS UMR6622 and Université <strong>de</strong> Nice-Sophia Antipolis<br />
Parc Valrose, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x, France<br />
3 Intern<strong>at</strong>ional Institute of Earthquake Prediction Theory and M<strong>at</strong>hem<strong>at</strong>ical Geophysics<br />
Russian Ac. Sci. Warshavskoye sh., 79, kor. 2, Moscow 113556, Russia<br />
4 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095<br />
e-mails: Yannick.Malevergne@unice.fr, Vlad@sirus.mitp.ru and sorn<strong>et</strong>te@unice.fr<br />
Abstract<br />
A large consensus now seems to take for granted th<strong>at</strong> the distributions of empirical r<strong>et</strong>urns of financial<br />
time series are regularly varying, with a tail exponent close to 3. First, we show by synth<strong>et</strong>ic tests performed<br />
on time series with long-range time <strong>de</strong>pen<strong>de</strong>nce in the vol<strong>at</strong>ility with both Par<strong>et</strong>o and Str<strong>et</strong>ched-<br />
Exponential distributions th<strong>at</strong> standard generalized extreme value (GEV) and Generalized Par<strong>et</strong>o Distribution<br />
(GPD) estim<strong>at</strong>ors are quite inefficient and cannot distinguish reliably b<strong>et</strong>ween the two classes<br />
of distributions, in contradiction with previous results. Then, we use a param<strong>et</strong>ric represent<strong>at</strong>ion of the<br />
distribution of r<strong>et</strong>urns of 100 years of daily r<strong>et</strong>urn of the Dow Jones Industrial Average and over 1 years<br />
of 5-minutes r<strong>et</strong>urns of the Nasdaq Composite in<strong>de</strong>x, encompassing both a regularly varying distribution<br />
in one limit of the param<strong>et</strong>ers and rapidly varying distributions of the class of the Str<strong>et</strong>ched-Exponential<br />
(SE) distributions in other limits. Using the m<strong>et</strong>hod of nested hypothesis testing (Wilk theorem), we<br />
conclu<strong>de</strong> th<strong>at</strong> both the SE distributions and Par<strong>et</strong>o distributions provi<strong>de</strong> reliable <strong>de</strong>scriptions of the d<strong>at</strong>a<br />
and cannot be distinguished for sufficiently high thresholds. Then, we introduced a novel encompassing<br />
test based on the discovery th<strong>at</strong> the SE contains the Par<strong>et</strong>o law in a certain limit th<strong>at</strong> confirms th<strong>at</strong> the SE<br />
encompasses any power law tail distribution. However, our best estim<strong>at</strong>ion of the param<strong>et</strong>ers of the SE<br />
mo<strong>de</strong>l suggests th<strong>at</strong> the tails <strong>de</strong>cay probably more slowly th<strong>at</strong> a pure SE. Summing up all the evi<strong>de</strong>nce<br />
provi<strong>de</strong>d by our b<strong>at</strong>tery of tests, it seems th<strong>at</strong> the tails ultim<strong>at</strong>ely <strong>de</strong>cay slower than any SE but probably<br />
faster than power laws with reasonable exponents. We discuss the implic<strong>at</strong>ions of our results on the<br />
“moment condition failure” and for risk estim<strong>at</strong>ion and management.<br />
1<br />
65
66 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
1 Motiv<strong>at</strong>ion of the study<br />
The d<strong>et</strong>ermin<strong>at</strong>ion of the precise shape of the tail of the distribution of r<strong>et</strong>urns is a major issue both from<br />
a practical and from an aca<strong>de</strong>mic point of view. For practitioners, it is crucial to accur<strong>at</strong>ely estim<strong>at</strong>e the<br />
low value quantiles of the distribution of r<strong>et</strong>urns (profit and loss) because they are involved in almost all<br />
the mo<strong>de</strong>rn risk management m<strong>et</strong>hods. From an aca<strong>de</strong>mic perspective, many economic and financial theories<br />
rely on a specific param<strong>et</strong>eriz<strong>at</strong>ion of the distributions whose param<strong>et</strong>ers are inten<strong>de</strong>d to represent the<br />
“macroscopic” variables the agents are sensitive to.<br />
The distribution of r<strong>et</strong>urns is one of the most basic characteristics of the mark<strong>et</strong>s and many papers have been<br />
<strong>de</strong>voted to it. Contrarily to the average or expected r<strong>et</strong>urn, for which economic theory provi<strong>de</strong>s gui<strong>de</strong>lines<br />
to assess them in rel<strong>at</strong>ion with risk premium, firm size or book-to-mark<strong>et</strong> equity (see for instance Fama<br />
and French (1996)), the functional form of the distribution of r<strong>et</strong>urns, and especially of extreme r<strong>et</strong>urns, is<br />
much less constrained and still a topic of active <strong>de</strong>b<strong>at</strong>e. Naively, the central limit theorem would lead to<br />
a Gaussian distribution for sufficiently large time intervals over which the r<strong>et</strong>urn is estim<strong>at</strong>ed. Taking the<br />
continuous time limit such th<strong>at</strong> any finite time interval is seen as the sum of an infinite number of increments<br />
thus leads to the paradigm of log-normal distributions of prices and equivalently of Gaussian distributions<br />
of r<strong>et</strong>urns, based on the pioneering work of Bachelier (1900) l<strong>at</strong>er improved by Samuelson (1965). The lognormal<br />
paradigm has been the starting point of many financial theories such as Markovitz (1959)’s portfolio<br />
selection m<strong>et</strong>hod, Sharpe (1964)’s mark<strong>et</strong> equilibrium mo<strong>de</strong>l or Black and Scholes (1973)’s r<strong>at</strong>ional option<br />
pricing theory. However, for real financial d<strong>at</strong>a, the convergence in distribution to a Gaussian law is very<br />
slow (Campbell <strong>et</strong> al. 1997, Bouchaud and Potters 2000, for instance), much slower th<strong>at</strong> predicted for<br />
in<strong>de</strong>pen<strong>de</strong>nt r<strong>et</strong>urns. As table 1 shows, the excess kurtosis (which is zero for a normal distribution) remains<br />
large even for monthly r<strong>et</strong>urns, testifying of significant <strong>de</strong>vi<strong>at</strong>ions from normality, of the heavy tail behavior<br />
of the distributions of r<strong>et</strong>urns and of significant <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween r<strong>et</strong>urns (Campbell <strong>et</strong> al. 1997).<br />
Another i<strong>de</strong>a rooted in economic theory consists in invoking the “Gibr<strong>at</strong> principle” (Simon 1957) initially<br />
used to account for the growth of cities and of wealth through a mechanism combining stochastic multiplic<strong>at</strong>ive<br />
and additive noises (Levy <strong>et</strong> al. 1996, Sorn<strong>et</strong>te and Cont 1997, Biham <strong>et</strong> al 1998, Sorn<strong>et</strong>te 1998)<br />
leading to a Par<strong>et</strong>o distribution of sizes (Champenowne 1953, Gabaix 1999). R<strong>at</strong>ional bubble mo<strong>de</strong>ls a<br />
la Blanchard and W<strong>at</strong>son (1982) can also be cast in this m<strong>at</strong>hem<strong>at</strong>ical framework of stochastic recurrence<br />
equ<strong>at</strong>ions and leads to distribution with power law tails, albeit with a strong constraint on the tail exponent<br />
(Lux and Sorn<strong>et</strong>te 2002, Malevergne and Sorn<strong>et</strong>te 2001). These frameworks suggest th<strong>at</strong> an altern<strong>at</strong>ive and<br />
n<strong>at</strong>ural way to capture the heavy tail character of the distributions of r<strong>et</strong>urns is to use distributions with<br />
power-like tails (Par<strong>et</strong>o, Generalized Par<strong>et</strong>o, stable laws) or more generally, regularly-varying distributions<br />
(Bingham <strong>et</strong> al 1987) 1 , the l<strong>at</strong>er encompassing all the former.<br />
In the early 1960s, Man<strong>de</strong>lbrot (1963) and Fama (1965) presented evi<strong>de</strong>nce th<strong>at</strong> distributions of r<strong>et</strong>urns can<br />
be well approxim<strong>at</strong>ed by a symm<strong>et</strong>ric Levy stable law with tail in<strong>de</strong>x b about 1.7. These estim<strong>at</strong>es of the<br />
power tail in<strong>de</strong>x have recently been confirmed by Mittnik <strong>et</strong> al. (1998), and slightly different indices of the<br />
stable law (b = 1.4) were suggested by Mantegna and Stanley (1995, 2000).<br />
On the other hand, there are numerous evi<strong>de</strong>nces of a larger value of the tail in<strong>de</strong>x b ∼ = 3 (Longin 1996,<br />
Guillaume <strong>et</strong> al. 1997, Gopikrishnan <strong>et</strong> al. 1998, Müller <strong>et</strong> al. 1998, Farmer 1999, Lux 2000). See also the<br />
various altern<strong>at</strong>ive param<strong>et</strong>eriz<strong>at</strong>ions in term of the Stu<strong>de</strong>nt distribution (Bl<strong>at</strong>tberg and Gonne<strong>de</strong>s 1974, Kon<br />
1984), hyperbolic distributions (Eberlein <strong>et</strong> al. 1998, Prause 1998), normal inverse Gaussian distributions<br />
(Barndorff-Nielsen 1997), and Pearson type-VII distributions (Nagahara and Kitagawa 1999), which all have<br />
an asymptotic power law tail and are regularly varying. Thus, a general conclusion of this group of authors<br />
1 The general represent<strong>at</strong>ion of a regularly varying distribution is given by ¯F(x) = L(x) · x −α , where L(·) is a slowly varying<br />
function, th<strong>at</strong> is, limx→∞L(tx)/L(x) = 1 for any finite t.<br />
2
concerning tail f<strong>at</strong>ness can be formul<strong>at</strong>ed as follows: tails of the distribution of r<strong>et</strong>urns are heavier than a<br />
Gaussian tail and heavier than an exponential tail; they certainly admit the existence of a finite variance<br />
(b > 2), whereas the existence of the third (skewness) and the fourth (kurtosis) moments is questionable.<br />
These apparent contradictory results actually do not apply to the same quantiles of the distributions of<br />
r<strong>et</strong>urns. In<strong>de</strong>ed, Mantegna and Stanley (1995) have shown th<strong>at</strong> the distribution of r<strong>et</strong>urns can be <strong>de</strong>scribed<br />
accur<strong>at</strong>ely by a Lévy law only within a range of approxim<strong>at</strong>ely nine standard <strong>de</strong>vi<strong>at</strong>ions, while a faster <strong>de</strong>cay<br />
of the distribution is observed beyond. This almost-but-not-quite Lévy stable <strong>de</strong>scription explains (in part)<br />
the slow convergence of the r<strong>et</strong>urns distribution to the Gaussian law un<strong>de</strong>r time aggreg<strong>at</strong>ion (Sorn<strong>et</strong>te 2000).<br />
And it is precisely outsi<strong>de</strong> this range where the Lévy law applies th<strong>at</strong> a tail in<strong>de</strong>x of about three have been<br />
estim<strong>at</strong>ed. This can be seen from the fact th<strong>at</strong> most authors who have reported a tail in<strong>de</strong>x b ∼ = 3 have used<br />
some optimality criteria for choosing the sample fractions (i.e., the largest values) for the estim<strong>at</strong>ion of the<br />
tail in<strong>de</strong>x. Thus, unlike the authors supporting stable laws, they have used only a fraction of the largest<br />
(positive tail) and smallest (neg<strong>at</strong>ive tail) sample values.<br />
It would thus seem th<strong>at</strong> all has been said on the distributions of r<strong>et</strong>urns. However, there are dissenting<br />
views in the liter<strong>at</strong>ure. In<strong>de</strong>ed, the class of regularly varying distributions is not the sole one able to account<br />
for the large kurtosis and f<strong>at</strong>-tailness of the distributions of r<strong>et</strong>urns. Some recent works suggest altern<strong>at</strong>ive<br />
<strong>de</strong>scriptions for the distributions of r<strong>et</strong>urns. For instance, Gouriéroux and Jasiak (1998) claim th<strong>at</strong> the<br />
distribution of r<strong>et</strong>urns on the French stock mark<strong>et</strong> <strong>de</strong>cays faster than any power law. Cont <strong>et</strong> al. (1997)<br />
have proposed to use exponentially trunc<strong>at</strong>ed stable distributions (Cont <strong>et</strong> al. 1997) while Laherrère and<br />
Sorn<strong>et</strong>te (1999) suggest to fit the distributions of stock r<strong>et</strong>urns by the Str<strong>et</strong>ched-Exponential (SE) law. These<br />
results, challenging the traditional hypothesis of power-like tail, offer a new represent<strong>at</strong>ion of the r<strong>et</strong>urns<br />
distributions and need to be tested rigorously on a st<strong>at</strong>istical ground.<br />
A priori, one could assert th<strong>at</strong> Longin (1996)’s results should rule out the exponential and Str<strong>et</strong>ched-<br />
Exponential hypotheses. In<strong>de</strong>ed, his results, based on extreme value theory, show th<strong>at</strong> the distributions<br />
of log-r<strong>et</strong>urns belong to the maximum domain of <strong>at</strong>traction of the Fréch<strong>et</strong> distribution, so th<strong>at</strong> they are necessarily<br />
regularly varying power-like laws. However, his study, like almost all others on this subject, has<br />
been performed un<strong>de</strong>r the assumption th<strong>at</strong> (1) financial time series are ma<strong>de</strong> of in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically<br />
distributed r<strong>et</strong>urns and (2) the corresponding distributions of r<strong>et</strong>urns belong to one of only three possible<br />
maximum domains of <strong>at</strong>traction. However, these assumptions are not fulfilled in general. While Smith<br />
(1985)’s results indic<strong>at</strong>e th<strong>at</strong> the <strong>de</strong>pen<strong>de</strong>nce of the d<strong>at</strong>a does not constitute a major problem in the limit<br />
of large samples, we shall see th<strong>at</strong> it can significantly bias standard st<strong>at</strong>istical m<strong>et</strong>hods for samples of size<br />
commonly used in extreme tails studies. Moreover, Longin’s conclusions are essentially based on an aggreg<strong>at</strong>ion<br />
procedure which stresses the central part of the distribution while smoothing the characteristics of<br />
the tail, which are essential in characterizing the tail behavior.<br />
In addition, real financial time series exhibit GARCH effects (Bollerslev 1986, Bollerslev <strong>et</strong> al. 1994) leading<br />
to h<strong>et</strong>eroskedasticity and to clustering of high threshold exceedances due to a long memory of the<br />
vol<strong>at</strong>ility. These r<strong>at</strong>her complex <strong>de</strong>pen<strong>de</strong>nt structures make difficult if not questionable the blind applic<strong>at</strong>ion<br />
of standard st<strong>at</strong>istical tools for d<strong>at</strong>a analysis. In particular, the existence of significant <strong>de</strong>pen<strong>de</strong>nce in the<br />
r<strong>et</strong>urn vol<strong>at</strong>ility leads to dram<strong>at</strong>ic un<strong>de</strong>restim<strong>at</strong>es of the true standard <strong>de</strong>vi<strong>at</strong>ion of the st<strong>at</strong>istical estim<strong>at</strong>ors<br />
of tail indices. In<strong>de</strong>ed, there are now many examples showing th<strong>at</strong> <strong>de</strong>pen<strong>de</strong>nces and long memories as well<br />
as nonlinearities mislead standard st<strong>at</strong>istical tests (An<strong>de</strong>rsson <strong>et</strong> al. 1999, Granger and Teräsvirta 1999, for<br />
instance). Consi<strong>de</strong>r the Hill’s and Pickand’s estim<strong>at</strong>ors, which play an important role in the study of the<br />
tails of distributions. It is often overlooked th<strong>at</strong>, for <strong>de</strong>pen<strong>de</strong>nt time series, Hill’s estim<strong>at</strong>or remains only<br />
consistent but not asymptotically efficient (Rootzen <strong>et</strong> al. 1998). Moreover, for financial time series with a<br />
<strong>de</strong>pen<strong>de</strong>nce structure <strong>de</strong>scribed by a GARCH process, Kearns and Pagan (1997) have shown th<strong>at</strong> the standard<br />
<strong>de</strong>vi<strong>at</strong>ion of Hill’s estim<strong>at</strong>or obtained by a bootstrap m<strong>et</strong>hod can be seven to eight time larger than the<br />
standard <strong>de</strong>vi<strong>at</strong>ion <strong>de</strong>rived un<strong>de</strong>r the asymptotic normality assumption. These figures are even worse for<br />
3<br />
67
68 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Pickand’s estim<strong>at</strong>or.<br />
The question then arises wh<strong>et</strong>her the many results and seemingly almost consensus obtained by ignoring<br />
the limit<strong>at</strong>ions of usual st<strong>at</strong>istical tools could have led to erroneous conclusions about the tail behavior of<br />
the distributions of r<strong>et</strong>urns. Here, we propose to investig<strong>at</strong>e once more this <strong>de</strong>lic<strong>at</strong>e problem of the tail<br />
behavior of distributions of r<strong>et</strong>urns in or<strong>de</strong>r to shed new lights. To this aim, we investig<strong>at</strong>e two time series:<br />
the daily r<strong>et</strong>urns of the Dow Jones Industrial Average (DJ) In<strong>de</strong>x over a century and the five-minutes r<strong>et</strong>urns<br />
of the Nasdaq Composite in<strong>de</strong>x (ND) over one year from April 1997 to May 1998. These two s<strong>et</strong>s of d<strong>at</strong>a<br />
have been chosen since they are typical of the d<strong>at</strong>a s<strong>et</strong>s used in most previous studies. Their size (about<br />
20,000 d<strong>at</strong>a points), while significant compared with those used in investment and portfolio analysis, is<br />
however much smaller than recent d<strong>at</strong>a-intensive studies using ten of millions of d<strong>at</strong>a points (Gopikrishnan<br />
<strong>et</strong> al. 1998, M<strong>at</strong>ia <strong>et</strong> al. 2002, Mizuno <strong>et</strong> al. 2002).<br />
Our first conclusion is th<strong>at</strong> none of the standard param<strong>et</strong>ric family distributions (Par<strong>et</strong>o, exponential, str<strong>et</strong>che<strong>de</strong>xponential<br />
and incompl<strong>et</strong>e Gamma) fits s<strong>at</strong>isfactorily the DJ and ND d<strong>at</strong>a on the whole range of either<br />
positive or neg<strong>at</strong>ive r<strong>et</strong>urns. While this is also true for the family of str<strong>et</strong>ched exponential distribution, this<br />
family appears to be the best among the four consi<strong>de</strong>red param<strong>et</strong>ric families, in so far as it is able to fit the<br />
d<strong>at</strong>a over the largest interval. Our second and main conclusion comes from the discovery th<strong>at</strong> the Par<strong>et</strong>o<br />
distribution is a limit case of the Str<strong>et</strong>ched-Exponential distribution. This allows us to test the encompassing<br />
of these two mo<strong>de</strong>ls with respect to the true d<strong>at</strong>a, from which it seems th<strong>at</strong> the Str<strong>et</strong>ched-Exponential distribution<br />
is the most relevant mo<strong>de</strong>l of r<strong>et</strong>urns distributions. The regular <strong>de</strong>cay of the fractional exponent of<br />
the SE mo<strong>de</strong>l tog<strong>et</strong>her with the regular increase of the tail in<strong>de</strong>x of the Par<strong>et</strong>o mo<strong>de</strong>l lead us to think th<strong>at</strong> the<br />
extreme tail of true distribution of r<strong>et</strong>urns is f<strong>at</strong>ter th<strong>at</strong> any str<strong>et</strong>ched-exponential, strictly speaking -i.e., with<br />
a strickly positive fractional exponent- but thinner than any power law. Notwithstanding our best efforts, we<br />
cannot conclu<strong>de</strong> on the exact n<strong>at</strong>ure of the far-tail of distributions of r<strong>et</strong>urns. As already mentioned, other<br />
works have proposed the so-called inverse-cubic law (b = 3) based on the analysis of distributions of r<strong>et</strong>urns<br />
of high-frequency d<strong>at</strong>a aggreg<strong>at</strong>ed over hundreds up to thousands of stocks. This aggreg<strong>at</strong>ing procedure<br />
leads to novel problems of interpr<strong>et</strong><strong>at</strong>ion. We think th<strong>at</strong> the relevant question for most practical applic<strong>at</strong>ions<br />
is not to d<strong>et</strong>ermine wh<strong>at</strong> is the true asymptotic tail but wh<strong>at</strong> is the best effective <strong>de</strong>scription of the tails in<br />
the domain of useful applic<strong>at</strong>ions. As we shall show below, it may be th<strong>at</strong> the extreme asymptotic tail is<br />
a regularly varying function with tail in<strong>de</strong>x b = 3 for daily r<strong>et</strong>urns, but this is not very useful if this tail<br />
<strong>de</strong>scribes events whose recurrence time is a century or more. Our present work must thus be gauged as<br />
an <strong>at</strong>tempt to provi<strong>de</strong> a simple efficient effective <strong>de</strong>scription of the tails of distribution of r<strong>et</strong>urns covering<br />
most of the range of interest for practical applic<strong>at</strong>ions. We feel th<strong>at</strong> the efforts requested to go beyond the<br />
tails analyzed here, while of gre<strong>at</strong> interest from a scientific point of view to potentially help unravel mark<strong>et</strong><br />
mechanisms, may be too artificial and unreachable to have significant applic<strong>at</strong>ions.<br />
The paper is organized as follows.<br />
The next section is <strong>de</strong>voted to the present<strong>at</strong>ion of our two d<strong>at</strong>a s<strong>et</strong>s and to some of their basic st<strong>at</strong>istical<br />
properties, emphasizing their f<strong>at</strong> tailed behavior. We discuss, in particular, the importance of the so-called<br />
“lunch effect” for the tail properties of intraday r<strong>et</strong>urns. We then <strong>de</strong>monstr<strong>at</strong>e the presence of a significant<br />
temporal <strong>de</strong>pen<strong>de</strong>nce structure and study the possible non-st<strong>at</strong>ionary character of these time series.<br />
Section 3 <strong>at</strong>tempts to account for the temporal <strong>de</strong>pen<strong>de</strong>nce of our time series and investig<strong>at</strong>es its effect on<br />
the d<strong>et</strong>ermin<strong>at</strong>ion of the extreme behavior of the tails of the distribution of r<strong>et</strong>urns. In this goal, we build<br />
two simple long memory stochastic vol<strong>at</strong>ility processes whose st<strong>at</strong>ionary distributions are by construction<br />
either asymptotically regularly varying or exponential. We show th<strong>at</strong>, due to the long range <strong>de</strong>pen<strong>de</strong>nce<br />
on the vol<strong>at</strong>ility, the estim<strong>at</strong>ion with standard st<strong>at</strong>istical estim<strong>at</strong>ors is severely biased. This leads to a very<br />
unreliable estim<strong>at</strong>ion of the param<strong>et</strong>ers. These results justify our re-examin<strong>at</strong>ion of previous claims of<br />
regularly varying tails.<br />
4
To fit our two d<strong>at</strong>a s<strong>et</strong>s, section 4 proposes a general param<strong>et</strong>ric represent<strong>at</strong>ion of the distribution of r<strong>et</strong>urns<br />
encompassing both a regularly varying distribution in one limit of the param<strong>et</strong>ers and rapidly varying distributions<br />
of the class of str<strong>et</strong>ched exponential distributions in another limit. The use of regularly varying<br />
distributions have been justified above. From a theor<strong>et</strong>ical view point, the class of str<strong>et</strong>ched exponentials is<br />
motiv<strong>at</strong>ed in part by the fact th<strong>at</strong> the large <strong>de</strong>vi<strong>at</strong>ions of multiplic<strong>at</strong>ive processes are generically distributed<br />
with str<strong>et</strong>ched exponential distributions (Frisch and Sorn<strong>et</strong>te 1997). Str<strong>et</strong>ched exponential distributions are<br />
also parsimonious examples of the important subs<strong>et</strong> of sub-exponentials, th<strong>at</strong> is, of the general class of<br />
distributions <strong>de</strong>caying slower than an exponential. This class of sub-exponentials share several important<br />
properties of heavy-tailed distributions (Embrechts <strong>et</strong> al. 1997), not shared by exponentials or distributions<br />
<strong>de</strong>creasing faster than exponentials.<br />
The <strong>de</strong>scriptive power of these different hypotheses are compared in section 5. We first consi<strong>de</strong>r nested<br />
hypotheses and use Wilk’s test to this aim. It appears th<strong>at</strong> both the str<strong>et</strong>ched-exponential and the Par<strong>et</strong>o<br />
distributions are the most parsimonous mo<strong>de</strong>ls comp<strong>at</strong>ible with the d<strong>at</strong>a with a slight advantage in favor<br />
of the str<strong>et</strong>ched exponential mo<strong>de</strong>l. Then, in or<strong>de</strong>r to directly compare the <strong>de</strong>scriptive power of these two<br />
mo<strong>de</strong>ls, we perform encompassing tests, which prove the validity of the two represent<strong>at</strong>ions, but for different<br />
quantile ranges. Finally we show th<strong>at</strong> these two distributions can be s<strong>et</strong> within a single mo<strong>de</strong>l.<br />
Section 7 summarizes our results and conclu<strong>de</strong>s.<br />
2 Some basic st<strong>at</strong>istical fe<strong>at</strong>ures<br />
2.1 The d<strong>at</strong>a<br />
We use two s<strong>et</strong>s of d<strong>at</strong>a. The first sample consists in the daily r<strong>et</strong>urns 2 of the Dow Jones Industrial Average<br />
In<strong>de</strong>x (DJ) over the time interval from May 27, 1896 to May 31, 2000, which represents a sample size<br />
n = 28415. The second d<strong>at</strong>a s<strong>et</strong> contains the high-frequency (5 minutes) r<strong>et</strong>urns of Nasdaq Composite (ND)<br />
in<strong>de</strong>x for the period from April 8, 1997 to May 29, 1998 which represents n=22123 d<strong>at</strong>a points. The choice<br />
of these two d<strong>at</strong>a s<strong>et</strong>s is justified by their similarity with (1) the d<strong>at</strong>a s<strong>et</strong> of daily r<strong>et</strong>urns used by Longin<br />
(1996) particularly and (2) the high frequency d<strong>at</strong>a used by Guillaume <strong>et</strong> al. (1997), Lux (2000), Müller <strong>et</strong><br />
al. (1998) among others.<br />
For the intra-day Nasdaq d<strong>at</strong>a, there are two cave<strong>at</strong>s th<strong>at</strong> must be addressed. First, in or<strong>de</strong>r to remove the<br />
effect of overnight price jumps, we have d<strong>et</strong>ermined the r<strong>et</strong>urns separ<strong>at</strong>ely for each of 289 days contained in<br />
the Nasdaq d<strong>at</strong>a and have taken the union of all these 289 r<strong>et</strong>urn d<strong>at</strong>a s<strong>et</strong>s to obtain a global r<strong>et</strong>urn d<strong>at</strong>a s<strong>et</strong>.<br />
Second, the vol<strong>at</strong>ility of intra-day d<strong>at</strong>a are known to exhibit a U-shape, also called “lunch-effect”, th<strong>at</strong> is, an<br />
abnormally high vol<strong>at</strong>ility <strong>at</strong> the begining and the end of the trading day compared with a low vol<strong>at</strong>ility <strong>at</strong><br />
the approxim<strong>at</strong>e time of lunch. Such effect is present in our d<strong>at</strong>a, as <strong>de</strong>picted on figure 1, where the average<br />
absolute r<strong>et</strong>urns are shown as a function of the time within a trading day. It is <strong>de</strong>sirable to correct the d<strong>at</strong>a<br />
from this system<strong>at</strong>ic effect. This has been performed by renormalizing the 5 minutes-r<strong>et</strong>urns <strong>at</strong> a given<br />
moment of the trading day by the corresponding average absolute r<strong>et</strong>urn <strong>at</strong> the same moment. We shall refer<br />
to this time series as the corrected Nasdaq r<strong>et</strong>urns in constrast with the raw (incorrect) Nasdaq r<strong>et</strong>urns and<br />
we shall examine both d<strong>at</strong>a s<strong>et</strong>s for comparison.<br />
Although the distributions of positive and neg<strong>at</strong>ive r<strong>et</strong>urns are known to be very similar (Jon<strong>de</strong>au and<br />
Rockinger 2001, for instance), we have chosen to tre<strong>at</strong> them separ<strong>at</strong>ely. For the Dow Jones, this gives<br />
us 14949 positive and 13464 neg<strong>at</strong>ive d<strong>at</strong>a points while, for the Nasdaq, we have 11241 positive and 10751<br />
neg<strong>at</strong>ive d<strong>at</strong>a points.<br />
2 Throughout the paper, we will use compound r<strong>et</strong>urns, i.e., log-r<strong>et</strong>urns.<br />
5<br />
69
70 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Table 1 summarizes the main st<strong>at</strong>istical properties of these two time series (both for the raw and for the<br />
corrected Nasdaq r<strong>et</strong>urns) in terms of the average r<strong>et</strong>urns, their standard <strong>de</strong>vi<strong>at</strong>ions, the skewness and the<br />
excess kurtosis for four time scales of five minutes, an hour, one day and one month. The Dow Jones<br />
exhibits a significaltly neg<strong>at</strong>ive skewness, which can be ascribed to the impact of the mark<strong>et</strong> crashes. The<br />
raw Nasdaq r<strong>et</strong>urns are significantly positively skewed while the r<strong>et</strong>urns corrected for the “lunch effect” are<br />
neg<strong>at</strong>ively skewed, showing th<strong>at</strong> the lunch effect plays an important role in the shaping of the distribution<br />
of the intra-day r<strong>et</strong>urns. Note also the important <strong>de</strong>crease of the kurtosis after correction of the Nasdaq<br />
r<strong>et</strong>urns for lunch effect, confirming the strong impact of the lunch effect. In all cases, the excess-kurtosis are<br />
high and remains significant even after a time aggreg<strong>at</strong>ion of one month. Jarque-Bera’s test (Cromwell <strong>et</strong><br />
al. 1994), a joint st<strong>at</strong>istic using skewness and kurtosis coefficients, is used to reject the normality assumption<br />
for these time series.<br />
2.2 Existence of time <strong>de</strong>pen<strong>de</strong>nce<br />
It is well-known th<strong>at</strong> financial time series exhibit complex <strong>de</strong>pen<strong>de</strong>nce structures like h<strong>et</strong>eroskedasticity<br />
or non-linearities. These properties are clearly observed in our two times series. For instance, we have<br />
estim<strong>at</strong>ed the st<strong>at</strong>istical characteristic V (for positive random variables) called coefficient of vari<strong>at</strong>ion<br />
V = Std(X)<br />
E(X)<br />
, (1)<br />
which is often used as a testing st<strong>at</strong>istic of the randomness property of a time series. It can be applied to<br />
a sequence of points (or, intervals gener<strong>at</strong>ed by these points on the line). If these points are “absolutely<br />
random,” th<strong>at</strong> is, gener<strong>at</strong>ed by a Poissonian flow, then the intervals b<strong>et</strong>ween them are distributed according<br />
to an exponential distribution for which V = 1. If V > 1 are associ<strong>at</strong>ed with a clustering phenomenon. We estim<strong>at</strong>ed V = V (u) for extrema X > u<br />
and X < −u as function of threshold u (both for positive and for neg<strong>at</strong>ive extrema). The results are shown<br />
in figure 2 for the Dow Jones daily r<strong>et</strong>urns. As the results are essentially the same for the Nasdaq, we do not<br />
show them. Figure 2 shows th<strong>at</strong>, in the main range |X| < 0.02, containing ∼ 95% of the sample, V increases<br />
with u, indic<strong>at</strong>ing th<strong>at</strong> the “clustering” property becomes stronger as the threshold u increases.<br />
We have then applied several formal st<strong>at</strong>istical tests of in<strong>de</strong>pen<strong>de</strong>nce. We have first performed the Lagrange<br />
multiplier test proposed by Engle (1984) which leads to the T · R 2 test st<strong>at</strong>istic, where T <strong>de</strong>notes the sample<br />
size and R 2 is the d<strong>et</strong>ermin<strong>at</strong>ion coefficient of the regression of the squared centered r<strong>et</strong>urns xt on a constant<br />
and on q of their lags xt−1,xt−2,··· ,xt−q. Un<strong>de</strong>r the null hypothesis of homoskedastic time series, T · R 2<br />
follows a χ 2 -st<strong>at</strong>istic with q <strong>de</strong>grees of freedom. The test have been performed up to q = 10 and, in every<br />
case, the null hypothesis is strongly rejected, <strong>at</strong> any usual significance level. Thus, the time series are<br />
h<strong>et</strong>eroskedastics and exhibit vol<strong>at</strong>ility clustering. We have also performed a BDS test (Brock <strong>et</strong> al. 1987)<br />
which allows us to d<strong>et</strong>ect not only vol<strong>at</strong>ility clustering, like in the previous test, but also <strong>de</strong>parture from<br />
iid-ness due to non-linearities. Again, we strongly rejects the null-hypothesis of iid d<strong>at</strong>a, <strong>at</strong> any usual<br />
significance level, confirming the Lagrange multiplier test.<br />
3 Can long memory processes lead to misleading measures of extreme properties?<br />
Since the <strong>de</strong>scriptive st<strong>at</strong>istics given in the previous section have clearly shown the existence of a significant<br />
temporal <strong>de</strong>pen<strong>de</strong>nce structure, it is important to consi<strong>de</strong>r the possibility th<strong>at</strong> it can lead to erroneous conclusions<br />
on estim<strong>at</strong>ed param<strong>et</strong>ers. We first briefly recall the standard procedures used to investig<strong>at</strong>e extremal<br />
6
properties, stressing the problems and drawbacks arising from the existence of temporal <strong>de</strong>pen<strong>de</strong>nce. We<br />
then perform a numerical simul<strong>at</strong>ion to study the behavior of the estim<strong>at</strong>ors in presence of <strong>de</strong>pen<strong>de</strong>nce. We<br />
put particular emphasis on the possible appearance of significant biases due to <strong>de</strong>pen<strong>de</strong>nce in the d<strong>at</strong>a s<strong>et</strong>.<br />
Finally, we present the results on the extremal properties of our two DJ and ND d<strong>at</strong>a s<strong>et</strong>s in the light of the<br />
bootstrap results.<br />
3.1 Some theor<strong>et</strong>ical results<br />
Two limit theorems allow one to study the extremal properties and to d<strong>et</strong>ermine the maximum domain of<br />
<strong>at</strong>traction (MDA) of a distribution function in two forms.<br />
First, consi<strong>de</strong>r a sample of N iid realiz<strong>at</strong>ions X1,X2,··· ,XN. L<strong>et</strong> X ∧ <strong>de</strong>notes the maximum of this sample.<br />
Then, the Gne<strong>de</strong>nko theorem st<strong>at</strong>es th<strong>at</strong> if, after an a<strong>de</strong>qu<strong>at</strong>e centering and normaliz<strong>at</strong>ion, the distribution of<br />
X ∧ converges to a non-<strong>de</strong>gener<strong>at</strong>e distribution as N goes to infinity, this limit distribution is then necessarily<br />
the Generalized Extreme Value (GEV) distribution <strong>de</strong>fined by<br />
H ξ(x) = exp<br />
When ξ = 0, H ξ(x) should be un<strong>de</strong>rstood as<br />
Thus, for N large enough<br />
71<br />
<br />
−(1 + ξ · x) −1/ξ<br />
. (2)<br />
H ξ=0(x) = exp[−exp(−x)]. (3)<br />
Pr X ∧ < x = H ξ<br />
x − µ<br />
ψ<br />
<br />
, (4)<br />
for some value of the centering param<strong>et</strong>er µ, scale factor ψ and tail in<strong>de</strong>x ξ. It should be noted th<strong>at</strong> the<br />
existence of non-<strong>de</strong>gener<strong>at</strong>e limit distribution of properly centered and normalized X ∧ is a r<strong>at</strong>her strong limit<strong>at</strong>ion.<br />
There are a lot of distribution functions th<strong>at</strong> do not s<strong>at</strong>isfy this limit<strong>at</strong>ion, e.g., infinitely altern<strong>at</strong>ing<br />
functions b<strong>et</strong>ween a power-like and an exponential behavior.<br />
The second limit theorem is called after Gne<strong>de</strong>nko-Pickands-Balkema-<strong>de</strong> Haan (GPBH) and its formul<strong>at</strong>ion<br />
is as follows. In or<strong>de</strong>r to st<strong>at</strong>e the GPBH theorem, we <strong>de</strong>fine the right endpoint xF of a distribution function<br />
F(x) as xF = sup{x : F(x) < 1}. L<strong>et</strong> us call the function<br />
Pr{X − u ≥ x | X > u} ≡ ¯Fu(x) (5)<br />
the excess distribution function (DF). Then, this DF ¯Fu(x) belongs to the Maximum Domain of Attraction of<br />
H ξ(x) <strong>de</strong>fined by eq.(2) if and only if there exists a positive scale-function s(u), <strong>de</strong>pending on the threshold<br />
u, such th<strong>at</strong><br />
limu→xF sup 0≤x≤xF −u | ¯Fu(x) − ¯G(x/ξ,s(u))| = 0 , (6)<br />
where<br />
G(x/ξ,s) = 1 + ln(H ξ(x/s)) = 1 − (1 + ξ · x/s) −1/ξ . (7)<br />
By taking the limit ξ → 0, expression (7) leads to the exponential distribution. The support of the distribution<br />
function (7) is <strong>de</strong>fined as follows: <br />
0 x < ∞,<br />
0 x −d/ξ,<br />
if ξ 0<br />
if ξ < 0.<br />
(8)<br />
Thus, the Generalized Par<strong>et</strong>o Distribution has a finite support for ξ < 0.<br />
7
72 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
The form param<strong>et</strong>er ξ is of paramount importance for the form of the limiting distribution. Its sign d<strong>et</strong>ermines<br />
three possible limiting forms of the distribution of maximua: If ξ > 0, then the limit distribution is<br />
the Fréch<strong>et</strong> power-like distribution; If ξ = 0, then the limit distribution is the Gumbel (double-exponential)<br />
distribution; If ξ < 0, then the limit distribution has a support boun<strong>de</strong>d from above. All these three distributions<br />
are united in eq.(2) by this param<strong>et</strong>eriz<strong>at</strong>ion. The d<strong>et</strong>ermin<strong>at</strong>ion of the param<strong>et</strong>er ξ is the central<br />
problem of extreme value analysis. In<strong>de</strong>ed, it allows one to d<strong>et</strong>ermine the maximum domain of <strong>at</strong>traction<br />
of the un<strong>de</strong>rling distribution. When ξ > 0, the un<strong>de</strong>rlying distribution belongs to the Fréch<strong>et</strong> maximum<br />
domain of <strong>at</strong>traction and is regularly varying (power-like tail). When ξ = 0, it belongs to the Gumbel MDA<br />
and is rapidly varying (exponential tail), while if ξ < 0 it belongs to the Weibull MDA and has a finite right<br />
endpoint.<br />
3.2 Examples of slow convergence to limit GEV and GPD distributions<br />
There exist two ways of estim<strong>at</strong>ing ξ. First, if there is a sample of maxima (taken from sub-samples of<br />
sufficiently large size), then one can fit to this sample the GEV distribution, thus estim<strong>at</strong>ing the param<strong>et</strong>ers<br />
by Maximum Likelihood m<strong>et</strong>hod. Altern<strong>at</strong>ively, one can prefer the distribution of exceedance over a large<br />
threshold given by the GPD (7), whose tail in<strong>de</strong>x can be estim<strong>at</strong>ed with Pickands’ estim<strong>at</strong>or or by Maximum<br />
Likelihood, as previously. Hill’s estim<strong>at</strong>or cannot be used since it assumes ξ > 0, while the essence<br />
of extreme value analysis is, as we said, to test for the class of limit distributions without excluding any<br />
possibility, and not only to d<strong>et</strong>ermine the quantit<strong>at</strong>ive value of an exponent. Each of these m<strong>et</strong>hods has its<br />
advantages and drawbacks, especially when one has to study <strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a, as we show below.<br />
Given a sample of size N, one consi<strong>de</strong>r the q-maxima drawn from q sub-samples of size p (such th<strong>at</strong> p · q =<br />
N) to estim<strong>at</strong>e the param<strong>et</strong>ers (µ,ψ,ξ) in (4) by Maximum Likelihood. This procedure yields consistent and<br />
asymptotically Gaussian estim<strong>at</strong>ors, provi<strong>de</strong>d th<strong>at</strong> ξ > −1/2 (Smith 1985). The properties of the estim<strong>at</strong>ors<br />
still hold approxim<strong>at</strong>ely for <strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a, provi<strong>de</strong>d th<strong>at</strong> the inter<strong>de</strong>pen<strong>de</strong>nce of d<strong>at</strong>a is weak. However, it<br />
is difficult to choose an optimal value of q of the sub-samples. It <strong>de</strong>pends both on the size N of the entire<br />
sample and on the un<strong>de</strong>rlying distribution: the maxima drawn from an Exponential distribution are known<br />
to converge very quickly to Gumbel’s distribution (Hall and Wellnel 1979), while for the Gaussian law,<br />
convergence is particularly slow (Hall 1979).<br />
The second possibility is to estim<strong>at</strong>e the param<strong>et</strong>er ξ from the distribution of exceedances (the GPD) or<br />
Pickand’s estim<strong>at</strong>or. For this, one can use either the Maximum Likelihood estim<strong>at</strong>or or Pickands’ estim<strong>at</strong>or.<br />
Maximum Likelihood estim<strong>at</strong>ors are well-known to be the most efficient ones (<strong>at</strong> least for ξ > −1/2 and for<br />
in<strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a) but, in this particular case, Pickands’ estim<strong>at</strong>or works reasonably well. Given an or<strong>de</strong>red<br />
sample x1 ≤ x2 ≤ ···xN of size N, Pickands’ estim<strong>at</strong>or is given by<br />
ˆξk,N = 1<br />
ln2 ln xk − x2k<br />
. (9)<br />
x2k − x4k<br />
For in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically distributed d<strong>at</strong>a, this estim<strong>at</strong>or is consistent provi<strong>de</strong>d th<strong>at</strong> k is chosen so th<strong>at</strong><br />
k → ∞ and k/N → 0 as N → ∞. Morover, ˆ ξk,n is asymptotically normal with variance<br />
σ( ˆ ξk,N) 2 · k = ξ2 (22ξ+1 + 1)<br />
(2(2ξ . (10)<br />
− 1)ln2) 2<br />
In the presence of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween d<strong>at</strong>a, one can expect an increase of the standard <strong>de</strong>vi<strong>at</strong>ion, as reported<br />
by Kearns and Pagan (1997). For time <strong>de</strong>pen<strong>de</strong>nce of the GARCH class, Kearns and Pagan (1997) have<br />
in<strong>de</strong>ed <strong>de</strong>monstr<strong>at</strong>ed a significant increase of the standard <strong>de</strong>vi<strong>at</strong>ion of the tail in<strong>de</strong>x estim<strong>at</strong>or, such as<br />
Hill’s estim<strong>at</strong>or, by a factor more than seven with respect to their asymptotic properties for iid samples. This<br />
leads to very inaccur<strong>at</strong>e in<strong>de</strong>x estim<strong>at</strong>es for time series with this kind of temporal <strong>de</strong>pen<strong>de</strong>nce.<br />
8
Another problem lies in the d<strong>et</strong>ermin<strong>at</strong>ion of the optimal threshold u of the GPD, which is in fact rel<strong>at</strong>ed to<br />
the optimal d<strong>et</strong>ermin<strong>at</strong>ion of the sub-samples size q in the case of the estim<strong>at</strong>ion of the param<strong>et</strong>ers of the<br />
distribution of maximum.<br />
In sum, none of these m<strong>et</strong>hods seem really s<strong>at</strong>isfying and each one presents severe drawbacks. The estim<strong>at</strong>ion<br />
of the param<strong>et</strong>ers of the GEV distribution and of the GPD may be less sensitive to the <strong>de</strong>pen<strong>de</strong>nce of<br />
the d<strong>at</strong>a, but this property is only asymptotic, thus it requires a bootstrap investig<strong>at</strong>ion to be able to compare<br />
the real power of each estim<strong>at</strong>ion m<strong>et</strong>hod.<br />
As a first simple example illustr<strong>at</strong>ing the possibly very slow convergence to the limit distributions of extreme<br />
value theory mentioned above, l<strong>et</strong> us consi<strong>de</strong>r a simul<strong>at</strong>ed sample of iid Weibull random variables (we thus<br />
fulfill the most basic assumption of extreme values theory, i.e, iid-ness). For such a distribution, one must<br />
obtain in the limit N → ∞ the exponential distribution (ξ = 0). We took two values for the exponent of<br />
the Weibull distribution: c = 0.7 and c = 0.3, with d = 1 (scale param<strong>et</strong>er). Such the estim<strong>at</strong>ion of ξ by<br />
the distribution of the GPD of exceedance should give estim<strong>at</strong>ed values of ξ close to zero in the limit of<br />
large N. In or<strong>de</strong>r to use the GPD, we have taken the conditional Weibull distribution un<strong>de</strong>r condition X ><br />
Uk,k = 1...15, where the thresholds Uk are choosen as: U1 = 0.1; U2 = 0.3; U3 = 1; U4 = 3; U5 = 10; U6 =<br />
30; U7 = 100; U8 = 300; U9 = 1000; U10 = 3000; U11 = 10 4 ; U12 = 3 · 10 4 ; U13 = 10 5 ; U14 = 3 · 10 5 and<br />
U15 = 10 6 .<br />
For each simul<strong>at</strong>ion, the size of the sample above the consi<strong>de</strong>red threshold Uk is choosen equal to 50,000<br />
in or<strong>de</strong>r to g<strong>et</strong> small standard <strong>de</strong>vi<strong>at</strong>ions. The Maximum-Likelihood estim<strong>at</strong>es of the GPD form param<strong>et</strong>er<br />
ξ are shown in figure 3. For c = 0.7, the threshold U7 gives an estim<strong>at</strong>e ξ = 0.0123 with standard <strong>de</strong>vi<strong>at</strong>ion<br />
equal to 0.0045, i.e., the estim<strong>at</strong>e for ξ differs significantly from zero (which is the correct theor<strong>et</strong>ical limit<br />
value). This occurs notwithstanding the huge size of the implied d<strong>at</strong>a s<strong>et</strong>; in<strong>de</strong>ed, the probability PrX > U7<br />
for c = 0.7 is about 10 −9 , so th<strong>at</strong> in or<strong>de</strong>r to obtain a d<strong>at</strong>a s<strong>et</strong> of conditional samples from an unconditional<br />
d<strong>at</strong>a s<strong>et</strong> of the size studied here (50,000 realiz<strong>at</strong>ions above U7), the size of such unconditional sample should<br />
be approxim<strong>at</strong>ely 10 9 times larger than the number of “peaks over threshold”, i.e., it is practically impossible<br />
to have such a sample. For c = 0.3 the convergence to the theor<strong>et</strong>ical value zero is even slower. In<strong>de</strong>ed,<br />
even the largest financial d<strong>at</strong>as<strong>et</strong>s for a single ass<strong>et</strong>, drawn from high frequency d<strong>at</strong>a, are no larger than or<br />
of the or<strong>de</strong>r of one million points 3 . The situ<strong>at</strong>ion does not change even for d<strong>at</strong>a s<strong>et</strong>s one or two or<strong>de</strong>rs of<br />
magnitu<strong>de</strong>s larger as consi<strong>de</strong>red in (Gopikrishnan <strong>et</strong> al. 1998), obtained by aggreg<strong>at</strong>ing thousands of stocks<br />
4 . Thus, although the GPD form param<strong>et</strong>er should be zero theor<strong>et</strong>ically in the limit of large sample for the<br />
Weibull distribution, this limit cannot be reached for any available sample sizes.<br />
This is a clear illustr<strong>at</strong>ion th<strong>at</strong> a rapidly varying distribution, like the Weibull distribution with exponent<br />
smaller than one, i.e., a Str<strong>et</strong>ched-Exponential distribution, can be mistaken for a Par<strong>et</strong>o or any other regularly<br />
varying distribution for any practical applic<strong>at</strong>ions.<br />
3.3 Gener<strong>at</strong>ion of a long memory process with a well-<strong>de</strong>fined st<strong>at</strong>ionary distribution<br />
In or<strong>de</strong>r to study the performance of the various estim<strong>at</strong>ors of the tail in<strong>de</strong>x ξ and the influence of inter<strong>de</strong>pen<strong>de</strong>nce<br />
of sample values, we have gener<strong>at</strong>ed six samples with distinct properties. The first three samples<br />
are ma<strong>de</strong> of iid realiz<strong>at</strong>ions drawn respectively from a Par<strong>et</strong>o Distribution with tail in<strong>de</strong>x b = 3 and from<br />
a Str<strong>et</strong>ched-Exponential distribution with exponent c = 0.3 and c = 0.7. The three other samples contain<br />
realiz<strong>at</strong>ions exhibiting long-range memory with the same three distributions as for the first three samples:<br />
a regularly varying distribution with tail in<strong>de</strong>x b = 3 and a Str<strong>et</strong>ched-Exponential distribution with expo-<br />
3 One year of d<strong>at</strong>a sampled <strong>at</strong> the 1 minute time scale gives approxim<strong>at</strong>ely 1.2 · 10 5 d<strong>at</strong>a points<br />
4 In this case, another issue arises concerning the fact th<strong>at</strong> the aggreg<strong>at</strong>ion of r<strong>et</strong>urns from differents ass<strong>et</strong>s may distord the<br />
inform<strong>at</strong>ion and the very structure of the tails of the pdfs, if they exhibit some intrinsic variability (M<strong>at</strong>ia <strong>et</strong> al. 2002).<br />
9<br />
73
74 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
nent c = 0.3 and c = 0.7. Thus, the three first samples are the iid counterparts of the l<strong>at</strong>er ones. The<br />
sample with regularly varying iid distributions converges to the Fréch<strong>et</strong>’s maximum domain of <strong>at</strong>traction<br />
with ξ = 1/3 = 0.33, while the iid Str<strong>et</strong>ched-Exponential distribution converges to Gumbel’s maximum domain<br />
of <strong>at</strong>traction with ξ = 0. We now study how well can one distinguish b<strong>et</strong>ween these two distributions<br />
belonging to two different maximum domains of <strong>at</strong>traction.<br />
For the stochastic processes with long memory, we use a simple stochastic vol<strong>at</strong>ility mo<strong>de</strong>l. First, we<br />
construct a Gaussian process {Xt}t≥1 with correl<strong>at</strong>ion function<br />
C(t) =<br />
1<br />
(1+|t|) α<br />
if |t| ≤ T ,<br />
0 if |t| > T .<br />
It should be noted th<strong>at</strong>, in or<strong>de</strong>r for the stochastic process to be well-<strong>de</strong>fined, the correl<strong>at</strong>ion function must<br />
s<strong>at</strong>isfy a positivity condition. More precisely, the spectral <strong>de</strong>nsity (the Fourier transform of the correl<strong>at</strong>ion<br />
function) must remain positive. This condition imposes th<strong>at</strong> the dur<strong>at</strong>ion of the memory T be larger than a<br />
constant <strong>de</strong>pending on α.<br />
The next step consists in building the process {Ut}t≥1, <strong>de</strong>fined by<br />
(11)<br />
Ut = Φ(Xt) , (12)<br />
where Φ(·) is the Gaussian distribution function. The process {Ut}t≥1 exhibits exactly the same long range<br />
<strong>de</strong>pen<strong>de</strong>nce as the process {Xt}t≥1. This is ensured by the property of invariance of the copula un<strong>de</strong>r<br />
strickly increasing change of variables. L<strong>et</strong> us recall th<strong>at</strong> a copula is the m<strong>at</strong>hem<strong>at</strong>ical embodiement of the<br />
<strong>de</strong>pen<strong>de</strong>nce structure b<strong>et</strong>ween different random variables (Joe 1997, Nelsen 1998). The process {Ut}t≥1<br />
thus possesses a Gaussian copula <strong>de</strong>pen<strong>de</strong>nce structure with long memory and uniform marginals 5 .<br />
In the last step, we <strong>de</strong>fine the vol<strong>at</strong>ility process<br />
σt = σ0 ·U −1/b<br />
t , (13)<br />
which ensures th<strong>at</strong> the st<strong>at</strong>ionary distribution of the vol<strong>at</strong>ility is a Par<strong>et</strong>o distribution with tail in<strong>de</strong>x b. Such<br />
a distribution of the vol<strong>at</strong>ility is not realistic in the bulk which is found to be approxim<strong>at</strong>ely a lognormal<br />
distribution for not too large vol<strong>at</strong>ilities (Sorn<strong>et</strong>te <strong>et</strong> al. 2000), but is in agreement with the hypothesis of<br />
an asymptotic regularly varying distribution. A change of variable more complic<strong>at</strong>ed than (13) can provi<strong>de</strong><br />
a more realistic behavior of the vol<strong>at</strong>ility on the entire range of the distribution but our main goal is not to<br />
provi<strong>de</strong> a realistic stochastic vol<strong>at</strong>ility mo<strong>de</strong>l but only to exhibit a long memory process with well-<strong>de</strong>fined<br />
prescribed marginals in or<strong>de</strong>r to test the influence of a long range <strong>de</strong>pen<strong>de</strong>nce structure.<br />
The r<strong>et</strong>urn process is then given by<br />
rt = σt · εt , (14)<br />
where the εt are Gaussian random variables in<strong>de</strong>pen<strong>de</strong>nt from σt. The construction (14) ensures the <strong>de</strong>correl<strong>at</strong>ion<br />
of the r<strong>et</strong>urns <strong>at</strong> every time lag. The st<strong>at</strong>ionary distribution of rt admits the <strong>de</strong>nsity<br />
p(r) = 2 b−1<br />
<br />
b − 1 r2 b<br />
2 · Γ , , (15)<br />
2 2 rb+1 <br />
b−1 r2<br />
which is regularly varying <strong>at</strong> infinity since Γ 2 , 2 goes to Γ <br />
b−1<br />
2 . This compl<strong>et</strong>es the construction and<br />
characteriz<strong>at</strong>ion of our long memory process with regularly varying st<strong>at</strong>ionary distribution.<br />
5 Of course, one can make the correl<strong>at</strong>ion as small as one wants un<strong>de</strong>r an a<strong>de</strong>qu<strong>at</strong>e choice of a strickly increasing transform<strong>at</strong>ion<br />
but this does not change the fact th<strong>at</strong> the <strong>de</strong>pen<strong>de</strong>nce remains unchanged. This is another illustr<strong>at</strong>ion of the fact th<strong>at</strong> the correl<strong>at</strong>ion<br />
is not always a good and adapted measure of <strong>de</strong>pen<strong>de</strong>nce (Malevergne and Sorn<strong>et</strong>te 2002).<br />
10
In or<strong>de</strong>r to obtain a process with Str<strong>et</strong>ched-Exponential distribution with long range <strong>de</strong>pen<strong>de</strong>nce, we apply<br />
to {rt}t≥1 the following increasing mapping G : r → y<br />
⎧<br />
⎪⎨ (x0 + ln<br />
G(r) =<br />
⎪⎩<br />
r<br />
r0 )1/c r > r0<br />
sgn(r) · |r| 1/c |r| ≤ r0<br />
−(r0 + ln|r/r0|) 1/c (16)<br />
r < −r0 .<br />
This transform<strong>at</strong>ion gives a str<strong>et</strong>ched exponential of in<strong>de</strong>x c for all values of the r<strong>et</strong>urn larger than the scale<br />
factor r0. This <strong>de</strong>rives from the fact th<strong>at</strong> the process {rt}t≥1 admits a regularly varying distribution function,<br />
characterized by ¯Fr(r) = 1 − Fr(r) = L(r)|r| −b , for some slowly varying function L. As a consequence, the<br />
st<strong>at</strong>ionary distribution of {Yt}t≥1 is given by<br />
¯FY (y) = L r0e −x0 exp(y c ) e br0<br />
which is a Str<strong>et</strong>ched-Exponential distribution.<br />
x b 0<br />
75<br />
· e −b|y|c<br />
, ∀|y| > r0, (17)<br />
= L ′ (y) · e −b|y|c<br />
, L ′ is slowly varying <strong>at</strong> infinity, (18)<br />
To summarize, starting with a long memory Gaussian process, we have <strong>de</strong>fined a long memory process<br />
characterized by a st<strong>at</strong>ionary distribution function of our choice, thanks to the invariance of the temporal<br />
<strong>de</strong>pen<strong>de</strong>nce structure (the copula) un<strong>de</strong>r strictly increasing change of variable. In particular, this approach<br />
gives long memory processes with a regularly varying marginal distribution and with a str<strong>et</strong>ched-exponential<br />
distribution. Notwithstanding the difference in their marginals, these two processes possess by construction<br />
exactly the same time <strong>de</strong>pen<strong>de</strong>nce. This allows us to compare the impact of the same <strong>de</strong>pen<strong>de</strong>nce on these<br />
two classes of marginals.<br />
3.4 Results of numerical simul<strong>at</strong>ions<br />
We have gener<strong>at</strong>ed 1000 samples of each kind (iid Str<strong>et</strong>ched-Exponential, iid Par<strong>et</strong>o, long memory process<br />
with a Par<strong>et</strong>o distribution and with a Str<strong>et</strong>ched-Exponential distribution). Each sample contains 10,000<br />
realiz<strong>at</strong>ions, which is approxim<strong>at</strong>ely the number of points in each tail of our real samples. In or<strong>de</strong>r to<br />
gener<strong>at</strong>e the Gaussian process with correl<strong>at</strong>ion function (11), we have used the algorithm based on Fast<br />
Fourier Transform <strong>de</strong>scribed in Beran (1994). The param<strong>et</strong>er T has been s<strong>et</strong> to 250 and α to 0.5 (it can be<br />
checked th<strong>at</strong> for α = 0.5 the lower bound for T is equal to 23).<br />
Panel (a) of table 2 presents the mean values and standard <strong>de</strong>vi<strong>at</strong>ions of the Maximum Likelihood estim<strong>at</strong>es<br />
of ξ, using the Generalized Extreme Value distribution and the Generalized Par<strong>et</strong>o Distribution for the three<br />
samples of iid d<strong>at</strong>a. To estim<strong>at</strong>e the param<strong>et</strong>ers of the GEV distribution and study the influence of the<br />
sub-sample size, we have grouped the d<strong>at</strong>a in clusters of size q = 10,50,100 and 200. For the analysis in<br />
terms of the GPD, we have consi<strong>de</strong>red four different large thresholds u, corresponding to the quantiles 90%,<br />
95%, 99% and 99.5%. The estim<strong>at</strong>es obtained from the distribution of maxima are significantly different<br />
from the theor<strong>et</strong>ical ones: 0.2 in average over the four different size of sub-samples intead of 0.0 for the<br />
Str<strong>et</strong>ched-Exponential distribution with c = 0.7, 1.0 instead of 0.0 for c = 0.3 for the Str<strong>et</strong>ched-Exponential<br />
distribution and 0.40 instead of 0.33 for the Par<strong>et</strong>o Distribution. At the same time, the standard <strong>de</strong>vi<strong>at</strong>ion of<br />
these estim<strong>at</strong>or remains very low. This significant bias of the estim<strong>at</strong>or is a clear sign th<strong>at</strong> the distribution of<br />
the maximum has not y<strong>et</strong> converged to the asymptotic GEV distribution, even for subsamples of size 200.<br />
The results are b<strong>et</strong>ter with smaller biases for the Maximum Likelihood estim<strong>at</strong>es obtained from the GPD.<br />
However, the standard <strong>de</strong>vi<strong>at</strong>ions are significantly larger than in the previous case, which testifies of the high<br />
variability of this estim<strong>at</strong>or. Thus, for such sample sizes, the GEV and GPD Maximum Likelihood estim<strong>at</strong>es<br />
seem not very reliable due to an important bias for the former and large st<strong>at</strong>istical fluctu<strong>at</strong>ions for the l<strong>at</strong>er.<br />
11
76 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Nevertheless, an optimistic view point is th<strong>at</strong>, discarding the largest quantile of 0.995, the GPD estim<strong>at</strong>or is<br />
comp<strong>at</strong>ible with a value ξ = 0.32 ± 0.1 for the iid d<strong>at</strong>a with Par<strong>et</strong>o distribution while it is comp<strong>at</strong>ible with<br />
ξ = (−0.21) − (0.19) ± 0.6 − 0.8 for c = 0.3 and with ξ = (−0.13) − (0.11) ± 0.04 − 0.5 for c = 0.7. The<br />
consistency of the GPD estim<strong>at</strong>or in the case of the Par<strong>et</strong>o distribution in contrast with the wild vari<strong>at</strong>ions for<br />
the Str<strong>et</strong>ched-Exponential distributions suggests th<strong>at</strong> one could conclu<strong>de</strong> th<strong>at</strong> the first result qualifies (correctly)<br />
a regularly varying function while the second one either (correctly) disqualifies a regularly varying<br />
function or more conserv<strong>at</strong>ily is unable to conclu<strong>de</strong>. In other words, when the sample d<strong>at</strong>a is truely Par<strong>et</strong>o,<br />
the GPD estim<strong>at</strong>or seems to be able to r<strong>et</strong>rieve this inform<strong>at</strong>ion reliably, in constrast with the GEV estim<strong>at</strong>or<br />
which is quite unreliable in all cases.<br />
Panel (b) of table 2 presents the same results for <strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a. The GEV estim<strong>at</strong>es exhibit in each case a<br />
significant bias, either positive or neg<strong>at</strong>ive and a huge increase of the standard <strong>de</strong>vi<strong>at</strong>ion in the case of the<br />
Str<strong>et</strong>ched-Exponential with exponent c = 0.7. Interestingly, the GEV estim<strong>at</strong>or for the Par<strong>et</strong>o distribution is<br />
utterly wrong. The situ<strong>at</strong>ion is different for the GPD estim<strong>at</strong>es which show a weak bias not really sensitive to<br />
the quantile. In constrast, the standard <strong>de</strong>vi<strong>at</strong>ions of the GPD estim<strong>at</strong>ors strongly increase with the quantile,<br />
which is n<strong>at</strong>ural since the number of oberv<strong>at</strong>ions <strong>de</strong>creases accordingly. The GPD behaves surprisingly well<br />
and seems to be the only one able to perfom a reasonable estim<strong>at</strong>ion of the tail in<strong>de</strong>x.<br />
To summarize, the Maximum Likelihood estim<strong>at</strong>ors <strong>de</strong>rived form the GEV or GPD distributions are not very<br />
efficient in the presence of <strong>de</strong>pen<strong>de</strong>nce in the d<strong>at</strong>a and of non-asymptotic effects due to the slow convergence<br />
toward the asymptotic GEV or GPD distributions. The only positive note is th<strong>at</strong> the GPD estim<strong>at</strong>or correctly<br />
recovers the range of the in<strong>de</strong>x ξ with an uncertainty smaller than 20% for d<strong>at</strong>a with a pure Par<strong>et</strong>o distribution<br />
while it is cannot reject the hypothesis th<strong>at</strong> ξ = 0 when the d<strong>at</strong>a is gener<strong>at</strong>ed with a Str<strong>et</strong>ched-Exponential<br />
distribution, albeit with a very large uncertainty, in other words with very little power.<br />
Table 3 focuses on the results given by Pickands’ estim<strong>at</strong>or for the tail in<strong>de</strong>x of the GPD. For each threshods<br />
u, corresponding to the quantiles 90%, 95%, 99% and 99.5% respectively, the results of our simul<strong>at</strong>ions are<br />
given for two particular values of k (<strong>de</strong>fined in 9) corresponding to N/k = 4, which is the largest admissible<br />
value, and N/k = 10 corresponding to be sufficiently far in the tail of the GPD. Table 3 provi<strong>de</strong>s the mean<br />
value and the numerically estim<strong>at</strong>ed as well as the theor<strong>et</strong>ical (given by (10)) standard <strong>de</strong>vi<strong>at</strong>ion of ˆ ξk,N.<br />
Panel (a) gives the result for iid d<strong>at</strong>a. The mean values do not exhibit a significant bias for the Par<strong>et</strong>o<br />
distribution and the Str<strong>et</strong>ched-Exponential with c = 0.7, but are utterly wrong in the case c = 0.3 since<br />
the estim<strong>at</strong>es are comparable with those given for the Par<strong>et</strong>o distribution. In each case, we note a very<br />
good agreement b<strong>et</strong>ween the empirical and theor<strong>et</strong>ical standard <strong>de</strong>vi<strong>at</strong>ions, even for the larger quantiles<br />
(and thus the smaller samples). Panel (b) presents the results for <strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a. The estim<strong>at</strong>ed standard<br />
<strong>de</strong>vi<strong>at</strong>ions remains of the same or<strong>de</strong>r as the theor<strong>et</strong>ical ones, contrarily to results reported by Kearns and<br />
Pagan (1997) for GARCH processes. However, like these authors, we find th<strong>at</strong> the bias, either positive or<br />
neg<strong>at</strong>ive, becomes very significant and leads on to misclassify a Str<strong>et</strong>ched-Exponential distribution with<br />
c = 0.3 for a Par<strong>et</strong>o distribution with b = 3. Thus, in presence of <strong>de</strong>pen<strong>de</strong>nce, Pickands’ estim<strong>at</strong>or is<br />
unreliable.<br />
To summarize, the impact of the <strong>de</strong>pen<strong>de</strong>nce can add a severe sc<strong>at</strong>ter of estim<strong>at</strong>ors which increase their<br />
standard <strong>de</strong>vi<strong>at</strong>ion. The d<strong>et</strong>ermin<strong>at</strong>ion of the maximum domain of <strong>at</strong>traction with usual estim<strong>at</strong>ors does<br />
not appear to be a very efficient way to study the extreme properties of <strong>de</strong>pen<strong>de</strong>nt times series. Almost all<br />
the previous studies which have investig<strong>at</strong>ed the tail behavior of ass<strong>et</strong> r<strong>et</strong>urns distributions have focused on<br />
these m<strong>et</strong>hods (see the influencial works of Longin (1996) for instance) and may thus have led to spurious<br />
results on the d<strong>et</strong>ermin<strong>at</strong>ion of the tail behavior. In particular, our simul<strong>at</strong>ions show th<strong>at</strong> rapidly varying<br />
function may be mistaken for regularly varying functions. Thus, according to our simul<strong>at</strong>ions, this casts<br />
doubts on the strength of the conclusion of previous works th<strong>at</strong> the distributions of r<strong>et</strong>urns are regularly<br />
varying as seems to have been the consensus until now and suggests to re-examine the possibility th<strong>at</strong> the<br />
distribution of r<strong>et</strong>urns may be rapidly varying as suggested by Gouriéroux and Jasiak (1998) or Laherrère<br />
12
and Sorn<strong>et</strong>te (1999) for instance. We now turn to this question using the framework of GEV and GDP<br />
estim<strong>at</strong>ors just <strong>de</strong>scribed.<br />
3.5 GEV and GPD estim<strong>at</strong>ors of the Dow Jones and Nasdaq d<strong>at</strong>a s<strong>et</strong>s<br />
We have applied the same analysis as in the previous section on the real samples of the Dow Jones and<br />
Nasdaq (raw and corrected) r<strong>et</strong>urns. To this aim, we have randomly gener<strong>at</strong>ed one thousand sub-samples,<br />
each sub-sample being constituted of ten thousand d<strong>at</strong>a points in the positive or neg<strong>at</strong>ive parts of the samples<br />
respectively (without replacement). Obviously, among the one thousand sub-samples, many of them are<br />
inter<strong>de</strong>pen<strong>de</strong>nt as they contain parts of the same observed values. With this d<strong>at</strong>abase, we have estim<strong>at</strong>ed the<br />
mean value and standard <strong>de</strong>vi<strong>at</strong>ions of Pickands’ estim<strong>at</strong>or for the GPD <strong>de</strong>rived from the upper quantiles of<br />
these distributions, and of ML-estim<strong>at</strong>ors for the distribution of maximum and for the GPD. The results are<br />
given in tables 4 and 5.<br />
These results confirm the confusion about the tail behavior of the r<strong>et</strong>urns distributions and it seems impossible<br />
to exclu<strong>de</strong> a rapidly varying behavior of their tails. In<strong>de</strong>ed, even the estim<strong>at</strong>ions performed by Maximum<br />
Likelihood with the GPD tail in<strong>de</strong>x, which have appeared as the less unreliable estim<strong>at</strong>or in our previous<br />
tests, does not allow us to reject the hypothesis th<strong>at</strong> the tails of the empirical distributions of r<strong>et</strong>urns are<br />
rapidly varying.<br />
For the Nasdaq d<strong>at</strong>as<strong>et</strong>, accounting for the lunch effect does not yield a significant change in the estim<strong>at</strong>ions,<br />
except for a very strong increase of the standard vari<strong>at</strong>ion of the GPD Maximum Likelihood estim<strong>at</strong>or. This<br />
results from the fact th<strong>at</strong> extremes are no more domin<strong>at</strong>ed by the few largest realiz<strong>at</strong>ions of the r<strong>et</strong>urns <strong>at</strong><br />
the begining or the end of trading days. In<strong>de</strong>ed, panel (b) of table 4 shows th<strong>at</strong> the sample variance of the<br />
GPD maximum likelihood estim<strong>at</strong>e vanishes for quantile 99% and 99.5%. It is due to the important overlap<br />
of the sub-samples tog<strong>et</strong>her with the impact of the extreme realiz<strong>at</strong>ions of the r<strong>et</strong>urns <strong>at</strong> the open or close<br />
trading days. In panel (c), which corresponds to the Nasdaq d<strong>at</strong>a corrected for the lunch effect, the sample<br />
variance vanishes only in one case (instead of four), clearly showing th<strong>at</strong> extremes are less domin<strong>at</strong>ed by<br />
the large r<strong>et</strong>urns of the beginning and <strong>at</strong> the end of each day.<br />
As a last non-param<strong>et</strong>ric <strong>at</strong>tempt to distinguish b<strong>et</strong>ween a regularly varying tail and a rapidly varying tail<br />
of the exponential or Str<strong>et</strong>ched-Exponential families, we study the Mean Excess Function which is one of<br />
the known m<strong>et</strong>hods th<strong>at</strong> often can help in <strong>de</strong>ciding wh<strong>at</strong> param<strong>et</strong>ric family is appropri<strong>at</strong>e for approxim<strong>at</strong>ion<br />
(see for d<strong>et</strong>ails Embrechts <strong>et</strong> al. (1997)). The Mean Excess Function MEF(u) of a random value X (also<br />
called “shortfall” when applied to neg<strong>at</strong>ive r<strong>et</strong>urns in the context of financial risk management) is <strong>de</strong>fined as<br />
MEF(u) = E(X − u|X > u) . (19)<br />
The Mean Excess Function MEF(u) is obviously rel<strong>at</strong>ed to the GPD for sufficiently large threshold u and<br />
its behavior can be <strong>de</strong>rived in this limit for the three maximum domains of <strong>at</strong>traction. In addition, more<br />
precise results can be given for particular radom variables, even in a non-asymptotic regime. In<strong>de</strong>ed, for an<br />
exponential random variable X, the MEF(u) is just a constant. For a Par<strong>et</strong>o random variable, the MEF(u)<br />
is a straight increasing line, whereas for the Streched-Exponential and the Gauss distributions the MEF(u)<br />
is a <strong>de</strong>creasing function. We evalu<strong>at</strong>ed the sample analogues of the MEF(u) (Embrechts <strong>et</strong> al. 1997, p.296)<br />
which are shown in figure 4. All <strong>at</strong>tempts to find a constant or a linearly increasing behavior of the MEF(u)<br />
on the main central part of the range of r<strong>et</strong>urns were ineffective. In the central part of the range of neg<strong>at</strong>ive<br />
r<strong>et</strong>urns (|X| > 0.002; q ∼ = 98% for ND d<strong>at</strong>a, and |X| > 0.025 ; q ∼ = 96% for DJ d<strong>at</strong>a), the MEF(u) behaves<br />
like a convex function which exclu<strong>de</strong> both exponential and power (Par<strong>et</strong>o) distributions. Thus, the MEF(u)<br />
tool does not support using any of these two distributions.<br />
In view of the stalem<strong>at</strong>e reached with the above non-param<strong>et</strong>ric approaches and in particular with the stan-<br />
13<br />
77
78 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
dard extreme value estim<strong>at</strong>ors, the sequel of this paper is <strong>de</strong>voted to the investig<strong>at</strong>ion of a param<strong>et</strong>ric approach<br />
in or<strong>de</strong>r to <strong>de</strong>ci<strong>de</strong> which class of extreme value distributions, rapidly versus regularly varying, accounts<br />
best for the empirical distributions of r<strong>et</strong>urns.<br />
4 Fitting distributions of r<strong>et</strong>urns with param<strong>et</strong>ric <strong>de</strong>nsities<br />
Since our previous results lead to doubt the validity of the rejection of the hypothesis th<strong>at</strong> the distribution of<br />
r<strong>et</strong>urns are rapidly varying, we now propose to pit a param<strong>et</strong>ric champion for this class of functions against<br />
the Par<strong>et</strong>o champion of regularly varying functions. To represent the class of rapidly varying functions, we<br />
propose the family of Str<strong>et</strong>ched-Exponentials. As discussed in the introduction, the class of str<strong>et</strong>ched exponentials<br />
is motiv<strong>at</strong>ed in part from a theor<strong>et</strong>ical view point by the fact th<strong>at</strong> the large <strong>de</strong>vi<strong>at</strong>ions of multiplic<strong>at</strong>ive<br />
processes are generically distributed with str<strong>et</strong>ched exponential distributions (Frisch and Sorn<strong>et</strong>te 1997).<br />
Str<strong>et</strong>ched exponential distributions are also parsimonious examples of sub-exponential distributions with f<strong>at</strong><br />
tails for instance in the sense of the asymptotic probability weight of the maximum compared with the sum<br />
of large samples (Feller 1971). Notwithstanding their f<strong>at</strong>-tailness, Str<strong>et</strong>ched Exponential distributions have<br />
all their moments finite 6 , in constrast with regularly varying distributions for which moments of or<strong>de</strong>r equal<br />
to or larger than the in<strong>de</strong>x b are not <strong>de</strong>fined. This property may provi<strong>de</strong> a substantial advantage to exploit in<br />
generaliz<strong>at</strong>ions of the mean-variance portfolio theory using higher-or<strong>de</strong>r moments (Rubinstein 1973, Fang<br />
and Lai 1997, Hwang and S<strong>at</strong>chell 1999, Sorn<strong>et</strong>te <strong>et</strong> al. 2000, An<strong>de</strong>rsen and Sorn<strong>et</strong>te 2001, Jurczenko and<br />
Maill<strong>et</strong> 2002, Malevergne and Sorn<strong>et</strong>te 2002, for instance ). Moreover, the existence of all moments is<br />
an important property allowing for an efficient estim<strong>at</strong>ion of any high-or<strong>de</strong>r moment, since it ensures th<strong>at</strong><br />
the estim<strong>at</strong>ors are asymptotically Gaussian. In particular, for Str<strong>et</strong>ched-Exponentially distributed random<br />
variables, the variance, skewness and kurtosis can be well estim<strong>at</strong>ed, contrarily to random variables with<br />
regularly varying distribution with tail in<strong>de</strong>x in the range 3 − 5.<br />
4.1 Definition of a general 3-param<strong>et</strong>ers family of distributions<br />
We thus consi<strong>de</strong>r a general 3-param<strong>et</strong>ers family of distributions and its particular restrictions corresponding<br />
to some fixed value(s) of two (one) param<strong>et</strong>ers. This family is <strong>de</strong>fined by its <strong>de</strong>nsity function given by:<br />
<br />
A(b,c,d,u) x<br />
fu(x|b,c,d) =<br />
−(b+1) exp − <br />
x c<br />
d if x u > 0<br />
(20)<br />
0 if x < u.<br />
Here, b,c,d are unknown param<strong>et</strong>ers, u is a known lower threshold th<strong>at</strong> will be varied for the purposes of<br />
our analysis and A(b,c,d,u) is a normalizing constant given by the expression:<br />
A(b,c,d,u) =<br />
db c<br />
Γ(−b/c,(u/d) c , (21)<br />
)<br />
where Γ(a,x) <strong>de</strong>notes the (non-normalized) incompl<strong>et</strong>e Gamma function. The param<strong>et</strong>er b ranges from<br />
minus infinity to infinity while c and d range from zero to infinity. In the particular case where c = 0,<br />
the param<strong>et</strong>er b also needs to be positive to ensure the normaliz<strong>at</strong>ion of the probability <strong>de</strong>nsity function<br />
(pdf). The interval of <strong>de</strong>finition of this family is the positive semi-axis. Neg<strong>at</strong>ive log-r<strong>et</strong>urns will be studied<br />
by taking their absolute values. The family (20) inclu<strong>de</strong>s several well-known pdf’s often used in different<br />
applic<strong>at</strong>ions. We enumer<strong>at</strong>e them.<br />
6 However, they do not admit an exponential moment, which leads to problems in the reconstruction of the distribution from the<br />
knowledge of their moments (Stuart and Ord 1994).<br />
14
1. The Par<strong>et</strong>o distribution:<br />
79<br />
Fu(x) = 1 − (u/x) b , (22)<br />
which corresponds to the s<strong>et</strong> of param<strong>et</strong>ers (b > 0,c = 0) with A(b,c,d,u) = b·u b . Several works have<br />
<strong>at</strong>tempted to <strong>de</strong>rive or justified the existence of a power tail of the distribution of r<strong>et</strong>urns from agentbased<br />
mo<strong>de</strong>ls (Chall<strong>et</strong> and Marsili 2002), from optimal trading of large funds with sizes distributed<br />
according to the Zipf law (Gabaix <strong>et</strong> al. 2002) or from stochastic processes (Biham <strong>et</strong> al 1998, 2002).<br />
2. The Weibull distribution:<br />
<br />
Fu(x) = 1 − exp −<br />
<br />
x<br />
c +<br />
d<br />
<br />
u<br />
c , (23)<br />
d<br />
with param<strong>et</strong>er s<strong>et</strong> (b = −c,c > 0,d > 0) and normaliz<strong>at</strong>ion constant A(b,c,d,u) = c<br />
dc exp <br />
u c<br />
d .<br />
This distribution is said to be a “Str<strong>et</strong>ched-Exponential” distribution when the exponent c is smaller<br />
than 1, namely when the distribution <strong>de</strong>cays more slowly than an exponential distribution.<br />
3. The exponential distribution:<br />
<br />
Fu(x) = 1 − exp − x<br />
d<br />
u<br />
<br />
+ , (24)<br />
d<br />
with param<strong>et</strong>er s<strong>et</strong> (b = −1, c = 1, d > 0) and normaliz<strong>at</strong>ion constant A(b,c,d,u) = 1<br />
d exp− u<br />
<br />
d .<br />
The exponential family can for instance <strong>de</strong>rive from a simple mo<strong>de</strong>l where stock price dynamics is<br />
governed by a geom<strong>et</strong>rical (multiplic<strong>at</strong>ive) Brownian motion with stochastic variance. Dragulescu<br />
and Yakovenko (2002) have found an excellent fit of this mo<strong>de</strong>l with the Dow-Jones in<strong>de</strong>x for time<br />
lags from 1 to 250 trading days, within an asymptotic exponential tail of the distribution of log-r<strong>et</strong>urns<br />
with a time-<strong>de</strong>pen<strong>de</strong>nt exponent.<br />
4. The incompl<strong>et</strong>e Gamma distribution:<br />
Fu(x) = 1 − Γ(−b,x/d)<br />
Γ(−b,u/d)<br />
with param<strong>et</strong>er s<strong>et</strong> (b, c = 1, d > 0) and normaliz<strong>at</strong>ion A(b,c,d,u) =<br />
d b<br />
Γ(−b,u/d) .<br />
Thus, the Par<strong>et</strong>o distribution (PD) and exponential distribution (ED) are one-param<strong>et</strong>er families, whereas<br />
the str<strong>et</strong>ched exponential (SE) and the incompl<strong>et</strong>e Gamma distribution (IG) are two-param<strong>et</strong>er families. The<br />
comprehensive distribution (CD) given by equ<strong>at</strong>ion (20) contains three unknown param<strong>et</strong>ers.<br />
Interesting links b<strong>et</strong>ween these different mo<strong>de</strong>ls reveal themselves un<strong>de</strong>r specific asymptotic conditions.<br />
For instance, in the limit b → +∞, the Par<strong>et</strong>o mo<strong>de</strong>l becomes the Exponential mo<strong>de</strong>l (Bouchaud and Potters<br />
2000). In<strong>de</strong>ed, provi<strong>de</strong>d th<strong>at</strong> the scale param<strong>et</strong>er u of the power law is simultaneously scaled as u b = (b/α) b ,<br />
we can write the tail of the cumul<strong>at</strong>ive distribution function of the PD as u b /(u + x) b which is in<strong>de</strong>ed of the<br />
form u b /x b for large x. Then, u b /(u + x) b = (1 + αx/b) −b → exp(−αx) for b → +∞. This shows th<strong>at</strong> the<br />
Exponential mo<strong>de</strong>l can be approxim<strong>at</strong>ed with any <strong>de</strong>sired accuracy on an arbitrary interval (u > 0,U) by<br />
the (PD) mo<strong>de</strong>l with param<strong>et</strong>ers (β,u) s<strong>at</strong>isfying u b = (b/α) b . Although the value b → +∞ does not give<br />
strickly speaking a Exponential distribution, the limit b → +∞ provi<strong>de</strong>s any <strong>de</strong>sired approxim<strong>at</strong>ion to the<br />
Exponential distribution, uniformly on any finite interval (u,U).<br />
More interesting for our present study is the behavior of the (SE) mo<strong>de</strong>l when c → 0. In this limit, and<br />
provi<strong>de</strong>d th<strong>at</strong><br />
<br />
u<br />
c c · → β,<br />
d<br />
as c → 0 . (26)<br />
15<br />
(25)
80 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
the (SE) mo<strong>de</strong>l goes to the Par<strong>et</strong>o mo<strong>de</strong>l. In<strong>de</strong>ed, we can write<br />
c<br />
dc · xc−1 <br />
· exp − xc − uc dc <br />
<br />
u<br />
c = c ·<br />
d<br />
xc−1<br />
<br />
u<br />
c <br />
x<br />
c <br />
exp − · − 1 ,<br />
uc d u<br />
β · x −1 <br />
u<br />
c exp −c · ln<br />
d<br />
x<br />
<br />
, as c → 0<br />
u<br />
β · x −1 <br />
exp −β · ln x<br />
<br />
,<br />
u<br />
β uβ<br />
, (27)<br />
xβ+1 which is the pdf of the (PD) mo<strong>de</strong>l with tail in<strong>de</strong>x β. The condition (26) comes n<strong>at</strong>urally from the properties<br />
of the maximum-likelihood estim<strong>at</strong>or of the scale param<strong>et</strong>er d given by equ<strong>at</strong>ion (47) and un<strong>de</strong>rlined by<br />
equ<strong>at</strong>ion (90) in the appendices <strong>at</strong> the end of the paper. It implies th<strong>at</strong>, as c → 0, the characteristic scale d<br />
of the (SE) mo<strong>de</strong>l must also go to zero with c to ensure the convergence of the (SE) mo<strong>de</strong>l towards the (PD)<br />
mo<strong>de</strong>l.<br />
This shows th<strong>at</strong> the Par<strong>et</strong>o mo<strong>de</strong>l can be approxim<strong>at</strong>ed with any <strong>de</strong>sired accuracy on an arbitrary interval<br />
(u > 0,U) by the (SE) mo<strong>de</strong>l with paramters (c,d) s<strong>at</strong>isfying equ<strong>at</strong>ion (26) where the arrow is replaced by an<br />
equality. Although the value c = 0 does not give strickly speaking a Str<strong>et</strong>ched-Exponential distribution, the<br />
limit c → 0 provi<strong>de</strong>s any <strong>de</strong>sired approxim<strong>at</strong>ion to the Par<strong>et</strong>o distribution, uniformly on any finite interval<br />
(u,U). This <strong>de</strong>ep rel<strong>at</strong>ionship b<strong>et</strong>ween the SE and PD mo<strong>de</strong>ls allows us to un<strong>de</strong>rstand why it can be very<br />
difficult to <strong>de</strong>ci<strong>de</strong>, on a st<strong>at</strong>istical basis, which of these mo<strong>de</strong>ls fits the d<strong>at</strong>a best.<br />
4.2 M<strong>et</strong>hodology<br />
We start with fitting our two d<strong>at</strong>a s<strong>et</strong>s (DJ and ND) by the five distributions enumer<strong>at</strong>ed above (20) and (22-<br />
25). Our first goal is to show th<strong>at</strong> no single param<strong>et</strong>ric represent<strong>at</strong>ion among any of the cited pdf’s fits the<br />
whole range of the d<strong>at</strong>a s<strong>et</strong>s. Recall th<strong>at</strong> we analyze separ<strong>at</strong>ely positive and neg<strong>at</strong>ive r<strong>et</strong>urns (the l<strong>at</strong>er being<br />
converted to the positive semi-axis). We shall use in our analysis a movable lower threshold u, restricting by<br />
this threshold our sample to observ<strong>at</strong>ions s<strong>at</strong>isfying to x > u.<br />
In addition to estim<strong>at</strong>ing the param<strong>et</strong>ers involved in each represent<strong>at</strong>ion (20,22-25) by maximum likelihood<br />
for each particular threshold u 7 , we need a characteriz<strong>at</strong>ion of the goodness-of-fit. For this, we propose<br />
to use a distance measure b<strong>et</strong>ween the estim<strong>at</strong>ed distribution and the sample distribution. Many distances<br />
can be used: mean-squared error, Kullback-Liebler distance 8 , Kolmogorov distance, Spearman distance (as<br />
in Longin (1996)) or An<strong>de</strong>rson-Darling distance, to cite a few. We can also use one of these distances<br />
to d<strong>et</strong>ermine the param<strong>et</strong>ers of each pdf according to the criterion of minimizing the distance b<strong>et</strong>ween the<br />
estim<strong>at</strong>ed distribution and the sample distribution. The chosen distance is thus useful both for characterizing<br />
and for estim<strong>at</strong>ing the param<strong>et</strong>ric pdf. In the l<strong>at</strong>er case, once an estim<strong>at</strong>ion of the param<strong>et</strong>ers of particular<br />
distribution family has been obtained according to the selected distance, we need to quantify the st<strong>at</strong>istical<br />
significance of the fit. This requires to <strong>de</strong>rive the st<strong>at</strong>istics associ<strong>at</strong>ed with the chosen distance. These<br />
st<strong>at</strong>istics are known for most of the distances cited above, in the limit of large sample.<br />
We have chosen the An<strong>de</strong>rson-Darling distance to <strong>de</strong>rive our estim<strong>at</strong>ed param<strong>et</strong>ers and perform our tests<br />
of goodness of fit. The An<strong>de</strong>rson-Darling distance b<strong>et</strong>ween a theor<strong>et</strong>ical distribution function F(x) and its<br />
7 The estim<strong>at</strong>ors and their asymptotic properties are <strong>de</strong>rived in appendix A.<br />
8 This distance (or divergence, strictly speaking) is the n<strong>at</strong>ural distance associ<strong>at</strong>ed with maximum-likelihood estim<strong>at</strong>ion since<br />
it is for these values of the estim<strong>at</strong>ed param<strong>et</strong>ers th<strong>at</strong> the distance b<strong>et</strong>ween the true mo<strong>de</strong>l and the assumed mo<strong>de</strong>l reaches its<br />
minimum.<br />
16
empirical analog FN(x), estim<strong>at</strong>ed from<br />
· a sample of N realiz<strong>at</strong>ions, is evalu<strong>at</strong>ed as follows:<br />
[FN(x) − F(x)]<br />
ADS = N 2<br />
dF(x) (28)<br />
F(x)(1 − F(x))<br />
= −N − 2<br />
N<br />
∑<br />
1<br />
{wk log(F(yk)) + (1 − wk)log(1 − F(yk))}, (29)<br />
where wk = 2k/(2N + 1), k = 1...N and y1 ... yN is its or<strong>de</strong>red sample. If the sample is drawn<br />
from a popul<strong>at</strong>ion with distribution function F(x), the An<strong>de</strong>rson-Darling st<strong>at</strong>istics (ADS) has a standard<br />
AD-distribution free of the theor<strong>et</strong>ical df F(x) (An<strong>de</strong>rson and Darling 1952), similarly to the χ 2 for the<br />
χ 2 -st<strong>at</strong>istic, or the Kolmogorov distribution for the Kolmogorov st<strong>at</strong>istic. It should be noted th<strong>at</strong> the ADS<br />
weights the squared difference in eq.(28) by 1/F(x)(1 − F(x)) which is nothing but the inverse of the variance<br />
of the difference in square brack<strong>et</strong>s. The AD distance thus emphasizes more the tails of the distribution<br />
than, say, the Kolmogorov distance which is d<strong>et</strong>ermined by the maximum absolute <strong>de</strong>vi<strong>at</strong>ion of Fn(x) from<br />
F(x) or the mean-squared error, which is mostly controlled by the middle of range of the distribution. Since<br />
we have to insert the estim<strong>at</strong>ed param<strong>et</strong>ers into the ADS, this st<strong>at</strong>istic does not obey any more the standard<br />
AD-distribution: the ADS <strong>de</strong>creases because the use of the fitting param<strong>et</strong>ers ensures a b<strong>et</strong>ter fit to the sample<br />
distribution. However, we can still use the standard quantiles of the AD-distribution as upper boundaries<br />
of the ADS. If the observed ADS is larger than the standard quantile with a high significance level (1 − ε),<br />
we can then conclu<strong>de</strong> th<strong>at</strong> the null hypothesis F(x) is rejected with significance level larger than (1 − ε). If<br />
we wish to estim<strong>at</strong>e the real significance level of the ADS in the case where it does not exceed the standard<br />
quantile of a high significance level, we are forced to use some other m<strong>et</strong>hod of estim<strong>at</strong>ion of the significance<br />
level of the ADS, such as the bootstrap m<strong>et</strong>hod.<br />
In the following, the estim<strong>at</strong>es minimizing the An<strong>de</strong>rson-Darling distance will be refered to as AD-estim<strong>at</strong>es.<br />
The maximum likelihood estim<strong>at</strong>es (ML-estim<strong>at</strong>es) are asymptotically more efficient than AD-estim<strong>at</strong>es for<br />
in<strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a and un<strong>de</strong>r the condition th<strong>at</strong> the null hypothesis (given by one of the four distributions (22-<br />
25), for instance) corresponds to the true d<strong>at</strong>a gener<strong>at</strong>ing mo<strong>de</strong>l. When this is not the case, the AD-estim<strong>at</strong>es<br />
provi<strong>de</strong> a b<strong>et</strong>ter practical tool for approxim<strong>at</strong>ing sample distributions compared with the ML-estim<strong>at</strong>es.<br />
We have d<strong>et</strong>ermined the AD-estim<strong>at</strong>es for 18 standard significance levels q1 ...q18 given in table 6. The<br />
corresponding sample quantiles corresponding to these significance levels or thresholds u1 ...u18 for our<br />
samples are also shown in table 6. Despite the fact th<strong>at</strong> thresholds uk vary from sample to sample, they<br />
always correspon<strong>de</strong>d to the same fixed s<strong>et</strong> of significance levels qk throughout the paper and allows us to<br />
compare the goodness-of-fit for samples of different sizes.<br />
4.3 Empirical results<br />
The An<strong>de</strong>rson-Darling st<strong>at</strong>istics (ADS) for five param<strong>et</strong>ric distributions (Weibull or Str<strong>et</strong>ched-Exponential,<br />
Generalized Par<strong>et</strong>o, Gamma, exponential and Par<strong>et</strong>o) are shown in table 7 for two quantile ranges, the first<br />
top half of the table corresponding to the 90% lowest thresholds while the second bottom half corresponds<br />
to the 10% highest ones. For the lowest thresholds, the ADS rejects all distributions, except the Str<strong>et</strong>ched-<br />
Exponential for the Nasdaq. Thus, none of the consi<strong>de</strong>red distributions is really a<strong>de</strong>qu<strong>at</strong>e to mo<strong>de</strong>l the d<strong>at</strong>a<br />
over such large ranges. For the 10% highest quantiles, only the exponential mo<strong>de</strong>l is rejected <strong>at</strong> the 95%<br />
confi<strong>de</strong>nce level. The Str<strong>et</strong>ched-Exponential distribution is the best, just before the Par<strong>et</strong>o distribution and<br />
the Incompl<strong>et</strong>e Gamma th<strong>at</strong> cannot be rejected. We now present an analysis of each case in more d<strong>et</strong>ails.<br />
17<br />
81
82 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
4.3.1 Par<strong>et</strong>o distribution<br />
Figure 5a shows the cumul<strong>at</strong>ive sample distribution function 1 − F(x) for the Dow Jones Industrial Average<br />
in<strong>de</strong>x, and in figure 5b the cumul<strong>at</strong>ive sample distribution function for the Nasdaq Composite in<strong>de</strong>x. The<br />
mism<strong>at</strong>ch b<strong>et</strong>ween the Par<strong>et</strong>o distribution and the d<strong>at</strong>a can be seen with the naked eye: if samples were<br />
taken from a Par<strong>et</strong>o popul<strong>at</strong>ion, the graph in double log-scale should be a straight line. Even in the tails,<br />
this is doubtful. To formalize this impression, we calcul<strong>at</strong>e the Hill and AD estim<strong>at</strong>ors for each threshold u.<br />
Denoting y1 ... ynu the or<strong>de</strong>red sub-sample of values exceeding u where Nu is size of this sub-sample,<br />
the Hill maximum likelihood estim<strong>at</strong>e of param<strong>et</strong>er b is (Hill 1975)<br />
<br />
1<br />
ˆbu =<br />
Nu<br />
The standard <strong>de</strong>vi<strong>at</strong>ions of ˆbu can be estim<strong>at</strong>ed as<br />
Nu<br />
∑ 1<br />
log(yk/u)<br />
−1<br />
. (30)<br />
Std(ˆbu) = ˆbu/ √ Nu, (31)<br />
un<strong>de</strong>r the assumption of iid d<strong>at</strong>a, but very severely un<strong>de</strong>restim<strong>at</strong>e the true standard <strong>de</strong>vi<strong>at</strong>ion when samples<br />
exhibit <strong>de</strong>pen<strong>de</strong>nce, as reported by Kearns and Pagan (1997).<br />
Figure 6a and 6b shows the Hill estim<strong>at</strong>es ˆbu as a function of u for the Dow Jones and for the Nasdaq.<br />
Instead of an approxim<strong>at</strong>ely constant exponent (as would be the case for true Par<strong>et</strong>o samples), the tail in<strong>de</strong>x<br />
estim<strong>at</strong>or increases until u ∼ = 0.04, beyond which it seems to slow its growth and oscill<strong>at</strong>es around a value<br />
≈ 3 − 4 up to the threshold u ∼ = .08. It should be noted th<strong>at</strong> interval [0,0.04] contains 99.12% of the sample<br />
whereas interval [0.04,0.08] contains only 0.64% of the sample. The behavior of ˆbu for the ND shown<br />
in figure 6b is similar: Hill’s estim<strong>at</strong>es ˆbu seem to slow its growth already <strong>at</strong> u ∼ = 0.0013 corresponding<br />
to the 95% quantile. Are these slowdowns of the growth of ˆbu genuine sign<strong>at</strong>ures of a possible constant<br />
well-<strong>de</strong>fined asymptotic value th<strong>at</strong> would qualify a regularly varying function?<br />
As a first answer to this question, table 8 compares the AD-estim<strong>at</strong>es of the tail exponent b with the corresponding<br />
maximum likelihood estim<strong>at</strong>es for the 18 intervals u1 ...u18. Both maximum likelihood and<br />
An<strong>de</strong>rson-Darling estim<strong>at</strong>es of b steadily increase with the threshold u (excepted for the highest quantiles<br />
of the positive tail of the Nasdaq). The corresponding figures for positive and neg<strong>at</strong>ive r<strong>et</strong>urns are very<br />
close each to other and almost never significantly different <strong>at</strong> the usual 95% confi<strong>de</strong>nce level. Some slight<br />
non-monotonicity of the increase for the highest thresholds can be explained by small sample sizes. One can<br />
observe th<strong>at</strong> both MLE and ADS estim<strong>at</strong>es continue increasing as the interval of estim<strong>at</strong>ion is contracting to<br />
the extreme values. It seems th<strong>at</strong> their growth potential has not been exhausted even for the largest quantile<br />
u18, except for the positive tail of the Nasdaq sample. This st<strong>at</strong>ement might be however not very strong<br />
as the standard <strong>de</strong>vi<strong>at</strong>ions of the tail in<strong>de</strong>x estim<strong>at</strong>or also grow when exploring the largest quantiles.<br />
However, the non-exhausted growth is observed for three samples out of four. Moreover, this effect is<br />
seen for several threshold values and we can add th<strong>at</strong> random fluctu<strong>at</strong>ions would distort the b-curve<br />
in a random manner, i.e, now up now down, whearas we note in three cases an increasing curve.<br />
Assuming th<strong>at</strong> the observ<strong>at</strong>ion, th<strong>at</strong> the sample distribution can be approxim<strong>at</strong>ed by a Par<strong>et</strong>o distribution<br />
with a growing in<strong>de</strong>x b, is correct, an important question arises: how far beyond the sample this growth will<br />
continue? Judging from table 8, we can think this growth is still not exhausted. Figure 7 suggests a specific<br />
form of this growth, by plotting the hill estim<strong>at</strong>or ˆbu for all four d<strong>at</strong>a s<strong>et</strong>s (positive and neg<strong>at</strong>ive branches<br />
of the distribution of r<strong>et</strong>urns for the DJ and for the ND) as a function of the in<strong>de</strong>x n = 1,...,18 of the 18<br />
quantiles or standard significance levels q1 ...q18 given in table 6. Similar results are obtained with the AD<br />
estim<strong>at</strong>es. Apart from the positive branch of the ND d<strong>at</strong>a s<strong>et</strong>, all other three branches suggest a continuous<br />
growth of the Hill estim<strong>at</strong>or ˆbu as a function of n = 1,...,18. Since the quantiles q1 ...q18 given in table 6<br />
18
have been chosen to converge to 1 approxim<strong>at</strong>ely exponentially as<br />
1 − qn = 3.08 e −0.342n , (32)<br />
the linear fit of ˆbu as a function of n shown as the dashed line in figure 7 corresponds to<br />
83<br />
ˆbu(qn) = 0.08 + 0.626ln 3.08<br />
. (33)<br />
1 − qn<br />
Expression (33) suggests an unbound logarithmic growth of ˆbu as the quantile approaches 1. For instance,<br />
for a quantile 1 −q = 0.1%, expression (33) predicts ˆbu(1 −q = 10 −3 ) = 5.1. For a quantile 1 −q = 0.01%,<br />
expression (33) predicts ˆbu(1 − q = 10 −4 ) = 6.5, and so on. Each time the quantile 1 − q is divi<strong>de</strong>d by<br />
a factor 10, the apparent exponent ˆbu(q) is increased by the additive constant ∼ = 1.45: ˆbu((1 − q)/10) =<br />
ˆbu(1 − q) + 1.45. This very slow growth uncovered here may be an explan<strong>at</strong>ion for the belief and possibly<br />
mistaken conclusion th<strong>at</strong> the Hill and other estim<strong>at</strong>ors of the tail in<strong>de</strong>x tends to a constant for high quantiles.<br />
In<strong>de</strong>ed, it is now clear th<strong>at</strong> the slowdowns of the growth of ˆbu seen in figures 6 <strong>de</strong>cor<strong>at</strong>ed by large fluctu<strong>at</strong>ions<br />
due to small size effects is mostly the result of a dil<strong>at</strong><strong>at</strong>ion of the d<strong>at</strong>a expressed in terms of threshold u.<br />
When recast in the more n<strong>at</strong>ural logarithm scale of the quantiles q1 ...q18, this slowdown disappears. Of<br />
course, it is impossible to know how long this growth given by (33) may go on as the quantile q tends to 1.<br />
In other words, how can we escape from the sample range when estim<strong>at</strong>ing quantiles? How can we estim<strong>at</strong>e<br />
the so-called “high quantiles” <strong>at</strong> the level q > 1 − 1/T where T is the total number of sampled points.<br />
Embrechts <strong>et</strong> al. (1997) have summarized the situ<strong>at</strong>ion in this way: “there is no free lunch when it comes to<br />
high quantiles estim<strong>at</strong>ion!” It is possible th<strong>at</strong> ˆbu(q) will grow without limit as would be the case if the true<br />
un<strong>de</strong>rlying distribution was rapidly varying. Altern<strong>at</strong>ively, ˆbu(q) may s<strong>at</strong>ur<strong>at</strong>e to a large value, as predicted<br />
for instance by the traditional GARCH mo<strong>de</strong>l which yields tails indices which can reach 10 − 20 (Engle<br />
and P<strong>at</strong>ton 2001, Starica and Pict<strong>et</strong> 1999) or by the recent multifractal random walk (MRW) mo<strong>de</strong>l which<br />
gives an asymptotic tail exponent in the range 20 − 50 (Muzy <strong>et</strong> al. 2000, Muzy <strong>et</strong> al. 2001). According to<br />
(33), a value ˆbu ≈ 20 (respectively 50) would be <strong>at</strong>tained for 1 − q ≈ 10 −13 (respectively 1 − q ≈ 10 −34 )! If<br />
one believes in the prediction of the MRW mo<strong>de</strong>l, the tail of the distribution of r<strong>et</strong>urns is regularly varying<br />
but this insight is compl<strong>et</strong>ely useless for all practical purposes due to the astronomically high st<strong>at</strong>istics th<strong>at</strong><br />
would be nee<strong>de</strong>d to sample this regime. In this context, we cannot hope to g<strong>et</strong> access to the true n<strong>at</strong>ure of the<br />
pdf of r<strong>et</strong>urns but only strive to <strong>de</strong>fine the best effective or apparent most parsimonious and robust mo<strong>de</strong>l.<br />
We do not discuss here the new class of estim<strong>at</strong>ion issues raised by the MRW mo<strong>de</strong>l, which is interesting in<br />
itself but requires a specific analysis of its own left for another work.<br />
The question of the exhaustion of the growth of the tail in<strong>de</strong>x is really crucial. In<strong>de</strong>ed, if it is unboun<strong>de</strong>dly<br />
increasing, it is the sign<strong>at</strong>ure th<strong>at</strong> the tails of the distributions of r<strong>et</strong>urns <strong>de</strong>cay faster than any power-law, and<br />
thus cannot be regularly varying. We revisit this question of the growth of the apparent exponent b, using<br />
the notion of local exponent, in Appendix C as an <strong>at</strong>tempt to b<strong>et</strong>ter constraint this growth. The analysis<br />
<strong>de</strong>veloped in this Appendix C basically confirms the first indic<strong>at</strong>ion shown in figure 7 .<br />
4.3.2 Weibull distributions<br />
L<strong>et</strong> us now fit our d<strong>at</strong>a with the Weibull (SE) distribution (23). The An<strong>de</strong>rson-Darling st<strong>at</strong>istics (ADS) for<br />
this case are shown in table 7. The ML-estim<strong>at</strong>es and AD-estim<strong>at</strong>es of the form param<strong>et</strong>er c are represented<br />
in table 9. Table 7 shows th<strong>at</strong>, for the higest quantiles, the ADS for the Str<strong>et</strong>ched-Exponential is the smallest<br />
of all ADS, suggesting th<strong>at</strong> the SE is the best mo<strong>de</strong>l of all. Moreover, for the lowest quantiles, it is the sole<br />
mo<strong>de</strong>l not system<strong>at</strong>ically rejected <strong>at</strong> the 95% level.<br />
The c-estim<strong>at</strong>es are found to <strong>de</strong>crease when increasing the or<strong>de</strong>r q of the threshold uq beyond which the<br />
estim<strong>at</strong>ions are performed. In addition, the c-estim<strong>at</strong>e is i<strong>de</strong>ntically zero for u18. However, this does not<br />
19
84 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
autom<strong>at</strong>ically imply th<strong>at</strong> the SE mo<strong>de</strong>l is not the correct mo<strong>de</strong>l for the d<strong>at</strong>a even for these highest quantiles.<br />
In<strong>de</strong>ed, numerical simul<strong>at</strong>ions show th<strong>at</strong>, even for synth<strong>et</strong>ic samples drawn from genuine Str<strong>et</strong>ched-<br />
Exponential distributions with exponent c smaller than 0.5 and whose size is comparable with th<strong>at</strong> of our<br />
d<strong>at</strong>a, in about one case out of three (<strong>de</strong>pending on the exact value of c) the estim<strong>at</strong>ed value of c is zero. This<br />
a priori surprising result comes from condition (51) in appendix A which is not fulfilled with certainty even<br />
for samples drawn for SE distributions.<br />
Notwithstanding this cautionary remark, note th<strong>at</strong> the c-estim<strong>at</strong>e of the positive tail of the Nasdaq d<strong>at</strong>a equal<br />
zero for all quantiles higher than q14 = 0.97%. In fact, in every cases, the estim<strong>at</strong>ed c is not significantly<br />
different from zero - <strong>at</strong> the 95% significance level - for quantiles higher than q12-q14. In addition, table 10<br />
gives the values of the estim<strong>at</strong>ed scale parem<strong>et</strong>er d, which are found very small - particularly for the Nasdaq<br />
- beyond q12 = 95%. In constrast, the Dow Jones keeps significant scale factors until q16 − q17.<br />
These evi<strong>de</strong>nces taken all tog<strong>et</strong>her provi<strong>de</strong> a clear indic<strong>at</strong>ion on the existence of a change of behavior of<br />
the true pdf of these four distributions: while the bulks of the distributions seem r<strong>at</strong>her well approxim<strong>at</strong>ed<br />
by a SE mo<strong>de</strong>l, a f<strong>at</strong>ter tailed distribution than th<strong>at</strong> of the (SE) mo<strong>de</strong>l is required for the highest quantiles.<br />
Actually, the fact th<strong>at</strong> both c and d are extremely small may be interpr<strong>et</strong>ed according to the asymptotic<br />
correspon<strong>de</strong>nce given by (26) and (27) as the existence of a possible power law tail.<br />
4.3.3 Exponential and incompl<strong>et</strong> Gamma distribution<br />
L<strong>et</strong> us now fit our d<strong>at</strong>a with the exponential distribution (24). The average ADS for this case are shown in<br />
table 7. The maximum likelihood- and An<strong>de</strong>rson-Darling estim<strong>at</strong>es of the scale param<strong>et</strong>er d are given in<br />
table 11. Note th<strong>at</strong> they always <strong>de</strong>crease as the threshold uq increases. Comparing the mean ADS-values<br />
of table 7 with the standard AD quantiles, we can conclu<strong>de</strong> th<strong>at</strong>, on the whole, the exponential distribution<br />
(even with moving scale param<strong>et</strong>er d) does not fit our d<strong>at</strong>a: this mo<strong>de</strong>l is system<strong>at</strong>ically rejected <strong>at</strong> the 95%<br />
confi<strong>de</strong>nce level for the lowest and highest quantiles - excepted for the neg<strong>at</strong>ive tail of the Nasdaq.<br />
Finally, we fit our d<strong>at</strong>a by the IG-distribution (25). The mean ADS for this class of functions are shown in<br />
table 7. The Maximum likelihood and An<strong>de</strong>rson Darling estim<strong>at</strong>es of the power in<strong>de</strong>x b are represented in<br />
table 12. Comparing the mean ADS-values of table 7 with the standard AD quantiles, we can again conclu<strong>de</strong><br />
th<strong>at</strong>, on the whole, the IG-distribution does not fit our d<strong>at</strong>a. The mo<strong>de</strong>l is rejected <strong>at</strong> the 95% confi<strong>de</strong>nce<br />
level excepted for the neg<strong>at</strong>ive tail of the Nasdaq for which it is not rejected marginally (significance level:<br />
94.13%). However, for the largest quantiles, this mo<strong>de</strong>l becomes again revelant since it cannot be rejected<br />
<strong>at</strong> the 95% level.<br />
4.4 Summary<br />
At this stage, two conclusions can be drawn. First, it appears th<strong>at</strong> none of the consi<strong>de</strong>red distributions fit the<br />
d<strong>at</strong>a over the entire range, which is not a surprise. Second, for the highest quantiles, three mo<strong>de</strong>ls seem to<br />
be able to represent to d<strong>at</strong>a, the Gamma mo<strong>de</strong>l, the Par<strong>et</strong>o mo<strong>de</strong>l and the Str<strong>et</strong>ched-Exponential mo<strong>de</strong>l. This<br />
last one has the lowest An<strong>de</strong>rson-Darling st<strong>at</strong>istic and thus seems to be the most reasonable mo<strong>de</strong>l among<br />
the three mo<strong>de</strong>ls comp<strong>at</strong>ible with the d<strong>at</strong>a.<br />
20
5 Comparison of the <strong>de</strong>scriptive power of the different families<br />
As we have seen by comparing the An<strong>de</strong>rson-Darling st<strong>at</strong>istics corresponding to the four param<strong>et</strong>ric families<br />
(22-25), the best mo<strong>de</strong>l in the sense of minimizing the An<strong>de</strong>rson-Darling distance is the Str<strong>et</strong>ched-<br />
Exponential distribution.<br />
We now compare these four distributions with the comprehensive distribution (20) using Wilks’ theorem<br />
(Wilks 1938) of nested hypotheses to check wh<strong>et</strong>her or not some of the four distributions are sufficient<br />
compared with the comprehensive distribution to <strong>de</strong>scribe the d<strong>at</strong>a. We then turn to the Wald encompassing<br />
test for non-nested hypotheses which provi<strong>de</strong>s a pairwise comparison of the different mo<strong>de</strong>ls.<br />
5.1 Comparison b<strong>et</strong>ween the four param<strong>et</strong>ric families and the comprehensive distribution<br />
According to Wilk’s theorem, the doubled generalized log-likelihood r<strong>at</strong>io Λ:<br />
85<br />
Λ = 2 log maxL(CD,X,Θ)<br />
, (34)<br />
maxL(z,X,θ)<br />
has asymptotically (as the size N of the sample X tends to infinity) the χ 2 -distribution. Here L <strong>de</strong>notes the<br />
likelihood function, θ and Θ are param<strong>et</strong>ric spaces corresponding to hypotheses z and CD correspondingly<br />
(hypothesis z is one of the four hypotheses (22-25) th<strong>at</strong> are particular cases of the CD un<strong>de</strong>r some param<strong>et</strong>er<br />
rel<strong>at</strong>ions). The st<strong>at</strong>ement of the theorem is valid un<strong>de</strong>r the condition th<strong>at</strong> the sample X obeys hypothesis z<br />
for some particular value of its param<strong>et</strong>er belonging to the space θ. The number of <strong>de</strong>grees of freedom of<br />
the χ 2 -distribution equals to the difference of the dimensions of the two spaces Θ and θ. Since dim(Θ) = 3<br />
and dim(θ) = 2 for the Str<strong>et</strong>ched-Exponential and Incompl<strong>et</strong> Gamma distributions; dim(θ) = 1 for the<br />
Par<strong>et</strong>o and the Exponential distributions, we have one <strong>de</strong>gree of freedom for the formers and two <strong>de</strong>grees of<br />
freedom for the l<strong>at</strong>ers. The maximum of the likelihood in the numer<strong>at</strong>or of (34) is taken over the space Θ,<br />
whereas the maximum of the likelihood in the <strong>de</strong>nomin<strong>at</strong>or of (34) is taken over the space θ. Since we have<br />
always θ ⊂ Θ, the likelihood r<strong>at</strong>io is always larger than 1, and the log-likelihood r<strong>at</strong>io is non-neg<strong>at</strong>ive. If the<br />
observed value of Λ does not exceed some high-confi<strong>de</strong>nce level (say, 99% confi<strong>de</strong>nce level) of the χ 2 , we<br />
then reject the hypothesis CD in favour of the hypothesis z, consi<strong>de</strong>ring the space Θ redundant. Otherwise,<br />
we accept the hypothesis CD, consi<strong>de</strong>ring the space θ insufficient.<br />
The doubled log-likelihood r<strong>at</strong>ios (34) are shown in figures 8 for the positive and neg<strong>at</strong>ive branches of the<br />
distribution of r<strong>et</strong>urns of the Nasdaq and in figures 9 for the Dow Jones. The 95% χ 2 confi<strong>de</strong>nce levels for<br />
1 and 2 <strong>de</strong>grees of freedom are given by the horizontal lines.<br />
For the Nasdaq d<strong>at</strong>a, figure 8 clearly shows th<strong>at</strong> Exponential distribution is compl<strong>et</strong>ely insufficient: for all<br />
lower thresholds, the Wilks log-likelihood r<strong>at</strong>io exceeds the 95% χ2 1 level 3.84. The Par<strong>et</strong>o distribution is<br />
insufficient for thresholds u1 − u11 (92.5% of the or<strong>de</strong>red sample) and becomes comparable with the Comprehensive<br />
distribution in the tail u12 − u18 (7.5% of the tail probability). It is n<strong>at</strong>ural th<strong>at</strong> two-param<strong>et</strong>ric<br />
families Incompl<strong>et</strong>e Gamma and Str<strong>et</strong>ched-Exponential have higher goodness-of-fit than the one-param<strong>et</strong>ric<br />
Exponential and Par<strong>et</strong>o distributions. The Incompl<strong>et</strong>e Gamma distribution is comparable with the Comprehensive<br />
distribution starting with u10 (90%), whereas the Str<strong>et</strong>ched-Exponential is somewh<strong>at</strong> b<strong>et</strong>ter (u9 or<br />
u8 , i.e., 70%). For the tails representing 7.5% of the d<strong>at</strong>a, all param<strong>et</strong>ric families except for the Exponential<br />
distribution fit the sample distribution with almost the same efficiency. The results obtained for the Dow<br />
Jones d<strong>at</strong>a shown in figure 9 are similar. The Str<strong>et</strong>ched-Exponential is comparable with the Comprehensive<br />
distribution starting with u8 (70%). On the whole, one can say th<strong>at</strong> the Str<strong>et</strong>ched-Exponential distribution<br />
performs b<strong>et</strong>ter than the three other param<strong>et</strong>ric families.<br />
21
86 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
We should stress th<strong>at</strong> each log-likelihood r<strong>at</strong>io represented in figures 8 and 9, so-to say “acts on its own<br />
ground,” th<strong>at</strong> is, the corresponding χ 2 -distribution is valid un<strong>de</strong>r the assumption of the validity of each<br />
particular hypothesis whose likelihood stands in the numer<strong>at</strong>or of the double log-likelihood (34). It would<br />
be <strong>de</strong>sirable to compare all combin<strong>at</strong>ions of pairs of hypotheses directly, in addition to comparing each<br />
of them with the comprehensive distribution. Unfortun<strong>at</strong>ely, the Wilks theorem can not be used in the<br />
case of pair-wise comparison because the problem is not more th<strong>at</strong> of comparing nested hypothesis (th<strong>at</strong><br />
is, one hypothesis is a particular case of the comprehensive mo<strong>de</strong>l). As a consequence, our results on the<br />
comparison of the rel<strong>at</strong>ive merits of each of the four distributions using the generalized log-likelihood r<strong>at</strong>io<br />
should be interpr<strong>et</strong>ed with a care, in particular, in a case of contradictory conclusions. Fortun<strong>at</strong>ely, the main<br />
conclusion of the comparison (an advantage of the Str<strong>et</strong>ched-Exponential distribution over the three other<br />
distribution) does not contradict our earlier results discussed above.<br />
5.2 Pair-wise comparison of the Par<strong>et</strong>o mo<strong>de</strong>l with the Str<strong>et</strong>ched-Exponential mo<strong>de</strong>l<br />
In or<strong>de</strong>r to compare formally the <strong>de</strong>scriptive power of the Str<strong>et</strong>ched-Exponential distribution (the best twoparam<strong>et</strong>ers<br />
ditribution) with th<strong>at</strong> of the Par<strong>et</strong>o distribution (the best one-param<strong>et</strong>er distribution), we need to<br />
use the m<strong>et</strong>hods for testing non-nested hypotheses. There are in fact many ways to perform such a test (see<br />
Gouriéroux and Monfort (1994) for a review). Concerning the log-likelihood r<strong>at</strong>io test used in the previous<br />
section for nested-hypotheses testing, a direct generaliz<strong>at</strong>ion for non-nested hypotheses has been provi<strong>de</strong>d<br />
by Cox’s test (1961, 1962). However, such a test requires th<strong>at</strong> the true distribution of the sample be nested in<br />
one of the two consi<strong>de</strong>red mo<strong>de</strong>ls. Our previous investig<strong>at</strong>ions have shown th<strong>at</strong> it is not the case, so we need<br />
to use another testing procedure. In<strong>de</strong>ed, when comparing the Par<strong>et</strong>o mo<strong>de</strong>l with the Str<strong>et</strong>ched-Exponential<br />
mo<strong>de</strong>l, we have found th<strong>at</strong> using the m<strong>et</strong>hodology of non-nested hypothesis leads to inconsistencies such<br />
as neg<strong>at</strong>ive variances of estim<strong>at</strong>ors. This is another (indirect) confirm<strong>at</strong>ion th<strong>at</strong> neither the Par<strong>et</strong>o nor the<br />
Str<strong>et</strong>ched-Exponential distributions are the true distribution.<br />
In the case where none of the tested hypotheses contain the true distribution, it can be useful to consi<strong>de</strong>r the<br />
encompassing principle introduced by Mizon and Richard (1986). A mo<strong>de</strong>l, (SE) say, is said to encompass<br />
another mo<strong>de</strong>l, (PD) for instance, if the represent<strong>at</strong>ive of (PD), which is the closest to the best represent<strong>at</strong>ive<br />
of (SE), is also the best represent<strong>at</strong>ive of (PD) per se. Here, the best represent<strong>at</strong>ive of a mo<strong>de</strong>l is the<br />
distribution which is the nearest to the true distribution for the consi<strong>de</strong>red mo<strong>de</strong>l. The d<strong>et</strong>ailled testing<br />
procedure is based on the Wald and Score encompassing tests (Gouriéroux and Monfort 1994), which are<br />
d<strong>et</strong>ailed in appendix D.<br />
Table 13 presents the results of the tests for the null hypothesis “(SE) encompasses (PD)”. In every cases,<br />
the null hypothesis cannot be rejected <strong>at</strong> the 95% significance level for quantiles higher than q6 = 0.5 and <strong>at</strong><br />
the 99% significance level for quantiles higher than q10 = 0.9. The unfilled entries for the largest quantiles<br />
correspond to MLE giving c → 0. In this case, as shown in Appendix D.1.2, b † has a non-trivial and well<strong>de</strong>fined<br />
limit which is nothing but the true value ˆb. Thus, the Wald tests is autom<strong>at</strong>ically verified <strong>at</strong> any<br />
confi<strong>de</strong>nce levels. Thus, in the tail, the Str<strong>et</strong>ched-Eponential mo<strong>de</strong>l encompasses the Par<strong>et</strong>o mo<strong>de</strong>l. We can<br />
then conclu<strong>de</strong> th<strong>at</strong> it provi<strong>de</strong>s a <strong>de</strong>scription of the d<strong>at</strong>a which is <strong>at</strong> least as good as th<strong>at</strong> given by the l<strong>at</strong>er.<br />
In or<strong>de</strong>r to test wh<strong>et</strong>her the (SE) mo<strong>de</strong>l is superior to the Par<strong>et</strong>o mo<strong>de</strong>l, we should perform the reverse<br />
test, namely the encompassing of the former mo<strong>de</strong>l into the l<strong>at</strong>er. This task is difficult since the pseudotrue<br />
values of the param<strong>et</strong>ers (c,d) are not always well-<strong>de</strong>fined as exposed in appendix D. Thus, Wald<br />
encompassing test cannot be performed as in the previous case. As a remedy and altern<strong>at</strong>ive, we propose<br />
a test of the null hypothesis H0 th<strong>at</strong> the Par<strong>et</strong>o distribution is the true un<strong>de</strong>rlying distribution. This test is<br />
22
ased on the fact th<strong>at</strong> the quantity<br />
ˆηT = ĉ<br />
<br />
ĉ<br />
u<br />
+ 1<br />
dˆ<br />
goes to b un<strong>de</strong>r the null hypothesis, wh<strong>at</strong>ever c being positive or neg<strong>at</strong>ive. This can be seen from the<br />
asymptotic correspon<strong>de</strong>nce given by (26) and (27). Moreover, it is proved in appendix D th<strong>at</strong> the variable<br />
<br />
ˆηT<br />
ζT = T − 1 , (36)<br />
ˆb<br />
asymptoticaly follows a χ 2 -distribution with one <strong>de</strong>gree of freedom.<br />
The results of this test are given in table 14. They show th<strong>at</strong> H0 is more often rejected for the Dow Jones<br />
than for the Nasdaq. In<strong>de</strong>ed, beyond quantile q12 = 95%, H0 cannot be rejected <strong>at</strong> the 95% confi<strong>de</strong>nce level<br />
for the Nasdaq d<strong>at</strong>a. For the Dow Jones, we must consi<strong>de</strong>r quantiles higher than q18 = 99% in or<strong>de</strong>r not to<br />
reject H0 <strong>at</strong> the 95% significance level. These results are in agreement with the central limit theorem: the<br />
power-law regime (if it really exists) is pushed back to higher quantiles due to time agreg<strong>at</strong>ion (recall th<strong>at</strong><br />
our Dow Jones d<strong>at</strong>a is <strong>at</strong> the daily scale while our Nasdaq d<strong>at</strong>a is <strong>at</strong> the 5 minutes time scale).<br />
In summary, the (SE) mo<strong>de</strong>l encompasses the Par<strong>et</strong>o mo<strong>de</strong>l as soon as one consi<strong>de</strong>rs quantiles higher than<br />
q6 = 50%. On the other hand, the null hypothesis th<strong>at</strong> the true distribution is the Par<strong>et</strong>o distribution is<br />
strongly rejected until quantiles 90% − 95% or so. Thus, within this range, the (SE) mo<strong>de</strong>l seems the best.<br />
But, for the very highest quantiles (above 95% − 98%) we cannot any more reject the hypothesis th<strong>at</strong> the<br />
Par<strong>et</strong>o mo<strong>de</strong>l is the right one. Thus, for these extreme quantiles, this last mo<strong>de</strong>l seems to slightly outperform<br />
the (SE) mo<strong>de</strong>l.<br />
However, recall th<strong>at</strong> section 4.1 has shown th<strong>at</strong> the Par<strong>et</strong>o mo<strong>de</strong>l is r<strong>et</strong>rieved as a limiting case of the (SE)<br />
mo<strong>de</strong>l when the fractional exponent c and the scale factor d go to zero <strong>at</strong> an appropri<strong>at</strong>ed r<strong>at</strong>e. Thus, all the<br />
results above are comp<strong>at</strong>ible with the (SE) mo<strong>de</strong>l in a generalized version. In<strong>de</strong>ed, <strong>de</strong>fining a generalized<br />
Str<strong>et</strong>ched-Exponential mo<strong>de</strong>l as<br />
<br />
¯Fu(x) = exp − xc−uc dc <br />
, c > 0<br />
¯Fu(x) = <br />
u b<br />
x , c = 0 and b = limc→0 c <br />
u c (37)<br />
d = b,<br />
our tests show the relevance of this represent<strong>at</strong>ion and its superiority over all the mo<strong>de</strong>ls consi<strong>de</strong>red here.<br />
In<strong>de</strong>ed, we have shown th<strong>at</strong> it is the best (i.e, the most parcimonious) represent<strong>at</strong>ion of the d<strong>at</strong>a for all<br />
quantiles above q9 = 80%.<br />
6 Discussion and Conclusions<br />
We have presented a st<strong>at</strong>istical analysis of the tail behavior of the distributions of the daily log-r<strong>et</strong>urns of<br />
the Dow Jones Industrial Average and of the 5-minutes log-r<strong>et</strong>urns of the Nasdaq Composite in<strong>de</strong>x. We<br />
have emphasized practical aspects of the applic<strong>at</strong>ion of st<strong>at</strong>istical m<strong>et</strong>hods to this problem. Although the<br />
applic<strong>at</strong>ion of st<strong>at</strong>istical m<strong>et</strong>hods to the study of empirical distributions of r<strong>et</strong>urns seems to be an obvious<br />
approach, it is necessary to keep in mind the existence of necessary conditions th<strong>at</strong> the empirical d<strong>at</strong>a must<br />
obey for the conclusions of the st<strong>at</strong>istical study to be valid. Maybe the most important condition in or<strong>de</strong>r to<br />
speak meaningfully about distribution functions is the st<strong>at</strong>ionarity of the d<strong>at</strong>a, a difficult issue th<strong>at</strong> we have<br />
not consi<strong>de</strong>red here. In particular, the importance of regime switching is now well established (Ramcham<br />
and Susmel 1998, Ang and Bekeart 2001) and should be accounted for.<br />
23<br />
87<br />
(35)
88 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Our purpose here has been to revisit a generally accepted fact th<strong>at</strong> the tails of the distributions of r<strong>et</strong>urns<br />
present a power-like behavior. Although there are some disagreements concerning the exact value of the<br />
power indices (the majority of previous workers accepts in<strong>de</strong>x values b<strong>et</strong>ween 3 and 3.5, <strong>de</strong>pending on the<br />
particular ass<strong>et</strong> and the investig<strong>at</strong>ed time interval), the power-like character of the tails of distributions of<br />
r<strong>et</strong>urns is not subjected to doubts. Often, the conviction of the existence of a power-like tail is based on<br />
the Gne<strong>de</strong>nko theorem st<strong>at</strong>ing the existence of only three possible types of limit distributions of normalized<br />
maxima (a finite maximum value, an exponential tail, and a power-like tail) tog<strong>et</strong>her with the exclusion of<br />
the first two types by experimental evi<strong>de</strong>nce. The power-like character of the log-r<strong>et</strong>urn tail ¯F(x) follows<br />
then simply from the power-like distribution of maxima. However, in this chain of arguments, the conditions<br />
nee<strong>de</strong>d for the fulfillment of the corresponding m<strong>at</strong>hem<strong>at</strong>ical theorems are often omitted and not discussed<br />
properly. In addition, wi<strong>de</strong>ly used arguments in favor of power law tails invoke the self-similarity of the d<strong>at</strong>a<br />
but are often assumptions r<strong>at</strong>her than experimental evi<strong>de</strong>nce or consequences of economic and financial<br />
laws.<br />
Here, we have shown th<strong>at</strong> standard st<strong>at</strong>istical estim<strong>at</strong>ors of heavy tails are much less efficient th<strong>at</strong> often<br />
assumed and cannot in general clearly distinguish b<strong>et</strong>ween a power law tail and a Str<strong>et</strong>ched Exponential<br />
tail even in the absence of long-range <strong>de</strong>pen<strong>de</strong>nce in the vol<strong>at</strong>ility. In fact, this can be r<strong>at</strong>ionalized by our<br />
discovery th<strong>at</strong>, in a certain limit where the exponent c of the str<strong>et</strong>ched exponential pdf goes to zero (tog<strong>et</strong>her<br />
with condition (26) as seen in the <strong>de</strong>riv<strong>at</strong>ion (27)), the str<strong>et</strong>ched exponential pdf tends to the Par<strong>et</strong>o distribution.<br />
Thus, the Par<strong>et</strong>o (or power law) distribution can be approxim<strong>at</strong>ed with any <strong>de</strong>sired accuracy on an<br />
arbitrary interval by a suitable adjustment of the pair (c,d) of the param<strong>et</strong>ers of the str<strong>et</strong>ched exponential<br />
pdf. We have then turned to param<strong>et</strong>ric tests which indic<strong>at</strong>e th<strong>at</strong> the class of Str<strong>et</strong>ched Exponential distribution<br />
provi<strong>de</strong>s a significantly b<strong>et</strong>ter fit to empirical r<strong>et</strong>urns than the Par<strong>et</strong>o, the exponential or the incompl<strong>et</strong>e<br />
Gamma distributions. All our tests are consistent with the conclusion th<strong>at</strong> the Str<strong>et</strong>ched Exponential mo<strong>de</strong>l<br />
provi<strong>de</strong>s the best effective apparent and parsimonious mo<strong>de</strong>l to account for the empirical d<strong>at</strong>a.<br />
However, this does not mean th<strong>at</strong> the str<strong>et</strong>ched exponential (SE) mo<strong>de</strong>l is the correct <strong>de</strong>scription of the tails<br />
of empirical distributions of r<strong>et</strong>urns. Again, as already mentioned, the strength of the SE mo<strong>de</strong>l comes<br />
from the fact th<strong>at</strong> it encompasses the Par<strong>et</strong>o mo<strong>de</strong>l in the tail and offers a b<strong>et</strong>ter <strong>de</strong>scription in the bulk of<br />
the distribution. To see where the problem arises, we report in table 6 our best ML-estim<strong>at</strong>es for the SE<br />
param<strong>et</strong>ers c (form param<strong>et</strong>er) and d (scale param<strong>et</strong>er) restricted to the quantile level q12 = 95%, which<br />
offers a good compromise b<strong>et</strong>ween a sufficiently large sample size and a restricted tail range leading to an<br />
accur<strong>at</strong>e approxim<strong>at</strong>ion in this range.<br />
Sample c d<br />
ND positive r<strong>et</strong>urns 0.039 (0.138) 4.54 · 10 −52 (2.17· 10 −49 )<br />
ND neg<strong>at</strong>ive r<strong>et</strong>urns 0.273 (0.155) 1.90 · 10 −7 (1.38· 10 −6 )<br />
DJ positive r<strong>et</strong>urns 0.274 (0.111) 4.81 · 10 −6 (2.49· 10 −5 )<br />
DJ neg<strong>at</strong>ive r<strong>et</strong>urns 0.362 (0.119) 1.02 · 10 −4 (2.87· 10 −4 )<br />
One can see th<strong>at</strong> c is very small (and all the more so for the scale param<strong>et</strong>er d) for the tail of positive r<strong>et</strong>urns<br />
of the Nasdaq d<strong>at</strong>a suggesting a convergence to a power law tail. The exponents c for the three other tails<br />
are an or<strong>de</strong>r of magnitu<strong>de</strong> larger but our tests show th<strong>at</strong> they are not incomp<strong>at</strong>ible with an asymptotic power<br />
tail either.<br />
Note also th<strong>at</strong> the exponents c seem larger for the daily DJ d<strong>at</strong>a than for the 5-minutes ND d<strong>at</strong>a, in agreement<br />
with an expected (slow) convergence to the Gaussian law according to the central limit theory (see Sorn<strong>et</strong>te<br />
<strong>et</strong> al. (2000) and figures 3.6-3.8 pp. 68 of Sorn<strong>et</strong>te (2000) where it is shown th<strong>at</strong> SE distributions are<br />
approxim<strong>at</strong>ely stable in family and the effect of aggreg<strong>at</strong>ion can be seen to slowly increase the exponent c).<br />
However, a t-test does not allow us to reject the hypotheses th<strong>at</strong> the exponents c remains the same for a given<br />
24
tail (positive or neg<strong>at</strong>ive) of the Dow Jones d<strong>at</strong>a. Thus, we confirm previous results (Lux 1996, Jon<strong>de</strong>au and<br />
Rockinger 2001, for instance) according to which the extreme tails can be consi<strong>de</strong>red as symm<strong>et</strong>ric, <strong>at</strong> least<br />
for the Dow Jones d<strong>at</strong>a.<br />
These are the evi<strong>de</strong>nce in favor of the existence of an asymptotic power law tail. Balancing this, many<br />
of our tests have shown th<strong>at</strong> the power law mo<strong>de</strong>l is not as powerful compared with the SE mo<strong>de</strong>l, even<br />
arbitrarily far in the tail (as far as the available d<strong>at</strong>a allows us to probe). In addition, our <strong>at</strong>tempts for a<br />
direct estim<strong>at</strong>ion of the exponent b of a possible power law tail has failed to confirm the existence of a<br />
well-converged asymptotic value (except maybe for the positive tail of the Nasdaq). In constrast, we have<br />
found th<strong>at</strong> the exponent b of the power law mo<strong>de</strong>l system<strong>at</strong>ically increases when going <strong>de</strong>eper and <strong>de</strong>eper<br />
in the tails, with no visible sign of exhausting this growth. We have proposed tent<strong>at</strong>ive param<strong>et</strong>eriz<strong>at</strong>ion of<br />
this growth of the apparent power law exponent. We note again th<strong>at</strong> this behavior is expected from mo<strong>de</strong>ls<br />
such as the GARCH or the Multifractal Random Walk mo<strong>de</strong>ls which predict asymptotic power law tails but<br />
with exponents of the or<strong>de</strong>r of 20 or larger, th<strong>at</strong> would be sampled <strong>at</strong> un<strong>at</strong>tainable quantiles.<br />
Attempting to wrap up the different results obtained by the b<strong>at</strong>tery of tests presented here, we can offer<br />
the following conserv<strong>at</strong>ive conclusion: it seems th<strong>at</strong> the four tails examined here are <strong>de</strong>caying faster than<br />
any (reasonable) power law but slower than any str<strong>et</strong>ched exponentials. Maybe log-normal or log-Weibull<br />
distributions could offer a b<strong>et</strong>ter effective <strong>de</strong>scription of the distribution of r<strong>et</strong>urns 9 . Such a mo<strong>de</strong>l has<br />
already been suggested by (Serva <strong>et</strong> al. 2002).<br />
The correct <strong>de</strong>scription of the distribution of r<strong>et</strong>urns has important implic<strong>at</strong>ions for the assessment of large<br />
risks not y<strong>et</strong> sampled by historical time series. In<strong>de</strong>ed, the whole purpose of a characteriz<strong>at</strong>ion of the functional<br />
form of the distribution of r<strong>et</strong>urns is to extrapol<strong>at</strong>e currently available historical time series beyond<br />
the range provi<strong>de</strong>d by the empirical reconstruction of the distributions. For risk management, the d<strong>et</strong>ermin<strong>at</strong>ion<br />
of the tail of the distribution is crucial. In<strong>de</strong>ed, many risk measures, such as the Value-<strong>at</strong>-Risk<br />
or the Expected-Shartfall, are based on the properties of the tail of the distributions of r<strong>et</strong>urns. In or<strong>de</strong>r to<br />
assess risk <strong>at</strong> probability levels of 95% or more, non-param<strong>et</strong>ric m<strong>et</strong>hods have merits. However, in or<strong>de</strong>r to<br />
estim<strong>at</strong>e risks <strong>at</strong> high probability level such as 99% or larger, non-param<strong>et</strong>ric estim<strong>at</strong>ions fail by lack of d<strong>at</strong>a<br />
and param<strong>et</strong>ric mo<strong>de</strong>ls become unavoidable. This shift in str<strong>at</strong>egy has a cost and replaces sampling errors<br />
by mo<strong>de</strong>l errors. The consi<strong>de</strong>red distribution can be too thin-tailed as when using normal laws, and risk will<br />
be un<strong>de</strong>restim<strong>at</strong>ed, or it is too f<strong>at</strong>-tailed and risk will be over estim<strong>at</strong>ed as with Lévy law and possibly with<br />
Par<strong>et</strong>o tails according to the present study. In each case, large amounts of money are <strong>at</strong> stake and can be lost<br />
due to a too conserv<strong>at</strong>ive or too optimistic risk measurement.<br />
Our present study suggests th<strong>at</strong> the Par<strong>et</strong>ian paradigm leads to an overestim<strong>at</strong>ion of the probability of large<br />
events and therefore leads to the adoption of too conserv<strong>at</strong>ive positions. Generalizing to larger time scales,<br />
the overly pessimistic view of large risks <strong>de</strong>riving from the Par<strong>et</strong>ian paradigm should be all the more revised,<br />
due to the action of the central limit theorem. Finally, an additional note of caution is in or<strong>de</strong>r. This study has<br />
focused on the marginal distributions of r<strong>et</strong>urns calcul<strong>at</strong>ed <strong>at</strong> fixed time scales and thus neglects the possible<br />
occurrence of runs of <strong>de</strong>pen<strong>de</strong>ncies, such as in cumul<strong>at</strong>ive drawdowns. In the presence of <strong>de</strong>pen<strong>de</strong>ncies<br />
b<strong>et</strong>ween r<strong>et</strong>urns, and especially if the <strong>de</strong>pen<strong>de</strong>nce is non st<strong>at</strong>ionary and increases in time of stress, the<br />
characteriz<strong>at</strong>ion of the marginal distributions of r<strong>et</strong>urns is not sufficient. As an example, Johansen and<br />
Sorn<strong>et</strong>te (2002) have recently shown th<strong>at</strong> the recurrence time of very large drawdowns cannot be predicted<br />
from the sole knowledge of the distribution of r<strong>et</strong>urns and th<strong>at</strong> transient <strong>de</strong>pen<strong>de</strong>nce eff<strong>et</strong>s occurring in time<br />
of stress make very large drawdowns more frequent, qualifying them as abnormal “outliers.”<br />
9 L<strong>et</strong> us stress th<strong>at</strong> we are speaking of a log-normal distribution of r<strong>et</strong>urns, not of price! In<strong>de</strong>ed, the standard Black and Scholes<br />
mo<strong>de</strong>l of a log-normal distribution of prices is equivalent to a Gaussian distribution of r<strong>et</strong>urns. Thus, a log-normal distribution of<br />
r<strong>et</strong>urns is much more f<strong>at</strong> tailed, and in fact brack<strong>et</strong>ed by power law tails and str<strong>et</strong>ched exponential tails.<br />
25<br />
89
90 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
A Maximum likelihood estim<strong>at</strong>ors<br />
In this appendix, we give the expressions of the maximum likelihood estim<strong>at</strong>ors <strong>de</strong>rived from the four<br />
distributions (22-25)<br />
A.1 The Par<strong>et</strong>o distribution<br />
According to expression (22), the Par<strong>et</strong>o distribution is given by<br />
<br />
u<br />
b Fu(x) = 1 − ,<br />
x<br />
x ≥ u (38)<br />
and its <strong>de</strong>nsity is<br />
L<strong>et</strong> us <strong>de</strong>note by<br />
L PD<br />
fu(x|b) = b ub<br />
x b+1<br />
T (ˆb) = max<br />
b<br />
T<br />
∑<br />
i=1<br />
(39)<br />
ln fu(xi|b) (40)<br />
the maximum of log-likekihood function <strong>de</strong>rived un<strong>de</strong>r hypothesis (PD). ˆb is the maximum likelihood estim<strong>at</strong>or<br />
of the tail in<strong>de</strong>x b un<strong>de</strong>r such hyptothesis.<br />
The maximum of the likelihood function is solution of<br />
which yields<br />
ˆb =<br />
<br />
1<br />
T<br />
T<br />
∑<br />
i=1<br />
lnxi − lnu<br />
1 1<br />
+ lnu −<br />
b T ∑lnxi = 0, (41)<br />
−1<br />
, and<br />
1<br />
T LPD T (ˆb) = ln ˆb<br />
u −<br />
<br />
1 + 1<br />
<br />
. (42)<br />
ˆb<br />
Moreover, one easily shows th<strong>at</strong> ˆb is asymptotically normally distributed:<br />
√ N(ˆb − b) ∼ N (0,b). (43)<br />
A.2 The Weibull distribution<br />
The Weibull distribution is given by equ<strong>at</strong>ion (23) and its <strong>de</strong>nsity is<br />
fu(x|c,d) = c<br />
d c · e( u d ) c<br />
x c−1 · exp<br />
The maximum of the log-likelihood function is<br />
L SE<br />
T (ĉ, d) ˆ = max<br />
T ∑Ti=1 T<br />
∑<br />
i=1<br />
c,d<br />
T<br />
∑<br />
i=1<br />
<br />
−<br />
x<br />
d<br />
c<br />
, x ≥ u. (44)<br />
ln fu(xi|c,d) (45)<br />
Thus, the maximum likehood estim<strong>at</strong>ors (ĉ, d) ˆ are solution of<br />
1<br />
c =<br />
1<br />
T ∑T xi c xi<br />
i=1 u ln<br />
u<br />
1 xi c −<br />
u − 1 1 T<br />
T ∑ ln<br />
i=1<br />
xi<br />
d<br />
,<br />
u<br />
(46)<br />
c = uc<br />
<br />
xi<br />
c<br />
− 1.<br />
T u<br />
(47)<br />
26
Equ<strong>at</strong>ion (46) <strong>de</strong>pends on c only and must be solved numerically. Then, the resulting value of c can be<br />
reinjected in (47) to g<strong>et</strong> d. The maximum of the log-likelihood function is<br />
1<br />
T LSE T (ĉ, d) ˆ = ln ĉ ĉ − 1<br />
+<br />
dˆ ĉ T<br />
T<br />
∑<br />
i=1<br />
91<br />
lnxi − 1. (48)<br />
Since c > 0, the vector √ N(ĉ − c, ˆ<br />
d − d) is asymptotically normal, with a covariance m<strong>at</strong>rix whose<br />
expression is given in appendix D by the inverse of K11.<br />
It should be noted th<strong>at</strong> maximum likelihood equ<strong>at</strong>ions (46-47) admit a solution with positive c not for<br />
all possible samples (x1,··· ,xN). In<strong>de</strong>ed, the function<br />
h(c) = 1<br />
c −<br />
1<br />
T ∑T xi c xi<br />
i=1 u ln<br />
u<br />
c +<br />
− 1 1 T<br />
ln<br />
T<br />
xi<br />
, (49)<br />
u<br />
1<br />
T ∑T xi<br />
i=1 u<br />
which is the total <strong>de</strong>riv<strong>at</strong>ive of LSE T (c, d(c)), ˆ is a <strong>de</strong>creasing function of c. It means, as one can expect,<br />
th<strong>at</strong> the likelihood function is concave. Thus, a necessary and sufficient condition for equ<strong>at</strong>ion (46) to<br />
admit a solution is th<strong>at</strong> h(0) is positive. After some calcul<strong>at</strong>ions, we find<br />
which is positive if and only if<br />
h(0) = 2 1<br />
T<br />
xi ∑ln u<br />
2 xi<br />
T ∑ln u<br />
∑<br />
i=1<br />
2 1 xi − T ∑ln2 u<br />
, (50)<br />
2 1 xi<br />
2<br />
T<br />
∑ln −<br />
u<br />
1 2 xi<br />
T<br />
∑ln > 0. (51)<br />
u<br />
However, the probability of occurring a sample providing a neg<strong>at</strong>ive maximum-likelyhood estim<strong>at</strong>e of<br />
c tends to zero (un<strong>de</strong>r Hypothesis of SE with a positive c) as<br />
<br />
Φ − c√ <br />
N σ<br />
√ e<br />
σ 2π Nc − c2N 2σ2 , (52)<br />
i.e. exponentially with respect to N. Here σ2 is the variance of the limit Gaussian distribution of<br />
maximum-likelihood c-estim<strong>at</strong>or th<strong>at</strong> can be <strong>de</strong>rived explicitly. If h(0) is neg<strong>at</strong>ive, LSE T reaches its<br />
maximum <strong>at</strong> c = 0 and in such a case<br />
1<br />
T LSE<br />
<br />
1 xi<br />
T (c = 0) = −ln<br />
T<br />
∑ln −<br />
u<br />
1<br />
T ∑lnxi − 1. (53)<br />
Now, if we apply maximum likelihood estim<strong>at</strong>ion based on SE assumption to samples distributed<br />
differently from SE, then we can g<strong>et</strong> neg<strong>at</strong>ive c-estim<strong>at</strong>e with some positive probability not tending<br />
to zero with N → ∞. If sample is distributed according to Par<strong>et</strong>o distribution, for instance, then<br />
maximum-likelihood c-estim<strong>at</strong>e converges in probability to a Gaussian random variable with zero<br />
mean, and thus the probability of neg<strong>at</strong>ive c-estim<strong>at</strong>es converges to 0.5.<br />
A.3 The Exponential distribution<br />
The Exponential distribution function is given by equ<strong>at</strong>ion (24), and its <strong>de</strong>nsity is<br />
fu(x|d) = exp <br />
u <br />
d exp −<br />
d<br />
x<br />
<br />
,<br />
d<br />
x ≥ u. (54)<br />
27
92 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
The maximum of the log-likelihood function is reach <strong>at</strong><br />
and is given by<br />
T<br />
d ˆ = 1<br />
T ∑ xi − u, (55)<br />
i=1<br />
1<br />
T LED T ( d) ˆ = −(1 + lnd). ˆ<br />
(56)<br />
The random variable √ N( ˆ<br />
d − d) is asymptotically normally distributed with zero mean and variance d 2 /N.<br />
A.4 The Incompl<strong>et</strong>e Gamma distribution<br />
The expression of the Incompl<strong>et</strong>e Gamma distribution function is given by (25) and its <strong>de</strong>nsity is<br />
fu(x|b,d) =<br />
d b<br />
Γ −b, u<br />
d<br />
· x −(b+1) <br />
exp −<br />
<br />
x<br />
<br />
, x ≥ u. (57)<br />
d<br />
L<strong>et</strong> us introduce the partial <strong>de</strong>riv<strong>at</strong>ive of the logarithm of the incompl<strong>et</strong>e Gamma function:<br />
Ψ(a,x) = ∂<br />
∞ 1<br />
lnΓ(a,x) = dt lnt t<br />
∂a Γ(a,x) x<br />
a−1 e −t . (58)<br />
The maximum of the log-likelihood function is reached <strong>at</strong> the point (ˆb, ˆ<br />
d) solution of<br />
and is equal to<br />
1<br />
T<br />
T<br />
∑<br />
i=1<br />
1<br />
T<br />
1<br />
T LIG<br />
<br />
T (ˆb, d) ˆ = −lndˆ − lnΓ<br />
ln xi<br />
d<br />
T<br />
xi<br />
∑<br />
i=1 d =<br />
−b, u<br />
d<br />
<br />
= Ψ<br />
−b, u<br />
d<br />
1<br />
Γ −b, u<br />
<br />
d<br />
<br />
<br />
+ (b + 1) · Ψ<br />
28<br />
<br />
, (59)<br />
<br />
u<br />
−b e<br />
d<br />
− u d − b, (60)<br />
−b, u<br />
d<br />
<br />
+ b −<br />
1<br />
Γ −b, u<br />
d<br />
<br />
<br />
u<br />
−b e<br />
d<br />
− u d . (61)
B Minimum An<strong>de</strong>rson-Darling Estim<strong>at</strong>ors<br />
We <strong>de</strong>rive in this appendix the expressions allowing the calcul<strong>at</strong>ion of the param<strong>et</strong>ers which minimize the<br />
An<strong>de</strong>rson-Darling distance b<strong>et</strong>ween the assumed distribution and the true distribution.<br />
Given the or<strong>de</strong>red sample x1 ≤ x2 ≤ ··· ≤ xN, the AD-distance is given by<br />
ADN = −N − 2<br />
N<br />
∑<br />
k=1<br />
[wk logF(xk|α) + (1 − wk)log(1 − F(xk|α))], (62)<br />
where α represents the vector of param<strong>at</strong>ers and wk = 2k/(2N + 1). It is easy to show th<strong>at</strong> the minimun is<br />
reached <strong>at</strong> the point ˆα solution of<br />
B.1 The Par<strong>et</strong>o distribution<br />
N<br />
∑<br />
k=1<br />
Applying equ<strong>at</strong>ion (63) to the Par<strong>et</strong>o distribution yields<br />
<br />
1 − wk<br />
<br />
log(1 − F(xk|α)) = 0. (63)<br />
F(xk|α)<br />
N<br />
∑<br />
k=1<br />
1 −<br />
wk<br />
<br />
u<br />
xk<br />
b ln u<br />
xk<br />
This equ<strong>at</strong>ion always admits a unique solution, and can easily be solved numerically.<br />
B.2 Str<strong>et</strong>ched-Exponential distribution<br />
In the Str<strong>et</strong>ched-Exponential case, we obtain the two following equ<strong>at</strong>ions<br />
with<br />
N<br />
∑<br />
k=1<br />
<br />
1 − wk<br />
ln<br />
Fk<br />
u<br />
d<br />
∑<br />
k=1<br />
=<br />
N<br />
∑<br />
k=1<br />
93<br />
ln u<br />
. (64)<br />
xk<br />
<br />
u<br />
c − ln<br />
d<br />
xk<br />
<br />
xk c<br />
d d<br />
= 0, (65)<br />
N <br />
1 − wk<br />
<br />
(u<br />
Fk<br />
c − x c k ) = 0, (66)<br />
<br />
Fk = 1 − exp − uc − xc k<br />
dc <br />
. (67)<br />
After some simple algebraic manipul<strong>at</strong>ions, the first equ<strong>at</strong>ion can be slightly simplified, to finally yields<br />
N<br />
∑<br />
k=1<br />
N<br />
∑<br />
k=1<br />
<br />
1 − wk<br />
<br />
ln<br />
Fk<br />
xk<br />
u<br />
<br />
1 − wk<br />
Fk<br />
xk<br />
u<br />
<br />
xk<br />
c<br />
u<br />
c <br />
− 1<br />
= 0, (68)<br />
= 0. (69)<br />
However, these two equ<strong>at</strong>ions remain coupled. Moreover, we have not y<strong>et</strong> been able to prove the unicity of<br />
the solution.<br />
29
94 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
B.3 Exponential distribution<br />
In the exponential case, equ<strong>at</strong>ion(63) becomes<br />
with<br />
N <br />
wk<br />
∑<br />
k=1<br />
Fk<br />
Here again, we can show th<strong>at</strong> this equ<strong>at</strong>ion admits a unique solution.<br />
<br />
− 1 (u − xk) = 0, (70)<br />
<br />
u − xk<br />
Fk = 1 − exp − . (71)<br />
d<br />
30
C Local exponent<br />
In this Appendix, we come back to the notion of a local Par<strong>et</strong>o exponent discussed above in section 4.3.1,<br />
with figure 7 and expression (33). We call the local exponent β.<br />
Generally speaking, any positive smooth function g(x) can be represented in the form<br />
g(x) = 1/x β(x) , x 1, (72)<br />
by <strong>de</strong>fining β(x) = −ln(g(x))/lnx. Thus, the local in<strong>de</strong>x β(x) of a distribution F(x) will be <strong>de</strong>fined as<br />
β(x) = −<br />
95<br />
ln(1 − F(x))<br />
. (73)<br />
ln(x)<br />
For example, the log-normal distribution can be mistaken for a power law over several <strong>de</strong>ca<strong>de</strong>s with very<br />
slowly (logarithmically) varying exponents if its variance is large (see figures 4.2 and 4.3 in section 4.1.3<br />
of (Sorn<strong>et</strong>te 2000)). When we approxim<strong>at</strong>e a sample distribution by some param<strong>et</strong>ric family with moving<br />
in<strong>de</strong>x β(x), it is important th<strong>at</strong> β(x) should have as small a vari<strong>at</strong>ion as possible, i.e., th<strong>at</strong> the represent<strong>at</strong>ion<br />
(72) be parsimonious. Given a sample x1,x2,··· ,xN, drawn from a distribution function F(x), x ≥ 1, the tail<br />
in<strong>de</strong>x β(x) is consistently estim<strong>at</strong>ed by<br />
ˆβ(x) = lnN − ln <br />
∑i 1 {xi>x}<br />
, (74)<br />
ln(x)<br />
where 1 {·} is the indic<strong>at</strong>or function which equals one if its argument is true and zero otherwhise. The<br />
asymptotic distribution of the estim<strong>at</strong>or is easily <strong>de</strong>rived and reads:<br />
with<br />
N 1/2 ·<br />
<br />
e ln(x)·ˆ β(x) − e ln(x)·β(x) d<br />
−→ N (0,σ 2 ), (75)<br />
σ 2 =<br />
1<br />
. (76)<br />
F(x) · (1 − F(x))<br />
As an example, l<strong>et</strong> us illustr<strong>at</strong>e the properties of the local in<strong>de</strong>x β(x) for regularly varying distributions. The<br />
general represent<strong>at</strong>ion of any regularly varying distribution is given by ¯F(x) = L(x) · x −α , where L(·) is a<br />
slowly varying function. In such a case, the local power in<strong>de</strong>x can be written as<br />
β(x) = α − lnL(x)<br />
, (77)<br />
ln(x)<br />
which goes to α as x goes to infinity, as expected from the build-in slow vari<strong>at</strong>ion of L(x). The upper panel<br />
of figure 10 shows the local in<strong>de</strong>x β(x) for a simul<strong>at</strong>ed Par<strong>et</strong>o distribution with power in<strong>de</strong>x b = 1.2 (X > 1).<br />
The estim<strong>at</strong>e β(x) oscill<strong>at</strong>es very closely to the true value b = 1.2.<br />
L<strong>et</strong> us now assume th<strong>at</strong> we observe a regularly varying local exponent<br />
β(x) = L(x) · x c , with c > 0, (78)<br />
it would be clearly the char<strong>at</strong>eriz<strong>at</strong>ion of a Str<strong>et</strong>ched-Exponential distribution<br />
¯F(x) = exp −L ′ (x) · x c , (79)<br />
31
96 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
where L ′ (x) = ln(x) L(x). The lower panel of figure 10 shows the local in<strong>de</strong>x β(x) for a simul<strong>at</strong>ed Str<strong>et</strong>ched-<br />
Exponential distribution with c = 0.3. In this case, the local in<strong>de</strong>x β(x) continuously increases, which corresponds<br />
to our intuition th<strong>at</strong> an Str<strong>et</strong>ched-Exponential behaves like a power-law with an always increasing<br />
exponent.<br />
Figure 11 shows the local in<strong>de</strong>x β(x) for a distribution constructed by joining two Par<strong>et</strong>o distributions with<br />
exponents b1 = 0.70 and b2 = 1.5 <strong>at</strong> the cross-over point u1 = 10. In this case, the local in<strong>de</strong>x β(x) again<br />
increases but not so quickly as in previous Str<strong>et</strong>ched-Exponential case. Even for such large sample size<br />
(n = 15000), the “final” β(x) is about 1.3 which is still less than the true b-value for the second Par<strong>et</strong>o part<br />
(b = 1.5), showing the existence of strong cross-over effects. Such very strong cross-over effects occurring<br />
in the presence of a transition from a power law to another power law have already been noticed in (Sorn<strong>et</strong>te<br />
<strong>et</strong> al. 1996).<br />
Similarly to the local in<strong>de</strong>x β(x) based on the Par<strong>et</strong>o distribution taken as a reference, we <strong>de</strong>fine the notion<br />
of a local exponent c(x) taking Str<strong>et</strong>ched-Exponential distributions as a reference. In<strong>de</strong>ed, given any<br />
sufficiently smooth positive function g(x), one can always find a function c(x) such th<strong>at</strong><br />
<br />
g(x) = exp 1 − x c(x)<br />
, x ≥ 1, (80)<br />
with c(x) = ln(1 − ln(g(x)))/ln(x). Obviously, for any Str<strong>et</strong>ched-Exponential distribution with exponent c,<br />
the local exponent c(x) converges to c as x goes to infinity. This property is the same as for the local in<strong>de</strong>x<br />
β(x) which goes to the true tail in<strong>de</strong>x β for regularly varying distributions.<br />
Figure 12 shows the sample tail (continuous line), the local in<strong>de</strong>x β(x) (dashed line) and the local exponent<br />
c(x) (dash-dotted line) for the neg<strong>at</strong>ive tail of the Nasdaq five-minutes r<strong>et</strong>urns. The local exponent c(x)<br />
clearly reaches an asymptotic value ∼ = 0.32 for large enough values of the r<strong>et</strong>urns. In contrast, the local exponent<br />
β(x) remains continuously increasing. The lower panel shows in double logarithmic scale the local<br />
in<strong>de</strong>x β(x). Over a large range, β(x) increases approxim<strong>at</strong>ely a power law of in<strong>de</strong>x 0.77 while beyond the<br />
quantile 99% (see the ins<strong>et</strong>) it behaves like a power law with smaller in<strong>de</strong>x equal to 0.54 implying a <strong>de</strong>celer<strong>at</strong>ing<br />
growth. The goodness of fit of the regression of lnβ(x) on lnx has been qualified by a χ 2 test which<br />
does not allow to reject this mo<strong>de</strong>l <strong>at</strong> any usual confi<strong>de</strong>nce level. Note th<strong>at</strong> a power law <strong>de</strong>pen<strong>de</strong>nce of the<br />
local Par<strong>et</strong>o exponent β(x) as a function of x qualifies a Str<strong>et</strong>ched-Exponential distribution, according to (78)<br />
and (79). The second regime fitted with the exponent c = 0.54 seems still perturbed by a cross-over effect<br />
as it does not r<strong>et</strong>rieve the value c ∼ = 0.32 which characterizes the tail of the distributions of r<strong>et</strong>urns according<br />
to the Str<strong>et</strong>ched-Exponential mo<strong>de</strong>l. These fits quantifying the growth of the local Par<strong>et</strong>o exponent are not<br />
in contradiction with figure 7 and expression (33). In<strong>de</strong>ed, a <strong>de</strong>creasing positive exponent is an altern<strong>at</strong>ive<br />
<strong>de</strong>scription for a logarithmic growth and vice-versa. These fits provi<strong>de</strong> an improved quantific<strong>at</strong>ion of the<br />
previously rough characteriz<strong>at</strong>ion of the growth of the exponent estim<strong>at</strong>ed per quantile shown with 7 and<br />
expression (33) but using a b<strong>et</strong>ter characteriz<strong>at</strong>ion of the “local” exponent.<br />
The positive tail of the Nasdaq and both tails of the Dow Jones exhibit exactly the same continuously increasing<br />
behaviour with the same characteristics and we are not showing them. Taken tog<strong>et</strong>her, these oberv<strong>at</strong>ions<br />
suggest th<strong>at</strong> the Str<strong>et</strong>ched-Exponential represent<strong>at</strong>ion provi<strong>de</strong>s a b<strong>et</strong>ter mo<strong>de</strong>l of the tail behavior of large<br />
r<strong>et</strong>urns than does a regularly varying distribution.<br />
32
D Testing non-nested hypotheses with the encompassing principle<br />
D.1 Testing the Par<strong>et</strong>o mo<strong>de</strong>l against the (SE) mo<strong>de</strong>l<br />
D.1.1 Pseudo-true value<br />
L<strong>et</strong> us consi<strong>de</strong>r the two mo<strong>de</strong>ls, str<strong>et</strong>ched exponential (SE) and Par<strong>et</strong>o (P). The pdf’s associ<strong>at</strong>ed with these<br />
two mo<strong>de</strong>ls are f1(x|c,d) and f2(x|b) respectively . Un<strong>de</strong>r the true distribution f0(x), we will d<strong>et</strong>ermine the<br />
pseudo-true values of the maximum likelihood estim<strong>at</strong>ors ˆb, ĉ and d, ˆ namely, the values of these param<strong>et</strong>ers<br />
which minimize the (Kullback-Leibler) distance b<strong>et</strong>ween the consi<strong>de</strong>red mo<strong>de</strong>l and the true distribution<br />
(Gouriéroux and Monfort 1994). Thus, the pseudo-true values b∗ , c∗ and d∗ of ˆb, ĉ and dˆ appear as the<br />
expected values of the estim<strong>at</strong>ors un<strong>de</strong>r f0. For instance<br />
b ∗ = arginf<br />
b E0<br />
<br />
ln f0(x)<br />
<br />
, (81)<br />
f2(x|b)<br />
where, in all wh<strong>at</strong> follows, E0[·] <strong>de</strong>notes the expect<strong>at</strong>ion un<strong>de</strong>r the probability measure associ<strong>at</strong>ed with the<br />
true distribution f0. Thus, b∗ is simply solution of<br />
∂<br />
∂b E0<br />
<br />
ln f0(x)<br />
<br />
= 0, (82)<br />
f2(x|b)<br />
which yields<br />
b ∗ = (E0[lnx] − lnu) −1 , (83)<br />
and is consistently estim<strong>at</strong>ed by the maximum likelihood estim<strong>at</strong>or<br />
<br />
ˆb<br />
1<br />
=<br />
T ∑lnxi<br />
−1 − lnu . (84)<br />
In fact, the maximum likelihood estim<strong>at</strong>or ˆb converges to its pseudo-true value, with (ˆb−b ∗ ) asymptotically<br />
normally distributed with zero mean.<br />
Similarly, the maximum likelihood estim<strong>at</strong>ors ĉ and ˆ<br />
d converge to their pseudo-true values c ∗ and d ∗ .<br />
D.1.2 Binding functions and encompassing<br />
L<strong>et</strong> us now ask wh<strong>at</strong> is the value b † (c,d) of the param<strong>et</strong>er b for which f2(x|b) is the nearest to f1(x|c,d), for<br />
a given (c,d). Such a value b † is the binding function and is solution of<br />
b † (c,d) = arginf<br />
b E1<br />
<br />
ln f1(x|c,d)<br />
<br />
, (85)<br />
f2(x|b)<br />
which only involves f1 and f2 but not the true distribution f0. The binding function is given by<br />
b † <br />
(c,d) = E1 ln x<br />
<br />
− ln<br />
d<br />
u<br />
−1 . (86)<br />
d<br />
After some calcul<strong>at</strong>ions, we find<br />
<br />
E1 ln x<br />
<br />
d<br />
97<br />
= c<br />
dc e( u d ) c<br />
·∞<br />
dx ln<br />
u<br />
x<br />
d xc−1 e −( x d ) c<br />
, (87)<br />
= ln u<br />
d + e( u d ) c<br />
c Γ<br />
<br />
u<br />
c 0, ,<br />
d<br />
(88)<br />
33
98 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
so th<strong>at</strong> the binding function can be expressed as<br />
b † (c,d) = c · e −( u d ) c<br />
<br />
−1 u<br />
c Γ 0, , when c > 0 (89)<br />
d<br />
In the case when c goes to zero and c and d are rel<strong>at</strong>ed by equ<strong>at</strong>ion (47), we can still calcul<strong>at</strong>e b † . In<strong>de</strong>ed,<br />
we can first show th<strong>at</strong><br />
<br />
u<br />
c ∼<br />
d<br />
1<br />
<br />
1<br />
c T<br />
T<br />
ln xi<br />
−1 =<br />
u<br />
1<br />
c · ˆb. (90)<br />
Now, using the asymptoptic rel<strong>at</strong>ion<br />
we conclu<strong>de</strong> th<strong>at</strong><br />
∑<br />
i=1<br />
e x · Γ(0,x) ∼ x −1 , as x → ∞, (91)<br />
b † (ĉ, ˆ<br />
d) = ˆb, as ĉ → 0. (92)<br />
This result is in fact n<strong>at</strong>ural because the PD mo<strong>de</strong>l can be seen formally as the limit of the SE mo<strong>de</strong>l for<br />
c → 0 un<strong>de</strong>r the condition<br />
as previously exposed in section 4.1<br />
<br />
u<br />
c c · → β,<br />
d<br />
as c → 0 , (93)<br />
Now, following Mizon and Richard (1986), the mo<strong>de</strong>l (SE) with pdf f1 is said to encompass the mo<strong>de</strong>l<br />
(PD) with pdf f2 if the best represent<strong>at</strong>ive of (PD) -with param<strong>et</strong>er b ∗ - is also the distribution nearest to the<br />
best represent<strong>at</strong>ive of (SE) -with param<strong>et</strong>ers (c ∗ ,d ∗ ). Thus, (SE) is said to encompass (PD) if and only if<br />
b ∗ = b † (c ∗ ,d ∗ ).<br />
The reverse situ<strong>at</strong>ion can be consi<strong>de</strong>red in or<strong>de</strong>r to study the encompassing of the mo<strong>de</strong>l (SE) by the mo<strong>de</strong>l<br />
(PD). Such situ<strong>at</strong>ion occurs if and only if<br />
c ∗<br />
D.1.3 Wald encompassing test<br />
d ∗<br />
<br />
c † (b∗ )<br />
=<br />
d † (b∗ )<br />
<br />
. (94)<br />
We first test the encompassing of (PD) into (SE), namely the null hypothesis H0 = {b ∗ = b † (c ∗ ,d ∗ )}.<br />
Un<strong>de</strong>r this null hypothesis, it can be shown (Gouriéroux and Monfort 1994) th<strong>at</strong> the random variable<br />
√ T ˆb − b † (ĉ, ˆ<br />
d) is asymptotically normally distributed with zero mean and variance V given by<br />
V = K −1<br />
C12]K −1<br />
22 [C22 −C21C −1<br />
11<br />
+ K −1 −1<br />
22<br />
[C21C11 − ˜C21K −1<br />
11<br />
22 +<br />
]C11[C −1<br />
11 C12 − K −1<br />
11<br />
˜C12]K −1<br />
22<br />
where the expression of the coefficients involved in V will be given below. Thus, un<strong>de</strong>r H0, the random variable<br />
ξT = T ˆb − b † (ĉ, d) ˆ<br />
ˆV −1 ˆb − b † (ĉ, d) ˆ<br />
, where ˆV −1 is a consistent estim<strong>at</strong>or of V −1 , follows asymptotically<br />
a χ2-distribution with one <strong>de</strong>grees of freedom.<br />
The m<strong>at</strong>rix K11 is given by<br />
K11(i, j) = −E0<br />
∂ 2 ln f1(x|c ∗ ,d ∗ )<br />
∂αi∂α j<br />
34<br />
(95)<br />
<br />
, i, j = 1,2, (96)
where α = (c ∗ ,d ∗ ). It can be consistently estim<strong>at</strong>ed by<br />
The coefficient K22 is given by<br />
ˆK11(1,1) = 1<br />
ĉ u u<br />
− ln2<br />
+<br />
ĉ2 dˆ<br />
dˆ<br />
1 T<br />
T ∑ ln<br />
i=1<br />
2<br />
ĉ xi xi<br />
, (97)<br />
dˆ<br />
dˆ<br />
ˆK11(1,2) = ˆK11(2,1) = c<br />
ĉ u u<br />
ln<br />
−<br />
d dˆ<br />
dˆ<br />
1 T <br />
ĉ<br />
xi xi<br />
T ∑ ln<br />
, (98)<br />
ˆ<br />
i=1 d dˆ<br />
2 ˆK11(2,2)<br />
ĉ<br />
= . (99)<br />
dˆ<br />
K22 = −E0<br />
which is consistently estim<strong>at</strong>ed by ˆK22 = ˆb −2 .<br />
∂ 2 ln f (x|b ∗ )<br />
∂b ∗2<br />
<br />
99<br />
= 1<br />
, (100)<br />
b∗2 Now, we have to calcul<strong>at</strong>e the two components of the vector ˜C12 = ˜C t 21 . Its first component is<br />
<br />
˜C12(1)<br />
∂ln f1(x|c<br />
= E1<br />
∗ ,d ∗ )<br />
∂c∗ · ∂ln f2(x|b∗ )<br />
∂b∗ <br />
, (101)<br />
<br />
1 u<br />
= E1 + ln<br />
c∗ d∗ <br />
u<br />
d∗ c∗ <br />
+ ln x x<br />
− ln<br />
d∗ d∗ <br />
x<br />
d∗ c∗ <br />
1 u<br />
· + ln<br />
b∗ d∗ <br />
− ln x<br />
d∗ <br />
,(102)<br />
<br />
1 u<br />
= + ln<br />
c∗ d∗ <br />
u<br />
d∗ c∗ <br />
1 u 1 u<br />
· + ln − + ln<br />
b∗ d c∗ d∗ <br />
u<br />
d∗ c∗ <br />
· E1 ln x<br />
d∗ <br />
<br />
1 u<br />
+ + ln<br />
b∗ d∗ <br />
· E1 ln x<br />
d∗ <br />
− E1 ln x<br />
d∗ <br />
x<br />
d∗ c∗ <br />
2 x<br />
− E1 ln<br />
d∗ <br />
2 x<br />
+ E1 ln<br />
d∗ <br />
x<br />
d∗ c∗ (103) .<br />
Some simple calcul<strong>at</strong>ions show th<strong>at</strong><br />
E1<br />
E1<br />
<br />
ln x<br />
d ·<br />
<br />
x<br />
c d<br />
<br />
2 x<br />
ln<br />
d ·<br />
<br />
x<br />
c d<br />
= 1 u<br />
+ ln<br />
c d ·<br />
<br />
u<br />
c <br />
+ E1 ln<br />
d<br />
x<br />
<br />
, (104)<br />
d<br />
2 u<br />
= ln<br />
d ·<br />
<br />
u<br />
c +<br />
d<br />
2<br />
c E1<br />
<br />
ln x<br />
<br />
2 x<br />
<br />
+ E1 ln , (105)<br />
d d<br />
which allows us to show th<strong>at</strong> the first and third terms cancel out, and it remains<br />
˜C12(1) = 1<br />
<br />
E1 ln<br />
c∗ x<br />
d∗ <br />
− ln u<br />
d∗ <br />
u<br />
d∗ c∗ <br />
E1 ln x<br />
d∗ <br />
− ln u<br />
d∗ <br />
. (106)<br />
The second component is<br />
˜C12(2) =<br />
<br />
∂ln f1(x|c<br />
E1<br />
∗ ,d ∗ )<br />
∂d∗ · ∂ln f2(x|b∗ )<br />
∂b∗ =<br />
<br />
,<br />
<br />
E1 −<br />
(107)<br />
c∗<br />
d∗ <br />
u<br />
1 +<br />
d∗ c∗ <br />
x<br />
−<br />
d∗ c∗ <br />
1 u<br />
· + ln<br />
b∗ d∗ <br />
− ln x<br />
d∗ =<br />
<br />
,<br />
−<br />
(108)<br />
c∗<br />
d∗ <br />
u<br />
1 +<br />
d∗ c∗ 1 u<br />
+ ln<br />
b∗ d∗ <br />
u<br />
− 1 +<br />
d∗ c∗ <br />
E1 ln x<br />
d∗ −<br />
<br />
<br />
1 u<br />
+ ln<br />
b∗ d∗ x<br />
E1<br />
d∗ c∗ <br />
+ E1 ln x<br />
d∗ <br />
x<br />
d∗ c∗ . (109)<br />
35
100 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Now, accounting for the rel<strong>at</strong>ion E1[xc ] = uc +d c , the first and third term within the brack<strong>et</strong>s cancel out, and<br />
accounting for equ<strong>at</strong>ion (104) yields<br />
˜C12(2) = − c∗<br />
d∗ <br />
1<br />
<br />
u<br />
+<br />
c∗ d∗ c∗ <br />
· ln u<br />
<br />
− E ln<br />
d∗ x<br />
d∗ . (110)<br />
Finally, ˜C12 can be consistently estim<strong>at</strong>ed by replacing (c ∗ ,d ∗ ) by (ĉ, ˆ<br />
d) and using the rel<strong>at</strong>ion<br />
L<strong>et</strong> us now <strong>de</strong>fine<br />
g1(x|c ∗ ,d ∗ ) =<br />
<br />
E<br />
ln x<br />
d<br />
∂ln f1(x|c ∗ ,d ∗ )<br />
∂c ∗<br />
∂ln f1(x|c ∗ ,d ∗ )<br />
∂d ∗<br />
g2(x|b ∗ ) = ∂ln f2(x|b ∗ )<br />
∂b ∗<br />
The m<strong>at</strong>rices Ci j are <strong>de</strong>fined as<br />
which can be consistently estim<strong>at</strong>ed by<br />
<br />
= ln u 1<br />
+<br />
d c e( u d ) c <br />
· Γ 0,<br />
<br />
=<br />
<br />
1<br />
c∗ + ln u<br />
d∗ − c∗<br />
d∗ <br />
u<br />
c . (111)<br />
d<br />
<br />
u<br />
d∗ c∗ <br />
+ ln x<br />
d∗ − ln x<br />
d∗ 1 + u<br />
d∗ c∗ − x<br />
d∗ c∗ <br />
x c∗ d<br />
<br />
(112)<br />
= 1<br />
+ lnu − lnx. (113)<br />
b∗<br />
<br />
Ci j = E0 gi(x) · g j(x) t , i, j = 1,2, (114)<br />
Ĉi j = 1<br />
T<br />
T<br />
∑<br />
i=1<br />
D.2 Testing the (SE) mo<strong>de</strong>l against the Par<strong>et</strong>o mo<strong>de</strong>l<br />
gi(x) · g j(x) t , i, j = 1,2. (115)<br />
L<strong>et</strong> us now assume th<strong>at</strong>, beyond a given high threshold u, the true mo<strong>de</strong>l is the Par<strong>et</strong>o mo<strong>de</strong>l, th<strong>at</strong> is, the true<br />
r<strong>et</strong>urns distribution is a power law with pdf<br />
This will be our null hypothesis H0.<br />
f0(x|b) = b ub<br />
, x ≥ u. (116)<br />
xb+1 Now consi<strong>de</strong>r the maximum likelihood estim<strong>at</strong>ors (ĉ, d) ˆ of the mo<strong>de</strong>l (SE). They are solution of equ<strong>at</strong>ions<br />
(46-47)<br />
1<br />
c =<br />
d c = uc<br />
T<br />
1<br />
T ∑T xi<br />
i=1 u<br />
1<br />
T ∑T xi<br />
i=1 u<br />
T <br />
xi<br />
∑<br />
i=1 u<br />
Un<strong>de</strong>r H0, (ĉ, ˆ<br />
d) converges to the pseudo-true values, solutions of<br />
1<br />
c<br />
<br />
x<br />
E0 u = <br />
x<br />
E0 u<br />
c xi ln<br />
u<br />
c −<br />
− 1 1 T<br />
T ∑ ln<br />
i=1<br />
xi<br />
,<br />
u<br />
(117)<br />
c − 1. (118)<br />
c <br />
x ln<br />
u<br />
c − 1 − E0<br />
<br />
ln x<br />
<br />
, (119)<br />
u<br />
d c = E0 [x c ] − u c , (120)<br />
36
where E0[·] <strong>de</strong>notes the expect<strong>at</strong>ion with respect to the power-law distribution f0. We have<br />
<br />
x<br />
c E0<br />
u<br />
<br />
E0 ln<br />
=<br />
b<br />
, and c < b,<br />
b − c<br />
(121)<br />
x<br />
<br />
u<br />
= 1<br />
<br />
x<br />
c E0 ln<br />
u<br />
,<br />
b<br />
(122)<br />
x<br />
<br />
u<br />
=<br />
b<br />
.<br />
(b − c) 2 (123)<br />
Thus, we easily obtain th<strong>at</strong> the unique solution of (119) is c = 0, and equ<strong>at</strong>ion (120) does not make sense<br />
any more. So, un<strong>de</strong>r H0, ĉ goes to zero and dˆ is not well <strong>de</strong>fined. Thus, Wald test cannot be performed un<strong>de</strong>r<br />
such a null hypothesis. We must find another way to test (SE) against H0.<br />
In this goal, we remark th<strong>at</strong> the quantity<br />
ˆηT = ĉ<br />
<br />
ĉ<br />
u<br />
+ 1<br />
dˆ<br />
101<br />
(124)<br />
is still well <strong>de</strong>fined. Using (120), it is easy to show th<strong>at</strong>, as T goes to infinity, this quantity goes to b,<br />
wh<strong>at</strong>ever ĉ being positive or equal to zero. For positive c, this is obvious from (120) and (121), while for<br />
c = 0, expanding (118) around ĉ = 0 yields<br />
ĉ<br />
ĉ u<br />
dˆ<br />
=<br />
<br />
1<br />
T<br />
T<br />
∑<br />
i=1<br />
ln xi<br />
u<br />
−1<br />
(125)<br />
= ˆb → b. (126)<br />
In or<strong>de</strong>r to test the <strong>de</strong>scriptive power of the (SE) mo<strong>de</strong>l against the null hypothesis th<strong>at</strong> the true mo<strong>de</strong>l is the<br />
Par<strong>et</strong>o mo<strong>de</strong>l, we can consi<strong>de</strong>r the st<strong>at</strong>istic<br />
<br />
ˆηT<br />
ζT = T − 1 , (127)<br />
ˆb<br />
which asymptoticaly follows a χ 2 -distribution with one <strong>de</strong>gree of freedom. In<strong>de</strong>ed expanding the quantity<br />
(xi/u) ĉ in power series around c = 0 gives<br />
which allows us to g<strong>et</strong><br />
where<br />
<br />
xi<br />
ĉ<br />
∼= 1 + ĉ · log(<br />
u<br />
xi<br />
u<br />
<br />
) + ĉ2<br />
2 · log2 <br />
xi<br />
+<br />
u<br />
ĉ3<br />
6 · log3 <br />
xi<br />
+ ··· , as ĉ → 0, (128)<br />
u<br />
1<br />
T ∑ <br />
xi<br />
ĉ<br />
∼= 1 + ĉ · S1 +<br />
u<br />
ĉ2<br />
2 · S2 + ĉ3<br />
3 S3, (129)<br />
1<br />
T ∑ <br />
xi<br />
c <br />
xi<br />
log ∼= S1 + ĉ · S2 +<br />
u u<br />
ĉ2<br />
2 S3, (130)<br />
S1 = 1<br />
T<br />
S2 = 1<br />
T<br />
S3 = 1<br />
T<br />
T<br />
∑<br />
i=1<br />
T<br />
∑<br />
i=1<br />
T<br />
∑<br />
i=1<br />
log<br />
xi<br />
u<br />
log 2 xi<br />
u<br />
log 3 xi<br />
u<br />
37<br />
<br />
, (131)<br />
<br />
, (132)<br />
<br />
. (133)<br />
(134)
102 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Thus, we g<strong>et</strong> from equ<strong>at</strong>ion (117):<br />
and from equ<strong>at</strong>ions (124) and (118):<br />
ĉ <br />
1<br />
2 S2 − S 2 1<br />
1<br />
2S1S2 − 1<br />
3S3 ˆηT<br />
ˆb − 1 = ĉ · S2 1<br />
1 − 2S2 + ĉ<br />
2 [S1S2 − 1<br />
S1 + ĉ<br />
2S2 + ĉ2<br />
6 S3<br />
, (135)<br />
3 S3]<br />
. (136)<br />
Now, accounting for the fact th<strong>at</strong> the variables ξ1 = S1 − b−1 , ξ2 = S2 − 2b−2 and ξ3 = S3 − 6b−3 are<br />
asymptoticaly Gaussian random variables with zero mean and variance of or<strong>de</strong>r T −1/2 , we obtain, <strong>at</strong> the<br />
lowest or<strong>de</strong>r in T −1/2 :<br />
ĉ = b 2<br />
<br />
2ξ1 − b<br />
2 ξ2<br />
<br />
, (137)<br />
and<br />
which shows th<strong>at</strong><br />
S2 1<br />
1 − 2S2 + ĉ<br />
2 [S1S2 − 1<br />
3S3] S1 + ĉ<br />
2 S2 + ĉ2<br />
6 S3<br />
ˆηT<br />
ˆb<br />
− 1 = b2<br />
<br />
<br />
= 2ξ1 − b<br />
2 ξ2<br />
<br />
, (138)<br />
2ξ1 − b<br />
2 ξ2<br />
2<br />
. (139)<br />
We now use the fact th<strong>at</strong> ξ1, ξ2 are asymptotically Gaussian random variables with zero mean. We find their<br />
variances:<br />
Var(ξ1) = 1<br />
T b2 , Var(ξ2) = 20<br />
T b4 , and Cov(ξ1,ξ2) = 4<br />
.<br />
T b3 (140)<br />
Using equ<strong>at</strong>ion (139), the random variable b(2ξ1 − bξ2/2) is in the limit of large T a Gaussian rv with zero<br />
mean and variance 1/T , i.e., ζT is asymptoticaly distributed according to a χ 2 distribution with one <strong>de</strong>gree<br />
of freedom.<br />
Thus, the test consists in accepting H0 if ζT ≤ χ2 1−ε (1) and rejecting it otherwise. (ε <strong>de</strong>notes the asymptotic<br />
level of the test).<br />
38
References<br />
An<strong>de</strong>rsen, J.V. and D. Sorn<strong>et</strong>te, 2001, Have your cake and e<strong>at</strong> it too: increasing r<strong>et</strong>urns while lowering large<br />
risks! Journal of Risk Finance 2 (3), 70-82 .<br />
An<strong>de</strong>rson, T.W. and D.A. Darling, 1952, Asymptotic theory of certain “goodness of fit” criteria, annals of<br />
M<strong>at</strong>hem<strong>at</strong>ical st<strong>at</strong>istics 23, 193-212.<br />
An<strong>de</strong>rsson, M., B. Eklund and J. Lyhagen, 1999, A simple linear time series mo<strong>de</strong>l with misleading nonlinear<br />
properties, Economics L<strong>et</strong>ter 65, 281-285.<br />
Ang, A. and G. Bekeart, 2001, Intern<strong>at</strong>ional ass<strong>et</strong> alloc<strong>at</strong>ion with regime shifts, Review of Financial Studies.<br />
Bachelier, L., 1900, Théorie <strong>de</strong> la spécul<strong>at</strong>ion, Annales <strong>de</strong> l’Ecole Normale Supérieure 17, 21-86.<br />
Barndorff-Nielsen, O.E., 1997, Normal inverse Gaussian distributions and the mo<strong>de</strong>lling of stock r<strong>et</strong>urns<br />
Scandinavian, J. St<strong>at</strong>istics 24, 1-13.<br />
Beran, J., 1994, St<strong>at</strong>istics for long-memory processes, Monographs on St<strong>at</strong>istics and Applied Probabilty 61,<br />
Chapman & Hall.<br />
Biham, O., Malcai, O., Levy, M. and Solomon, S., 1998, Generic emergence of power law distributions and<br />
Levy-Stable intermittent fluctu<strong>at</strong>ions in discr<strong>et</strong>e logistic systems, Phys. Rev. E 58, 1352-1358.<br />
Biham, O., Z-F. Huang, O. Malcai and S. Solomon, 2002, Long-Time Fluctu<strong>at</strong>ions in a Dynamical Mo<strong>de</strong>l<br />
of Stock Mark<strong>et</strong> Indices, preprint <strong>at</strong> http://arXiv.org/abs/cond-m<strong>at</strong>/0208464<br />
Bingham, N.H., C.M. Goldie and J.L. Teugel, 1987, Regular vari<strong>at</strong>ion, Cambridge University Press.<br />
Black, F. and M. Scholes, 1973, the pricing of options and corpor<strong>at</strong>e liabilities, Journal of Political Economy<br />
81, 637-653.<br />
Blanchard, O.J. and M.W. W<strong>at</strong>son, 1982, Bubbles, R<strong>at</strong>ional Expect<strong>at</strong>ions and Specul<strong>at</strong>ive Mark<strong>et</strong>s, in:<br />
Wachtel, P. ,eds., Crisis in Economic and Financial Structure: Bubbles, Bursts, and Shocks. Lexington<br />
Books: Lexington.<br />
Bl<strong>at</strong>tberg, R. and Gonne<strong>de</strong>s, N., 1974, A comparison of stable and Stu<strong>de</strong>nt distribution as st<strong>at</strong>istical mo<strong>de</strong>ls<br />
for stock prices, J. Business 47, 244-280.<br />
Bollerslev, T., 1986, Generalized autoregressive conditional h<strong>et</strong>eroskedasticity, Journal of Econom<strong>et</strong>rics 31,<br />
307-327.<br />
Bollerslev T, R.F. Engle and D.B. Nelson, 1994, ARCH mo<strong>de</strong>ls, Handbook of Econom<strong>et</strong>rics IV, 2959-3038.<br />
Bouchaud, J.-P. and M. Potters, Theory of financial risks: from st<strong>at</strong>istical physics to risk management<br />
(Cambridge [England]; New York: Cambridge University Press, 2000).<br />
Brock, W.A., W.D. Dechert and J.A. Scheinkman, 1987, A Test for in<strong>de</strong>pen<strong>de</strong>nce Based on the Correl<strong>at</strong>ion<br />
Dimension, unpublished manuscript, Department of Economics, University of Wisconsin, Madison.<br />
Campbell, J.Y., Lo, A.W. and MacKinlay, A.C., 1997, The econom<strong>et</strong>rics of financial mark<strong>et</strong>s (Princ<strong>et</strong>on,<br />
N.J. : Princ<strong>et</strong>on University Press).<br />
Chall<strong>et</strong>, D. and M. Marsili, 2002, Criticality and finite size effects in a simple realistic mo<strong>de</strong>l of stock<br />
mark<strong>et</strong>, preprint <strong>at</strong> http://arXiv.org/abs/cond-m<strong>at</strong>/0210549<br />
39<br />
103
104 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Champenowne, D.G, 1953, Economic Journal 63, 318-51.<br />
Cont, R., Potters, M. and J.-P. Bouchaud, 1997, Scaling in stock mark<strong>et</strong> d<strong>at</strong>a: stable laws and beyond, in<br />
Scale Invariance and Beyond (Proc. CNRS Workshop on Scale Invariance, Les Houches, 1997), eds.<br />
B. Dubrulle, F. Graner and D. Sorn<strong>et</strong>te (Berlin: Springer).<br />
Cox, D.R., 1961, Test of separ<strong>at</strong>e families of hypotheses, in : Proceedings of the Fourth Berkeley Symposium<br />
on M<strong>at</strong>hem<strong>at</strong>ical St<strong>at</strong>istics and Probabilty 1, 105-123, University of California Press: Berkeley .<br />
Cox, D.R., 1962, Further results of separ<strong>at</strong>e families of hypotheses, journal of the Royal St<strong>at</strong>istical Soci<strong>et</strong>y<br />
B 24, 406-424.<br />
Cromwell, J.B., W.C. Labys and M. Terraza, 1994, Univari<strong>at</strong>e Tests for Time Series Mo<strong>de</strong>ls (Sage, Thousand<br />
Oaks, CA, pp. 20-22).<br />
Dacorogna M.M., U.A. Müller, O.V. Pict<strong>et</strong> and C.G. <strong>de</strong> Vries, 1992, The distribution of extremal foreign<br />
exchange r<strong>at</strong>e r<strong>et</strong>urns in large d<strong>at</strong>e s<strong>et</strong>s, Working Paper, Olsen and Associ<strong>at</strong>es Internal Documents<br />
UAM, 19921022.<br />
Dragulescu, A.A. and V.M. Yakovenko, 2002, Probability distribution of r<strong>et</strong>urns for a mo<strong>de</strong>l with stochastic<br />
vol<strong>at</strong>ility, preprint <strong>at</strong> http://xxx.lanl.gov/abs/cond-m<strong>at</strong>/0203046<br />
Eberlein, E., Keller, U. and Prause, K., 1998, New insights into smile, mispricing and value <strong>at</strong> risk: the<br />
hyperbolic mo<strong>de</strong>l J. Business 71, 371-405.<br />
Embrechts P., C.P. Klüppelberg and T. Mikosh, 1997, Mo<strong>de</strong>lling Extremal Events (Springer-Verlag).<br />
Engle, R.F., 1984, Wald, Likelihood R<strong>at</strong>io, and Lagrange Multiplier Tests in Econom<strong>et</strong>rics , in: Z. Griliches<br />
and M.D. Intrilig<strong>at</strong>or, eds., Handbook of Econom<strong>et</strong>rics, Vol. II. North-Holland: Amsterdam.<br />
Engle, R.F. and A.J. P<strong>at</strong>ton, 2001, Wh<strong>at</strong> good is a vol<strong>at</strong>ility mo<strong>de</strong>l ?, Quantit<strong>at</strong>ive Finance 1, 237-245.<br />
Fama E.F., 1965, The Behavior of Stock Mark<strong>et</strong> Prices, J. Business 38, 34-105.<br />
Fama E.F. and K.R. French, 1996, Multifactor explan<strong>at</strong>ions of ass<strong>et</strong> pricing anomalies, Journal of Finance<br />
51.<br />
Fang, H. and T. Lai, 1997, Co-kurtosis and capital ass<strong>et</strong> pricing, Financial Review 32, 293-307.<br />
Farmer J.D., 1999, Physicists Attempt to Scale the Ivory Towers of Finance, Computing in Science and<br />
Engineering, Nov/Dec 1999, 26-39.<br />
Feller, W., 1971, An Introduction to Probability Theory and its Applic<strong>at</strong>ions, vol. II (John Wiley and<br />
sons, New York).<br />
Frisch, U. and D. Sorn<strong>et</strong>te, 1997, Extreme <strong>de</strong>vi<strong>at</strong>ions and applic<strong>at</strong>ions, J. Phys. I France 7, 1155-1171.<br />
Gabaix, X., 1999, Zipf’s law for cities: An explan<strong>at</strong>ion, Qu<strong>at</strong>erly J. Econ. 114, 739-767.<br />
Gabaix, X., P. Gopikrishnan, V. Plerou and H.E. Stanley, 2002, Un<strong>de</strong>rstanding Large Movements in Stock<br />
Mark<strong>et</strong> Activity, in press in N<strong>at</strong>ure<br />
Gopikrishnan P., M. Meyer, L.A.N. Amaral and H.E. Stanley, 1998, Inverse Cubic Law for the Distribution<br />
of Stock Price Vari<strong>at</strong>ions, European Physical Journal B 3, 139 –140.<br />
Gouriéroux C. and J. Jasiak, 1998, Trunc<strong>at</strong>ed maximum likelihood, goodness of fit tests and tail analysis,<br />
Working paper, CREST.<br />
40
Gouriéroux C. and A. Monfort, 1994, Testing non nested hypothesis, Handbook of Econom<strong>et</strong>rics 4, 2585-<br />
2637.<br />
Granger, C.W.J. and T. Teräsvirta, 1999, A simple nonlinear mo<strong>de</strong>l with misleading properties, Economics<br />
L<strong>et</strong>ter 62, 741-782.<br />
Guillaume D.M., M.M. Dacorogna, R.R. Davé, U.A. Müller, R.B. Olsen and O.V. Pict<strong>et</strong>, 1997, From the<br />
bird’s eye to the microscope: a survey of new stylized facts of the intra-day foreign exchange mark<strong>et</strong>s,<br />
Finance and Stochastics 1, 95-130.<br />
Hall, P.G, 1979, On the r<strong>at</strong>e of convergence of normal extremes. Journal of Applied Probabilities 16, 433-<br />
439.<br />
Hall, W.J. and J.A. Wellnel, 1979, The r<strong>at</strong>e of convergence in law of the maximum of an exponential sample.<br />
St<strong>at</strong>istica Neerlandica 33, 151-154.<br />
Hill, B.M., 1975, A simple general approach to inference about the tailof a distribution, Annals of st<strong>at</strong>istics<br />
3, 1163-1174.<br />
Hwang, S. and S. S<strong>at</strong>chell, 1999, Mo<strong>de</strong>lling emerging mark<strong>et</strong> risk premia using higher moments, Intern<strong>at</strong>ional<br />
Journal of Finance and Economics 4, 271-296.<br />
Joe, H, 1997, Multivari<strong>at</strong>e mo<strong>de</strong>ls and <strong>de</strong>pen<strong>de</strong>nce concepts, Chapman & Hall, London.<br />
Johansen, A. and D. Sorn<strong>et</strong>te, 2002, Large Stock Mark<strong>et</strong> Price Drawdowns Are Outliers, Journal of Risk 4<br />
(2), 69-110.<br />
E. Jon<strong>de</strong>au and M. Rockinger, 2001, Testing for differences in the tails of stock-mark<strong>et</strong> r<strong>et</strong>urns. Working<br />
Paper available <strong>at</strong> htt p : //papers.ssrn.com/paper.ta f ?abstract id = 291399<br />
Jurcenko, E. and B. Maill<strong>et</strong>, 2002, The four-moment capital ass<strong>et</strong> pricing mo<strong>de</strong>l: some basic results, Working<br />
Paper.<br />
Kearns P. and A. Pagan, 1997, Estim<strong>at</strong>ing the <strong>de</strong>nsity tail in<strong>de</strong>x for financial time series, Review of Economics<br />
and St<strong>at</strong>istics 79, 171-175.<br />
Kon, S., 1984, Mo<strong>de</strong>ls of stock r<strong>et</strong>urns: a comparison J. Finance XXXIX, 147-165.<br />
Laherrère J. and D. Sorn<strong>et</strong>te, 1999, Str<strong>et</strong>ched exponential distributions in n<strong>at</strong>ure and economy: F<strong>at</strong> tails with<br />
characteristic scales, European Physical Journal B 2, 525-539.<br />
Levy, M., S. Solomon and G. Ram, 1996, Dynamical explan<strong>at</strong>ion of the emergence of power law in a stock<br />
mark<strong>et</strong> mo<strong>de</strong>l, Int. J. Mod. Phys. C 7, 65-72.<br />
Longin F.M., 1996, The asymptotic distribution of extreme stock mark<strong>et</strong> r<strong>et</strong>urns, Journal of Business 96,<br />
383-408.<br />
Lux T., 1996, The stable Par<strong>et</strong>ian hypothesis and the frequency of large r<strong>et</strong>urns: an examin<strong>at</strong>ion of major<br />
German stocks, Applied Financial Economics 6, 463-475.<br />
Lux T., 2000, The Limiting Extreme Behavior of Specul<strong>at</strong>ive R<strong>et</strong>urns: An Analysis of Intra-Daily D<strong>at</strong>a from<br />
the Francfurt Stock Exchange, preprint,March 2000, 29 p.<br />
Lux, L. and D. Sorn<strong>et</strong>te, 2002, On R<strong>at</strong>ional Bubbles and F<strong>at</strong> Tails, J. Money Credit and Banking, Part 1, 34,<br />
589-610.<br />
41<br />
105
106 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2001, Multi-dimensional R<strong>at</strong>ional Bubbles and f<strong>at</strong> tails, Quantit<strong>at</strong>ive Finance<br />
1, 533-541.<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2002, Multi-Moments M<strong>et</strong>hod for Portfolio Management: Generalized<br />
Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>l in Homogeneous and H<strong>et</strong>erogeneous mark<strong>et</strong>s, working paper (htt p :<br />
//papers.ssrn.com/paper.ta f ?abstract id = 319544)<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2002, Investig<strong>at</strong>ing Extreme Depen<strong>de</strong>nces: Concepts and Tools, working<br />
paper (htt p : //papers.ssrn.com/paper.ta f ?abstractid = 303465)<br />
Man<strong>de</strong>lbrot B., 1963, The vari<strong>at</strong>ion of certain specul<strong>at</strong>ive prices, Journal of Business 36, 392-417.<br />
Mantegna R.N. and H.E. Stanley, 1995, Scaling behavior of an economic in<strong>de</strong>x, N<strong>at</strong>ure 376, 46-55.<br />
Mantegna R.N. and H.E.Stanley, 2000, An Introduction to Econophysics, Correl<strong>at</strong>ions and Complexity in<br />
Finance, Cambridge Univ. Press, Cambridge UK.<br />
Markovitz, H., 1959, Portfolio Selection: Efficient diversific<strong>at</strong>ion of investments, john Wiley and Sons, New<br />
York.<br />
M<strong>at</strong>ia, K., L.A.N. Amaral, S.P. Goodwin and H.E. Stanley, Non-Lévy Distribution of Commodity Price<br />
Fluctu<strong>at</strong>ions, preprint cond-m<strong>at</strong>/0202028<br />
Mittnik S., S.T.Rachev, M.S.Paolella,1998, Stable Par<strong>et</strong>ian Mo<strong>de</strong>ling in Finance: Some Empirical and Theor<strong>et</strong>ical<br />
Aspects, In: A Practical Gui<strong>de</strong> to Heavy Tails, pp.79-110, Eds. R.J.Adler, R.E.Feldman,<br />
M.S.Taqqu, Birkhauser, Boston.<br />
Mizon, G.E. and J.F. Richard, 1984, The encompassing principle and its applic<strong>at</strong>ions to testing non-nested<br />
hypotheses, Econom<strong>et</strong>rica 54, 675-678.<br />
Müller U.A., M.M.Dacarogna, O.V.Pickt<strong>et</strong>, 1998, Heavy Tails in High-Frequency Financial D<strong>at</strong>a, In: A<br />
Practical Gui<strong>de</strong> to Heavy Tails, pp.55-78, Eds. R.J.Adler, R.E.Feldman, M.S.Taqqu, Birkhauser,<br />
Boston.<br />
Muzy, J.-F., Delour, J. and Bacry, E., 2000, Mo<strong>de</strong>lling fluctu<strong>at</strong>ions of financial time series: from casca<strong>de</strong><br />
process to stochastic vol<strong>at</strong>ility mo<strong>de</strong>l, European Physical Journal 17, 537-548.<br />
Muzy, J.-F., D. Sorn<strong>et</strong>te, J. Delour and A. Arneodo, 2001, Multifractal r<strong>et</strong>urns and Hierarchical Portfolio<br />
Theory, Quantit<strong>at</strong>ive Finance 1, 131-148.<br />
Nagahara, Y. and G. Kitagawa, 1999, A non-Gaussian stochastic vol<strong>at</strong>ility mo<strong>de</strong>l, J. Comput<strong>at</strong>ional Finance<br />
2, 33-47.<br />
Nelsen, R., 1998, An introduction to copulas, Lectures notes in st<strong>at</strong>istics 139, Springer Verlag, New York.<br />
Pagan A., 1996, The econom<strong>et</strong>rics of financial mark<strong>et</strong>s, Journal of Empirical Finance 3, 15-102.<br />
Prause, K., 1998, The generalized hyperbolic mo<strong>de</strong>l, PhD Dissert<strong>at</strong>ion, University of Freiburg.<br />
Ramshand, L. and R. Susmel, 1998, Vol<strong>at</strong>ility and cross correl<strong>at</strong>ion across major stock mark<strong>et</strong>s, Journal of<br />
Empirical Finance 5, 397-416.<br />
Rootzen H., M.R. Leadb<strong>et</strong>ter and L. <strong>de</strong> Haan, 1998, On the distribution of tail array sums for strongly<br />
mixing st<strong>at</strong>ionnary sequences, Annals of Applied Probability 8, 868-885.<br />
42
Rubinstein, M., 1973, The fundamental theorem of param<strong>et</strong>er-preference security valu<strong>at</strong>ion. Journal of Financial<br />
and Quantit<strong>at</strong>ive Analysis 8, 61-69.<br />
Samuelson, P.A., 1965, Proof th<strong>at</strong> Properly Anticip<strong>at</strong>ed Prices Fluctu<strong>at</strong>e Randomly, Industrial Management<br />
Review 6, 41-49.<br />
Serva, M., U.L. Fulco, M.L. Lyra and G.M. Viswan<strong>at</strong>han, 2002, Kinem<strong>at</strong>ics of stock prices, preprint <strong>at</strong><br />
http://arXiv.org/abs/cond-m<strong>at</strong>/0209103<br />
Sharpe, W., 1964, Capital ass<strong>et</strong> prices: A theory of mark<strong>et</strong> equilirium un<strong>de</strong>r condition of risk, journal of<br />
finance 19, 425-442.<br />
Simon, H.A., 1957, Mo<strong>de</strong>ls of man: social and r<strong>at</strong>ional; m<strong>at</strong>hem<strong>at</strong>ical essays on r<strong>at</strong>ional human behavior in<br />
a social s<strong>et</strong>ting (New York, Wiley)<br />
Smith, R.L., 1985, Maximum likelihood estim<strong>at</strong>ion in a class of non-regular cases. Biom<strong>et</strong>rika 72, 67-90.<br />
Sorn<strong>et</strong>te, D., 1998, Linear stochastic dynamics with nonlinear fractal properties, Physica A 250, 295-314.<br />
Sorn<strong>et</strong>te, D., 2000, Critical Phenomena in N<strong>at</strong>ural Sciences, Chaos, Fractals, Self-organiz<strong>at</strong>ion and Disor<strong>de</strong>r:<br />
Concepts and Tools, (Springer Series in Synerg<strong>et</strong>ics, Hei<strong>de</strong>lberg).<br />
Sorn<strong>et</strong>te, D. and R. Cont, 1997, Convergent multiplic<strong>at</strong>ive processes repelled from zero: power laws and<br />
trunc<strong>at</strong>ed power laws, J. Phys. I France 7, 431-444.<br />
Sorn<strong>et</strong>te, D., L. Knopoff, Y.Y. Kagan and C. Vanneste, 1996, Rank-or<strong>de</strong>ring st<strong>at</strong>istics of extreme events:<br />
applic<strong>at</strong>ion to the distribution of large earthquakes, J.Geophys.Res. 101, 13883-13893.<br />
Sorn<strong>et</strong>te, D., P. Simon<strong>et</strong>ti and J.V. An<strong>de</strong>rsen, 2000, φ q -field theory for Portfolio optimiz<strong>at</strong>ion: “f<strong>at</strong> tails” and<br />
non-linear correl<strong>at</strong>ions, Physics Report 335 (2), 19-92.<br />
Starica C. and O. Pict<strong>et</strong>, 1999, The tales the tails of GARCH(1,1) process tell. Working Paper, Univ. of<br />
Pennsylvania.<br />
Stuart A. and K. Ord, 1994, Kendall’s advances theory of st<strong>at</strong>istics (John Wiley and Sons).<br />
Mizuno, T., S. Kurihara, M. Takayasu and H. Takayasu, 2002, Analysis of high-resolution foreign exchange<br />
d<strong>at</strong>a of USD-JPY for 13 years, working paper http://xxx.lanl.gov/abs/cond-m<strong>at</strong>/0211162<br />
Vries, <strong>de</strong> C.G., 1994, Stylized facts of nominal exchange r<strong>at</strong>e r<strong>et</strong>urns, in The Handbook of Intern<strong>at</strong>ional<br />
Macroeconomics, F. van <strong>de</strong>r Ploeg (ed.), 348-389 (Blackwell).<br />
Wilks, S. S., 1938, The Large Sample Distribution of the Likelihood R<strong>at</strong>io for Testing Composite Hypotheses<br />
, Annals of M<strong>at</strong>hemurical St<strong>at</strong>istics 9, 60-62.<br />
43<br />
107
108 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Mean St. Dev. Skewness Ex. Kurtosis Jarque-Bera<br />
Nasdaq (5 minutes) † 1.80 ·10 −6 6.61 ·10 −4 0.0326 11.8535 1.30 ·10 5 (0.00)<br />
Nasdaq (1 hour) † 2.40 ·10 −5 3.30 ·10 −3 1.3396 23.7946 4.40 ·10 4 (0.00)<br />
Nasdaq (5 minutes) ‡ - 6.33 ·10 −9 3.85 ·10 −4 -0.0562 6.9641 4.50 ·10 4 (0.00)<br />
Nasdaq (1 hour) ‡ 1.05 ·10 −6 1.90 ·10 −3 -0.0374 4.5250 1.58 ·10 3 (0.00)<br />
Dow Jones (1 day) 8.96·10 −5 4.70 ·10 −3 -0.6101 22.5443 6.03 ·10 5 (0.00)<br />
Dow jones (1 month) 1.80 ·10 −3 2.54 ·10 −2 -0.6998 5.3619 1.28 ·10 3 (0.00)<br />
Table 1: Descriptive st<strong>at</strong>istics for the Dow Jones r<strong>et</strong>urns calcul<strong>at</strong>ed over one day and one month and for<br />
the Nasdaq r<strong>et</strong>urns calcul<strong>at</strong>ed over five minutes and one hour. The numbers within parenthesis represent<br />
the p-value of Jarque-Bera’s normality test. (†) raw d<strong>at</strong>a, (‡) d<strong>at</strong>a corrected for the U-shape of the intra-day<br />
vol<strong>at</strong>ility due to the lunch effect.<br />
44
(a) In<strong>de</strong>pen<strong>de</strong>nt D<strong>at</strong>a<br />
Str<strong>et</strong>ched-Exponential c=0.7 Str<strong>et</strong>ched-Exponential c=0.3 Par<strong>et</strong>o Distribution b=3<br />
Maximum Maximum Maximum<br />
cluster 10 50 100 200 cluster 10 50 100 200 cluster 10 50 100 200<br />
mean 0.1870 0.1734 0.2032 0.2290 mean 0.9313 0.9910 1.0617 1.1011 mean 0.3669 0.4007 0.4455 0.4769<br />
Emp Std 0.0262 0.0340 0.0382 0.0415 Emp Std 0.0380 0.0464 0.0492 0.0508 Emp Std 0.0267 0.0353 0.0384 0.0407<br />
GPD GPD GPD<br />
quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995<br />
mean 0.1108 0.0758 -0.1374 -0.0509 mean -0.1234 0.1947 -0.2147 -0.2681 mean 0.3336 0.3271 0.3193 0.2585<br />
Emp Std 0.0346 0.1579 0.4648 0.4147 Emp Std 0.8139 0.6197 0.6625 0.597 Emp Std 0.0454 0.0604 0.1271 0.2991<br />
(b) Depen<strong>de</strong>nt D<strong>at</strong>a<br />
Str<strong>et</strong>ched-Exponential c=0.7 with long memory Str<strong>et</strong>ched-Exponential c=0.3 with long memory Par<strong>et</strong>o with long memory b=3<br />
Maximum Maximum Maximum<br />
cluster 10 50 100 200 cluster 10 50 100 200 cluster 10 50 100 200<br />
mean -0.2169 -0.2226 -0.2027 -0.1918 mean 0.1230 0.1143 0.1359 0.1569 mean 0.0566 0.0606 0.0794 0.0978<br />
Emp Std 0.1553 0.1512 0.1553 0.1718 Emp Std 0.0328 0.0414 0.0450 0.0443 Emp Std 0.0422 0.0495 0.0533 0.0560<br />
45<br />
GPD GPD GPD<br />
quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995<br />
mean -0.1005 -0.0593 -0.0189 -0.0314 mean 0.0732 0.0740 0.0031 -0.0559 mean 0.2952 0.3203 0.3153 0.2570<br />
Emp Std 0.0386 0.0503 0.1097 0.1700 Emp Std 0.0942 0.1804 0.3744 0.4822 Emp Std 0.0548 0.0694 0.1812 0.3382<br />
Table 2: Mean values and standard <strong>de</strong>vi<strong>at</strong>ions of the Maximum Likelihood estim<strong>at</strong>es of the param<strong>et</strong>er ξ (inverse of the Par<strong>et</strong>o exponent) for the distribution<br />
of maxima (cf. equ<strong>at</strong>ion 4) when d<strong>at</strong>a are clustered in samples of size 10,50,100 and 200 and for the Generalized Par<strong>et</strong>o Distribution (7) for thresholds<br />
u corresponding to quantiles 90%,95%,99% ans 99.5%. In panel (a), we have used iid samples of size 10000 drawn from a Str<strong>et</strong>ched-Exponential<br />
distribution with c = 0.7 and c = 0.3 and a Par<strong>et</strong>o distribution with tail in<strong>de</strong>x b = 3, while in panel (b) the samples are drawn from a long memory process<br />
with Str<strong>et</strong>ched-Exponential marginals and regularly-varying marginal as explained in the text.<br />
109
110 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
(a) In<strong>de</strong>pen<strong>de</strong>nt d<strong>at</strong>a<br />
Str<strong>et</strong>ched-Exponential c=0.7 Str<strong>et</strong>ched-Exponential c=0.3 Par<strong>et</strong>o Distribution b=3<br />
quantile 0.005 0.01 0.05 0.1 quantile 0.005 0.01 0.05 0.1 quantile 0.005 0.01 0.05 0.1<br />
N/k=4 N/k=4 N/k=4<br />
mean 0.0791 0.0689 0.1172 0.1432 mean 0.3843 0.4340 0.6301 0.7768 mean 0.3460 0.3496 0.3359 0.3334<br />
emp. Std 0.5156 0.3614 0.1652 0.1154 emp. Std 0.5643 0.3925 0.1800 0.1340 emp. Std 0.5384 0.3631 0.1760 0.1175<br />
th. Std 0.5148 0.3635 0.1636 0.1161 th. Std 0.5393 0.3847 0.1786 0.1302 th. Std 0.5358 0.3791 0.1691 0.1195<br />
N/k=10 N/k=10 N/k=10<br />
mean 0.0726 0.0652 0.0880 0.1083 mean 0.3613 0.3819 0.5028 0.5950 mean 0.3134 0.3379 0.3213 0.3382<br />
emp. Std 0.8218 0.5925 0.2602 0.1839 emp. Std 0.8961 0.6141 0.2741 0.1963 emp. Std 0.9003 0.6056 0.2551 0.1933<br />
th. Std 0.8133 0.5745 0.2577 0.1827 th. Std 0.8493 0.6027 0.2755 0.1983 th. Std 0.8425 0.5982 0.2668 0.1892<br />
(b) Depen<strong>de</strong>nt d<strong>at</strong>a<br />
Str<strong>et</strong>ched-Exponential c=0.7 with long memory Str<strong>et</strong>ched-Exponential c=0.3 with long memory Par<strong>et</strong>o b=3 with long memory<br />
quantile 0.005 0.01 0.05 0.1 quantile 0.005 0.01 0.05 0.1 quantile 0.005 0.01 0.05 0.1<br />
N/k=4 N/k=4 N/k=4<br />
mean -0.0439 -0.0736 -0.2134 -0.3395 mean 0.1297 0.1215 0.0410 0.0235 mean 0.1458 0.1261 -0.0023 -0.0790<br />
emp. Std 0.5396 0.3599 0.1643 0.1128 emp. Std 0.5557 0.3756 0.1633 0.1130 emp. Std 0.5572 0.3765 0.1633 0.1125<br />
th. Std 0.5073 0.3576 0.1579 0.1108 th. Std 0.5183 0.3661 0.1620 0.1143 th. Std 0.5194 0.3663 0.1612 0.1130<br />
46<br />
N/k=10 N/k=10 N/k=10<br />
mean -0.0371 -0.0307 -0.0928 -0.1781 mean 0.0961 0.1275 0.0682 0.0521 mean 0.1239 0.1461 0.0584 0.0196<br />
emp. Std 0.8325 0.5754 0.2561 0.1821 emp. Std 0.8314 0.5796 0.2607 0.1801 emp. Std 0.8352 0.5811 0.2608 0.1801<br />
th. Std 0.8028 0.5680 0.2524 0.1771 th. Std 0.8158 0.5793 0.2570 0.1814 th. Std 0.8188 0.5808 0.2567 0.1807<br />
Table 3: Pickands estim<strong>at</strong>es (9) of the param<strong>et</strong>er ξ for the Generalized Par<strong>et</strong>o Distribution (7) for thresholds u corresponding to quantiles 90%,95%,99%<br />
ans 99.5% and two different values of the r<strong>at</strong>io N/k respectively equal to 4 and 10. In panel (a), we have used iid samples of size 10000 drawn from a<br />
Str<strong>et</strong>ched-Exponential distribution with c = 0.7 and c = 0.3 and a Par<strong>et</strong>o distribution with tail in<strong>de</strong>x b = 3, while in panel (b) the samples are drawn from<br />
a long memory process with Str<strong>et</strong>ched-Exponential marginals and regularly-varying marginal.
(a) Dow Jones<br />
Positive Tail Neg<strong>at</strong>ive Tail<br />
Maximum Maximum<br />
cluster 10 50 100 200 cluster 10 50 100 200<br />
mean 0.2195 0.2150 0.2429 0.2884 mean 0.1346 0.1012 0.1222 0.1700<br />
Emp Std 0.0850 0.1163 0.1342 0.1238 Emp Std 0.1318 0.1437 0.1549 0.1671<br />
GPD GPD<br />
quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995<br />
mean 0.2479 0.2493 0.1808 0.3458 mean 0.2121 0.0244 0.2496 0.3118<br />
Emp Std 0.0224 0.0436 0.0654 0.1209 Emp Std 0.0294 0.0344 0.0979 0.1452<br />
(b) Nasdaq (Raw d<strong>at</strong>a)<br />
Positive Tail Neg<strong>at</strong>ive Tail<br />
Maximum Maximum<br />
cluster 10 50 100 200 cluster 10 50 100 200<br />
mean 1.2961 1.1992 0.9858 0.7844 mean 1.2105 0.4707 0.4146 0.4200<br />
Emp Std 0.2837 0.2955 0.4377 0.4679 Emp Std 0.2780 0.4625 0.4117 0.3980<br />
GPD GPD<br />
quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995<br />
mean 0.2123 0.3025 0.5973 0.5973 mean 0.1477 0.2189 0.5973 0.5973<br />
Emp Std 0.0693 0.0712 0 0 Emp Std 0.0468 0.0798 0 0<br />
(c) Nasdaq (Corr<strong>et</strong>ed d<strong>at</strong>a)<br />
Positive Tail Neg<strong>at</strong>ive Tail<br />
Maximum Maximum<br />
cluster 10 50 100 200 cluster 10 50 100 200<br />
mean 0.2163 0.2652 0.2738 0.2771 mean 0.2530 0.2276 0.3173 0.3681<br />
Emp Std 0.2461 0.2825 0.2364 0.2604 Emp Std 0.3345 0.3273 0.4303 0.4289<br />
GPD GPD<br />
quantile 0.9 0.95 0.99 0.995 quantile 0.9 0.95 0.99 0.995<br />
mean 0.2569 0.4945 0.5943 0.5973 mean 0.2560 0.3185 0.5929 0.5847<br />
Emp Std 0.1287 0.1683 0.0304 0.0000 Emp Std 0.1794 0.2137 0.0446 0.0883<br />
Table 4: Mean values and standard <strong>de</strong>vi<strong>at</strong>ions of the Maximum Likelihood estim<strong>at</strong>es of the param<strong>et</strong>er ξ for<br />
the distribution of maximum (cf. equ<strong>at</strong>ion 4) when d<strong>at</strong>a are clustered in samples of size 10,50,100 and 200<br />
and for the Generalized Par<strong>et</strong>o Distribution (7) for thresholds u corresponding to quantiles 90%,95%,99%<br />
ans 99.5%. In panel (a), are presented the results for the Dow Jones, in panel (b) for the Nasdaq for raw d<strong>at</strong>a<br />
and in panel (c) the Nasdaq corrected for the “lunch effect”.<br />
47<br />
111
112 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
(a) Dow Jones<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
quantile 0.995 0.99 0.95 0.90 quantile 0.995 0.99 0.95 0.90<br />
N/k =4 N/k =4<br />
mean 0.2491 -0.0366 0.2721 0.2268 mean 0.6427 -0.3035 0.4006 0.2338<br />
emp. Std 0.3561 0.2382 0.0974 0.0639 emp. Std 0.3913 0.2233 0.0902 0.0695<br />
th. Std 0.5274 0.3590 0.1674 0.1175 th. Std 0.5663 0.3510 0.1710 0.1177<br />
N/k =10 N/k=10<br />
mean 0.4449 -0.2539 0.0279 0.2905 mean -0.2228 0.8057 0.3622 0.3107<br />
emp. Std 0.5354 0.5563 0.1314 0.0943 emp. Std 1.3601 0.5420 0.1301 0.1227<br />
th. Std 0.8619 0.5568 0.2558 0.1877 th. Std 0.7891 0.6552 0.2686 0.1883<br />
(b) Nasdaq (Raw d<strong>at</strong>a)<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
quantile 0.995 0.99 0.95 0.90 quantile 0.995 0.99 0.95 0.90<br />
N/k =4 N/k=4<br />
mean 0.3189 0.0006 0.1055 0.0284 mean 1.0709 0.2538 0.1090 0.0386<br />
emp. Std 0.2489 0.1295 0.0649 0.043 emp. Std 0.3461 0.2190 0.0626 0.0384<br />
th. Std 0.5333 0.3605 0.1634 0.1144 th. Std 0.6222 0.3732 0.1634 0.1145<br />
N/k =10 N/k =10<br />
mean 0.2034 -0.1098 0.1634 0.2445 mean -0.8666 1.1228 0.1601 0.2846<br />
emp. Std 0.6173 0.4596 0.0871 0.0571 emp. Std 0.6405 0.3499 0.1157 0.0762<br />
th. Std 0.8281 0.5634 0.2604 0.1863 th. Std 0.7835 0.7042 0.2602 0.1875<br />
(c) Nasdaq (Corrected d<strong>at</strong>a)<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
quantile 0.995 0.99 0.95 0.90 quantile 0.995 0.99 0.95 0.90<br />
N/k =4 N/k=4<br />
mean 0.1910 0.3435 0.0616 0.2070 mean 1.4051 -0.1313 0.0255 0.2259<br />
emp. Std 0.2238 0.1441 0.0487 0.0333 emp. Std 0.2299 0.1424 0.0739 0.0390<br />
th. Std 0.5228 0.3787 0.1624 0.1172 th. Std 0.6742 0.3556 0.1617 0.1175<br />
N/k =10 N/k =10<br />
mean 0.0630 0.0543 0.4537 -0.0699 mean -0.3357 1.3961 0.4034 0.0770<br />
emp. Std 0.4899 0.1004 0.0780 0.0614 emp. Std 0.6830 0.2481 0.0891 0.0643<br />
th. Std 0.5137 0.3629 0.1727 0.1150 th. Std 0.7835 0.7521 0.2705 0.1820<br />
Table 5: Pickands estim<strong>at</strong>es (9) of the param<strong>et</strong>er ξ for the Generalized Par<strong>et</strong>o Distribution (7) for thresholds u<br />
corresponding to quantiles 90%,95%,99% ans 99.5% and two different values of the r<strong>at</strong>io N/k respectiveley<br />
equal to 4 and 10. In panel (a), are presented the results for the Dow Jones, in panel (b) for the Nasdaq for<br />
raw d<strong>at</strong>a and in panel (c) the Nasdaq corrected for the “lunch effect”.<br />
48
Nasdaq Dow Jones<br />
Pos. Tail Neg. Tail Pos. Tail Neg. Tail<br />
q 10 3 u nu 10 3 u nu 10 2 u nu 10 2 u nu<br />
q1=0 0.0053 11241 0.0053 10751 0.0032 14949 0.0028 13464<br />
q2=0.1 0.0573 10117 0.0571 9676 0.0976 13454 0.0862 12118<br />
q3=0.2 0.1124 8993 0.1129 8601 0.1833 11959 0.1739 10771<br />
q4=0.3 0.1729 7869 0.1723 7526 0.2783 10464 0.263 9425<br />
q5=0.4 0.238 6745 0.2365 6451 0.3872 8969 0.3697 8078<br />
q6=0.5 0.3157 5620 0.3147 5376 0.5055 7475 0.4963 6732<br />
q7=0.6 0.406 4496 0.412 4300 0.6426 5980 0.6492 5386<br />
q8=0.7 0.5211 3372 0.5374 3225 0.8225 4485 0.8376 4039<br />
q9=0.8 0.6901 2248 0.7188 2150 1.0545 2990 1.1057 2693<br />
q10=0.9 0.973 1124 1.0494 1075 1.4919 1495 1.6223 1346<br />
q11=0.925 1.1016 843 1.1833 806 1.6956 1121 1.8637 1010<br />
q12=0.95 1.2926 562 1.3888 538 1.9846 747 2.2285 673<br />
q13=0.96 1.3859 450 1.4955 430 2.1734 598 2.4197 539<br />
q14=0.97 1.53 337 1.639 323 2.413 448 2.7218 404<br />
q15=0.98 1.713 225 1.8557 215 2.7949 299 3.1647 269<br />
q16=0.99 2.1188 112 1.8855 108 3.5704 149 4.1025 135<br />
q17=0.9925 2.3176 84 2.4451 81 3.9701 112 4.3781 101<br />
q18=0.995 3.0508 56 2.7623 54 4.5746 75 5.0944 67<br />
Table 6: Significance levels qk and their corresponding lower thresholds uk for the four different samples.<br />
The number nu provi<strong>de</strong>s the size of the sub-sample beyond the threshold uk.<br />
49<br />
113
114 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Mean AD-st<strong>at</strong>istic for u1 – u9<br />
N-pos N-neg DJ-pos DJ-neg<br />
Weibull 1.86 (89.34%) 1.34 (79.18%) 5.43 (99.82%) 3.85 (98.91%)<br />
Gen. Par<strong>et</strong>o 4.45 (99.45%) 2.68 (96.01%) 12.47 (99.996%) 6.44 (99,99%)<br />
Gamma 3.59 (98.59%) 2.82 (96.62%) 8.76 (99.996%) 7.23 (99.996%)<br />
Exponential 3.64 (98.66%) 2.76 (96.36%) 13.96 (99.996%) 10.10 (99.996%)<br />
Par<strong>et</strong>o 475.2 (99.996%) 441.7 (99.996%) 691.7 (99.996%) 607.9 (99.996%)<br />
Mean AD-st<strong>at</strong>istic for u10 – u18<br />
Weibull 1.21 (74.29%) .988 (63.39%) .835 (53,52%) .849 (54.54%)<br />
Gen. Par<strong>et</strong>o 2.29 (93.57%) 1.88 (89.52%) 1.95 (90.28%) 1.36 (79.67%)<br />
Gamma 2.49 (95.00%) 1.90 (89.74%) 2.12 (92.01%) 1.63 (86.02%)<br />
Exponential 3.26 (97.97%) 1.93 (90.02%) 4.52 (99.10%) 2.70 (96.11%)<br />
Par<strong>et</strong>o 1.80 (88.60%) 1.77 (88.23%) 1.18 (73.01%) 1.65 (86.40%)<br />
Table 7: Mean An<strong>de</strong>rson-Darling distances in the range of thresholds u1-u9 and in the range u10-u18. The<br />
figures within parenthesis characterize the goodness of fit: they represent the significance levels with which<br />
the consi<strong>de</strong>red mo<strong>de</strong>l can be rejected.<br />
50
Nasdaq Dow Jones<br />
Pos. Tail Neg. Tail Pos. Tail Neg. Tail<br />
MLE ADE MLE ADE MLE ADE MLE ADE<br />
1 0.256 (0.002) 0.192 0.254 (0.002) 0.191 0.204 (0.002) 0.150 0.199 (0.002) 0.147<br />
2 0.555 (0.006) 0.443 0.548 (0.006) 0.439 0.576 (0.005) 0.461 0.538 (0.005) 0.431<br />
3 0.765 (0.008) 0.630 0.755 (0.008) 0.625 0.782 (0.007) 0.644 0.745 (0.007) 0.617<br />
4 0.970 (0.011) 0.819 0.945 (0.011) 0.800 0.989 (0.010) 0.833 0.920 (0.009) 0.777<br />
5 1.169 (0.014) 1.004 1.122 (0.014) 0.965 1.219 (0.013) 1.053 1.114 (0.012) 0.960<br />
6 1.400 (0.019) 1.227 1.325 (0.018) 1.157 1.447 (0.017) 1.279 1.327 (0.016) 1.169<br />
7 1.639 (0.024) 1.460 1.562 (0.024) 1.386 1.685 (0.022) 1.519 1.563 (0.021) 1.408<br />
8 1.916 (0.033) 1.733 1.838 (0.032) 1.655 1.984 (0.030) 1.840 1.804 (0.028) 1.659<br />
9 2.308 (0.049) 2.145 2.195 (0.047) 1.999 2.240 (0.041) 2.115 2.060 (0.040) 1.921<br />
10 2.759 (0.082) 2.613 2.824 (0.086) 2.651 2.575 (0.067) 2.474 2.436 (0.066) 2.315<br />
11 2.955 (0.102) 2.839 3.008 (0.106) 2.836 2.715 (0.081) 2.648 2.581 (0.081) 2.467<br />
12 3.232 (0.136) 3.210 3.352 (0.145) 3.259 2.787 (0.102) 2.707 2.765 (0.107) 2.655<br />
13 3.231 (0.152) 3.193 3.441 (0.166) 3.352 2.877 (0.118) 2.808 2.782 (0.120) 2.642<br />
14 3.358 (0.183) 3.390 3.551 (0.198) 3.479 2.920 (0.138) 2.841 2.903 (0.144) 2.740<br />
15 3.281 (0.219) 3.306 3.728 (0.254) 3.730 2.989 (0.173) 2.871 3.059 (0.186) 2.870<br />
16 3.327 (0.313) 3.472 3.990 (0.384) 3.983 3.226 (0.263) 3.114 3.690 (0.318) 3.668<br />
17 3.372 (0.366) 3.636 3.917 (0.435) 3.860 3.427 (0.322) 3.351 3.518 (0.350) 3.397<br />
18 3.136 (0.415) 3.326 4.251 (0.578) 4.302 3.818 (0.441) 3.989 4.168 (0.506) 4.395<br />
Table 8: Maximum Likelihood and An<strong>de</strong>rson-Darling estim<strong>at</strong>es of the Par<strong>et</strong>o param<strong>et</strong>er b. Figures within<br />
parentheses give the standard <strong>de</strong>vi<strong>at</strong>ion of the Maximum Likelihood estim<strong>at</strong>or.<br />
51<br />
115
116 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Nasdaq Dow Jones<br />
Pos. Tail Neg. Tail Pos. Tail Neg. Tail<br />
MLE ADE MLE ADE MLE ADE MLE ADE<br />
1 1.007 (0.008) 1.053 0.987 (0.008) 1.017 1.040 (0.007) 1.104 0.975 (0.007) 1.026<br />
2 0.983 (0.011) 1.051 0.953 (0.011) 0.993 0.973 (0.010) 1.075 0.910 (0.010) 0.989<br />
3 0.944 (0.014) 1.031 0.912 (0.014) 0.955 0.931 (0.013) 1.064 0.856 (0.012) 0.948<br />
4 0.896 (0.018) 0.995 0.876 (0.018) 0.916 0.878 (0.015) 1.038 0.821 (0.015) 0.933<br />
5 0.857 (0.021) 0.978 0.861 (0.021) 0.912 0.792 (0.019) 0.955 0.767 (0.018) 0.889<br />
6 0.790 (0.026) 0.916 0.833 (0.026) 0.891 0.708 (0.023) 0.873 0.698 (0.022) 0.819<br />
7 0.732 (0.033) 0.882 0.796 (0.033) 0.859 0.622 (0.028) 0.788 0.612 (0.028) 0.713<br />
8 0.661 (0.042) 0.846 0.756 (0.042) 0.834 0.480 (0.035) 0.586 0.531 (0.035) 0.597<br />
9 0.509 (0.058) 0.676 0.715 (0.059) 0.865 0.394 (0.047) 0.461 0.478 (0.047) 0.527<br />
10 0.359 (0.092) 0.631 0.522 (0.099) 0.688 0.304 (0.074) 0.346 0.403 (0.076) 0.387<br />
11 0.252 (0.110) 0.515 0.481 (0.120) 0.697 0.231 (0.087) 0.158 0.379 (0.091) 0.337<br />
12 0.039 (0.138) 0.177 0.273 (0.155) 0.275 0.269 (0.111) 0.207 0.357 (0.119) 0.288<br />
13 0.057 (0.155) 0.233 0.255 (0.177) 0.274 0.247 (0.127) 0.147 0.428 (0.136) 0.465<br />
14 0 0 0.215 (0.209) 0.194 0.283 (0.150) 0.174 0.448 (0.164) 0.641<br />
15 0 0 0.091 (0.260) 0 0.374 (0.192) 0.407 0.451 (0.210) 0.863<br />
16 0 0 0.064 (0.390) 0 0.372 (0.290) 0.382 0.022 (0.319) 0.110<br />
17 0 0 0.158 (0.452) 0.224 0.281 (0.346) 0.255 0.178 (0.367) 0.703<br />
18 0 0 0 0 0 0 0 0<br />
Table 9: Maximum Likelihood and An<strong>de</strong>rson-Darling estim<strong>at</strong>es of the form param<strong>et</strong>er c of the Weibull<br />
(Str<strong>et</strong>ched-Exponential) distribution.<br />
52
Nasdaq Dow Jones<br />
Pos. Tail Neg. Tail Pos. Tail Neg. Tail<br />
MLE ADE MLE ADE MLE ADE MLE ADE<br />
1 0.443 (0.004) 0.441 0.455 (0.005) 0.452 7.137 (0.060) 7.107 7.268 (0.068) 7.127<br />
2 0.429 (0.006) 0.440 0.436 (0.006) 0.443 6.639 (0.082) 6.894 6.726 (0.094) 6.952<br />
3 0.406 (0.008) 0.432 0.410 (0.009) 0.424 6.236 (0.113) 6.841 6.108 (0.131) 6.640<br />
4 0.372 (0.011) 0.414 0.383 (0.012) 0.402 5.621 (0.155) 6.655 5.656 (0.175) 6.515<br />
5 0.341 (0.015) 0.404 0.369 (0.016) 0.399 4.515 (0.215) 5.942 4.876 (0.235) 6.066<br />
6 0.283 (0.020) 0.364 0.345 (0.021) 0.383 3.358 (0.277) 5.081 3.801 (0.305) 5.220<br />
7 0.231 (0.026) 0.339 0.309 (0.028) 0.358 2.192 (0.326) 4.073 2.475 (0.366) 3.764<br />
8 0.166 (0.034) 0.311 0.269 (0.039) 0.336 0.682 (0.256) 1.606 1.385 (0.389) 2.149<br />
9 0.053 (0.030) 0.164 0.225 (0.057) 0.365 0.195 (0.163) 0.510 0.810 (0.417) 1.297<br />
10 0.005 (0.010) 0.128 0.058 (0.057) 0.184 0.019 (0.048) 0.065 0.276 (0.361) 0.207<br />
11 0.000 (0.001) 0.049 0.036 (0.053) 0.194 0.001 (0.003) 0.000 0.169 (0.316) 0.065<br />
12 0.000 (0.000) 0.000 0.000 (0.001) 0.000 0.005 (0.025) 0.000 0.103 (0.291) 0.012<br />
13 0.000 (0.000) 0.000 0.000 (0.001) 0.000 0.001 (0.010) 0.000 0.427 (0.912) 0.729<br />
14 - - 0.000 (0.000) 0.000 0.009 (0.055) 0.000 0.577 (1.357) 3.509<br />
15 - - 0.000 (0.000) - 0.149 (0.629) 0.282 0.613 (1.855) 9.640<br />
16 - - 0.000 (0.000) - 0.145 (0.960) 0.179 0.000 (0.000) 0.000<br />
17 - - 0.000 (0.000) 0.000 0.007 (0.109) 0.002 0.000 (0.000) 5.528<br />
18 - - - - - - - -<br />
Table 10: Maximum Likelihood and An<strong>de</strong>rson-Darling estim<strong>at</strong>es of the form param<strong>et</strong>er d(×10 3 ) of the<br />
Weibull (Str<strong>et</strong>ched-Exponential) distribution.<br />
53<br />
117
118 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Nasdaq Dow Jones<br />
Pos. Tail Neg. Tail Pos. Tail Neg. Tail<br />
MLE ADE MLE ADE MLE ADE MLE ADE<br />
1 0.441 (0.004) 0.441 0.458 (0.004) 0.451 7.012 (0.057) 7.055 7.358 (0.063) 7.135<br />
2 0.435 (0.004) 0.431 0.454 (0.005) 0.444 6.793 (0.059) 6.701 7.292 (0.066) 6.982<br />
3 0.431 (0.005) 0.424 0.452 (0.005) 0.438 6.731 (0.062) 6.575 7.275 (0.070) 6.890<br />
4 0.428 (0.005) 0.416 0.453 (0.005) 0.437 6.675 (0.065) 6.444 7.358 (0.076) 6.938<br />
5 0.429 (0.005) 0.415 0.458 (0.006) 0.443 6.607 (0.070) 6.264 7.429 (0.083) 6.941<br />
6 0.429 (0.006) 0.411 0.464 (0.006) 0.447 6.630 (0.077) 6.186 7.529 (0.092) 6.951<br />
7 0.436 (0.006) 0.413 0.472 (0.007) 0.453 6.750 (0.087) 6.207 7.700 (0.105) 7.005<br />
8 0.447 (0.008) 0.421 0.483 (0.009) 0.463 6.920 (0.103) 6.199 8.071 (0.127) 7.264<br />
9 0.462 (0.010) 0.425 0.503 (0.011) 0.482 7.513 (0.137) 6.662 8.797 (0.170) 7.908<br />
10 0.517 (0.015) 0.468 0.529 (0.016) 0.496 8.792 (0.227) 7.745 10.205 (0.278) 9.175<br />
11 0.540 (0.019) 0.479 0.551 (0.019) 0.514 9.349 (0.279) 8.148 10.835 (0.341) 9.751<br />
12 0.574 (0.024) 0.489 0.570 (0.025) 0.516 10.487 (0.383) 9.265 11.796 (0.454) 10.657<br />
13 0.615 (0.029) 0.526 0.594 (0.029) 0.537 11.017 (0.451) 9.722 12.598 (0.543) 11.581<br />
14 0.653 (0.035) 0.543 0.627 (0.035) 0.564 11.920 (0.563) 10.626 13.349 (0.664) 12.386<br />
15 0.750 (0.050) 0.625 0.671 (0.046) 0.594 13.251 (0.766) 12.062 14.462 (0.880) 13.521<br />
16 0.917 (0.086) 0.741 0.760 (0.073) 0.674 15.264 (1.246) 13.943 15.294 (1.316) 13.285<br />
17 0.991 (0.107) 0.783 0.827 (0.092) 0.744 15.766 (1.483) 14.210 17.140 (1.705) 15.327<br />
18 1.178 (0.156) 0.978 0.857 (0.117) 0.742 16.207 (1.871) 13.697 16.883 (2.047) 13.476<br />
Table 11: Maximum Likelihood- and An<strong>de</strong>rson-Darling estim<strong>at</strong>es of the scale param<strong>et</strong>er d = 10 −3 d ′ of the<br />
Exponential distribution.Figures within parentheses give the standard <strong>de</strong>vi<strong>at</strong>ion of the Maximum Likelihood<br />
estim<strong>at</strong>or.<br />
54
Nasdaq Dow Jones<br />
Pos. Tail Neg. Tail Pos. Tail Neg. Tail<br />
MLE ADE MLE ADE MLE ADE MLE ADE<br />
MLE ADE MLE ADE MLE ADE MLE ADE<br />
1 -1.03 -1.09 -1.00 -1.03 -1.12 -1.18 -0.100 -1.05<br />
2 -1.02 -1.13 -0.934 -1.01 -1.01 -1.19 -0.862 -1.01<br />
3 -0.931 -1.13 -0.821 -0.955 -0.921 -1.23 -0.710 -0.943<br />
4 -0.787 -1.09 -0.701 -0.887 -0.766 -1.24 -0.594 -0.944<br />
5 -0.655 -1.12 -0.636 -0.914 -0.458 -1.09 -0.397 -0.870<br />
6 -0.395 -1.01 -0.518 -0.911 -0.119 -0.929 -0.118 -0.715<br />
7 -0.142 -1.03 -0.351 -0.906 0.261 -0.763 0.251 -0.462<br />
8 0.206 -1.09 -0.149 -0.97 0.881 -0.202 0.619 -0.160<br />
9 0.971 -0.754 0.101 -1.35 1.31 0.127 0.930 -0.018<br />
10 1.83 -1.04 1.17 -1.33 1.82 0.408 1.40 0.435<br />
11 2.34 -0.441 1.45 -1.53 2.10 0.949 1.59 0.420<br />
12 3.12 -0.445 2.52 -0.435 2.04 0.733 1.78 0.403<br />
13 3.10 -0.444 2.63 -0.402 2.16 0.886 1.57 -0.375<br />
14 3.35 1.43 2.89 -0.419 2.07 0.786 1.58 -0.425<br />
15 3.27 1.57 3.36 1.35 1.82 -0.282 1.64 -2.75<br />
16 3.30 2.97 3.80 -0.411 1.88 -0.129 3.60 -0.428<br />
17 3.34 3.19 3.46 -0.412 2.35 -0.317 3.19 -0.433<br />
18 2.74 2.90 4.22 -0.408 3.73 3.27 4.11 0.374<br />
Table 12: Maximum Likelihood- and An<strong>de</strong>rson-Darling estim<strong>at</strong>es of the form param<strong>et</strong>er b of the Incompl<strong>et</strong>e<br />
Gamma distribution.<br />
55<br />
119
120 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Dow Jones Positive Tail Dow Jones Neg<strong>at</strong>ive Tail<br />
ˆb b † Wald Test Score Test ˆb b † Wald Test Score Test<br />
1 0.204 0.205 79.009% 78.793% 0.199 0.200 77.394% 77.166%<br />
2 0.576 0.580 52.580% 52.300% 0.538 0.541 45.609% 45.385%<br />
3 0.782 0.787 40.708% 40.474% 0.745 0.749 28.921% 28.792%<br />
4 0.989 0.995 29.416% 29.251% 0.920 0.924 21.447% 21.354%<br />
5 1.219 1.224 15.222% 15.159% 1.114 1.118 13.074% 13.027%<br />
6 1.447 1.452 7.202% 7.181% 1.327 1.330 6.456% 6.439%<br />
7 1.685 1.689 2.984% 2.978% 1.563 1.565 2.224% 2.221%<br />
8 1.984 1.986 0.449% 0.448% 1.804 1.805 0.559% 0.558%<br />
9 2.240 2.241 0.078% 0.078% 2.060 2.060 0.179% 0.179%<br />
10 2.575 2.575 0.005% 0.005% 2.436 2.437 0.020% 0.020%<br />
11 2.715 2.715 0.004% 0.004% 2.581 2.582 0.010% 0.010%<br />
12 2.787 2.787 0.004% 0.004% 2.765 2.765 0.009% 0.009%<br />
13 2.877 2.877 0.004% 0.004% 2.782 2.783 0.048% 0.048%<br />
14 2.920 2.920 0.005% 0.005% 2.903 2.905 0.082% 0.082%<br />
15 2.989 2.989 0.002% 0.002% 3.059 3.063 0.099% 0.099%<br />
16 3.226 3.226 0.001% 0.001% 3.690 3.690 0.000% 0.000%<br />
17 3.427 3.427 0.001% 0.001% 3.518 3.519 0.002% 0.002%<br />
18 3.818 - - - 4.168 - - -<br />
Nasdaq Positive Tail Nasdaq Neg<strong>at</strong>ive Tail<br />
ˆb b † Wald Test Score Test ˆb b † Wald Test Score Test<br />
1 0.256 0.256 38.328% 38.238% 0.254 0.255 39.906% 39.812%<br />
2 0.555 0.557 38.525% 38.352% 0.548 0.550 26.450% 26.369%<br />
3 0.765 0.769 27.629% 27.505% 0.755 0.757 14.636% 14.601%<br />
4 0.970 0.974 17.914% 17.844% 0.945 0.946 7.777% 7.763%<br />
5 1.169 1.174 13.144% 13.094% 1.122 1.124 6.269% 6.258%<br />
6 1.400 1.404 7.080% 7.060% 1.325 1.327 4.096% 4.090%<br />
7 1.639 1.643 4.221% 4.211% 1.562 1.564 2.468% 2.464%<br />
8 1.916 1.920 2.167% 2.162% 1.838 1.840 1.481% 1.480%<br />
9 2.308 2.311 0.420% 0.419% 2.195 2.199 1.061% 1.059%<br />
10 2.759 2.761 0.071% 0.071% 2.824 2.826 0.137% 0.136%<br />
11 2.955 2.956 0.013% 0.013% 3.008 3.010 0.089% 0.089%<br />
12 3.232 3.232 0.000% 0.000% 3.352 3.353 0.002% 0.002%<br />
13 3.231 3.231 0.000% 0.000% 3.441 3.441 0.001% 0.001%<br />
14 3.358 - - - 3.551 3.551 0.000% 0.000%<br />
15 3.281 - - - 3.728 3.728 0.000% 0.000%<br />
16 3.327 - - - 3.990 3.990 0.000% 0.000%<br />
17 3.372 - - - 3.917 3.917 0.000% 0.000%<br />
18 3.136 - - - 4.251 - - -<br />
Table 13: Wald and Score encompassing Test for non-nested hypotheses. The p-value gives the significance<br />
with which one can reject the null hypothesis: (SE) encompasses (PD). b † is the estim<strong>at</strong>or of the Par<strong>et</strong>o<br />
mo<strong>de</strong>l un<strong>de</strong>r the hypothesis for the SE distribution. ˆb is the estim<strong>at</strong>or un<strong>de</strong>r the true distribution for the<br />
Par<strong>et</strong>o mo<strong>de</strong>l.<br />
56
Dow Jones positive tail Dow Jones neg<strong>at</strong>ive tail<br />
ˆηT ˆb p-value ˆηT ˆb p-value<br />
1 1.044 0.204 100.000% 0.979 0.199 100.000%<br />
2 1.123 0.576 100.000% 1.050 0.538 100.000%<br />
3 1.229 0.782 100.000% 1.148 0.745 100.000%<br />
4 1.351 0.989 100.000% 1.259 0.920 100.000%<br />
5 1.493 1.219 100.000% 1.388 1.114 100.000%<br />
6 1.653 1.447 100.000% 1.538 1.327 100.000%<br />
7 1.835 1.685 100.000% 1.715 1.563 100.000%<br />
8 2.069 1.984 100.000% 1.912 1.804 100.000%<br />
9 2.294 2.240 100.000% 2.141 2.060 100.000%<br />
10 2.605 2.575 99.997% 2.489 2.436 100.000%<br />
11 2.732 2.715 99.165% 2.626 2.581 99.997%<br />
12 2.809 2.787 98.487% 2.803 2.765 99.766%<br />
13 2.895 2.877 94.723% 2.835 2.782 99.869%<br />
14 2.943 2.920 94.010% 2.960 2.903 99.514%<br />
15 3.027 2.989 95.032% 3.116 3.059 97.480%<br />
16 3.261 3.226 80.157% 3.690 3.690 5.634%<br />
17 3.447 3.427 58.002% 3.527 3.518 38.929%<br />
18 3.818 3.818 - 4.168 4.168 -<br />
Nasdaq positive tail Nasdaq neg<strong>at</strong>ive tail<br />
ˆηT ˆb p-value ˆηT ˆb p-value<br />
1 1.019 0.256 100.000% 1.000 0.254 100.000%<br />
2 1.118 0.555 100.000% 1.090 0.548 100.000%<br />
3 1.225 0.765 100.000% 1.194 0.755 100.000%<br />
4 1.347 0.970 100.000% 1.312 0.945 100.000%<br />
5 1.486 1.169 100.000% 1.447 1.122 100.000%<br />
6 1.650 1.400 100.000% 1.605 1.325 100.000%<br />
7 1.840 1.639 100.000% 1.796 1.562 100.000%<br />
8 2.069 1.916 100.000% 2.032 1.838 100.000%<br />
9 2.393 2.308 100.000% 2.353 2.195 100.000%<br />
10 2.799 2.759 99.994% 2.900 2.824 100.000%<br />
11 2.974 2.955 98.132% 3.070 3.008 99.996%<br />
12 3.232 3.232 22.570% 3.372 3.352 92.329%<br />
13 3.232 3.231 28.945% 3.457 3.441 85.305%<br />
14 3.358 3.358 - 3.563 3.551 69.909%<br />
15 3.281 3.281 - 3.730 3.728 27.179%<br />
16 3.327 3.327 - 3.991 3.990 13.117%<br />
17 3.372 3.372 - 3.923 3.917 27.655%<br />
18 3.136 3.136 - 4.251 4.251 -<br />
Table 14: Test of the (SE) mo<strong>de</strong>l against the null hypothesis th<strong>at</strong> the true mo<strong>de</strong>l is the Par<strong>et</strong>o mo<strong>de</strong>l. The<br />
p-value gives the significance with which one can reject the null hypothesis.<br />
57<br />
121
122 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
mean absolute r<strong>et</strong>urn |r|<br />
x 10−4<br />
10<br />
9<br />
8<br />
7<br />
6<br />
5<br />
4<br />
3<br />
2<br />
0 10 20 30 40<br />
time (by five minutes)<br />
50 60 70 80<br />
Figure 1: Average absolute r<strong>et</strong>urn, as a function of time within a trading day. The U-shape characterizes the<br />
so-called lunch effect.<br />
58
Vari<strong>at</strong>ion coefficient<br />
Vari<strong>at</strong>ion coefficient<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
Vari<strong>at</strong>ion coeff.=std/mean of time intervals b<strong>et</strong>ween pos. extremums of DJ<br />
0<br />
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08<br />
Threshold u, (for extremums X > u)<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
Vari<strong>at</strong>ion coeff.=std/mean of time intervals b<strong>et</strong>ween neg. extremums, DJ<br />
0<br />
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08<br />
Threshold, u (for extremums X < −u)<br />
Figure 2: Coefficient of vari<strong>at</strong>ion V for the Dow Jones daily r<strong>et</strong>urns. An increase of V characterizes the<br />
increase of “clustering”.<br />
59<br />
123
124 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
MLE of ξ<br />
1.8<br />
1.6<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
MLE of GPD form param<strong>et</strong>er ξ from SE samples, n=50000<br />
c=0.7<br />
c=0.3<br />
−0.2<br />
0 2 4 6 8 10 12 14 16<br />
Number of lower threshold h<br />
Figure 3: Maximum Likelihood estim<strong>at</strong>es of the GPD form param<strong>et</strong>er for Str<strong>et</strong>ched-Exponentail samples of<br />
size 50,000.<br />
60
Mean excess function<br />
Mean excess function<br />
0.06<br />
0.05<br />
0.04<br />
0.03<br />
0.02<br />
0.01<br />
0.5<br />
Mean excess functions for DJ−daily pos.(line) and neg.(pointwise)<br />
0<br />
0 0.05 0.1 0.15<br />
Lower threshold, u<br />
x 10−3<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
Mean excess functions for ND−5min pos.(line) and neg.(pointwise)<br />
0<br />
0 0.001 0.002 0.003 0.004 0.005 0.006 0.007 0.008 0.009 0.01<br />
Lower threshold, u<br />
Figure 4: Mean excess functions for the Dow Jones daily r<strong>et</strong>urns (upper panel) and the Nasdaq five minutes<br />
r<strong>et</strong>urns (lower panel). The plain line represents the positive r<strong>et</strong>urns and the dotted line the neg<strong>at</strong>ive ones<br />
61<br />
125
126 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Complementary sample DF<br />
Complementary sample DF<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −3<br />
10 −4<br />
10 −5<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −3<br />
10 −4<br />
10 −5<br />
10 −5<br />
10 −6<br />
Complementary DF of DJ−daily pos.(line),n=14949 and neg.(pointwise),n=13464<br />
10 −4<br />
10 −5<br />
10 −3<br />
10 −2<br />
Absolute log−r<strong>et</strong>urn, x<br />
10 −4<br />
10 −3<br />
10 −1<br />
Complementary DF of ND−5min pos.(line),n=11241 and neg.(pointwise) n=10751<br />
Absolute log−r<strong>et</strong>urn, x<br />
Figure 5: Cumul<strong>at</strong>ive sample distributions for the Dow Jones (a) and for the Nasdaq (b) d<strong>at</strong>a s<strong>et</strong>s.<br />
62<br />
10 0<br />
10 −2
Hill‘s estim<strong>at</strong>e b u<br />
Hill‘s estim<strong>at</strong>e b u<br />
4.5<br />
3.5<br />
2.5<br />
1.5<br />
1<br />
0.5<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
5<br />
4<br />
3<br />
2<br />
0<br />
10 −5<br />
10 −5<br />
Hill‘s estim<strong>at</strong>es of b u for DJ−daily pos.(line), n=14949, and neg.(pointwise),n=13464<br />
10 −4<br />
10 −3<br />
Lower threshold, u<br />
Hill‘s estim<strong>at</strong>es of b u for ND−5min pos.(line),n=11241, and neg.(pointwise),n=10751<br />
10 −4<br />
Lower threshold, u<br />
Figure 6: Hill estim<strong>at</strong>es ˆbu as a function of the threshold u for the Dow Jones (a) and for the Nasdaq (b).<br />
63<br />
10 −2<br />
10 −3<br />
127
128 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
b<br />
5<br />
4<br />
3<br />
2<br />
1<br />
ND<<br />
ND><br />
DJ><br />
DJ<<br />
0<br />
0 5 10 15 20<br />
In<strong>de</strong>x n of the quantile q n<br />
Figure 7: Hill estim<strong>at</strong>or ˆbu for all four d<strong>at</strong>a s<strong>et</strong>s (positive and neg<strong>at</strong>ive branches of the distribution of r<strong>et</strong>urns<br />
for the DJ and for the ND) as a function of the in<strong>de</strong>x n = 1,...,18 of the 18 quantiles or standard significance<br />
levels q1 ...q18 given in table 6. The dashed line is expression (33) with 1 − qn = 3.08 e −0.342n given by<br />
(32).<br />
64
Wilks st<strong>at</strong>istic (doubled log−likelihood r<strong>at</strong>io)<br />
Wilks st<strong>at</strong>istic (doubled log−likelihood r<strong>at</strong>io)<br />
100<br />
90<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
SE<br />
IG<br />
Wilks st<strong>at</strong>istics for CD vs 4 param<strong>et</strong>ric families, Npos, n=11241<br />
PD<br />
0 0.5 1 1.5 2 2.5 3<br />
x 10 −3<br />
0<br />
Lower threshold, u<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
SE<br />
ED<br />
2<br />
χ (.95)<br />
2<br />
Wilks st<strong>at</strong>istics for CD vs 4 param<strong>et</strong>ric families, Nneg, n=10751<br />
ED<br />
IG<br />
PD<br />
2<br />
χ (.95)<br />
1<br />
0 0.5 1 1.5 2 2.5 3<br />
x 10 −3<br />
0<br />
Lower threshold, u<br />
Figure 8: Wilks st<strong>at</strong>istic for the comprehensive distribution versus the four param<strong>et</strong>ric distributions : Par<strong>et</strong>o<br />
(PD), Weibull (SE), Exponential (ED) and Incompl<strong>et</strong>e Gamma (IG) for the Nasdaq five minutes r<strong>et</strong>urns.<br />
The upper panel refers to the positive r<strong>et</strong>urns and lower panel to the neg<strong>at</strong>ive ones.<br />
65<br />
2<br />
χ (.95)<br />
2<br />
2<br />
χ (.95)<br />
1<br />
129
130 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Wilks st<strong>at</strong>istic (doubled log−likelihood r<strong>at</strong>io)<br />
Wilks st<strong>at</strong>istic (doubled log−likelihood r<strong>at</strong>io)<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
SE<br />
Wilks st<strong>at</strong>istics for CD vs 4 param<strong>et</strong>ric families, DJpos, n=14949<br />
IG<br />
PD<br />
0<br />
0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0.05<br />
Lower threshold, u<br />
80<br />
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
SE<br />
IG<br />
ED<br />
2<br />
χ (.95)<br />
2<br />
Wilks st<strong>at</strong>istics for CD vs 4 param<strong>et</strong>ric families, DJneg, n=13464<br />
PD<br />
ED<br />
2<br />
χ (.95)<br />
2<br />
2<br />
χ (.95)<br />
1<br />
2<br />
χ (.95)<br />
1<br />
0<br />
0 0.01 0.02 0.03<br />
Lower threshold, u<br />
0.04 0.05 0.06<br />
Figure 9: Wilks st<strong>at</strong>istic for the comprehensive distribution versus the four param<strong>et</strong>ric distributions : Par<strong>et</strong>o<br />
(PD), Weibull (SE), Exponential (ED) and Incompl<strong>et</strong>e Gamma (IG) for the Dow Jones daily r<strong>et</strong>urns. The<br />
upper panel refers to the positive r<strong>et</strong>urns and the lower panel to the neg<strong>at</strong>ive ones.<br />
66
Tail 1−F(x) and param<strong>et</strong>er b<br />
Tail 1−F(x) and param<strong>et</strong>er b<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Sample tail 1−F(x)(thick) and "local" Par<strong>et</strong>o−b(thin), simul<strong>at</strong>ed Par<strong>et</strong>o−1.2,n=15000<br />
10 20 30 40 50<br />
x<br />
60 70 80 90 100<br />
Sample tail 1−F(x)(thick) and "local" Par<strong>et</strong>o−b(thin), simul<strong>at</strong>ed SE−0.3,n=15000<br />
100 200 300 400 500<br />
x<br />
600 700 800 900 1000<br />
Figure 10: Local in<strong>de</strong>x β(x) estim<strong>at</strong>ed for a Par<strong>et</strong>o distribution with tail in<strong>de</strong>x = 1.2 (upper panel) and a<br />
Str<strong>et</strong>ched exponential with exponent c = 0.3 (lower panel).<br />
67<br />
131
132 3. Distributions exponentielles étirées contre distributions régulièrement variables<br />
Tail and param<strong>et</strong>er b<br />
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Sample tail 1−F(x)(tick) and Par<strong>et</strong>o−b(thin), simul<strong>at</strong>ed crossover<br />
Simul<strong>at</strong>ed 2 Par<strong>et</strong>o sample<br />
b 1 =0.70; b 2 =1.5;<br />
Lower threshold u 0 =1;<br />
Crossover point u 1 =10;<br />
PX>10≅0.1;<br />
( )<br />
20 40 60 80 100<br />
x<br />
120 140 160 180 200<br />
Figure 11: Local in<strong>de</strong>x β(x) for a distribution constructed by joining two Par<strong>et</strong>o distributions with exponents<br />
b1 = 0.70 and b2 = 1.5 <strong>at</strong> the cross-over point u1 = 10.<br />
68
1.4<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0 1 2 3 4 5 6 7<br />
x 10 −3<br />
0<br />
log−r<strong>et</strong>urn x<br />
10 1<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −5<br />
10 1<br />
10 0<br />
10 −1<br />
10 −3<br />
y ∼ x 0.54<br />
10 −4<br />
10 −2<br />
log−r<strong>et</strong>urn x<br />
y ∼ x 0.77<br />
Figure 12: Upper panel: Sample tail (continuous line), local in<strong>de</strong>x β(x) (dashed line) and local exponent c(x)<br />
(dash-dotted line) for the neg<strong>at</strong>ive tail of the Nasdaq five minutes r<strong>et</strong>urns. Lower panel: doubled logarithmic<br />
plot of the local in<strong>de</strong>x β(x). Over most of the d<strong>at</strong>a s<strong>et</strong>, β(x) increases with the log-r<strong>et</strong>urn x as a power law<br />
of in<strong>de</strong>x 0.77 while beyond the quantile 99% (see the ins<strong>et</strong>) it behaves like another power law of a smaller<br />
in<strong>de</strong>x equal to 0.54. The goodness of fit with these two power laws has been checked by a χ 2 test.<br />
69<br />
10 −3<br />
10 −2<br />
133
134 3. Distributions exponentielles étirées contre distributions régulièrement variables
Chapitre 4<br />
Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
L’un <strong>de</strong>s enjeux <strong>de</strong> la <strong>théorie</strong> financière est <strong>de</strong> comprendre comment le flux incessant d’inform<strong>at</strong>ion<br />
arrivant sur les marchés financiers est incorporé dans le prix <strong>de</strong>s actifs. Toutes les nouvelles n’ayant<br />
pas le même impact, la question se pose <strong>de</strong> savoir si l’on peut distinguer les eff<strong>et</strong>s <strong>de</strong>s <strong>at</strong>tent<strong>at</strong>s du 11<br />
septembre 2001 ou du coup d’ét<strong>at</strong> contre Gorb<strong>at</strong>chev le 19 aout 1991 du crash <strong>de</strong> 1987 ou d’autre chocs<br />
<strong>de</strong> vol<strong>at</strong>ilité <strong>de</strong> plus faible amplitu<strong>de</strong> ? Utilisant un processus autorégressif à mémoire longue défini sur<br />
le logarithme <strong>de</strong> la vol<strong>at</strong>ilité - le processus <strong>de</strong> marche alé<strong>at</strong>oire multifractale (MRW) - nous prédisons <strong>de</strong>s<br />
fonctions <strong>de</strong> réponse <strong>de</strong> la vol<strong>at</strong>ilité <strong>de</strong>s prix totalement différentes aux grands chocs externes comparées<br />
à celles que nous prédisons pour les chocs endogènes, i.e, qui résultent d’une accumul<strong>at</strong>ion constructive<br />
(ou cohérente) d’un grand nombre <strong>de</strong> p<strong>et</strong>its chocs.<br />
Ces prédictions sont remarquablement bien confirmées empiriquement sur divers chocs <strong>de</strong> vol<strong>at</strong>ilité<br />
d’amplitu<strong>de</strong>s variées. Notre <strong>théorie</strong> perm<strong>et</strong> <strong>de</strong> distinguer <strong>de</strong>ux types d’évèments (endogènes <strong>et</strong> exogènes)<br />
avec <strong>de</strong>s sign<strong>at</strong>ures spécifiques <strong>et</strong> <strong>de</strong>s précurseurs caractéristiques pour la classe <strong>de</strong>s chocs endogènes.<br />
Cela explique aussi l’origine <strong>de</strong> ce type <strong>de</strong> chocs par l’accumul<strong>at</strong>ion cohérente <strong>de</strong> p<strong>et</strong>ites mauvaises nouvelles,<br />
<strong>et</strong> ainsi perm<strong>et</strong> d’unifier les précé<strong>de</strong>ntes explic<strong>at</strong>ions <strong>de</strong>s grands krachs, incluant celui d’octobre<br />
1987.<br />
135
136 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité
Vol<strong>at</strong>ility Fingerprints of Large Shocks:<br />
Endogeneous Versus Exogeneous ∗<br />
D. Sorn<strong>et</strong>te 1,2 , Y. Malevergne 1,3 and J.-F. Muzy 4<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095, USA<br />
3 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x, France<br />
4 Labor<strong>at</strong>oire Systèmes Physiques <strong>de</strong> l’Environemment, CNRS UMR 6134<br />
Université <strong>de</strong> Corse, Quartier Gross<strong>et</strong>ti, 20250 Corte, France<br />
email: sorn<strong>et</strong>te@unice.fr, Yannick.Malevergne@unice.fr and muzy@univ-corse.fr<br />
fax: (310) 206 30 51<br />
Forthcoming Risk<br />
Abstract<br />
Finance is about how the continuous stream of news g<strong>et</strong>s incorpor<strong>at</strong>ed into prices. But not all news<br />
have the same impact. Can one distinguish the effects of the Sept. 11, 2001 <strong>at</strong>tack or of the coup<br />
against Gorbachev on Aug., 19, 1991 from financial crashes such as Oct. 1987 as well as smaller<br />
vol<strong>at</strong>ility bursts? Using a parsimonious autoregressive process with long-range memory <strong>de</strong>fined on the<br />
logarithm of the vol<strong>at</strong>ility, we predict strikingly different response functions of the price vol<strong>at</strong>ility to gre<strong>at</strong><br />
external shocks compared to wh<strong>at</strong> we term endogeneous shocks, i.e., which result from the cooper<strong>at</strong>ive<br />
accumul<strong>at</strong>ion of many small shocks. These predictions are remarkably well-confirmed empirically on<br />
a hierarchy of vol<strong>at</strong>ility shocks. Our theory allows us to classify two classes of events (endogeneous<br />
and exogeneous) with specific sign<strong>at</strong>ures and characteristic precursors for the endogeneous class. It also<br />
explains the origin of endogeneous shocks as the coherent accumul<strong>at</strong>ions of tiny bad news, and thus<br />
unify all previous explan<strong>at</strong>ions of large crashes including Oct. 1987.<br />
1 Introduction<br />
A mark<strong>et</strong> crash occurring simultaneously on most of the stock mark<strong>et</strong>s of the world as witnessed in Oct.<br />
1987 would amount to the quasi-instantaneous evapor<strong>at</strong>ion of trillions of dollars. Mark<strong>et</strong> crashes are the<br />
extreme end members of a hierarchy of mark<strong>et</strong> shocks, which shake stock mark<strong>et</strong>s repe<strong>at</strong>edly. Among<br />
∗ We acknowledge helpful discussions and exchanges with E. Bacry and V. Pisarenko. This work was partially supported by the<br />
James S. Mc Donnell Found<strong>at</strong>ion 21st century scientist award/studying complex system.<br />
1<br />
137
138 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
recent events still fresh in memories are the Hong-Kong crash and the turmoil on US mark<strong>et</strong>s on oct. 1997,<br />
the Russian <strong>de</strong>fault in Aug. 1998 and the ensuing mark<strong>et</strong> turbulence in western stock mark<strong>et</strong>s and the<br />
collapse of the “new economy” bubble with the crash of the Nasdaq in<strong>de</strong>x in March 2000.<br />
In each case, a lot of work has been carried out to unravel the origin(s) of the crash, so as to un<strong>de</strong>rstand<br />
its causes and <strong>de</strong>velop possible remedies. However, no clear cause can usually be singled out. A case in<br />
point is the Oct. 1987 crash, for which many explan<strong>at</strong>ions have been proposed but none has been wi<strong>de</strong>ly<br />
accepted unambiguously. These proposed causes inclu<strong>de</strong> computer trading, <strong>de</strong>riv<strong>at</strong>ive securities, illiquidity,<br />
tra<strong>de</strong> and budg<strong>et</strong> <strong>de</strong>ficits, over-infl<strong>at</strong>ed prices gener<strong>at</strong>ed by specul<strong>at</strong>ive bubble during the earlier period, the<br />
auction system itself, the presence or absence of limits on price movements, regul<strong>at</strong>ed margin requirements,<br />
off-mark<strong>et</strong> and off-hours trading, the presence or absence of floor brokers, the extent of trading in the<br />
cash mark<strong>et</strong> versus the forward mark<strong>et</strong>, the i<strong>de</strong>ntity of tra<strong>de</strong>rs (i.e. institutions such as banks or specialized<br />
trading firms), the significance of transaction taxes, <strong>et</strong>c. More rigorous and system<strong>at</strong>ic analyses on univari<strong>at</strong>e<br />
associ<strong>at</strong>ions and multiple regressions of these various factors conclu<strong>de</strong> th<strong>at</strong> it is not <strong>at</strong> all clear wh<strong>at</strong> caused<br />
the crash (Barro <strong>et</strong> al. 1989). The most precise st<strong>at</strong>ement, albeit somewh<strong>at</strong> self-referencial, is th<strong>at</strong> the most<br />
st<strong>at</strong>istically significant explan<strong>at</strong>ory variable in the October crash can be ascribed to the normal response of<br />
each country’s stock mark<strong>et</strong> to a worldwi<strong>de</strong> mark<strong>et</strong> motion (Barro <strong>et</strong> al. 1989).<br />
In view of the stalem<strong>at</strong>e reached by the approaches <strong>at</strong>tempting to find a proximal cause of a mark<strong>et</strong> shock,<br />
several researchers have looked for more fundamental origins and have proposed th<strong>at</strong> a crash may be the climax<br />
of an endogeneous instability associ<strong>at</strong>ed with the (r<strong>at</strong>ional or irr<strong>at</strong>ional) imit<strong>at</strong>ive behavior of agents (see<br />
for instance (Orléan 1989, Orléan 1995, Johansen and Sorn<strong>et</strong>te 1999, Shiller 2000)). Are there qualifying<br />
sign<strong>at</strong>ures of such a mechanism? According to (Johansen and Sorn<strong>et</strong>te 1999, Sorn<strong>et</strong>te and Johansen 2001)<br />
for which a crash is a stochastic event associ<strong>at</strong>ed with the end of a bubble, the d<strong>et</strong>ection of such bubble would<br />
provi<strong>de</strong> a fingerprint. A large liter<strong>at</strong>ure has emerged on the empirical d<strong>et</strong>ectability of bubbles in financial<br />
d<strong>at</strong>a and in particular on r<strong>at</strong>ional expect<strong>at</strong>ion bubbles (see (Camerer 1989, Adam and Szafarz 1992) for a survey).<br />
Unfortun<strong>at</strong>ely, the present evi<strong>de</strong>nce for specul<strong>at</strong>ive bubbles is fuzzy and unresolved <strong>at</strong> best, according<br />
to the standard economic and econom<strong>et</strong>ric liter<strong>at</strong>ure. Other than the still controversial (Feigenbaum 2001)<br />
sug<strong>gestion</strong> th<strong>at</strong> super-exponential price acceler<strong>at</strong>ion (Sorn<strong>et</strong>te and An<strong>de</strong>rsen 2002) and log-periodicity may<br />
qualify a specul<strong>at</strong>ive bubble (Johansen and Sorn<strong>et</strong>te 1999, Sorn<strong>et</strong>te and Johansen 2001), there are no unambiguous<br />
sign<strong>at</strong>ures th<strong>at</strong> would allow one to qualify a mark<strong>et</strong> shock or a crash as specifically endogeneous.<br />
On the other end, standard economic theory holds th<strong>at</strong> the complex trajectory of stock mark<strong>et</strong> prices is the<br />
faithful reflection of the continuous flow of news th<strong>at</strong> are interpr<strong>et</strong>ed and digested by an army of analysts<br />
and tra<strong>de</strong>rs (Cutler <strong>et</strong> al. 1989). Accordingly, large shocks should result from really bad surprises. It is<br />
a fact th<strong>at</strong> exogeneous shocks exist, as epitomized by the recent events of Sept. 11, 2001 and the coup<br />
against Gorbachev on Aug., 19, 1991, and there is no doubt about the existence of utterly exogeneous bad<br />
news th<strong>at</strong> move stock mark<strong>et</strong> prices and cre<strong>at</strong>e strong bursts of vol<strong>at</strong>ility. However, some could argue th<strong>at</strong><br />
precursory fingerprints of these events were known to some elites, suggesting the possibility the action of<br />
these informed agents may have been reflected in part in stock mark<strong>et</strong>s prices. Even more difficult is the<br />
classific<strong>at</strong>ion (endogeneous versus exogeneous) of the hierarchy of vol<strong>at</strong>ility bursts th<strong>at</strong> continuously shake<br />
stock mark<strong>et</strong>s. While it is a common practice to associ<strong>at</strong>e the large mark<strong>et</strong> moves and strong bursts of<br />
vol<strong>at</strong>ility with external economic, political or n<strong>at</strong>ural events (White 1996), there is not convincing evi<strong>de</strong>nce<br />
supporting it.<br />
Here, we provi<strong>de</strong> a clear and novel sign<strong>at</strong>ure allowing us to distinguish b<strong>et</strong>ween an endogeneous and an<br />
exogeneous origin to a vol<strong>at</strong>ility shock. Tests on the Oct. 1987 crash, on a hierarchy of vol<strong>at</strong>ility shocks<br />
and on a few of the obvious exogeneous shocks valid<strong>at</strong>e the concept. Our theor<strong>et</strong>ical framework combines<br />
a r<strong>at</strong>her novel but really powerful and parsimonious so-called multifractal random walk with conditional<br />
probability calcul<strong>at</strong>ions.<br />
2
2 Long-range memory and distinction b<strong>et</strong>ween endogeneous and exogeneous<br />
shocks<br />
While r<strong>et</strong>urns do not exhibit discernable correl<strong>at</strong>ions beyond a time scale of a few minutes in liquid arbitraged<br />
mark<strong>et</strong>s, the historical vol<strong>at</strong>ility (measured as the standard <strong>de</strong>vi<strong>at</strong>ion of price r<strong>et</strong>urns or more<br />
generally as a positive power of the absolute value of centered price r<strong>et</strong>urns) exhibits a long-range <strong>de</strong>pen<strong>de</strong>nce<br />
characterized by a power law <strong>de</strong>caying two-point correl<strong>at</strong>ion function (Ding <strong>et</strong> al. 1993, Ding<br />
and Granger 1996, Arneodo <strong>et</strong> al. 1998) approxim<strong>at</strong>ely following a (t/T ) −ν <strong>de</strong>cay r<strong>at</strong>e with an exponent<br />
ν ≈ 0.2. A vari<strong>et</strong>y of mo<strong>de</strong>ls have been proposed to account for these long-range correl<strong>at</strong>ions (Granger and<br />
Ding 1996, Baillie 1996, Müller <strong>et</strong> al. 1997, Muzy <strong>et</strong> al. 2000, Muzy <strong>et</strong> al. 2001, Müller <strong>et</strong> al. 1997).<br />
In addition, not only are r<strong>et</strong>urns clustered in bursts of vol<strong>at</strong>ility exhibiting long-range <strong>de</strong>pen<strong>de</strong>nce, but they<br />
also exhibit the property of multifractal scale invariance (or multifractality), according to which moments<br />
mq ≡ 〈|rτ | q 〉 of the r<strong>et</strong>urns <strong>at</strong> time scale τ are found to scale as mq ∝ τ ζq , with the exponent ζq being a<br />
non-linear function of the moment or<strong>de</strong>r q (Man<strong>de</strong>lbrot 1997, Muzy <strong>et</strong> al. 2000).<br />
To make quantit<strong>at</strong>ive predictions, we use a flexible and parsimonious mo<strong>de</strong>l, the so-called multifractal random<br />
walk (MRW) (see Appendix A and (Muzy <strong>et</strong> al. 2000, Bacry <strong>et</strong> al. 2001)), which unifies these two<br />
empirical observ<strong>at</strong>ions by <strong>de</strong>riving n<strong>at</strong>urally the multifractal scale invariance from the vol<strong>at</strong>ility long range<br />
<strong>de</strong>pen<strong>de</strong>nce.<br />
The long-range n<strong>at</strong>ure of the vol<strong>at</strong>ility correl<strong>at</strong>ion function can be seen as the direct consequence of a<br />
slow power law <strong>de</strong>cay of the response function K∆(t) of the mark<strong>et</strong> vol<strong>at</strong>ility measured a time t after the<br />
occurrence of an external perturb<strong>at</strong>ion of the vol<strong>at</strong>ility <strong>at</strong> scale ∆t. We find th<strong>at</strong> the distinct difference<br />
b<strong>et</strong>ween exogeneous and endogeneous shocks is found in the way the vol<strong>at</strong>ility relaxes to its unconditional<br />
average value.<br />
The prediction of the MRW mo<strong>de</strong>l (see Appendix B for the technical <strong>de</strong>riv<strong>at</strong>ion) is th<strong>at</strong> the excess vol<strong>at</strong>ility<br />
Eexo[σ 2 (t) | ω0] − σ 2 (t), <strong>at</strong> scale ∆t, due to an external shock of amplitu<strong>de</strong> ω0 relaxes to zero according to<br />
the universal response<br />
139<br />
Eexo[σ 2 (t) | ω0] − σ2 2K0t−1/2 (t) ∝ e − 1 ≈ 1 √ , (1)<br />
t<br />
for not too small times, where σ 2 (t) = σ 2 ∆t is the unconditional average vol<strong>at</strong>ility. This prediction is<br />
nothing but the response function K∆(t) of the MRW mo<strong>de</strong>l to a single piece of very bad news th<strong>at</strong> is<br />
sufficient by itself to move the mark<strong>et</strong> significantly. This prediction is well-verified by the empirical d<strong>at</strong>a<br />
shown in figure 1.<br />
On the other hand, an “endogeneous” shock is the result of the cumul<strong>at</strong>ive effect of many small bad news,<br />
each one looking rel<strong>at</strong>ively benign taken alone, but when taken all tog<strong>et</strong>her collectively along the full p<strong>at</strong>h<br />
of news can add up coherently due to the long-range memory of the vol<strong>at</strong>ility dynamics to cre<strong>at</strong>e a large<br />
“endogeneous” shock. This term “endogeneous” is thus not exactly a<strong>de</strong>qu<strong>at</strong>e since prices and vol<strong>at</strong>ilities<br />
are always moved by external news. The difference is th<strong>at</strong> an endogeneous shock in the present sense is the<br />
sum of the contribution of many “small” news adding up according to a specific most probable trajectory.<br />
It is this s<strong>et</strong> of small bad news prior to the large shock th<strong>at</strong> not only led to it but also continues to influence<br />
the dynamics of the vol<strong>at</strong>ility time series and cre<strong>at</strong>es an anomalously slow relax<strong>at</strong>ion. Appendix C gives the<br />
<strong>de</strong>riv<strong>at</strong>ion of the specific relax<strong>at</strong>ion (21) associ<strong>at</strong>ed with endogeneous shocks.<br />
Figure 2 reports empirical estim<strong>at</strong>es of the conditional vol<strong>at</strong>ility relax<strong>at</strong>ion after local maxima of the S&P100<br />
intradaily series ma<strong>de</strong> of 5 minute close prices during the period from 04/08/1997 to 12/24/2001 (figure<br />
1(a)). The original intraday squared r<strong>et</strong>urns have been preprocessed in or<strong>de</strong>r to remove the U-shaped vol<strong>at</strong>ility<br />
modul<strong>at</strong>ion associ<strong>at</strong>ed with the intraday vari<strong>at</strong>ions of mark<strong>et</strong> activity. Figure 2(b) shows th<strong>at</strong> the MRW<br />
3
140 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
∫ dt E[σ(t) | ω 0 ] − E[σ(t)]<br />
10 −1<br />
10 −2<br />
10 0<br />
Nikkei 250 : Aug 19, 1991<br />
S&P 500 : Aug 19, 1991<br />
FT−SE 100 : Aug 19, 1991<br />
CAC 40 : Sep 11, 2001<br />
S&P 500 : Oct 19, 1987<br />
10 1<br />
Time (trading days)<br />
Slope α = 1/2<br />
August 19, 1991 : Putsh against Presi<strong>de</strong>nt Gorbachev<br />
September 11, 2001: Attack against the WTC<br />
Figure 1: Cumul<strong>at</strong>ive excess vol<strong>at</strong>ility <strong>at</strong> scale ∆t, th<strong>at</strong> is, integral over time of Eexo[σ 2 (t) | ω0] − σ 2 (t),<br />
due to the vol<strong>at</strong>ility shock induced by the coup against Presi<strong>de</strong>nt Gorbachev observed in three British,<br />
Japanese and USA indices and the shock induced by the <strong>at</strong>tack of September 11, 2002 against the World<br />
Tra<strong>de</strong> Center. The dashed line is the theor<strong>et</strong>ical prediction obtained by integr<strong>at</strong>ing (1), which gives a ∝ √ t<br />
time-<strong>de</strong>pen<strong>de</strong>nce. The cumul<strong>at</strong>ive excess vol<strong>at</strong>ility following the crash of October 1987 is also shown with<br />
circles. Notice th<strong>at</strong> the slope of the non-constant curve for the October 1987 crash is very different from the<br />
value 1/2 expected and observed for exogeneous shocks. This crash and the resulting vol<strong>at</strong>ility relax<strong>at</strong>ion<br />
can be interpr<strong>et</strong>ed as an endogeneous event.<br />
4<br />
10 2
Figure 2: Measuring the conditional vol<strong>at</strong>ility response exponent α(s) for S&P 100 intradaily time series<br />
as a function of the endogeneous shock amplitu<strong>de</strong> param<strong>et</strong>erized by s, <strong>de</strong>fined by (19). (a) The original<br />
5 minute intradaily time series from 04/08/1997 to 12/24/2001. The 5 minute <strong>de</strong>-seasonalized squared<br />
r<strong>et</strong>urns are aggreg<strong>at</strong>ed in or<strong>de</strong>r to estim<strong>at</strong>e the 40 minutes and daily vol<strong>at</strong>ilities. (b) 40 minute log-vol<strong>at</strong>ility<br />
covariance C40(τ) as a function of the logarithm of the lag τ. The MRW theor<strong>et</strong>ical curve with λ 2 = 0.018<br />
and T = 1 year (dashed line) provi<strong>de</strong>s an excellent fit of the d<strong>at</strong>a up to lags of one month. (c) Conditional<br />
vol<strong>at</strong>ility response ln(Eendo[σ 2 (t) | s]) as a function of ln(t) for three shocks with amplitu<strong>de</strong>s given by<br />
s = −1, 0, 1. (d) Estim<strong>at</strong>ed exponent α(s) for ∆t = 40 minutes (•) as a function of s. The solid line<br />
is the prediction corresponding to Eq. (22). The dashed line corresponds to the empirical MRW estim<strong>at</strong>e<br />
obtained by averaging over 500 Monte-Carlo trials. It fits more accur<strong>at</strong>ely for neg<strong>at</strong>ive s (vol<strong>at</strong>ility lower<br />
than normal) due to the fact th<strong>at</strong> the estim<strong>at</strong>ions of the variance by aggreg<strong>at</strong>ion over smaller scales is very<br />
noisy for small variance values. The error bars give the 95 % confi<strong>de</strong>nce intervals estim<strong>at</strong>ed by Monte-Carlo<br />
trials of the MRW process. In the ins<strong>et</strong>, α(s) is compared for ∆t = 40 minutes (•) and ∆t = 1 day (×).<br />
5<br />
141
142 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
mo<strong>de</strong>l provi<strong>de</strong>s a very good fit of the empirical vol<strong>at</strong>ility covariance in a range of time scales from 5 minutes<br />
to one month. Fig. 2(c) plots in a double logarithmic represent<strong>at</strong>ion, for the time scale ∆t = 40 minutes,<br />
the estim<strong>at</strong>ed conditional vol<strong>at</strong>ility responses for s = 1, 0, −1, where the endogeneous shocks are param<strong>et</strong>erized<br />
by e 2s σ 2 (t). A value s > 0 (resp. s < 0) corresponds to a positive bump (resp. neg<strong>at</strong>ive dip) of the<br />
vol<strong>at</strong>ility above (resp. below) the average level σ 2 (t). The straight lines are the predictions (Eqs. (24,22))<br />
of the MRW mo<strong>de</strong>l and qualify power law responses whose exponents α(s) are continuous function of the<br />
shock amplitu<strong>de</strong> s. Figure 1(d) plots the conditional response exponent α(s) as a function of s for the two<br />
time scales ∆t = 40 minutes and ∆t = 1 day (ins<strong>et</strong>). For ∆t = 40 minutes, we observe th<strong>at</strong> α varies<br />
b<strong>et</strong>ween −0.2 for the largest positive shocks to +0.2 for the largest neg<strong>at</strong>ive shocks, in excellent agreement<br />
with MRW estim<strong>at</strong>es (dashed line) and, for α ≥ 0, with Eq. (22) obtained without any adjustable param<strong>et</strong>ers.<br />
1 The error bars represent the 95 % confi<strong>de</strong>nce intervals estim<strong>at</strong>ed using 500 trials of synth<strong>et</strong>ic MRW<br />
with the same param<strong>et</strong>ers as observed for the S&P 100 series. By comparing α(s) for different ∆t (ins<strong>et</strong>),<br />
we can see the the MRW mo<strong>de</strong>l is thus able to recover not only the s-<strong>de</strong>pen<strong>de</strong>nce of the exponent α(s) of<br />
the conditional response function to endogeneous shocks but also its time scale ∆t vari<strong>at</strong>ions: this exponent<br />
increases as one goes from fine to coarse scales. Similar results are obtained for other intradaily time series<br />
(Nasdaq, FX-r<strong>at</strong>es, <strong>et</strong>c.). We also obtain the same results for 17 years of daily r<strong>et</strong>urn times series of various<br />
indices (French, German, canada, Japan, <strong>et</strong>c.).<br />
In summary, the most remarkable result is the qualit<strong>at</strong>ively different functional <strong>de</strong>pen<strong>de</strong>nce of the response<br />
(1) to an exogeneous compared to the response (24,22) to an endogeneous shock. The former gives a <strong>de</strong>cay<br />
of the burst of vol<strong>at</strong>ility ∝ 1/t 1/2 compared to 1/t α(s) for endogeneous shocks with amplitu<strong>de</strong> e 2s σ 2 (t),<br />
with an exponent α(s) being a linear function of s.<br />
3 Discussion<br />
Wh<strong>at</strong> is the source of endogeneous shocks characterized by the response function (21)? Appendix D and<br />
equ<strong>at</strong>ion (29) predict th<strong>at</strong> the expected p<strong>at</strong>h of the continuous inform<strong>at</strong>ion flow prior to the endogeneous<br />
shock grows proportionally to the response function K(tc − t) measured in backward time to the shock<br />
occuring <strong>at</strong> tc. In other words, conditioned on the observ<strong>at</strong>ion of a large endogeneous shock, there is<br />
specific s<strong>et</strong> of trajectories of the news flow th<strong>at</strong> led to it. This specific flow has an expect<strong>at</strong>ion given by (29).<br />
This result allows us to un<strong>de</strong>rstand the distinctive fe<strong>at</strong>ures of an endogeneous shock compared to an external<br />
shock. The l<strong>at</strong>er is a single piece of very bad news th<strong>at</strong> is sufficient by itself to move the mark<strong>et</strong> significantly<br />
according to (1). In contrast, an “endogeneous” shock is the result of the cumul<strong>at</strong>ive effect of many small<br />
bad news, each one looking rel<strong>at</strong>ively benign taken alone, but when taken all tog<strong>et</strong>her collectively along<br />
the full p<strong>at</strong>h of news can add up coherently due to the long-range memory of the log-vol<strong>at</strong>ility dynamics to<br />
cre<strong>at</strong>e a large “endogeneous” shock. This term “endogeneous” is thus not exactly a<strong>de</strong>qu<strong>at</strong>e since prices and<br />
vol<strong>at</strong>ilities are always moved by external news. The difference is th<strong>at</strong> an endogeneous shock in the present<br />
sense is the sum of the contribution of many “small” news adding up according to a specific most probable<br />
trajectory. It is this s<strong>et</strong> of small bad news prior to the large shock th<strong>at</strong> not only led to it but also continues to<br />
influence the dynamics of the vol<strong>at</strong>ility time series and cre<strong>at</strong>es the anomalously slow relax<strong>at</strong>ion (21).<br />
In this respect, this result allows us to r<strong>at</strong>ionalize and unify the many explan<strong>at</strong>ions proposed to account for<br />
the Oct. 1987 crash: according to the present theory, each of the explan<strong>at</strong>ions is insufficient to explain the<br />
crash; however, our theory suggests th<strong>at</strong> it is the cumul<strong>at</strong>ive effect of many such effects th<strong>at</strong> led to the crash.<br />
In a sense, the different comment<strong>at</strong>ors and analysts were all right in <strong>at</strong>tributing the origin of the Oct. 1987<br />
crash to many different factors but they missed the main point th<strong>at</strong> the crash was the extreme response of<br />
1 The <strong>de</strong>vi<strong>at</strong>ion of α(s) from expression (22) for neg<strong>at</strong>ive s, origin<strong>at</strong>es from the error in the vol<strong>at</strong>ility estim<strong>at</strong>ion using a sum of<br />
squared r<strong>et</strong>urns. The smaller the sum of squared r<strong>et</strong>urns, the larger the error is. As ∆t increases, this error becomes negligible<br />
6
the system to the accumul<strong>at</strong>ion of many tiny bad news contributions. To test this i<strong>de</strong>a, we note th<strong>at</strong> the<br />
<strong>de</strong>cay of the vol<strong>at</strong>ility response after the Oct. 1987 crash has been <strong>de</strong>scribed by a power law 1/t 0.3 (Lillo<br />
and Mantegna 2001), which is in line with the prediction of our MRW theory with equ<strong>at</strong>ion (22) for such<br />
a large shock (see also figure 2 panel d). This value of the exponent is still significantly smaller than 0.5.<br />
Figure 1 <strong>de</strong>monstr<strong>at</strong>es further the difference b<strong>et</strong>ween the relax<strong>at</strong>ion of the vol<strong>at</strong>ility after this event shown<br />
with circle and those following the exogenous coup against Gorbachev and the September 11 <strong>at</strong>tack. There<br />
is clearly a strong constrast which qualifies the Oct. 1987 crash as endogeneous, in the sense of our theory<br />
of “conditional response.” This provi<strong>de</strong>s an in<strong>de</strong>pen<strong>de</strong>nt confirm<strong>at</strong>ion of the concept advanced before in<br />
(Johansen and Sorn<strong>et</strong>te 1999, Sorn<strong>et</strong>te and Johansen 2001).<br />
It is also interesting to compare the prediction (21) with those obtained with a linear autoregressive mo<strong>de</strong>l<br />
of the type (5), in which ω(t) is replaced by σ(t). FIGARCH mo<strong>de</strong>ls fall in this general class. It is easy<br />
to show in this case th<strong>at</strong> this linear (in vol<strong>at</strong>ility) mo<strong>de</strong>l predicts the same exponent for the response of the<br />
vol<strong>at</strong>ility to endogeneous shocks, in<strong>de</strong>pen<strong>de</strong>ntly of their magnitu<strong>de</strong>. This prediction is in stark constrast<br />
with the prediction (21) of the log-vol<strong>at</strong>ility MRW mo<strong>de</strong>l. The l<strong>at</strong>er mo<strong>de</strong>l is thus strongly valid<strong>at</strong>ed by our<br />
empirical tests.<br />
Appendix A: The Multifractal Randow Walk (MRW) mo<strong>de</strong>l<br />
The multifractal random walk mo<strong>de</strong>l is the continuous time limit of a stochastic vol<strong>at</strong>ility mo<strong>de</strong>l where<br />
log-vol<strong>at</strong>ility 2 correl<strong>at</strong>ions <strong>de</strong>cay logarithmically. It possesses a nice “stability” property rel<strong>at</strong>ed to its scale<br />
invariance property: For each time scale ∆t ≤ T , the r<strong>et</strong>urns <strong>at</strong> scale ∆t, r∆t(t) ≡ ln[p(t)/p(t − ∆t)], can<br />
be <strong>de</strong>scribed as a stochastic vol<strong>at</strong>ility mo<strong>de</strong>l:<br />
143<br />
r∆t(t) = ɛ(t) · σ∆t(t) = ɛ(t) · e ω∆t(t) , (2)<br />
where ɛ(t) is a standardized Gaussian white noise in<strong>de</strong>pen<strong>de</strong>nt of ω∆t(t) and ω∆t(t) is a nearly Gaussian<br />
process with mean and covariance:<br />
µ∆t = 1<br />
2 ln(σ2∆t) − C∆t(0) (3)<br />
C∆t(τ) = Cov[ω∆t(t), ω∆t(t + τ)] = λ 2 <br />
T<br />
ln<br />
|τ| + e−3/2 <br />
. (4)<br />
∆t<br />
σ 2 ∆t is the r<strong>et</strong>urn variance <strong>at</strong> scale ∆t and T represents an “integral” (correl<strong>at</strong>ion) time scale. Such logarithmic<br />
<strong>de</strong>cay of log-vol<strong>at</strong>ility covariance <strong>at</strong> different time scales has been <strong>de</strong>monstr<strong>at</strong>ed empirically in<br />
(Arneodo <strong>et</strong> al. 1998, Muzy <strong>et</strong> al. 2000). Typical values for T and λ 2 are respectively 1 year and 0.02.<br />
According to the MRW mo<strong>de</strong>l, the vol<strong>at</strong>ility correl<strong>at</strong>ion exponent ν is rel<strong>at</strong>ed to λ 2 by ν = 4λ 2 .<br />
The MRW mo<strong>de</strong>l can be expressed in a more familiar form, in which the log-vol<strong>at</strong>ility ω∆t(t) obeys an<br />
auto-regressive equ<strong>at</strong>ion whose solution reads<br />
t<br />
ω∆t(t) = µ∆t + dτ η(τ) K∆t(t − τ) , (5)<br />
−∞<br />
where η(t) <strong>de</strong>notes a standardized Gaussian white noise and the memory kernel K∆t(·) is a causal function,<br />
ensuring th<strong>at</strong> the system is not anticip<strong>at</strong>ive. The process η(t) can be seen as the inform<strong>at</strong>ion flow. Thus ω(t)<br />
represents the response of the mark<strong>et</strong> to incoming inform<strong>at</strong>ion up to the d<strong>at</strong>e t. At time t, the distribution<br />
2 The log-vol<strong>at</strong>ilty is the n<strong>at</strong>ural quantity used in canonical stoch<strong>at</strong>ic vol<strong>at</strong>ility mo<strong>de</strong>ls (see (Kim <strong>et</strong> al. 1998, and references<br />
therein)).<br />
7
144 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
of ω∆t(t) is Gaussian with mean µ∆t and variance V∆t = ∞<br />
0 dτ K2 ∆t (τ) = λ2 ln<br />
which entirely specifies the random process, is given by<br />
∞<br />
C∆t(τ) =<br />
0<br />
T e 3/2<br />
∆t<br />
<br />
. Its covariance,<br />
dt K∆t(t)K∆t(t + |τ|) . (6)<br />
Performing a Fourier tranform, we obtain ˆ K∆t(f) 2 = Ĉ∆t(f) = 2λ2 f −1<br />
T f<br />
0<br />
which shows th<strong>at</strong> for τ small enough<br />
K∆t(τ) ∼ K0<br />
λ 2 T<br />
τ<br />
<br />
sin(t)<br />
t dt + O (f∆t ln(f∆t)) ,<br />
for ∆t
Appendix C: “Conditional response” to an endogeneous shock<br />
L<strong>et</strong> us consi<strong>de</strong>r the n<strong>at</strong>ural evolution of the system, without any large external shock, which nevertheless<br />
exhibits a large vol<strong>at</strong>ility burst ω(t = 0) = ω0 <strong>at</strong> t = 0. From the <strong>de</strong>finition (2) with (5) and (6), it is<br />
clear th<strong>at</strong> a large “endogeneous” shock requires a special s<strong>et</strong> of realiz<strong>at</strong>ion of the “small news” {η(t)}. To<br />
quantifies the response in such case, we can evalu<strong>at</strong>e Eendo[σ 2 (t) | ω0] = Eendo[e 2ω(t) | ω0]. Since ω(t) is a<br />
Gaussian process, the new process ω(t) conditional on ω0 remains Gaussian, so th<strong>at</strong><br />
145<br />
Eendo[σ 2 (t) | ω0] = Eendo[e 2ω(t) | ω0] (13)<br />
= exp (2E[w(t) | ω0] + 2Var[ω(t) | ω0]) . (14)<br />
Due to the still Gaussian n<strong>at</strong>ure of the condition log-vol<strong>at</strong>ility ω(t), we easily obtain using (3) and (4),<br />
and<br />
L<strong>et</strong> us s<strong>et</strong>:<br />
Eendo[ω(t) | ω0] = E[ω(t)] +<br />
Cov[ω(t), ω0]<br />
Var[ω0]<br />
· (ω0 − E[ω0]) , (15)<br />
= (ω0 − µ) · C(t)<br />
+ µ , (16)<br />
C(0)<br />
Cov[ω(t), ω0] 2<br />
Varendo[ω(t) | ω0] = Var[ω(t)] − , (17)<br />
Var[ω0]<br />
<br />
= C(0) 1 − C2 (t)<br />
C2 <br />
. (18)<br />
(0)<br />
e 2ω0 = e 2s σ 2 (t) ⇒ ω0 − µ = s + C(0) (19)<br />
By subsitution in (14), we obtain thanks to (3) and (4),<br />
Eendo[σ 2 (t) | ω0] = σ2 <br />
(t) exp 2(ω0 − µ) · C(t)<br />
C(0) − 2C2 =<br />
<br />
(t)<br />
,<br />
C(0)<br />
σ<br />
(20)<br />
2 α(s)+β(t) T<br />
(t)<br />
t<br />
(21)<br />
where<br />
α(s) =<br />
ln(<br />
2s<br />
, (22)<br />
T e3/2<br />
∆t )<br />
β(t) = 2λ 2 ln(t/∆t)<br />
ln(T e3/2 . (23)<br />
/∆t)<br />
Within the range ∆t < t
146 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
where<br />
It is thus easy to obtain the estim<strong>at</strong>e:<br />
β ′ (t) = 2λ 2 ln(T/t)<br />
ln(T e3/2 /∆t)<br />
<br />
Varendo [σ2 (t) | ω0] ≤ Eendo[σ 2 <br />
(t) | ω0] ( T<br />
∆t )2β(t) 1/2 <br />
− 1 6λ2 ln(te3/2 /∆t) Eendo[σ 2 (t) | ω0]<br />
We thus conclu<strong>de</strong>, th<strong>at</strong>, for s large enough (i.e., α(s) large enough):<br />
Eendo[σ 2 (t) | ω0] − σ 2 (t)<br />
Varendo [σ 2 (t) | ω0] <br />
1<br />
6λ 2 ln(te 3/2 /∆t)<br />
Over the first <strong>de</strong>ca<strong>de</strong> ∆t ≤ t ≤ 10∆t, the <strong>de</strong>vi<strong>at</strong>ion of the conditional mean vol<strong>at</strong>ility from the unconditional<br />
vol<strong>at</strong>ility σ 2 (t) is gre<strong>at</strong>er than the conditional variance, which ensures the existence of a strong d<strong>et</strong>erministic<br />
component of the conditional response above the stochastic components.<br />
Expressions (24,22) are our two main predictions. These equ<strong>at</strong>ions predict th<strong>at</strong> the conditional response<br />
function Eendo[σ2 (t) | ω0] of the vol<strong>at</strong>ility <strong>de</strong>cays as a power law ∼ 1/tα <br />
of the time since the endogeneous<br />
shock, with an exponent α ≈ 2ω0 − ln(σ2 <br />
λ2 (t))<br />
C(0) which <strong>de</strong>pends linearly upon the amplitu<strong>de</strong> ω0 of the<br />
shock. Note in particular, th<strong>at</strong> α changes sign: it is positive for w0 > 1<br />
2 ln(σ2 (t)) and neg<strong>at</strong>ive otherwise.<br />
Appendix D: D<strong>et</strong>ermin<strong>at</strong>ion of the sources of endogeneous shocks<br />
Wh<strong>at</strong> is the source of endogeneous shocks characterized by the response function (21)? To answer, l<strong>et</strong><br />
us consi<strong>de</strong>r the process W (t) ≡ t<br />
−∞ dτ η(τ), where η(t) is a standardized Gaussian white noise which<br />
captures the inform<strong>at</strong>ion flow impacting on the vol<strong>at</strong>ility, as <strong>de</strong>fined in (5). Extending the property (16), we<br />
find th<strong>at</strong><br />
t<br />
Cov[W (t), ω0]<br />
Eendo[W (t) | ω0] = · (ω0 − E[ω0]) ∝ (ω0 − E[ω0]) dτ K(−τ) . (29)<br />
Var[ω0]<br />
−∞<br />
Expression (29) predicts th<strong>at</strong> the expected p<strong>at</strong>h of the continuous inform<strong>at</strong>ion flow prior to the endogeneous<br />
shock (i.e., for t < 0) grows like ∆W (t) = η(t)∆t ∼ K(−t)∆t ∼ ∆t/ √ −t for t < 0 upon the approach<br />
to the time t = 0 of the large endogeneous shock. In other words, conditioned on the observ<strong>at</strong>ion of a large<br />
endogeneous shock, there is specific s<strong>et</strong> of trajectories of the news flow η(t) th<strong>at</strong> led to it. These conditional<br />
news flows have an expect<strong>at</strong>ion given by (29).<br />
References<br />
Adam, M.C. and A. Szafarz, 1992, Specul<strong>at</strong>ive Bubbles and Financial Mark<strong>et</strong>s, Oxford Economic Papers<br />
44, 626-640.<br />
Arneodo, A., J.F. Muzy and D. Sorn<strong>et</strong>te, 1998, Direct causal casca<strong>de</strong> in the stock mark<strong>et</strong>, The European<br />
physical Journal B 2, 277-282.<br />
Bacry, E., J. Delour and J.F. Muzy, 2001, Multifractal ranom walk, Physical Review E 64, 026103.<br />
10<br />
(26)<br />
(27)<br />
(28)
Baillie, R.T., 1996, Long memory processes and fractional integr<strong>at</strong>ion in econom<strong>et</strong>rics, Journal of Econom<strong>et</strong>rics<br />
73, 5-59.<br />
Barro, R.J., E.F. Fama, D.R. Fischel, A.H. Meltzer, R. Roll and L.G. Telser, 1989, Black monday and the<br />
future of financial mark<strong>et</strong>s, edited by R.W. Kamphuis, Jr., R.C. Kormendi and J.W.H. W<strong>at</strong>son (Mid<br />
American Institute for Public Policy Research, Inc. and Dow Jones-Irwin, Inc.).<br />
Camerer, C., 1989, Bubbles and Fads in Ass<strong>et</strong> Prices, Journal of Economic Surveys 3, 3-41.<br />
Cutler, D., J. Poterba and L. Summers, Wh<strong>at</strong> Moves Stock Prices? 1989, Journal of Portfolio Management,<br />
Spring, 4-12.<br />
Ding, Z., Granger, C.W.J. and Engle, R., 1993, A long memory property of stock r<strong>et</strong>urns and a new mo<strong>de</strong>l,<br />
Journal of Empirical Finance 1, 83-106.<br />
Ding, Z., Granger, C.W.J., 1996, Mo<strong>de</strong>ling vol<strong>at</strong>ility persistence of specul<strong>at</strong>ive r<strong>et</strong>urns: A new approach,<br />
Journal of Econom<strong>et</strong>rics 73, 185-215.<br />
Feigenbaum, J.A., 2001, A st<strong>at</strong>istical analysis of log-periodic precursors to financial crashes, Quantit<strong>at</strong>ive<br />
Finance 1(3), 346-360.<br />
Granger, C.W.J. and Ding, Z., 1996, Vari<strong>et</strong>ies of long memory mo<strong>de</strong>ls, Journal of Econom<strong>et</strong>rics 73, 61-77.<br />
Johansen, A. and Sorn<strong>et</strong>te, 1999, Critical Crashes, Risk 12 (1), 91-94.<br />
Lillo, F. and R.N. Mantegna, Power-law relax<strong>at</strong>ion in a complex system: the fluctu<strong>at</strong>ion <strong>de</strong>cay after a financial<br />
mark<strong>et</strong> crash, preprint cond-m<strong>at</strong>/0111257<br />
Man<strong>de</strong>lbrot, B.B., 1997, Fractals and scaling in finance : discontinuity, concentr<strong>at</strong>ion, risk : selecta volume<br />
E, New York : Springer.<br />
Müller, U.A., M.M. Dacorogna, R. Dav, R.B. Olsen, O.V. Pict<strong>et</strong> and J.E. von Weizs”cker, 1997, Vol<strong>at</strong>ilities<br />
of Different Time Resolutions - Analyzing the Dynamics of Mark<strong>et</strong> Components, Journal of<br />
Empirical Finance 4, No. 2-3, 213-240.<br />
Muzy, J.F., J. Delour and E. Bacry, 2000, Mo<strong>de</strong>lling fluctu<strong>at</strong>ions of financial time series: from casca<strong>de</strong><br />
process to stochastic vol<strong>at</strong>ility mo<strong>de</strong>l, The European physical Journal B 17, 537-548.<br />
Muzy, J.-F., D. Sorn<strong>et</strong>te, J. Delour and A. Arneodo, 2001, Multifractal r<strong>et</strong>urns and Hierarchical Portfolio<br />
Theory, Quantit<strong>at</strong>ive Finance 1 (1), 131-148.<br />
Orléan, A., 1989, Mim<strong>et</strong>ic contagion and specul<strong>at</strong>ive bubbles, Theory and Decision 27, 63-92.<br />
Orléan, A., 1995, Bayesian interactions and collective dynamics of opinion: Heard behavior and mim<strong>et</strong>ic<br />
contagion, journal of Economic Behavior & Organiz<strong>at</strong>ion 28, 257-274.<br />
Kim, S., N. Shepard, and S. Chib , 1998, Stochastic vol<strong>at</strong>ility: Likelyhood inference and comparison with<br />
ARCH mo<strong>de</strong>ls, Review of Economic Studies 65, 361-393.<br />
Shiller, R.J., 2000, Irr<strong>at</strong>ional exuberance (Princ<strong>et</strong>on University Press, Princ<strong>et</strong>on, NJ).<br />
Sorn<strong>et</strong>te, D. and A. Johansen, 2001, Significance of log-periodic precursors to financial crashes, Quantit<strong>at</strong>ive<br />
Finance 1 (4), 452-471.<br />
Sorn<strong>et</strong>te, D. and J.V. An<strong>de</strong>rsen, 2002, A Nonlinear Super-Exponential R<strong>at</strong>ional Mo<strong>de</strong>l of Specul<strong>at</strong>ive Financial<br />
Bubbles, Int. J. Mod. Phys. C 13 (2), 171-188.<br />
11<br />
147
148 4. Relax<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité<br />
White E.N., 1996, Stock mark<strong>et</strong> crashes and specul<strong>at</strong>ive manias. In The intern<strong>at</strong>ional library of macroeconomic<br />
and financial history, 13 (An Elgar Reference Collection, Cheltenham, UK; Brookfield, US).<br />
12
Chapitre 5<br />
Approche comportementale <strong>de</strong>s marchés<br />
financiers : l’apport <strong>de</strong>s modèles d’agents<br />
Depuis quelques années, les sciences économiques <strong>et</strong> financières abandonnent progressivement le modèle<br />
standard <strong>de</strong> l’agent économique représent<strong>at</strong>if dont on sait qu’il soutient notamment l’hypothèse d’efficience<br />
<strong>de</strong>s marchés, <strong>et</strong> ce dans le but <strong>de</strong> prendre en considér<strong>at</strong>ion le fait que les marchés financiers<br />
reflètent les actions <strong>et</strong> les émotions d’êtres réels. Ainsi s’est développé ce qu’il convient désormais d’appeler<br />
l’économie ou la finance cognitive (voir par exemple Thaller (1993), Sheffrin (2000), Shleiffer<br />
(2000) ou encore Goldberg <strong>et</strong> von Nitzsch (2001)). Le but <strong>de</strong> c<strong>et</strong>te nouvelle approche est d’intégrer<br />
la r<strong>at</strong>ionalité nécessairement limitée <strong>de</strong>s agents, qui peuvent certes tenter au mieux <strong>de</strong> leurs capacités<br />
d’être r<strong>at</strong>ionnels <strong>et</strong> d’agir <strong>de</strong> manière optimale, il n’en reste pas moins qu’ils seront toujours suj<strong>et</strong> à<br />
<strong>de</strong>s émotions ou <strong>de</strong>s influences extérieures qu’ils ne contrôlent pas. Ainsi, il est normal que les dimensions<br />
psychologique, sociologique <strong>et</strong> cognitive au sens large ayant une influence sur l’agent économique<br />
soient prises en compte. Pour autant <strong>et</strong> en dépit <strong>de</strong> c<strong>et</strong>te r<strong>at</strong>ionalité limitée <strong>de</strong>s agents, les marchés sont le<br />
plus souvent très proches <strong>de</strong> l’efficience. Ainsi il convient <strong>de</strong> se <strong>de</strong>man<strong>de</strong>r comment l’auto-organis<strong>at</strong>ion<br />
<strong>de</strong>s agents perm<strong>et</strong> d’<strong>at</strong>teindre c<strong>et</strong>te efficience <strong>et</strong> pourquoi / comment le système s’en écarte parfois <strong>de</strong><br />
manière spectaculaire.<br />
C’est en fait tout l’enjeu <strong>de</strong> l’étu<strong>de</strong> <strong>de</strong>s modèles d’agents en interaction que <strong>de</strong> tenter d’apporter <strong>de</strong>s<br />
réponses à ces questions. L’approche théorique nouvelle offerte par ces modèles perm<strong>et</strong> d’intégrer beaucoup<br />
<strong>de</strong>s limit<strong>at</strong>ions <strong>de</strong>s agents réels <strong>et</strong> à partir <strong>de</strong> la <strong>de</strong>scription - même très sommaire - du comportement<br />
<strong>de</strong>s acteurs agissant sur les marchés, il est possible d’en déduire certains mo<strong>de</strong>s <strong>de</strong> fonctionnement. Cela<br />
perm<strong>et</strong> notamment <strong>de</strong> mieux comprendre comment, partant d’agents individuels non r<strong>at</strong>ionnels au sens<br />
strict, peut émerger un comportement collectif conforme aux modèles r<strong>at</strong>ionnels, puis dans un <strong>de</strong>uxième<br />
temps, d’étudier les dévi<strong>at</strong>ions possibles <strong>et</strong> les relax<strong>at</strong>ions vers l’ét<strong>at</strong> d’efficience du marché prédit par<br />
ces modèles r<strong>at</strong>ionnels. Réciproquement, l’adéqu<strong>at</strong>ion ou non <strong>de</strong>s prévisions <strong>de</strong> ces modèles aux faits<br />
stylisés, que nous avons décrits au chapitre 1, perm<strong>et</strong> si ce n’est d’accepter ou <strong>de</strong> rej<strong>et</strong>er définitivement la<br />
pertinence <strong>de</strong> certains types <strong>de</strong> comportements <strong>de</strong>s agents vis-à-vis <strong>de</strong> la génér<strong>at</strong>ion <strong>de</strong> l’organis<strong>at</strong>ion <strong>de</strong>s<br />
marchés, du moins <strong>de</strong> cerner les mécanismes les plus importants quant à la <strong>de</strong>scription <strong>de</strong>s faits stylisés.<br />
De plus, à la différence notoire <strong>de</strong> la physique <strong>et</strong> <strong>de</strong>s sciences n<strong>at</strong>urelles en général, il est quasiment<br />
impossible <strong>de</strong> se livrer à <strong>de</strong>s expériences sur les marchés financiers afin <strong>de</strong> tester directement telle ou<br />
telle hypothèse ou l’impact d’une nouvelle réglement<strong>at</strong>ion. Certes, quelques simul<strong>at</strong>ions <strong>de</strong> marché ont<br />
été conduites en labor<strong>at</strong>oire (Smith 1994, Smith 1998), mais le nombre restreint d’acteurs qui y prennent<br />
part <strong>et</strong> le fait qu’ils sachent qu’ils participent à une expérience, en limite quelque peu la validité. Donc,<br />
le développement <strong>de</strong> modèles <strong>de</strong> marchés réalistes est particulièrement intéressant dans la mesure où<br />
149
150 5. Approche comportementale <strong>de</strong>s marchés financiers : l’apport <strong>de</strong>s modèles d’agents<br />
ils pourraient servir <strong>de</strong> substitut <strong>de</strong> labor<strong>at</strong>oire. En fait c<strong>et</strong>te approche est d’ores <strong>et</strong> déjà en train <strong>de</strong> se<br />
développer, tant sur le plan académique dans le but d’étudier l’eff<strong>et</strong> <strong>de</strong> l’introduction <strong>de</strong> la taxe Tobin<br />
ou l’influence <strong>de</strong>s ordres limites sur la vol<strong>at</strong>ilité du marché (Westerhoff 2001, Westerhoff 2002, par<br />
exemple), que sur le plan pr<strong>at</strong>ique pour tester divers scénarii <strong>et</strong> m<strong>et</strong>tre au point certaines str<strong>at</strong>égies.<br />
Comme nous venons <strong>de</strong> le dire, l’émergence <strong>de</strong>s modèles d’agents en interaction est en gran<strong>de</strong> partie<br />
due à l’échec <strong>de</strong>s modèles standards dont les principales conséquences sont les propriétés d’efficience<br />
<strong>de</strong>s marchés <strong>et</strong> le fait que les prix <strong>de</strong>s actifs financiers suivent <strong>de</strong>s marches alé<strong>at</strong>oires. Le défaut majeur,<br />
aujourd’hui bien établi, <strong>de</strong>s modèles économiques standards est <strong>de</strong> considérer un agent dit représent<strong>at</strong>if,<br />
ce qui revient à faire une approxim<strong>at</strong>ion <strong>de</strong> “champ moyen”, alors qu’il apparaît très clairement que<br />
l’hétérogénéité est un concept clé <strong>de</strong> la compréhension <strong>de</strong>s marchés financiers <strong>et</strong> que malgrè c<strong>et</strong>te<br />
hétérogénéité, une auto-organis<strong>at</strong>ion voit le jour.<br />
Il existe désormais un grand nombre <strong>de</strong> modèles d’agents perm<strong>et</strong>tant <strong>de</strong> rendre compte <strong>de</strong> tels ou tels<br />
faits stylisés. Notre objectif n’est pas <strong>de</strong> les décrire un par un <strong>de</strong> manière exhaustive, ce qui n’aurait<br />
d’ailleurs pas grand intérêt. Nous préférons nous <strong>at</strong>tacher à présenter quelques gran<strong>de</strong>s classes <strong>de</strong><br />
modèles d’agents parmi lesquelles nous distinguons d’une part les modèles d’opinion <strong>et</strong> les modèles<br />
<strong>de</strong> marchés <strong>et</strong> d’autre part les modèles à agents adapt<strong>at</strong>ifs <strong>et</strong> non adapt<strong>at</strong>ifs. D’autres c<strong>at</strong>égories peuvent<br />
évi<strong>de</strong>mment être établies, conduisant notamment à l’importante distinction entre modèles à hétérogénéité<br />
forte ou faible, mais nous n’irons pas aussi loin, d’autant que nous souhaitons m<strong>et</strong>tre en exergue les<br />
quelques mécanismes récurrents se r<strong>et</strong>rouvant dans la plupart <strong>de</strong>s modèles.<br />
Avant d’entrer dans le vif du suj<strong>et</strong>, il convient <strong>de</strong> spécifier un <strong>de</strong>rnier point, à savoir, étant donnée une<br />
popul<strong>at</strong>ion d’agents économiques s’échangeant un actif financier, comment se forme le prix <strong>de</strong> c<strong>et</strong> actif.<br />
Un grand nombre <strong>de</strong> modèles continuent <strong>de</strong> supporter l’idée que la form<strong>at</strong>ion <strong>de</strong>s prix sur les marchés<br />
financiers peut résulter d’un équilibre entre offre <strong>et</strong> <strong>de</strong>man<strong>de</strong>. Ce sont les modèles dits d’équilibres r<strong>at</strong>ionnels<br />
adapt<strong>at</strong>ifs, parmi lesquels on peut citer Brock <strong>et</strong> LeBaron (1996), Brock <strong>et</strong> Hommes (1997) ou<br />
Grandmont (1998) par exemple. Cependant, au vu <strong>de</strong> résult<strong>at</strong>s empiriques <strong>de</strong> plus en plus nombreux<br />
(Brown, Walsh <strong>et</strong> Yuen 1997, Chordia, Roll <strong>et</strong> Subrahmanyam 2002), il parait extrêmement difficile <strong>de</strong><br />
continuer à adm<strong>et</strong>tre c<strong>et</strong>te hypothèse <strong>et</strong> il <strong>de</strong>vient nécessaire <strong>de</strong> reconnaître que les prix observés sur les<br />
marchés financiers se forment en <strong>de</strong>hors <strong>de</strong> tout équilibre. Donc, avant <strong>de</strong> nous intéresser aux différentes<br />
classes <strong>de</strong> modèles d’agents, nous allons montrer comment définir un pris hors équilibre.<br />
5.1 Prix d’un actif <strong>et</strong> excès <strong>de</strong> <strong>de</strong>man<strong>de</strong><br />
L’un <strong>de</strong>s principaux enjeux <strong>de</strong> tout modèle d’agents est <strong>de</strong> rendre compte <strong>et</strong> d’expliquer l’évolution du<br />
prix <strong>de</strong>s actifs sur les marchés financiers. Dans le modèle du tâtonnement Walrasien, les agents signalent<br />
le prix auquel ils sont prêts à ach<strong>et</strong>er ou vendre, <strong>et</strong> après ajustement, la transaction est effectuée lorsqu’un<br />
prix d’équilibre est <strong>at</strong>teint (Samuelson 1941, par exemple). Si ce type d’approche semblait convenir à la<br />
modélis<strong>at</strong>ion <strong>de</strong>s marchés financiers au début du siècle lorsque les prix étaient fixés une fois par jour<br />
<strong>et</strong> non en continu, elle parait désormais très loin <strong>de</strong> refléter le fonctionnement <strong>de</strong>s places financières<br />
mo<strong>de</strong>rnes, où l’on peut penser que les prix se forment en <strong>de</strong>hors <strong>de</strong> tout équilibre. Il est donc nécessaire <strong>de</strong><br />
savoir comment relier l’évolution <strong>de</strong>s prix, hors <strong>de</strong> l’équilibre, à la répartition <strong>de</strong>s agents entre ach<strong>et</strong>eurs<br />
<strong>et</strong> ven<strong>de</strong>urs. Une simple applic<strong>at</strong>ion <strong>de</strong> la loi <strong>de</strong> l’offre <strong>et</strong> <strong>de</strong> la <strong>de</strong>man<strong>de</strong> nous indique que les prix doivent<br />
augmenter lorsque le nombre d’ach<strong>et</strong>eurs est supérieur au nombre <strong>de</strong> ven<strong>de</strong>urs <strong>et</strong> réciproquement, celuici<br />
doit diminuer lorsque les ven<strong>de</strong>urs sont en plus grand nombre que les ach<strong>et</strong>eurs.<br />
Donc, il apparaît qu’une quantité n<strong>at</strong>urelle dont dépend l’évolution <strong>de</strong>s prix est l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong>, que<br />
nous noterons ED dans la suite, <strong>et</strong> qui représente la différence entre le nombre d’ach<strong>et</strong>eurs <strong>et</strong> le nombre
5.1. Prix d’un actif <strong>et</strong> excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> 151<br />
<strong>de</strong> ven<strong>de</strong>urs. Plus précisément, l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> est la différence entre la <strong>de</strong>man<strong>de</strong> <strong>de</strong>s ach<strong>et</strong>eurs <strong>et</strong><br />
l’offre <strong>de</strong>s ven<strong>de</strong>urs. Ainsi, la vari<strong>at</strong>ion <strong>de</strong> prix dP (t), entre les instants t <strong>et</strong> t+dt, est fonction <strong>de</strong> l’excès<br />
<strong>de</strong> <strong>de</strong>man<strong>de</strong> ED(t), mais aussi a priori du prix lui-même P (t) :<br />
dP (t)<br />
dt<br />
= f(P (t), ED(t)), (5.1)<br />
où il suffit <strong>de</strong> choisir f(P (t), ED(t)) = P (t) · g(ED(t)), avec g(·) croissante <strong>et</strong> g(0) = 0 pour assurer<br />
la positivité du prix P (t) à tout instant :<br />
On peut noter que la quantité 1<br />
P (t)<br />
1<br />
P (t)<br />
dP (t)<br />
dt<br />
associé au prix P (t), si bien que r(t) est égal à g(ED(t)).<br />
= g(ED(t)). (5.2)<br />
· dP (t)<br />
dt n’est en fait rien d’autre que le ren<strong>de</strong>ment instantané r(t)<br />
Pour aller plus avant, il est nécessaire <strong>de</strong> spécifier un peu plus l’expression <strong>de</strong> la fonction g(·). En l’absence<br />
<strong>de</strong> toute autre inform<strong>at</strong>ion, l’approche la plus simple consiste à remarquer que pour les faibles<br />
excès <strong>de</strong> <strong>de</strong>man<strong>de</strong>, un développement au premier ordre <strong>de</strong> g(·), pourvu que g ′ (0) existe <strong>et</strong> soit non nulle,<br />
est déjà intéressant. En eff<strong>et</strong>, il vient alors<br />
1<br />
P (t)<br />
dP (t)<br />
dt<br />
ED(t)<br />
= , (5.3)<br />
λ<br />
où l’on a noté g ′ (0) = λ −1 . Ce facteur λ mesure la liquidité ou profon<strong>de</strong>ur du marché (Hausman, Lo <strong>et</strong><br />
MacKinlay 1992, Kempf <strong>et</strong> Korn 1999). Il représente l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> nécessaire pour faire bouger<br />
le ren<strong>de</strong>ment instantané d’une unité <strong>et</strong> mesure donc la sensibilité du prix aux fluctu<strong>at</strong>ions <strong>de</strong> l’excès <strong>de</strong><br />
<strong>de</strong>man<strong>de</strong>. La plupart <strong>de</strong>s modèles que nous rencontrerons dans la suite utilisent c<strong>et</strong>te représent<strong>at</strong>ion.<br />
Farmer (1998) a montré que c<strong>et</strong>te expression qui n’apparaît ici que comme une approxim<strong>at</strong>ion (linéaire)<br />
est en fait la rel<strong>at</strong>ion exacte entre le prix <strong>et</strong> l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> si l’on adm<strong>et</strong> que les ordres traités<br />
simultanément le sont au même prix. Ceci est réaliste pour <strong>de</strong>s marchés où le carn<strong>et</strong> d’ordre est traité en<br />
continu, comme sur EURONEXT par exemple, mais <strong>de</strong>vient peu crédible lorsque les ordres sont traités<br />
<strong>de</strong> manière décentralisée par différents mark<strong>et</strong>-makers comme sur le NYSE.<br />
En conclusion, la rel<strong>at</strong>ion (5.3), même si elle adm<strong>et</strong> certaines justific<strong>at</strong>ions théoriques, n’est généralement<br />
qu’approxim<strong>at</strong>ive, il est donc important <strong>de</strong> se <strong>de</strong>man<strong>de</strong>r quel est son domaine <strong>de</strong> validité. Les étu<strong>de</strong>s<br />
empiriques <strong>de</strong> Campbell <strong>et</strong> al. (1997), Chan <strong>et</strong> Fong (2000) ou Plerou, Gopikrishnan, Gabaix <strong>et</strong> Stanley<br />
(2001) ont montré que lorsque l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> <strong>de</strong>vient trop important, <strong>de</strong>s non-linéarités - traduisant<br />
un phénomène <strong>de</strong> s<strong>at</strong>ur<strong>at</strong>ion - <strong>de</strong>viennent appréciables. Ainsi, selon Plerou <strong>et</strong> al. (2001) la fonction g(·)<br />
peut être raisonnablement approchée par la fonction tangente hyperbolique. Notons, que dans ce cas,<br />
l’approxim<strong>at</strong>ion (5.3) reste valable pour les excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> modérés. Ceci n’est plus vrai dans la<br />
représent<strong>at</strong>ion <strong>de</strong> Zhang (1999) où la fonction g(·) est la fonction racine carrée, qui n’adm<strong>et</strong> pas <strong>de</strong><br />
dérivée en zéro. C<strong>et</strong>te représent<strong>at</strong>ion en terme <strong>de</strong> racine carrée apparaît lorsque l’on adm<strong>et</strong> qu’un ordre<br />
peut ne pas être exécuté instantanément, mais requière d’autant plus <strong>de</strong> temps que sa taille est importante.<br />
Son intérêt a récemment été confirmé par l’étu<strong>de</strong> empirique <strong>de</strong> Lillo, Farmer <strong>et</strong> Mantegna (2002).<br />
C<strong>et</strong>te <strong>de</strong>rnière remarque m<strong>et</strong> en lumière l’influence du temps d’exécution d’un ordre sur l’évolution du<br />
prix, ce qui constitue en fait une autre <strong>de</strong>s limit<strong>at</strong>ions <strong>de</strong> (5.3) (<strong>et</strong> <strong>de</strong> 5.1). En eff<strong>et</strong>, il est incorrect <strong>de</strong><br />
considérer que l’évolution du prix à l’instant t ne dépend que <strong>de</strong> l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> au même instant :<br />
Chordia <strong>et</strong> al. (2002) notamment ont montré que les gran<strong>de</strong>s fluctu<strong>at</strong>ions <strong>de</strong> prix étaient aussi affectées<br />
par les excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> passés. Plus précisément ceci semble là encore dépendre <strong>de</strong> l’organis<strong>at</strong>ion du<br />
marché, puisque selon Brown <strong>et</strong> al. (1997) c<strong>et</strong> eff<strong>et</strong> est plus important pour les marchés organisés autour
152 5. Approche comportementale <strong>de</strong>s marchés financiers : l’apport <strong>de</strong>s modèles d’agents<br />
<strong>de</strong> mark<strong>et</strong>-makers comme sur le NYSE que pour les marchés organisés autour d’un carn<strong>et</strong> d’ordres<br />
gérés inform<strong>at</strong>iquement comme sur EURONEXT ou sur le marché australien ASX.<br />
Malgré ces quelques limit<strong>at</strong>ions, la modélis<strong>at</strong>ion (5.3) est la plus souvent r<strong>et</strong>enue du fait <strong>de</strong> sa simplicité,<br />
ce qui est pleinement justifiable lorsque c<strong>et</strong>te simplific<strong>at</strong>ion perm<strong>et</strong> d’obtenir <strong>de</strong>s solutions analytiques<br />
pour le modèle considéré. Malheureusement, la plupart <strong>de</strong>s modèles d’agents que l’on rencontre dans<br />
la littér<strong>at</strong>ure doivent, du fait <strong>de</strong> leur complexité, être étudiés numériquement. Du coup, c<strong>et</strong>te approche<br />
simplific<strong>at</strong>rice, n’a <strong>de</strong> raison d’être que dans la mesure où elle facilite l’interprét<strong>at</strong>ion <strong>et</strong> la compréhension<br />
<strong>de</strong>s résult<strong>at</strong>s numériques en perm<strong>et</strong>tant <strong>de</strong> réduire l’ensemble <strong>de</strong>s paramètres du système à son minimum.<br />
5.2 Modèles d’opinion contre modèles <strong>de</strong> marché<br />
Outre la modélis<strong>at</strong>ion du prix <strong>de</strong>s actifs financiers, un <strong>de</strong>s objectifs <strong>de</strong> l’étu<strong>de</strong> <strong>de</strong>s marchés en terme <strong>de</strong><br />
modèle d’agents est <strong>de</strong> s’intéresser à la dynamique <strong>de</strong> répartition <strong>de</strong> ces agents en différents groupes :<br />
ach<strong>et</strong>eurs/ven<strong>de</strong>urs, fondamentalistes/chartistes ou encore imit<strong>at</strong>eurs/antagonistes C’est pour cela, qu’une<br />
fois fixées les règles <strong>de</strong> form<strong>at</strong>ion <strong>de</strong>s groupes, l’étu<strong>de</strong> <strong>de</strong>s modèles d’agents reprend les principes directeurs<br />
<strong>de</strong> l’étu<strong>de</strong> <strong>de</strong> n’importe quel système dynamique.<br />
Un point important dans la distinction entre les différents types <strong>de</strong> modèles d’agents est <strong>de</strong> savoir si la<br />
variable prix est endogène au modèle ou exogène. Par exogène, nous entendons que le prix est fixé par<br />
l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> <strong>et</strong> donc par la composition <strong>de</strong>s différents groupes d’agents, mais qu’il n’influe pas sur<br />
la form<strong>at</strong>ion <strong>de</strong> ces groupes. En ce sens, <strong>de</strong> tels modèles d’agents ne sont rien d’autres que <strong>de</strong>s modèles<br />
d’opinion, où le déséquilibre entre les différents groupes est traduit par un prix, à l’ai<strong>de</strong> <strong>de</strong> l’équ<strong>at</strong>ion<br />
(5.3). Au contraire, lorsque le prix <strong>de</strong>vient une variable endogène, la form<strong>at</strong>ion <strong>de</strong>s différents groupes<br />
d’agents dépend <strong>de</strong> celui-ci, notamment par l’intermédiaire <strong>de</strong> règles <strong>de</strong> transition entre groupes. Il y<br />
a donc, d’une part une action <strong>de</strong> la composition <strong>de</strong>s groupes sur le prix (via 5.3), <strong>et</strong> d’autre part une<br />
rétro-action du prix sur la composition <strong>de</strong>s groupes (via les règles <strong>de</strong> form<strong>at</strong>ion/transition). On est alors<br />
réellement en présence d’un modèle <strong>de</strong> marché.<br />
Parmi les nombreux modèles d’opinion ayant vu le jour, le modèle <strong>de</strong> Cont <strong>et</strong> Bouchaud (2000) présente<br />
l’avantage d’être d’une gran<strong>de</strong> généralité. En eff<strong>et</strong>, il ne spécifie nullement la n<strong>at</strong>ure <strong>de</strong>s interactions<br />
entre les agents - celle-ci étant extrêmement difficile à déterminer - <strong>et</strong> va même jusqu’à tirer profit <strong>de</strong><br />
c<strong>et</strong>te indétermin<strong>at</strong>ion en reprenant l’idée défendue par Kirman (1983) selon laquelle la communic<strong>at</strong>ion<br />
entre agents économiques peut être représentée comme un réseau <strong>de</strong> connexions alé<strong>at</strong>oires. Cont <strong>et</strong> Bouchaud<br />
(2000) formalisent alors le problème <strong>de</strong> sorte qu’il apparaisse exactement comme un problème <strong>de</strong><br />
percol<strong>at</strong>ion, phénomène critique bien connu <strong>de</strong>s m<strong>at</strong>hém<strong>at</strong>iciens <strong>et</strong> physiciens.<br />
Plus précisément, Cont <strong>et</strong> Bouchaud (2000) considèrent un ensemble <strong>de</strong> N agents choisissant d’être<br />
ach<strong>et</strong>eurs, ven<strong>de</strong>urs ou <strong>de</strong> ne pas prendre position sur le marché, avec <strong>de</strong>s probabilités respectivement<br />
égales à a, a <strong>et</strong> 1−2a, <strong>et</strong> a ∈]0, 1/2[. Ces agents créent <strong>de</strong>s coalitions au sein <strong>de</strong>squelles chaque individu<br />
possè<strong>de</strong> la même opinion, ce qui peut rendre compte par exemple <strong>de</strong> l’existence <strong>et</strong> <strong>de</strong> l’évolution <strong>de</strong><br />
groupes d’investisseurs associés au sein <strong>de</strong> fonds mutuels. Les coalitions se forment <strong>de</strong> manière alé<strong>at</strong>oire<br />
par appariement entre agents. Ainsi, on peut considérer que les agents sont situés sur les nœuds d’un<br />
graphe, <strong>et</strong> que <strong>de</strong>ux agents en rel<strong>at</strong>ion sont liés par une connexion entre les nœuds qu’ils occupent.<br />
Un groupe d’opinion est alors représenté par un amas <strong>de</strong> nœuds reliés entres eux par <strong>de</strong>s connexions<br />
binaires, <strong>et</strong> la distribution <strong>de</strong> taille <strong>de</strong> ces amas peut être obtenue à l’ai<strong>de</strong> <strong>de</strong> résult<strong>at</strong>s standards en <strong>théorie</strong><br />
<strong>de</strong>s graphes.<br />
En supposant que la probabilité que <strong>de</strong>ux agents soient appariés est p = c/N, ce qui garantit que le<br />
nombre moyen d’agents avec lequel un agent donné est relié reste fini (est égal à c) quand N tend
5.2. Modèles d’opinion contre modèles <strong>de</strong> marché 153<br />
vers l’infini, les résult<strong>at</strong>s <strong>de</strong> Erdös <strong>et</strong> Renyi (1960) perm<strong>et</strong>tent <strong>de</strong> montrer que la distribution <strong>de</strong> taille <strong>de</strong>s<br />
coalitions est une loi <strong>de</strong> puissance d’exposant 3/2 lorsque c = 1, <strong>et</strong> que c<strong>et</strong>te loi <strong>de</strong> puissance est tronquée<br />
exponentiellement lorsque c < 1. Ainsi, lorsque le système se trouve au seuil <strong>de</strong> percol<strong>at</strong>ion c = 1, ce<br />
modèle perm<strong>et</strong> <strong>de</strong> rendre compte <strong>de</strong> la décroissance en loi <strong>de</strong> puissance observée pour la distribution<br />
<strong>de</strong>s ren<strong>de</strong>ments <strong>de</strong>s actifs financiers. Cependant, l’accord n’est que qualit<strong>at</strong>if, puisque l’indice <strong>de</strong> queue<br />
généralement observé est plutôt <strong>de</strong> l’ordre <strong>de</strong> 3 − 4, alors que ce modèle prédit un exposant égal à 3/2.<br />
Une <strong>de</strong>s critiques élevée contre ce modèle est qu’a priori, le système n’a aucune raison <strong>de</strong> se trouver<br />
au seuil c = 1. C’est pour cela que Stauffer <strong>et</strong> Sorn<strong>et</strong>te (1999) ont proposé que le paramètre c puisse<br />
varier, ce qui perm<strong>et</strong> <strong>de</strong> rendre compte <strong>de</strong> l’évolution <strong>de</strong>s opinions au cours du temps. Moyennant cela,<br />
la distribution <strong>de</strong>s ren<strong>de</strong>ments est une loi <strong>de</strong> puissance dont l’exposant peut être ajusté pour s<strong>at</strong>isfaire les<br />
observ<strong>at</strong>ions expérimentales. De plus, c<strong>et</strong>te version dynamique du modèle <strong>de</strong> Cont <strong>et</strong> Bouchaud (2000)<br />
est comp<strong>at</strong>ible avec la lente décroissante <strong>de</strong> la corrél<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité <strong>de</strong>s ren<strong>de</strong>ments.<br />
Ce modèle est clairement un modèle d’opinion puisque les décisions <strong>de</strong>s agents sont indépendantes du<br />
prix <strong>de</strong> l’actif qu’ils s’échangent. En particulier, le fait que les agents choisissent <strong>de</strong> placer un ordre sur<br />
le marché ou non ne dépend pas du prix du marché, ce qui est assez peu réaliste. Cela montre comment<br />
rendre <strong>de</strong> façon simple le prix endogène dans ce type <strong>de</strong> modèle, perm<strong>et</strong>tant ainsi <strong>de</strong> d’obtenir un réel<br />
modèle <strong>de</strong> marché. Il suffit pour cela <strong>de</strong> rendre les probabilités d’entrer ou non dans le marché <strong>et</strong> d’être<br />
ach<strong>et</strong>eur ou ven<strong>de</strong>ur dépendantes du prix <strong>de</strong> l’actif. A notre connaissance cela n’a pas encore été fait<br />
pour ce type <strong>de</strong> modèle.<br />
C<strong>et</strong>te métho<strong>de</strong> d’endogénéis<strong>at</strong>ion du prix est généralement celle suivie dans tous les modèles <strong>de</strong> marché.<br />
Ainsi, dans les modèles proposés par Kirman (1991), Brock (1993), Lux (1995) ou encore Farmer (1998)<br />
parmi bien d’autres, il existe <strong>de</strong>ux classes d’agents : certains sont supposés fondamentalistes <strong>et</strong> prennent<br />
donc position en fonction <strong>de</strong> l’écart entre le prix du marché <strong>et</strong> un prix fondamental, tandis que d’autres<br />
sont chartistes <strong>et</strong> suivent la tendance qu’ils essaient <strong>de</strong> détecter dans les cours. Le seuil <strong>de</strong> déclenchement<br />
du placement d’ordre est donc fonction du prix du marché.<br />
Dans le même esprit, mais à un niveau macroscopique c<strong>et</strong>te fois, Bouchaud <strong>et</strong> Cont (1998) - dans une<br />
version linéaire - puis I<strong>de</strong> <strong>et</strong> Sorn<strong>et</strong>te (2002) - dans une généralis<strong>at</strong>ion non linéaire - ont repris c<strong>et</strong>te<br />
approche afin <strong>de</strong> décrire non plus l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> individuelle mais l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> agrégée <strong>de</strong>s<br />
agents, ce qui leur a permis - dans l’approche non linéaire - <strong>de</strong> rendre compte <strong>de</strong> la croissance superexponentielle<br />
<strong>de</strong>s bulles spécul<strong>at</strong>ives ainsi que <strong>de</strong>s oscill<strong>at</strong>ions log-périodiques précédant les krachs.<br />
Il convient <strong>de</strong> noter que dans tous ces modèles que nous venons <strong>de</strong> citer, les fondamentalistes jouent le<br />
rôle d’une force <strong>de</strong> rappel qui tend à ramener le prix <strong>de</strong> marché vers le prix fondamental, alors que les<br />
chartistes ten<strong>de</strong>nt à l’en écarter, <strong>et</strong> ce d’autant plus vite que le prix <strong>de</strong> marché augmente rapi<strong>de</strong>ment.<br />
Ceci donne alors naissance à une succession <strong>de</strong> phases calmes, où les fondamentalistes sont en majorité,<br />
alternant avec <strong>de</strong>s phases <strong>de</strong> forte vol<strong>at</strong>ilité, où les chartistes dominent le marché.<br />
Bien d’autres exemples <strong>de</strong> modèles, tel celui <strong>de</strong> Brock <strong>et</strong> Hommes (1997) ou celui <strong>de</strong> Lux (1998), puis<br />
Lux <strong>et</strong> Marchesi (1999), reposent sur ces mêmes idées, mais dans tous les cas, l’endogénéis<strong>at</strong>ion du prix<br />
s’effectue <strong>de</strong> la même manière : les seuils <strong>de</strong> déclenchement <strong>de</strong>s actions <strong>de</strong>s agents sur le marché sont<br />
rendus dépendants du prix, ce qui est à la fois la manière la plus simple <strong>et</strong> la plus n<strong>at</strong>urelle <strong>de</strong> concevoir<br />
la chose.
154 5. Approche comportementale <strong>de</strong>s marchés financiers : l’apport <strong>de</strong>s modèles d’agents<br />
5.3 Modèles à agents adapt<strong>at</strong>ifs contre modèles à agents non adapt<strong>at</strong>ifs<br />
Parmi les modèles que nous venons <strong>de</strong> présenter, certains sont dits à agents non adapt<strong>at</strong>ifs <strong>et</strong> d’autres à<br />
agents adapt<strong>at</strong>ifs. Pour illustrer c<strong>et</strong>te distinction, considérons par exemple le modèle <strong>de</strong> Farmer (1998).<br />
Les agents y sont soit fondamentalistes soit chartistes <strong>et</strong> le nombre d’agents appartenant à chacune <strong>de</strong><br />
ces c<strong>at</strong>égories, fixé au départ, n’évolue pas au cours du temps. Ainsi, les agents choisissent certes d’être<br />
ach<strong>et</strong>eur ou ven<strong>de</strong>ur selon le prix du marché (le modèle <strong>de</strong> Farmer (1998) est un modèle <strong>de</strong> marché<br />
<strong>et</strong> non un simple modèle d’opinion), mais ils ne peuvent pas adapter leur str<strong>at</strong>égie - fondamentaliste<br />
ou chartiste- aux modific<strong>at</strong>ions <strong>de</strong> leur environnement. Le modèle <strong>de</strong> Farmer (1998) est donc qualifié <strong>de</strong><br />
modèle à agents non adapt<strong>at</strong>ifs. Le modèle <strong>de</strong> Bouchaud <strong>et</strong> Cont (1998) <strong>et</strong> son extension proposée par I<strong>de</strong><br />
<strong>et</strong> Sorn<strong>et</strong>te (2002), <strong>de</strong> même que le modèle <strong>de</strong> Lévy, Lévy <strong>et</strong> Solomon (1995), par exemple, appartiennent<br />
eux aussi à c<strong>et</strong>te c<strong>at</strong>égorie.<br />
Au contraire, le modèle <strong>de</strong> Lux (1998) ou Lux <strong>et</strong> Marchesi (1999), fait partie <strong>de</strong>s modèles à agents adapt<strong>at</strong>ifs.<br />
En eff<strong>et</strong>, à la différence du modèle <strong>de</strong> Farmer (1998), les différents agents communiquent entre<br />
eux <strong>et</strong> comparent les résult<strong>at</strong>s <strong>de</strong>s différentes str<strong>at</strong>égies afin <strong>de</strong> se déterminer. Un agent peut donc passer<br />
d’une str<strong>at</strong>égie chartiste à fondamentaliste (ou vice-versa) s’il a l’espoir que cela lui perm<strong>et</strong>te <strong>de</strong> réaliser<br />
un gain plus important. Il est donc sensible à son environnement <strong>et</strong> tend notamment à imiter les agents<br />
avec lesquels il rentre en contact si cela lui est profitable. Ces changements <strong>de</strong> str<strong>at</strong>égie perm<strong>et</strong>tent <strong>de</strong><br />
rendre compte <strong>de</strong>s bouffées <strong>de</strong> vol<strong>at</strong>ilité caractéristiques <strong>de</strong>s séries financières. En eff<strong>et</strong>, lorsque les fondamentalistes<br />
sont majoritaires, le marché est calme <strong>et</strong> les cours ne présentent que <strong>de</strong>s fluctu<strong>at</strong>ions <strong>de</strong><br />
faibles amplitu<strong>de</strong>s autour <strong>de</strong> leur prix fondamental. Au contraire, dès que les chartistes sont en nombre<br />
suffisant, une tendance peut apparaître, puis s’amplifier, conduisant alors à une augment<strong>at</strong>ion <strong>de</strong> la vol<strong>at</strong>ilité,<br />
qui ne r<strong>et</strong>ombera que lorsque l’enthousiasme <strong>de</strong>s chartistes aura disparu <strong>et</strong> que les fondamentalistes<br />
reprendront le <strong>de</strong>ssus. Ainsi, ce modèle perm<strong>et</strong> d’interpréter en terme microstructurel pourquoi les séries<br />
financières présentent <strong>de</strong>s bouffées <strong>de</strong> vol<strong>at</strong>ilité, <strong>et</strong> justifie l’idée selon laquelle la dynamique <strong>de</strong>s marchés<br />
financiers présente <strong>de</strong>s changements <strong>de</strong> régime (Susmel 1996).<br />
Le modèle <strong>de</strong> Cont <strong>et</strong> Bouchaud (2000) dans la version généralisée <strong>de</strong> Stauffer <strong>et</strong> Sorn<strong>et</strong>te (1999) peut<br />
lui aussi être considéré, à la limite, comme un modèle à agents adapt<strong>at</strong>ifs. En eff<strong>et</strong>, même si le processus<br />
d’adapt<strong>at</strong>ion n’est pas clairement précisé <strong>et</strong> est spécifié <strong>de</strong> manière totalement ad hoc, les agents peuvent<br />
changer <strong>de</strong> coalition au cours du temps, marquant ainsi leur adhésion à telle ou telle opinion <strong>et</strong> conduisant<br />
à une évolution <strong>de</strong> l’opinion globale du système.<br />
Dans les <strong>de</strong>ux modèles précé<strong>de</strong>nts, les agents n’ont le choix qu’entre un p<strong>et</strong>it nombre <strong>de</strong> str<strong>at</strong>égies,<br />
l’hétérogénéité <strong>de</strong> ces modèles est donc faible. C’est pourquoi il est important <strong>de</strong> mentionner <strong>de</strong>ux autres<br />
types <strong>de</strong> modèles à agents adapt<strong>at</strong>ifs, complètement différents <strong>de</strong> tous ceux évoqués jusqu’ici, à savoir<br />
les jeux <strong>de</strong> minorité <strong>et</strong> le modèle d’agents “génétiques” développé au Santa-Fe Intitut (Palmer, Arthur,<br />
Holland, LeBaron <strong>et</strong> Taylor 1994, Arthur, Holland, LeBaron, Palmer <strong>et</strong> Taylor 1997, LeBaron, Arthur <strong>et</strong><br />
Palmer 1999). Dans ce <strong>de</strong>rnier modèle, les agents sont représentés par <strong>de</strong>s algorithmes génétiques, p<strong>et</strong>its<br />
programmes inform<strong>at</strong>iques imitant le processus <strong>de</strong> sélection n<strong>at</strong>urelle <strong>de</strong>s gènes en compétition pour leur<br />
survie <strong>et</strong> leur développement. Ainsi, ces agents, doués <strong>de</strong> facultés d’adapt<strong>at</strong>ion très poussées, forment<br />
<strong>de</strong>s prédictions sur les prix futurs <strong>et</strong> placent <strong>de</strong>s ordres sur le marché en fonction <strong>de</strong> leurs prévisions. Les<br />
règles <strong>de</strong> sélection <strong>et</strong> d’évolution <strong>de</strong>s agents sont ensuite déterminées par leur performance, <strong>de</strong> sorte qu’à<br />
chaque génér<strong>at</strong>ion, les agents les moins performants disparaissent au profit <strong>de</strong>s plus performants. Il est<br />
alors observé que dans certaines conditions, ces agents apprennent à coopérer <strong>et</strong> <strong>de</strong> leurs actions individuelles<br />
émerge un équilibre r<strong>at</strong>ionnel dynamique conforme à ce qui est prévu par la <strong>théorie</strong> économique.<br />
Le modèle <strong>de</strong> Farmer (1998) que nous avons évoqué plus haut constitue en fait une version simplifiée <strong>de</strong><br />
c<strong>et</strong>te approche, qui exploite l’analogie entre le marché <strong>et</strong> un écosystème où se développe une compétition<br />
entre diverses str<strong>at</strong>égies favorisant ainsi l’émergence <strong>de</strong> nouvelles métho<strong>de</strong>s tirant parti <strong>de</strong>s inefficacités
5.4. Conséquences <strong>de</strong>s phénomènes d’imit<strong>at</strong>ion <strong>et</strong> d’antagonisme... 155<br />
<strong>de</strong>s anciennes qu’elles finissent par remplacer.<br />
Les modèles <strong>de</strong> minorité, introduits par Chall<strong>et</strong> <strong>et</strong> Zhang (1997), procè<strong>de</strong> d’une toute autre approche.<br />
Considèrons N agents en compétition pour une ressource limitée <strong>et</strong> dont les choix se résument à <strong>de</strong>ux<br />
altern<strong>at</strong>ives mutuellement exclusives, que l’on pourra noter 0 ou 1. Chaque agent effectue un choix, <strong>et</strong> les<br />
gagnants sont ceux qui sont en minorité (on peut reconnaître ici le célèbre problème du bar d’El Farol introduit<br />
par Arthur (1994)). Un tel modèle est pertinent dans un contexte financier, car un agent qui déci<strong>de</strong><br />
<strong>de</strong> vendre, par exemple, ne va empocher un bénéfice que si la majorité <strong>de</strong>s autres agents sont ach<strong>et</strong>eurs,<br />
<strong>de</strong> sorte que l’excès <strong>de</strong> <strong>de</strong>man<strong>de</strong> soit positif <strong>et</strong> que le prix augmente (voir Chall<strong>et</strong>, Chessa, Marsili <strong>et</strong><br />
Zhang (2001) pour une discussion <strong>de</strong> l’applic<strong>at</strong>ion <strong>de</strong>s jeux <strong>de</strong> minorité aux marchés financiers).<br />
Concrètement, le jeu est répété un grand nombre <strong>de</strong> fois, si bien que chaque agent dispose <strong>de</strong> la même<br />
inform<strong>at</strong>ion : la succession <strong>de</strong>s décisions minoritaires, soit une séquence <strong>de</strong> 0 <strong>et</strong> <strong>de</strong> 1. En supposant que<br />
les agents ne disposent que d’une mémoire <strong>de</strong> taille limitée m, tous les agents connaissent donc les m<br />
<strong>de</strong>rniers choix gagnants. Une séquence <strong>de</strong> longueur m représente donc l’histoire <strong>de</strong>s événements <strong>et</strong> il y<br />
a évi<strong>de</strong>mment 2m histoires possibles. Chaque agent établit <strong>de</strong>s str<strong>at</strong>égies visant à prédire l’ét<strong>at</strong> futur du<br />
système en se fondant sur l’histoire passée. Il existe donc en tout 22m str<strong>at</strong>égies. Tous les agents possè<strong>de</strong>nt<br />
un même nombre <strong>de</strong> str<strong>at</strong>égies, ces str<strong>at</strong>égies étant différentes d’un agent à l’autre. C<strong>et</strong>te approche perm<strong>et</strong><br />
<strong>de</strong> rendre compte <strong>de</strong> l’extraordinaire diversité <strong>de</strong>s acteurs économiques présents sur les marchés<br />
réels. En eff<strong>et</strong>, ceux-ci ont <strong>de</strong>s objectifs différents, aussi bien en terme <strong>de</strong> gains espérés que d’horizon<br />
d’investissements mais aussi d’aversion au risque <strong>et</strong> présentent donc <strong>de</strong>s str<strong>at</strong>égies extrêmement variées.<br />
Les résult<strong>at</strong>s fournis par ce type <strong>de</strong> modèles sont s<strong>at</strong>isfaisants dans la mesure où ils perm<strong>et</strong>tent <strong>de</strong> rendre<br />
compte <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ment en loi <strong>de</strong> puissance <strong>et</strong> <strong>de</strong> la lente décroissance <strong>de</strong> la corrél<strong>at</strong>ion <strong>de</strong><br />
la vol<strong>at</strong>ilité lorsque le rapport m/N est voisin d’une certaine valeur critique. Cependant, comme relevé<br />
récemment par An<strong>de</strong>rsen <strong>et</strong> Sorn<strong>et</strong>te (2002b) <strong>et</strong> Giardina <strong>et</strong> Bouchaud (2002), le principe <strong>de</strong> minorité<br />
n’est pas toujours le plus réaliste. En eff<strong>et</strong>, un agent a intérêt à se trouver dans la minorité lorsqu’un<br />
renversement <strong>de</strong> tendance est observé sur les prix, mais lorsque la tendance se poursuit, il faut être dans<br />
la majorité pour profiter <strong>de</strong> celle-ci. Dans le cadre <strong>de</strong> ce modèle, la trajectoire <strong>de</strong> la richesse <strong>de</strong>s agents<br />
est notablement différente <strong>de</strong> celle observée dans les jeux <strong>de</strong> minorités standards, mais ce type <strong>de</strong> modèle<br />
rend tout aussi bien compte <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ment à queue épaisse que <strong>de</strong> la corrél<strong>at</strong>ion à longue<br />
portée <strong>de</strong> la vol<strong>at</strong>ilité.<br />
5.4 Conséquences <strong>de</strong>s phénomènes d’imit<strong>at</strong>ion <strong>et</strong> d’antagonisme sur la<br />
forme <strong>de</strong> la trajectoire <strong>de</strong>s prix d’un actif.<br />
Tous les modèles microscopiques que nous avons décrits jusqu’à présent perm<strong>et</strong>tent <strong>de</strong> rendre compte<br />
<strong>de</strong> la distribution <strong>de</strong>s ren<strong>de</strong>ments ainsi que <strong>de</strong> la corrél<strong>at</strong>ion à longue portée <strong>de</strong> la vol<strong>at</strong>ilité, qui sont les<br />
faits stylisés les plus marquants observés au suj<strong>et</strong> <strong>de</strong>s rentabilités boursières. Pour autant, aucun <strong>de</strong> ces<br />
modèles n’est capable d’expliquer la croissance super-exponentielle <strong>de</strong>s prix que l’on observe lorsqu’une<br />
bulle spécul<strong>at</strong>ive se développe.<br />
Afin <strong>de</strong> pouvoir modéliser ce phénomène, il convient <strong>de</strong> comprendre ce que recouvre une croissance<br />
super-exponentielle du prix d’un actif. Lorsque le ren<strong>de</strong>ment d’un actif est constant, son prix augmente<br />
n<strong>at</strong>urellement <strong>de</strong> façon exponentielle, puisque l’actualis<strong>at</strong>ion est un processus multiplic<strong>at</strong>if. Une croissance<br />
super-exponentielle traduit donc le fait que le ren<strong>de</strong>ment <strong>de</strong> l’actif considéré augmente au fur <strong>et</strong><br />
à mesure que son prix lui-même augmente. Un mécanisme possible est le suivant : lorsqu’une tendance<br />
haussière suffisamment forte apparaît, le nombre d’agents ach<strong>et</strong>eurs augmente, <strong>at</strong>tiré par la promesse <strong>de</strong><br />
gains importants, ce qui en r<strong>et</strong>our intensifie la tendance du fait <strong>de</strong> la forte pression ach<strong>et</strong>euse, validant du
156 5. Approche comportementale <strong>de</strong>s marchés financiers : l’apport <strong>de</strong>s modèles d’agents<br />
même coup la croyance en un gain important, <strong>et</strong> <strong>at</strong>tire en conséquence toujours plus d’ach<strong>et</strong>eurs.<br />
Nous pouvons donc conclure qu’un eff<strong>et</strong> d’aubaine ou un eff<strong>et</strong> <strong>de</strong> foule semble tout à fait à même <strong>de</strong> justifier<br />
la croissance super-exponentielle <strong>de</strong>s prix observés en présence <strong>de</strong> bulles spécul<strong>at</strong>ives. Ces eff<strong>et</strong>s <strong>de</strong><br />
foules ont été observés sur les marchés spécul<strong>at</strong>ifs (Arthur 1987, Orléan 1992, Shiller 2000, notamment)<br />
<strong>et</strong> traduisent le phénomène d’imit<strong>at</strong>ion entre agents économiques. Il est à noter que ce phénomène d’imit<strong>at</strong>ion<br />
est parfaitement r<strong>at</strong>ionnel, comme le souligne Orléan (1989), <strong>et</strong> découle simplement <strong>de</strong>s règles<br />
d’apprentissage bayésien.<br />
Si l’on s’en tient au comportement mimétique <strong>et</strong> auto-référentiel exposé ci-<strong>de</strong>ssus, la bulle spécul<strong>at</strong>ive<br />
n’a aucune raison <strong>de</strong> prendre fin. Il est donc raisonnable <strong>de</strong> penser qu’à un certain moment, les agents<br />
réalisent que les prix sont <strong>de</strong>venus trop éloignés <strong>de</strong>s fondamentaux <strong>de</strong> l’économie ou que les liquidités<br />
vont venir à manquer <strong>et</strong> donc choisissent <strong>de</strong> vendre pour engranger leurs bénéfices tant que l’afflux<br />
d’ach<strong>et</strong>eurs <strong>de</strong>meure suffisamment important pour pouvoir absorber leurs ordres <strong>de</strong> vente sans que les<br />
cours ne chutent. Ainsi, durant c<strong>et</strong>te phase, certains agents choisissent <strong>de</strong> prendre le marché à contrecourant,<br />
<strong>et</strong> présentent un comportement antagoniste par rapport à l’opinion majoritaire à ce moment<br />
là. On peut noter que dans l’esprit, ceci est assez proche du modèle <strong>de</strong> An<strong>de</strong>rsen <strong>et</strong> Sorn<strong>et</strong>te (2002b)<br />
évoqué plus haut, <strong>et</strong> il apparaît clairement pourquoi le principe <strong>de</strong> minorité ne saurait parvenir à décrire<br />
complètement le fonctionnement <strong>de</strong>s marchés financiers.<br />
C<strong>et</strong>te approche heuristique perm<strong>et</strong> <strong>de</strong> comprendre que les comportements mimétiques <strong>et</strong> antagonistes<br />
<strong>de</strong>s agents peuvent suffire à expliquer la croissance super-exponentielle <strong>de</strong>s bulles spécul<strong>at</strong>ives, ce qui<br />
nous a conduit à développer le modèle simple exposé dans Corcos, Eckmann, Malaspinas, Malevergne <strong>et</strong><br />
Sorn<strong>et</strong>te (2002). Dans ce modèle, nous considérons un marché où s’échange un seul actif <strong>et</strong> un ensemble<br />
<strong>de</strong> N agents pouvant être soit ach<strong>et</strong>eurs soit ven<strong>de</strong>urs. Chacun interroge m <strong>de</strong> ses voisins <strong>et</strong> peut choisir<br />
<strong>de</strong> suivre l’opinion majoritaire qu’il a relevée ou bien d’en prendre le contre-pied. Par exemple, un agent<br />
ach<strong>et</strong>eur changera d’avis <strong>et</strong> <strong>de</strong>viendra ven<strong>de</strong>ur si la proportion d’agents ven<strong>de</strong>urs parmi les m agents<br />
qu’il a interrogés dépasse un certain seuil ρhb, ce qui correspond à un comportement mimétique, ou si la<br />
proportion d’agents ach<strong>et</strong>eurs est jugée trop importante <strong>et</strong> dépasse un autre seuil ρhh, ce qui correspond<br />
à un comportement antagoniste, tandis qu’il conserve son opinion si aucun <strong>de</strong> ces seuils n’est dépassé.<br />
Nous montrons alors que, comme <strong>at</strong>tendu, le prix présente <strong>de</strong>s bulles spécul<strong>at</strong>ives dont la croissance est<br />
d’abord exponentielle, puis accélère <strong>et</strong> tend à présenter une singularité en temps fini. En fait, à cause<br />
du mécanisme <strong>de</strong> ré-injection lié à l’existence <strong>de</strong> comportements antagonistes, la bulle écl<strong>at</strong>e avant que<br />
la divergence ait lieu. Ainsi donc ce modèle m<strong>et</strong> en évi<strong>de</strong>nce l’importance <strong>de</strong>s phénomènes d’imit<strong>at</strong>ion<br />
<strong>et</strong> d’antagonisme, qui à eux seuls suffisent à expliquer la croissance super-exponentielle <strong>de</strong>s bulles<br />
spécul<strong>at</strong>ives suivies par <strong>de</strong>s krachs. Par ailleurs, ce modèle rend aussi compte, <strong>de</strong> manière qualit<strong>at</strong>ive,<br />
<strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments à queues épaisses. Cependant, il conduit à <strong>de</strong>s ren<strong>de</strong>ments beaucoup<br />
trop corrélés pour être réalistes.<br />
Ce défaut est dû au caractère faiblement stochastique <strong>de</strong> notre modèle <strong>et</strong> peut être facilement corrigé.<br />
En fait, d’après notre classific<strong>at</strong>ion, ce modèle n’est pas un modèle <strong>de</strong> marché mais un simple modèle<br />
d’opinion. Il convient donc d’en endogénéiser le prix, ce qui est réalisable par une généralis<strong>at</strong>ion simple<br />
du modèle précé<strong>de</strong>nt. Pour cela, nous pouvons considérer que les N agents ne sont pas tous en position<br />
sur le marché à un instant donné. Plus précisément, chaque agent estime la valeur du prix fondamental<br />
<strong>de</strong> l’actif <strong>et</strong> déci<strong>de</strong> d’entrer sur le marché si (<strong>et</strong> seulement si) l’écart entre le prix <strong>de</strong> marché <strong>de</strong> c<strong>et</strong> actif<br />
<strong>et</strong> son fondamental dépasse un certain seuil. Etant entré sur le marché, l’agent interroge m agents euxmêmes<br />
sur le marché <strong>et</strong> peut alors se comporter comme un fondamentaliste s’il ne note aucune majorité<br />
forte perm<strong>et</strong>tant <strong>de</strong> soutenir une tendance ou au contraire, peut choisir <strong>de</strong> suivre c<strong>et</strong>te tendance s’il pense<br />
qu’elle est viable <strong>et</strong> donc qu’il peut en tirer profit.<br />
Par rapport au modèle originel, c<strong>et</strong>te généralis<strong>at</strong>ion apporte un couplage entre le volume <strong>et</strong> la vol<strong>at</strong>ilité,
5.5. Conclusion 157<br />
ce qui induit une beaucoup plus gran<strong>de</strong> variabilité dans la dynamique <strong>de</strong>s prix. Nous espérons donc que<br />
ceci perm<strong>et</strong>tra <strong>de</strong> guérir les carences observées sur le modèle d’origine.<br />
5.5 Conclusion<br />
Les modèles d’agents en interaction se sont développés pour palier aux insuffisances <strong>de</strong> l’approche standard<br />
en terme d’agent économique représent<strong>at</strong>if, qui n’est rien d’autre qu’une approxim<strong>at</strong>ion <strong>de</strong> champ<br />
moyen m<strong>et</strong>tant l’accent sur l’équilibre. C<strong>et</strong>te nouvelle façon d’abor<strong>de</strong>r la modélis<strong>at</strong>ion <strong>de</strong>s marchés financiers<br />
semble prom<strong>et</strong>teuse <strong>et</strong> à notre avis présente un double intérêt.<br />
D’une part, elle offre un moyen <strong>de</strong> cerner les comportements minimaux <strong>de</strong>s agents nécessaires à la<br />
restitution <strong>de</strong>s faits stylisés <strong>de</strong>s séries financières. En cela, elle perm<strong>et</strong> d’abor<strong>de</strong>r l’aspect comportemental<br />
<strong>de</strong>s marchés financiers, aspect qui nous semble essentiel à la détermin<strong>at</strong>ion <strong>de</strong> certaines composantes du<br />
risque associé à l’activité financière. En eff<strong>et</strong>, l’un <strong>de</strong>s apports <strong>de</strong>s modèles d’agents en interaction est <strong>de</strong><br />
pouvoir traiter c<strong>et</strong>te dimension comportementale non plus du seul point <strong>de</strong> vue qualit<strong>at</strong>if mais également<br />
<strong>de</strong> manière quantit<strong>at</strong>ive.<br />
D’autre part, une fois un modèle crédible établi, il peut perm<strong>et</strong>tre <strong>de</strong> réaliser <strong>de</strong>s expériences, visant par<br />
exemple à étudier l’influence <strong>de</strong> l’évolution <strong>de</strong> l’environnement réglementaire, mais aussi <strong>de</strong> simuler<br />
divers scénarii <strong>et</strong> donc apporter <strong>de</strong> nouveaux moyens <strong>de</strong> <strong>gestion</strong> <strong>de</strong>s risques.
158 5. Approche comportementale <strong>de</strong>s marchés financiers : l’apport <strong>de</strong>s modèles d’agents
Chapitre 6<br />
Comportements mimétiques <strong>et</strong><br />
antagonistes : bulles hyperboliques, krachs<br />
<strong>et</strong> chaos<br />
Les comportements mimétiques <strong>et</strong> antagonistes sont <strong>de</strong>ux <strong>at</strong>titu<strong>de</strong>s opposées typiques <strong>de</strong>s investisseurs<br />
sur les marchés financiers. Nous introduisons un modèle simple nous perm<strong>et</strong>tant d’étudier le rôle <strong>de</strong><br />
chacun <strong>de</strong> ces comportements sur un marché financier où les agents ne peuvent choisir qu’entre <strong>de</strong>ux<br />
altern<strong>at</strong>ives : être ach<strong>et</strong>eur ou ven<strong>de</strong>ur. Chaque agent ach<strong>et</strong>eur (resp. ven<strong>de</strong>ur) interroge m “amis” <strong>et</strong><br />
change d’opinion pour <strong>de</strong>venir ven<strong>de</strong>ur (resp. ach<strong>et</strong>eur) si<br />
– au moins m · ρhb (resp. m · ρbh) parmi les m agents consultés sont ven<strong>de</strong>urs (resp. ach<strong>et</strong>eurs),<br />
– ou si au moins m · ρhh > m · ρhb (resp. m · ρbb > m · ρbh) parmi les m agents consultés sont ach<strong>et</strong>eurs<br />
(resp. ven<strong>de</strong>urs).<br />
Ces conditions correspon<strong>de</strong>nt respectivement à <strong>de</strong>s comportements mimétique <strong>et</strong> antagoniste.<br />
Dans la limite où le nombre N d’agents est infini, la dynamique <strong>de</strong> la fraction d’agents ach<strong>et</strong>eurs est<br />
déterministe <strong>et</strong> présente un comportement chaotique dans un domaine signific<strong>at</strong>if <strong>de</strong> l’espace <strong>de</strong>s paramètres<br />
{ρhb, ρbh, ρhh, ρbb, m}. Une trajectoire chaotique typique est caractérisée par <strong>de</strong>s phases intermittentes<br />
<strong>de</strong> chaos, <strong>de</strong> comportement quasi-périodique <strong>et</strong> <strong>de</strong> développement super-exponentiel <strong>de</strong> bulles<br />
ponctuées par <strong>de</strong>s krachs. Une bulle commence par croître initialement à un rythme exponentiel puis<br />
accélère jusqu’à conduire à une singularité en temps fini. Le mécanisme <strong>de</strong> réinjection dû à la présence<br />
d’agents antagonistes introduit un eff<strong>et</strong> <strong>de</strong> taille finie, prévenant c<strong>et</strong>te singularité <strong>et</strong> conduisant au chaos.<br />
Nous documentons les principaux faits stylisés <strong>de</strong> ce modèle dans les cas symétrique <strong>et</strong> asymétrique. Il<br />
est à noter que ce modèle est l’un <strong>de</strong>s rares modèles d’agents donnant naissance à une dynamique non<br />
triviale dans la limite “thermodynamique” - c’est-à-dire lorsque le nombre N d’agents tend vers l’infini.<br />
Nous discutons aussi du cas où le nombre d’agents est fini, ce qui introduit une source endogène <strong>de</strong> bruit<br />
qui se superpose à la dynamique chaotique.<br />
Reprint from : A. Corcos, J.-P. Eckmann, A. Malaspinas, Y. Malevergne <strong>et</strong> D. Sorn<strong>et</strong>te (2002), “Imit<strong>at</strong>ion<br />
and contrarian behavior : hyperbolic bubbles, crashes and chaos”, Quantit<strong>at</strong>ive Finance 2, 264-281.<br />
Err<strong>at</strong>um : page 164, l’équ<strong>at</strong>ion Q = prob({x < m(1 − ρbh)} ∪ {x > mρbb}) doit être remplacée par<br />
Q = prob({x < m(1 − ρbb)} ∪ {x > mρbh}).<br />
159
160 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos
RESEARCH PAPER Q UANTITATIVE F INANCE V OLUME 2 (2002) 264–281<br />
quant.iop.org I NSTITUTE OF P HYSICS P UBLISHING<br />
Imit<strong>at</strong>ion and contrarian behaviour:<br />
hyperbolic bubbles, crashes and chaos<br />
A Corcos 1 , J-P Eckmann 2,3 , A Malaspinas 2 , Y Malevergne 4,5<br />
and D Sorn<strong>et</strong>te 4,6<br />
1 CRIISEA, Université <strong>de</strong> Picardie, BP 2716, 80027 Amiens, France<br />
2 Dépt. <strong>de</strong> Physique Théorique, Université <strong>de</strong>Genève, CH-1211 Genève 4,<br />
Switzerland<br />
3 Section <strong>de</strong> M<strong>at</strong>hém<strong>at</strong>iques, Université <strong>de</strong>Genève, CH-1211 Genève 4,<br />
Switzerland<br />
4 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée, CNRS UMR6622 and<br />
Université <strong>de</strong> Nice-Sophia Antipolis, BP 71, Parc Valrose,<br />
06108 Nice Ce<strong>de</strong>x 2, France<br />
5 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances—Université Lyon I, 43,<br />
Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x, France<br />
6 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and<br />
Space Science, University of California, Los Angeles, CA 90095, USA<br />
Received 25 September 2001, in final form 26 November 2001<br />
Published 2 August 2002<br />
Online <strong>at</strong> stacks.iop.org/Quant/2/264<br />
Abstract<br />
Imit<strong>at</strong>ive and contrarian behaviours are the two typical opposite <strong>at</strong>titu<strong>de</strong>s of<br />
investors in stock mark<strong>et</strong>s. We introduce a simple mo<strong>de</strong>l to investig<strong>at</strong>e their<br />
interplay in a stock mark<strong>et</strong> where agents can take only two st<strong>at</strong>es, bullish or<br />
bearish. Each bullish (bearish) agent polls m ‘friends’ and changes her<br />
opinion to bearish (bullish) if (i) <strong>at</strong> least mρhb (mρbh) among the m agents<br />
inspected are bearish (bullish) or (ii) <strong>at</strong> least mρhh >mρhb (mρbb >mρbh)<br />
among the m agents inspected are bullish (bearish). The condition (i) ((ii))<br />
corresponds to imit<strong>at</strong>ive (antagonistic) behaviour. In the limit where the<br />
number N of agents is infinite, the dynamics of the fraction of bullish agents<br />
is d<strong>et</strong>erministic and exhibits chaotic behaviour in a significant domain of the<br />
param<strong>et</strong>er space {ρhb,ρbh,ρhh,ρbb,m}. A typical chaotic trajectory is<br />
characterized by intermittent phases of chaos, quasi-periodic behaviour and<br />
super-exponentially growing bubbles followed by crashes. A typical bubble<br />
starts initially by growing <strong>at</strong> an exponential r<strong>at</strong>e and then crosses over to a<br />
nonlinear power-law growth r<strong>at</strong>e leading to a finite-time singularity. The<br />
reinjection mechanism provi<strong>de</strong>d by the contrarian behaviour introduces a<br />
finite-size effect, rounding off these singularities and leads to chaos. We<br />
document the main stylized facts of this mo<strong>de</strong>l in the symm<strong>et</strong>ric and<br />
asymm<strong>et</strong>ric cases. This mo<strong>de</strong>l is one of the rare agent-based mo<strong>de</strong>ls th<strong>at</strong> give<br />
rise to interesting non-periodic complex dynamics in the ‘thermodynamic’<br />
limit (of an infinite number N of agents). We also discuss the case of a finite<br />
number of agents, which introduces an endogenous source of noise<br />
superimposed on the chaotic dynamics.<br />
264 1469-7688/02/040264+18$30.00 © 2002 IOP Publishing Ltd PII: S1469-7688(02)29456-7<br />
161
162 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
‘Human behaviour is a main factor in how mark<strong>et</strong>s<br />
act. In<strong>de</strong>ed, som<strong>et</strong>imes mark<strong>et</strong>s act quickly, violently<br />
with little warning. [...] Ultim<strong>at</strong>ely, history tells us<br />
th<strong>at</strong> there will be a correction of some significant dimension.<br />
I have no doubt th<strong>at</strong>, human n<strong>at</strong>ure being<br />
wh<strong>at</strong> it is, th<strong>at</strong> it is going to happen again and again.’<br />
Alan Greenspan, Chairman of the Fe<strong>de</strong>ral Reserve<br />
of the USA, before the Committee on Banking and<br />
Financial Services, US House of Represent<strong>at</strong>ives, 24<br />
July 1998.<br />
1. Introduction<br />
In recent economic and finance research, there is a growing<br />
interest in incorpor<strong>at</strong>ing i<strong>de</strong>as from social sciences to account<br />
for the fact th<strong>at</strong> mark<strong>et</strong>s reflect the thoughts, emotions and<br />
actions of real people as opposed to the i<strong>de</strong>alized economic<br />
investor whose behaviour un<strong>de</strong>rlies the efficient mark<strong>et</strong> and<br />
random walk hypothesis. This was captured by the now<br />
famous pronouncement of Keynes (1936) th<strong>at</strong> most investors’<br />
<strong>de</strong>cisions ‘can only be taken as a result of animal spirits—of<br />
a spontaneous urge to action r<strong>at</strong>her than inaction, and not the<br />
outcome of a weighed average of benefits multiplied by the<br />
quantit<strong>at</strong>ive probabilities’. A real investor may intend to be<br />
r<strong>at</strong>ional and may try to optimize his actions, but th<strong>at</strong> r<strong>at</strong>ionality<br />
tends to be hampered by cognitive biases, emotional quirks<br />
and social influences. ‘Behavioural finance’ is a growing<br />
research field (Thaler 1993, De Bondt and Thaler 1995, Shefrin<br />
2000, Shleifer 2000, Goldberg and von Nitzsch 2001), which<br />
uses psychology, sociology and other behavioural theories<br />
to <strong>at</strong>tempt to explain the behaviour of investors and money<br />
managers. The behaviour of financial mark<strong>et</strong>s is thought to<br />
result from varying <strong>at</strong>titu<strong>de</strong>s towards risk, the h<strong>et</strong>erogeneity in<br />
the framing of inform<strong>at</strong>ion, from cognitive errors, self-control<br />
and lack thereof, from regr<strong>et</strong> in financial <strong>de</strong>cision-making and<br />
from the influence of mass psychology. Assumptions about the<br />
frailty of human r<strong>at</strong>ionality and the acceptance of such drives<br />
as fear and greed are un<strong>de</strong>rlying the recipes <strong>de</strong>veloped over<br />
<strong>de</strong>ca<strong>de</strong>s by so-called technical analysts.<br />
There is growing empirical evi<strong>de</strong>nce for the existence<br />
of herd or ‘crowd’ behaviour in specul<strong>at</strong>ive mark<strong>et</strong>s (Arthur<br />
1987, Bikhchandani <strong>et</strong> al 1992, Johansen <strong>et</strong> al 1999, 2000,<br />
Orléan 1986, 1990, 1992, Shiller 1984, 2000, Topol 1991,<br />
West 1988). Herd behaviour is often said to occur when<br />
many people take the same action, because some mimic the<br />
actions of others. Herding has been linked to many economic<br />
activities, such as investment recommend<strong>at</strong>ions (Graham and<br />
Dodd 1934, Scharfstein and Stein 1990), price behaviour of<br />
IPOs (Initial Public Offering) (Welch 1992) fads and customs<br />
(Bikhchandani <strong>et</strong> al 1992), earnings forecasts (Trueman 1994),<br />
corpor<strong>at</strong>e conserv<strong>at</strong>ism (Zwiebel 1995) and <strong>de</strong>leg<strong>at</strong>ed portfolio<br />
management (Maug and Naik 1995).<br />
Here, we introduce arguably the simplest mo<strong>de</strong>l capturing<br />
the interplay b<strong>et</strong>ween mim<strong>et</strong>ic and contrarian behaviour in<br />
a popul<strong>at</strong>ion of N agents taking only two possible st<strong>at</strong>es,<br />
‘bullish’ or ‘bearish’ (buying or selling). In the limit of<br />
an infinite number N → ∞ of agents, the key variable,<br />
which is the fraction p of bullish agents, follows a chaotic<br />
d<strong>et</strong>erministic dynamics on a subspace of positive measure in<br />
the param<strong>et</strong>er space. Before explaining and analysing the<br />
mo<strong>de</strong>l in subsequent sections, we compare it in three respects<br />
to standard theories of economic behaviour.<br />
1. Since in the limit N →∞, the mo<strong>de</strong>l oper<strong>at</strong>es on<br />
a purely d<strong>et</strong>erministic basis, it actually challenges the purely<br />
external and unpredictable origin of mark<strong>et</strong> prices. Our mo<strong>de</strong>l<br />
exploits the continuous mimicry of financial mark<strong>et</strong>s to show<br />
th<strong>at</strong> the disor<strong>de</strong>red and random aspect of the time series of<br />
prices can be in part explained not only by the advent of<br />
‘random’ news and events, but can also be gener<strong>at</strong>ed by the<br />
behaviour of the agents fixing the prices.<br />
In the limit N → ∞, the dynamics of prices in our<br />
mo<strong>de</strong>l is d<strong>et</strong>erministic and <strong>de</strong>rives from the theory of chaotic<br />
dynamical systems, which have the fe<strong>at</strong>ure of exhibiting<br />
endogenously perturbed motion. After the first papers on<br />
the theory of chaotic systems, such as Lorenz (1963), May<br />
(1976), (see, e.g. Coll<strong>et</strong> and Eckmann (1980) for an early<br />
exposition), a series of economic papers <strong>de</strong>alt with mo<strong>de</strong>ls<br />
mostly of growth (Benhabib and Day 1981, Day 1982, 1983,<br />
Stutzer 1980). L<strong>at</strong>er, a vast and varied number of fields of<br />
economics were analysed in the light of the theory of chaos<br />
(Grandmont 1985, 1987, Grandmont and Malgrange 1986).<br />
They extend from macro-economics—business cycles, mo<strong>de</strong>ls<br />
of class struggles, political economy—to micro-economics—<br />
mo<strong>de</strong>ls with overlapping gener<strong>at</strong>ions, optimizing behaviour—<br />
and touch subjects such as game theory and the theory<br />
of finance. The applicability of these theories has been<br />
thoroughly tested on the stock mark<strong>et</strong> prices (Brock 1988,<br />
Brock <strong>et</strong> al 1987, 1991, Brock and Dechert 1988, LeBaron<br />
1988, Hsieh 1989, Scheinkman and LeBaron 1989a, 1989b)<br />
in studies which tried to d<strong>et</strong>ect signs of nonlinear effects and<br />
to nail down the d<strong>et</strong>erministic n<strong>at</strong>ure of these prices. While<br />
the theor<strong>et</strong>ical mo<strong>de</strong>ls (Van Der Ploeg 1986, De Grauwe and<br />
Vansanten 1990, De Grauwe <strong>et</strong> al 1993) seem to agree on<br />
the relevance of chaotic d<strong>et</strong>erministic dynamics, the empirical<br />
studies (Eckmann <strong>et</strong> al 1988, Hsieh and LeBaron 1988, Hsieh<br />
1989, 1991, 1992, LeBaron 1988, Scheinkman and LeBaron<br />
1989a, 1989b) are less clear-cut, mostly because of lack of<br />
sufficiently long time series (Eckmann and Ruelle 1992) or<br />
because the d<strong>et</strong>erministic component of mark<strong>et</strong> behaviour is<br />
necessarily overshadowed by the inevitable external effects.<br />
An additional source of ‘noise’ is found to result from the<br />
finiteness of the number N of agents. For finite N, the<br />
d<strong>et</strong>erministically chaotic dynamics of the price is replaced by a<br />
stochastic dynamics shadowing the corresponding trajectories<br />
obtained for N →∞.<br />
The mo<strong>de</strong>l presented here shows a mechanism of price<br />
fixing—<strong>de</strong>cisions to buy or sell dict<strong>at</strong>ed by comparison with<br />
other agents—which is <strong>at</strong> the origin of an instability of<br />
prices. From one period to the next, and in the absence of<br />
inform<strong>at</strong>ion other than the anticip<strong>at</strong>ions of other agents, prices<br />
can continuously exhibit err<strong>at</strong>ic behaviour and never stabilize,<br />
without diverging. Thus, the mo<strong>de</strong>l questions the fundamental<br />
hypothesis th<strong>at</strong> equilibrium prices have to converge to the<br />
intrinsic value of an ass<strong>et</strong>.<br />
2. We can also consi<strong>de</strong>r our mo<strong>de</strong>l in the context of the<br />
excess mark<strong>et</strong> vol<strong>at</strong>ility of financial mark<strong>et</strong>s. The vol<strong>at</strong>ility of<br />
265
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
prices gener<strong>at</strong>ed by our chaotic mo<strong>de</strong>l could give the beginning<br />
of an explan<strong>at</strong>ion of the excess vol<strong>at</strong>ility observed on financial<br />
mark<strong>et</strong>s (Grossman and Shiller 1981, Fama 1965, Flavin 1983,<br />
Shiller 1981, West 1988) which traditional mo<strong>de</strong>ls, such as<br />
ARCH, try to incorpor<strong>at</strong>e (Engle 1982, Bollerslev <strong>et</strong> al 1991,<br />
Bollerslev 1987).<br />
3. Finally, we can see specul<strong>at</strong>ive bubbles in our mo<strong>de</strong>l<br />
as a n<strong>at</strong>ural consequence of mim<strong>et</strong>ism. We can compare<br />
this to the two basic trends in explaining the problem of<br />
bubbles. The first makes reference to r<strong>at</strong>ional anticip<strong>at</strong>ions<br />
(Muth 1961) and rests on the hypothesis of efficient mark<strong>et</strong>s.<br />
With fixed inform<strong>at</strong>ion, and knowing the dynamics of prices,<br />
the recurrence rel<strong>at</strong>ion for the price is seen to <strong>de</strong>pend on the<br />
fundamental value and a self-referential component, which<br />
tends to cause a <strong>de</strong>vi<strong>at</strong>ion from the fundamental value: this<br />
is a specul<strong>at</strong>ive bubble (Blanchard and W<strong>at</strong>son 1982). This<br />
theory of r<strong>at</strong>ional specul<strong>at</strong>ive bubbles fails to explain the birth<br />
of such events, and even less their collapse, which it does<br />
not predict either. Recent <strong>de</strong>velopments improve on these<br />
traditional approaches by combining the r<strong>at</strong>ional agents in<br />
the economy with irr<strong>at</strong>ional ‘noise’ tra<strong>de</strong>rs (Johansen <strong>et</strong> al<br />
1999, 2000, Sorn<strong>et</strong>te and Johansen 2001). These noise tra<strong>de</strong>rs<br />
are imit<strong>at</strong>ive investors who resi<strong>de</strong> on an interaction n<strong>et</strong>work.<br />
Neighbours of an agent on this n<strong>et</strong>work can be viewed as the<br />
agent’s friends or contacts, and an agent will incorpor<strong>at</strong>e his<br />
neighbours’ views regarding the stock into his own view. These<br />
noise tra<strong>de</strong>rs are responsible for triggering crashes. Sorn<strong>et</strong>te<br />
and An<strong>de</strong>rsen (2001) <strong>de</strong>velop a similar mo<strong>de</strong>l in which the<br />
noise tra<strong>de</strong>rs induce a nonlinear positive feedback in the stock<br />
price dynamics with an interplay b<strong>et</strong>ween nonlinearity and<br />
multiplic<strong>at</strong>ive noise. The <strong>de</strong>rived hyperbolic stochastic finit<strong>et</strong>ime<br />
singularity formula transforms a Gaussian white noise<br />
into a rich time series possessing all the stylized facts of<br />
empirical prices, as well as acceler<strong>at</strong>ed specul<strong>at</strong>ive bubbles<br />
preceding crashes.<br />
The second trend purports to explain specul<strong>at</strong>ive bubbles<br />
by a limit<strong>at</strong>ion of r<strong>at</strong>ionality (Shiller 1984, 2000, West 1988,<br />
Topol 1991). It allows us to incorpor<strong>at</strong>e notions which the<br />
neo-classical analysis does not take into account: asymm<strong>et</strong>ry<br />
of inform<strong>at</strong>ion, inefficiency of prices, h<strong>et</strong>erogeneity of<br />
anticip<strong>at</strong>ions (Grossman 1977, Grossman and Stiglitz 1980,<br />
Grossman 1981, Radner 1972, 1979). In our approach, which<br />
follows the second trend, the agents act without knowing the<br />
actual effect of their behaviour: this contrasts with the position<br />
of a mo<strong>de</strong>l-buil<strong>de</strong>r (Orléan 1986, 1989, 1990, 1992). This, in<br />
turn, can lead to prices which disconnect from the fundamental<br />
indic<strong>at</strong>ors of economics.<br />
In the present paper we show th<strong>at</strong> self-referred behaviour<br />
in financial mark<strong>et</strong>s can gener<strong>at</strong>e chaos and specul<strong>at</strong>ive<br />
bubbles. They will be seen to be caused by mim<strong>et</strong>ic behaviour:<br />
bubbles will form due to imit<strong>at</strong>ive behaviour and collapse when<br />
certain agents believe in the advent of a turn of trend, while<br />
they observe the behaviour of their peers.<br />
Our work can also be seen as a dynamical generaliz<strong>at</strong>ion<br />
of Galam and Moscovici (1991) and Galam (1997) who<br />
have introduced the i<strong>de</strong>a of a universal behaviour in group<br />
<strong>de</strong>cision making, in<strong>de</strong>pen<strong>de</strong>ntly of the n<strong>at</strong>ure of the <strong>de</strong>cision, in<br />
situ<strong>at</strong>ions where two opposite choices are proposed. Using the<br />
266<br />
163<br />
formalism of the Ising mo<strong>de</strong>l with the condition of minimal<br />
conflict specifying the interactions b<strong>et</strong>ween members of the<br />
group, Galam and Moscovici (1991) and Galam (1997) have<br />
established a st<strong>at</strong>ic phase diagram of possible behaviours.<br />
Local pressure and anticip<strong>at</strong>ion have been taken into account<br />
by a local random field and a mean field respectively, in the<br />
context of the Ising formalism.<br />
Section 2 <strong>de</strong>fines the mo<strong>de</strong>l. Section 3 provi<strong>de</strong>s<br />
a qualit<strong>at</strong>ive un<strong>de</strong>rstanding and analysis of its dynamical<br />
properties. Section 4 extends it with a quantit<strong>at</strong>ive analysis<br />
of the phases of specul<strong>at</strong>ive bubbles in the symm<strong>et</strong>ric case.<br />
Section 5 <strong>de</strong>scribes the st<strong>at</strong>istical properties of the price r<strong>et</strong>urns<br />
<strong>de</strong>rived from its dynamics in the symm<strong>et</strong>ric case. Section 6<br />
discusses the asymm<strong>et</strong>ric case. Section 7 explores some effects<br />
introduced by the finiteness N
164 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
value investors (I<strong>de</strong> and Sorn<strong>et</strong>te 2001, Sorn<strong>et</strong>te and I<strong>de</strong> 2001).<br />
We build on this insight and construct a very simple mo<strong>de</strong>l<br />
of price dynamics, which puts emphasis on the fundamental<br />
nonlinear behaviour of both classes of agents.<br />
These well-known principles gener<strong>at</strong>e different kinds of<br />
risks b<strong>et</strong>ween which agents choose by arbitrage. The former<br />
is a comp<strong>et</strong>ing risk (Keynes 1936, Orléan 1989) which leads<br />
agents to imit<strong>at</strong>e the collective point of view since the mark<strong>et</strong><br />
price inclu<strong>de</strong>s it. Thus, it is assumed th<strong>at</strong> Keynes’ animal<br />
spirits may exist. More simply, there is the risk of mistaken<br />
expect<strong>at</strong>ion: agents believe in a price different from the mark<strong>et</strong><br />
price. Keynes uses his famous beauty contest as a parable<br />
for stock mark<strong>et</strong>s. In or<strong>de</strong>r to predict the winner of a beauty<br />
contest, objective beauty is not very important, but knowledge<br />
or prediction of others’ prediction of beauty is. In Keynes’<br />
view, the optimal str<strong>at</strong>egy is not to pick those faces the player<br />
thinks the pr<strong>et</strong>tiest, but those the other players are likely to<br />
think the average opinion will be, or those the other players<br />
will think the others will think the average opinion will be, or<br />
even further along this iter<strong>at</strong>ive loop.<br />
On the other hand, in the l<strong>at</strong>ter case, the emerging<br />
price is not necessarily in harmony with economic reality<br />
and fundamental value. Self-referred <strong>de</strong>cisions and selfvalid<strong>at</strong>ion<br />
phenomena can then in<strong>de</strong>ed lead to specul<strong>at</strong>ive<br />
bubbles or sunspots (i.e. external random events) (Azariadis<br />
1981, Azariadis and Guesnerie 1982, Blanchard and W<strong>at</strong>son<br />
1982, Jevons 1871, Kreps 1977). Thus, the l<strong>at</strong>ter risk is<br />
the result of precaution. It addresses the fitting of mark<strong>et</strong><br />
price to fundamental value and, by extension, collapse of the<br />
specul<strong>at</strong>ive bubble.<br />
Both <strong>at</strong>titu<strong>de</strong>s are likely to be important and are integr<strong>at</strong>ed<br />
in <strong>de</strong>cision rules. Agents realize an arbitrage b<strong>et</strong>ween the two<br />
kinds of risk we have <strong>de</strong>scribed. Th<strong>at</strong> is why they have both a<br />
mim<strong>et</strong>ic behaviour and an antagonistic one: they either follow<br />
the collective point of view or they have reversed expect<strong>at</strong>ions.<br />
We are now going to put these assumptions into the<br />
simplest possible m<strong>at</strong>hem<strong>at</strong>ical form. We assume th<strong>at</strong>, <strong>at</strong><br />
any given time t, the popul<strong>at</strong>ion is divi<strong>de</strong>d into two parts.<br />
Agents are explicitly differenti<strong>at</strong>ed as being bullish or bearish<br />
in proportions pt and qt = 1 − pt respectively. The first ones<br />
expect an increase of the price, while the bearish ones expect<br />
a <strong>de</strong>crease. The agents then form their opinion for time t +1<br />
by sampling the expect<strong>at</strong>ions of m other agents <strong>at</strong> time t, and<br />
modifying their own expect<strong>at</strong>ions accordingly. The number m<br />
of agents polled by a given agent to form her opinion <strong>at</strong> time<br />
t + 1 is the first important param<strong>et</strong>er in our mo<strong>de</strong>l.<br />
We then introduce threshold <strong>de</strong>nsities ρhb and ρhh. We<br />
assume 0 ρhb ρhh 1. A bullish agent will change<br />
opinion if <strong>at</strong> least one of the following propositions is true.<br />
(1) At least m · ρhb among the m agents inspected are bearish.<br />
(2) At least m · ρhh among the m agents inspected are bullish.<br />
The first case corresponds to ‘following the crowd’, while<br />
the second case corresponds to the ‘antagonistic behaviour’.<br />
The quantity ρhb is thus the threshold for a bullish agent<br />
(‘haussier’) to become bearish (‘baissier’) for mim<strong>et</strong>ic reasons,<br />
and similarly, ρhh is the threshold for a bullish agent to<br />
become bearish because there are ‘too many’ bullish agents.<br />
One reason for this behaviour is, as we said, th<strong>at</strong> the <strong>de</strong>vi<strong>at</strong>ion<br />
of the mark<strong>et</strong> price from the fundamental value is felt to be<br />
unsustainable. Another reason is th<strong>at</strong> if many managers tell<br />
you th<strong>at</strong> they are bullish, it is probable th<strong>at</strong> they have large<br />
‘long’ positions in the mark<strong>et</strong>: they therefore tell you to buy,<br />
hoping to be able to unfold in part their position in favourable<br />
conditions with a good profit.<br />
The <strong>de</strong>vi<strong>at</strong>ion of the threshold ρhb above the symm<strong>et</strong>ric<br />
value 1/2 is a measure of the ‘stubbornness’ (or ‘buy-and-hold’<br />
ten<strong>de</strong>ncy) of the agent to keep her position. For ρhb = 1/2,<br />
the agent strictly endorses without <strong>de</strong>lay the opinion of the<br />
majority and believes in any weak trend. This corresponds to a<br />
reversible dynamics. A value ρhb > 1/2 expresses a ten<strong>de</strong>ncy<br />
towards conserv<strong>at</strong>ism: a large ρhb means th<strong>at</strong> the agent will<br />
rarely change opinion. She is risk-adverse and would like to<br />
see an almost unanimity appearing before changing her mind.<br />
Her future behaviour has thus a strong memory of her past<br />
position. ρhb − 1/2 can be called the bullish ‘buy-and-hold’<br />
in<strong>de</strong>x.<br />
The <strong>de</strong>vi<strong>at</strong>ion of the threshold ρhh below 1 quantifies the<br />
strength of disbelief of the agent in the sustainability of a<br />
specul<strong>at</strong>ive trend. For ρhh = 1, she always follows the crowd<br />
and is never contrarian. For ρhh close to 1/2, she has little faith<br />
in trend-following str<strong>at</strong>egies and is closer to a fundamentalist,<br />
expecting the price to revert rapidly to its fundamental value.<br />
1 − ρhh can be called the bullish reversal in<strong>de</strong>x.<br />
Putting the above rules into m<strong>at</strong>hem<strong>at</strong>ical equ<strong>at</strong>ions we<br />
see th<strong>at</strong> the probability P for an agent who is bullish <strong>at</strong> time t<br />
to change his opinion <strong>at</strong> time t + 1 is:<br />
P = prob({x mρhh}), (1)<br />
where x is the number of bullish agents found in the sample of<br />
m agents.<br />
In an entirely similar way, we introduce thresholds ρbh,<br />
and ρbb. The thresholds ρbh and ρbb have compl<strong>et</strong>ely<br />
symm<strong>et</strong>ric roles when the agent is initially bearish. ρbh − 1/2<br />
can be called the bearish ‘buy-and-hold’ in<strong>de</strong>x. 1 − ρbb can<br />
be called the bearish reversal in<strong>de</strong>x. The probability Q for a<br />
bearish agent <strong>at</strong> time t to become bullish <strong>at</strong> time t + 1 is:<br />
Q = prob({x mρbb}).<br />
We can combine these two rules into a dynamical law<br />
governing the time evolution of the popul<strong>at</strong>ions. Denoting pt<br />
the proportion of bullish agents in the popul<strong>at</strong>ion <strong>at</strong> time t,we<br />
can find the new proportion, pt+1, <strong>at</strong> time t + 1, by taking into<br />
account those agents which have changed opinion according to<br />
the d<strong>et</strong>erministic law given above. To simplify not<strong>at</strong>ion, we l<strong>et</strong><br />
pt+1 = p ′ and pt = p. Then, the above st<strong>at</strong>ements are easily<br />
used to express p ′ in terms of p, by using the probability of<br />
finding j bullish people among m (Corcos 1993):<br />
p ′ = p − p<br />
<br />
<br />
m<br />
p<br />
j<br />
m−j (1 − p) j<br />
+ (1 − p)<br />
jm·ρhb<br />
or j
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
Figure 1. The family of functions Fρ,m(p) for ρhb = ρbh = 0.72<br />
and ρhh = ρbb = 0.85. The curves are for<br />
m = 13 + j · 26,j = 0,...,13. Note the convergence to the<br />
function Gρ, (indic<strong>at</strong>ed by m =∞).<br />
where ρ ={ρhb,ρbh,ρhh,ρbb}. Thus, the function Fρ,m(p)<br />
compl<strong>et</strong>ely characterizes the dynamics of the proportion of<br />
bullish and bearish popul<strong>at</strong>ions.<br />
3. Qualit<strong>at</strong>ive analysis of the dynamical<br />
properties<br />
3.1. The limit m →∞<br />
The law given by equ<strong>at</strong>ion (2) is not easy to analyse, and we<br />
give in figure1afewsample curves Fρ,m. We see th<strong>at</strong> as m g<strong>et</strong>s<br />
larger, the curves seem to tend to a limiting curve. Using this<br />
observ<strong>at</strong>ion, our conceptual un<strong>de</strong>rstanding of the dynamics can<br />
be drastically simplified if we consi<strong>de</strong>r the problem for a large<br />
number m of polled partners. In<strong>de</strong>ed, it is most convenient<br />
to first study the unrealistic problem m =∞and to view the<br />
large m case as a perturb<strong>at</strong>ion of this limiting case. The main<br />
ingredient in the study of the case m =∞is the Law of Large<br />
Numbers, which we use in a form given in Feller (1966):<br />
Lemma. L<strong>et</strong> g be a continuous function on [0, 1]. Then, for<br />
p ∈ [0, 1],<br />
m<br />
<br />
m<br />
lim p<br />
m→∞ j<br />
j (1 − p) m−j g(j/m) = g(p). (3)<br />
j=0<br />
We apply this lemma to the (piecewise continuous)<br />
function g = fh, where fh is the indic<strong>at</strong>or function of the<br />
s<strong>et</strong> <strong>de</strong>fining P :<br />
<br />
1, if x ρhb or x
166 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
p t<br />
p t<br />
p t<br />
p t<br />
1.0<br />
0.5<br />
0.0<br />
1.0<br />
0.5<br />
0.0<br />
1.0<br />
0.5<br />
0.0<br />
1.0<br />
0.5<br />
0.0<br />
r bb = 0.75<br />
r bb = 0.76<br />
r bb = 0.77<br />
r bb = 0.85<br />
Figure 3. The time series for the same param<strong>et</strong>er values as in<br />
figure 2. Note th<strong>at</strong>, for ρhh = ρbb = 0.75, one has convergence to a<br />
bullish equilibrium, for 0.76 a bullish period 2, for 0.77 a bullish, but<br />
chaotic behaviour. The most interesting case is ρhh = ρbb = 0.85,<br />
where calm periods altern<strong>at</strong>e in a seemingly random fashion with<br />
specul<strong>at</strong>ive bubbles ending in crashes and anti-crashes.<br />
(2) The next, more interesting case, is the appearance of a limit<br />
cycle (of period 2): <strong>at</strong> successive times, the popul<strong>at</strong>ion<br />
of bullish and bearish agents will oscill<strong>at</strong>e b<strong>et</strong>ween two<br />
different values. This happens, e.g. for ρhh = ρbb = 0.76,<br />
with the other param<strong>et</strong>ers as before (see second panel of<br />
figure 3).<br />
(3) But for certain values of the param<strong>et</strong>ers, e.g. ρhh = ρbb =<br />
0.85, the sequence of values of pt is a chaotic sequence,<br />
with positive Lyapunov exponent (compare with Eckmann<br />
and Ruelle 1985). The mechanism for this is really a<br />
combin<strong>at</strong>ion of sufficiently strong buy-and-hold in<strong>de</strong>x<br />
ρhb − 1/2 and of sufficiently weak reversal in<strong>de</strong>x 1 − ρhh.<br />
This regime thus occurs when the opinion of a tra<strong>de</strong>r has<br />
a strong memory of her past positions and changes it only<br />
when a strong majority appears. This regime also requires<br />
a weak belief of the agent in fundamental valu<strong>at</strong>ion, as<br />
she will believe until very l<strong>at</strong>e th<strong>at</strong> a strong bullish or<br />
bearish specul<strong>at</strong>ive trend is sustainable. Fundamentally, it<br />
is this self-referential behaviour of the anticip<strong>at</strong>ions alone<br />
which is responsible for a d<strong>et</strong>erministic, but seemingly<br />
err<strong>at</strong>ic evolution of the popul<strong>at</strong>ion of bullish and bearish<br />
agents. No external noise is nee<strong>de</strong>d to make this happen,<br />
and in general, we view external stimuli as acting on top of<br />
the intrinsic mechanism which we exhibit here (Eckmann<br />
1981). Note th<strong>at</strong> the s<strong>et</strong> of param<strong>et</strong>er values ρ for which<br />
chaos is expected (say, near the values used <strong>at</strong> the bottom<br />
of figure 3) has positive Lebesgue measure.<br />
t<br />
t<br />
t<br />
t<br />
Wh<strong>at</strong> should be a reasonable value for ρhh = ρbb to<br />
<strong>de</strong>scribe empirical d<strong>at</strong>a? We believe th<strong>at</strong> this is an illposed<br />
question and, conditioned on the use of the present<br />
very simplified mo<strong>de</strong>l, ρhh = ρbb is probably not constant<br />
in time but is a (possibly slowly varying) function of time.<br />
As a consequence, the mark<strong>et</strong> dynamics may explore all the<br />
possible different regimes <strong>de</strong>scribed above: stability of price,<br />
oscill<strong>at</strong>ions or cycles, chaotic behaviour (of course each of<br />
them being <strong>de</strong>cor<strong>at</strong>ed by noise capturing a finite N as discussed<br />
below as well as other external sources of uncertainties not<br />
captured by the mo<strong>de</strong>l). The i<strong>de</strong>a of a time-varying control<br />
param<strong>et</strong>er ρhh = ρbb is similar to th<strong>at</strong> advoc<strong>at</strong>ed in Stauffer and<br />
Sorn<strong>et</strong>te (1999) to cure the percol<strong>at</strong>ion mo<strong>de</strong>l of stock mark<strong>et</strong><br />
prices (Cont and Bouchaud 2000), which nee<strong>de</strong>d to be tuned<br />
very artificially to close to the critical percol<strong>at</strong>ion threshold<br />
in or<strong>de</strong>r to exhibit interesting reasonable st<strong>at</strong>istics. Here, we<br />
shall not explore further this possibility but will instead go on<br />
investig<strong>at</strong>ing the properties of the mo<strong>de</strong>l with fixed param<strong>et</strong>ers.<br />
We next consi<strong>de</strong>r in more d<strong>et</strong>ail the time evolution of<br />
pt for the param<strong>et</strong>er values of the last frame of figure 3,<br />
which are typical for the abundant s<strong>et</strong> of ‘chaotic’ param<strong>et</strong>er<br />
values, and we will show how the time evolution exhibits<br />
‘specul<strong>at</strong>ive bubbles’. This phenomenon is akin to the<br />
notion of intermittency (of ‘type I’) as known to physicists,<br />
see e.g. Manneville (1991) for an exposition. In<strong>de</strong>ed, we<br />
can distinguish two distinct behaviours in the last frame of<br />
figure 3, which occur repe<strong>at</strong>edly with more or less pronounced<br />
separ<strong>at</strong>ion. The first process is the ‘laminar phase’, which<br />
is seen to occur when the popul<strong>at</strong>ion pt is near 0.5. Then,<br />
the evolution of the popul<strong>at</strong>ion is slow, and the popul<strong>at</strong>ion<br />
grows slowly away from 0.5, either monotonically or through<br />
an oscill<strong>at</strong>ion of period 2, <strong>de</strong>pending on ρ. This motion is<br />
slower when the inspected sample size (m) is larger, reflecting<br />
a more stable evolution for less in<strong>de</strong>pen<strong>de</strong>nt agents. When<br />
the distance from 0.5 is large, err<strong>at</strong>ic behaviour s<strong>et</strong>s in, which<br />
persists until the popul<strong>at</strong>ion reaches again a value of about 0.5,<br />
<strong>at</strong> which point the whole scenario repe<strong>at</strong>s. The d<strong>et</strong>erminism of<br />
the mo<strong>de</strong>l is reflected by ‘equal causes lead to equal effects’,<br />
while its chaotic n<strong>at</strong>ure is reflected by the err<strong>at</strong>ic length of the<br />
laminar periods, as well as of the bubbles of wild behaviour.<br />
Having analysed qualit<strong>at</strong>ively the evolution of the number<br />
of bullish agents, we next <strong>de</strong>scribe how the price πt+1 of an<br />
ass<strong>et</strong> <strong>at</strong> time t + 1 is rel<strong>at</strong>ed to the proportion pt of bullish<br />
agents. One can argue (Corcos 1993, Bouchaud and Cont<br />
1998, Farmer 1998) th<strong>at</strong> the price change πt+1 − πt from one<br />
period to the next is a monotonic function of pt (and, perhaps,<br />
of πt). This function is positive when pt > 1/2 and neg<strong>at</strong>ive<br />
when pt < 1/2. If the reaction to a change in pt is reflected<br />
in the prices in the next period, then a bubble in pt will lead to<br />
a specul<strong>at</strong>ive bubble in the prices in the next period. Thus, our<br />
mo<strong>de</strong>l predicts the occurrence of bubbles from the behaviour<br />
of the agents alone. Furthermore, for quite general laws of the<br />
form<br />
πt+1 = H(πt,pt), (7)<br />
a simple applic<strong>at</strong>ion of the chain rule of differenti<strong>at</strong>ion leads<br />
to the observ<strong>at</strong>ion th<strong>at</strong> the variable πt has the same Lyapunov<br />
exponent as pt. In fact, this will be the case if 0
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
p t<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000<br />
t<br />
Figure 4. Evolution of the system over 10 000 time steps for<br />
N =∞, m = 60 polled agents and the param<strong>et</strong>ers<br />
ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85. Note the existence of<br />
acceler<strong>at</strong>ions ending in crashes and anti-crashes (fast jumps).<br />
and ∂pH >c>0, where λ is the Lyapunov exponent for pt,as<br />
follows from δπt+1 = ∂πH · δπt + ∂pH · δpt. This condition is,<br />
in particular, s<strong>at</strong>isfied for a law of the form πt+1 = πt + G(pt),<br />
where G is strictly monotonic. Thus, chaotic behaviour of<br />
bullish agents leads to chaotic behaviour of prices.<br />
In the following, we shall take the simplest form of a<br />
log-difference of the price linearly proportional to the or<strong>de</strong>r<br />
unbalance (Farmer 1998), leading to<br />
ln πt+1 − ln πt ≡ rt+1 = γ(pt − 1<br />
), (8)<br />
2<br />
showing th<strong>at</strong> the r<strong>et</strong>urn rt calcul<strong>at</strong>ed over one period is<br />
proportional to the imbalance pt − 1<br />
. Thus, the properties<br />
2<br />
of the r<strong>et</strong>urn time series can be <strong>de</strong>rived directly from those of<br />
pt as we document below. This linear function of (pt − 1<br />
) can 2<br />
be improved for instance by the hyperbolic tangent function<br />
tanh(pt − 1<br />
) documented in Plerou <strong>et</strong> al (2001).<br />
2<br />
To summarize this qualit<strong>at</strong>ive analysis of the case of an<br />
infinite number N of agents, we observe a time evolution<br />
which, while s<strong>at</strong>isfying certain criteria of randomness (such<br />
as possessing an absolutely continuous invariant measure<br />
and exhibiting a positive Lyapunov exponent (compare with<br />
Eckmann and Ruelle 1985)) <strong>at</strong> the same time exhibits some<br />
regularities on short time scales, since it is d<strong>et</strong>erministic.<br />
Our mo<strong>de</strong>l thus establishes th<strong>at</strong> straightforward fundamental<br />
conditions may suffice to gener<strong>at</strong>e chaotic stock mark<strong>et</strong><br />
behaviour, <strong>de</strong>pending on the param<strong>et</strong>er values. If the mark<strong>et</strong><br />
adjusts present mark<strong>et</strong> price on the basis of expect<strong>at</strong>ions and<br />
mimicry—self-referred behaviour—then chaotic evolution of<br />
the popul<strong>at</strong>ion will also imply chaotic evolution of prices.<br />
4. Quantit<strong>at</strong>ive analysis of the<br />
specul<strong>at</strong>ive bubbles within the chaotic<br />
regime in the symm<strong>et</strong>ric case<br />
For an infinite number N of agents and in the symm<strong>et</strong>ric case<br />
ρhb = ρbh ≡ ρ1 and ρhh = ρbb ≡ ρ2, l<strong>et</strong> us rewrite the<br />
270<br />
p t –1/2<br />
p t –1/2<br />
0.25<br />
167<br />
0.20<br />
0.15<br />
0.10<br />
0.05<br />
0.00<br />
0 100 200 300 400 500 600<br />
10 0<br />
t<br />
10 –1<br />
10 –2<br />
10 –3<br />
10 –4<br />
0 100 200 300<br />
t<br />
400 500 600<br />
Figure 5. The first bubble of figure 4 for N =∞agents with<br />
m = 60 polled agents and param<strong>et</strong>ers ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85.<br />
dynamical evolution (2) of the system as<br />
p ′ m<br />
<br />
m<br />
= p − p p<br />
j j=0<br />
m−j (1 − p) j <br />
j<br />
f<br />
m<br />
m<br />
<br />
m<br />
+ (1 − p) (1 − p)<br />
j j=0<br />
m−j p j <br />
j<br />
f ,<br />
m<br />
where<br />
(9)<br />
<br />
1,<br />
f(x)=<br />
0,<br />
if x ρ1 or x
168 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
Fr,m(p)–p<br />
10 0<br />
10 –1<br />
10 –2<br />
10 –3<br />
10 –4<br />
10 –5<br />
10 –6<br />
10 –7<br />
m = 30<br />
m = 60<br />
m = 100<br />
10 –3 10 –2<br />
slope m(m = 60)=3<br />
p–1/2<br />
10 –1<br />
Figure 6. The logarithm of Fm(p) − p versus the logarithm of<br />
p − 1/2 for three different values of m = 30, 60 and 100, with<br />
ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85. Note the transition from a<br />
slope 1 to a large effective slope before the reinjection due to the<br />
contrarian mechanism.<br />
κ>0) followed by a super-exponential growth acceler<strong>at</strong>ing<br />
so much as to give the impression of reaching a singularity in<br />
finite-time.<br />
To un<strong>de</strong>rstand this phenomenon, we plot the logarithm of<br />
Fm(p)−p versus the logarithm of p −1/2 in figure 6 for three<br />
different values of m = 30, 60 and 100. Two regimes can be<br />
observed.<br />
(1) For small p − 1/2, the slope of log 10 (Fm(p) − p) versus<br />
log 10 (p − 1/2) is 1, i.e.<br />
10 0<br />
p ′ − p ≡ Fm(p) − p α(m)(p − 1<br />
). (13)<br />
2<br />
This expression (13) explains the exponential growth<br />
observed <strong>at</strong> early times in figure 5.<br />
(2) For larger p − 1/2, the slope of log 10 (Fm(p) − p) versus<br />
log 10 (p − 1/2) increases above 1 and stabilizes to a<br />
value µ(m) before <strong>de</strong>creasing again due to the reinjection<br />
produced by the contrarian mechanism. The interval in<br />
p − 1/2 in which the slope is approxim<strong>at</strong>ely stabilized <strong>at</strong><br />
the value µ(m) enables us to write<br />
Fm(p)−p β(m)(p − 1<br />
2 )µ(m) , with µ>1. (14)<br />
These two regimes can be summarized in the following<br />
phenomenological expression for Fm(p):<br />
Fm(p) = 1<br />
+ (1 − 2gm(1/2)<br />
2<br />
− g ′ 1<br />
1<br />
m (1/2))(p − ) + β(m)(p − 2 2 )µ(m) , (15)<br />
= 1<br />
2<br />
+ (p − 1<br />
2<br />
) + α(m)(p − 1<br />
2 )<br />
+ β(m)(p − 1<br />
2 )µ(m) with µ>1, (16)<br />
and<br />
α(m) =−2gm(1/2) − g ′ m (1/2). (17)<br />
Figure 7. Approxim<strong>at</strong>ion of the function Fm(p) − 1<br />
2<br />
function f(p)= [1 + α](p + 1<br />
2<br />
by the<br />
) + β(p + 1<br />
2 )µ over different<br />
p-intervals, for m = 60 interacting agents and param<strong>et</strong>ers<br />
ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85.<br />
Table 1. Optimized param<strong>et</strong>ers α, β and µ for several optimiz<strong>at</strong>ion<br />
intervals with m = 60 interacting agents and ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85.<br />
Optimiz<strong>at</strong>ion domain α β µ<br />
0 p − 1<br />
0.05 2 0.011 11.67 3.27<br />
0 p − 1<br />
0.10 2 0.013 43.66 3.77<br />
0 p − 1<br />
0.15 2<br />
0 p −<br />
0.014 60.32 3.91<br />
1<br />
0.20 2 0.004 30.64 3.54<br />
This expression can be obtained as an approxim<strong>at</strong>ion of the<br />
exact expansion <strong>de</strong>rived in the appendix.<br />
In or<strong>de</strong>r to check the hypothesis (16), we numerically solve<br />
the following problem<br />
min<br />
{α,β,µ} Fm(p) − 1<br />
2<br />
1<br />
1<br />
− [1 + α](p − ) − β(p − 2 2 )µ 2 , (18)<br />
which amounts to constructing the best approxim<strong>at</strong>ion of<br />
the exact map Fm(p) in terms of an effective power-law<br />
acceler<strong>at</strong>ion (see (20) below). The results obtained for m =<br />
60 interacting agents and ρhb = ρbh = 0.72 and ρhh =<br />
ρbb = 0.85 are given in table 1 and shown in figure 7.<br />
The numerical values of α are in good agreement with the<br />
theor<strong>et</strong>ical prediction: α(m) = F ′ m (1/2) − 1 which yields<br />
α(m) 0.011 in the present case (m = 60, ρhb = ρbh = 0.72<br />
and ρhh = ρbb = 0.85). As a first approxim<strong>at</strong>ion, we<br />
can consi<strong>de</strong>r th<strong>at</strong> the exponent µ is fixed over the interval<br />
of interest, which is reasonable according to the very good<br />
quality of the fits shown in figure 7. We can conclu<strong>de</strong> from<br />
this numerical investig<strong>at</strong>ion th<strong>at</strong> µ(m) ∈ [3, 4]. A finer<br />
analysis shows, however, th<strong>at</strong> the exponent µ is in fact not<br />
perfectly constant but shifts slowly from about 3 to 4 as p<br />
increases. This should be expected as the function Fm(p)<br />
contains many higher-or<strong>de</strong>r terms. We can also note th<strong>at</strong> the<br />
param<strong>et</strong>er pc = (β/α) −1/µ , which <strong>de</strong>fines the typical scale<br />
of the crossover remains constant and equal to pc 0.70 for<br />
271
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
all the fits (except for the largest interval p − 1/2 < 0.2,<br />
for which pc = 0.8). In sum, the procedure (18) and its<br />
results show th<strong>at</strong> the effective power-law represent<strong>at</strong>ion (16) is<br />
a crossover phenomenon: it is not domin<strong>at</strong>ed by the ‘critical’<br />
value ρhb = ρbh of the jump of the map obtained in the limit<br />
of large m.<br />
Introducing the not<strong>at</strong>ion ɛ = p − 1/2, the dynamics<br />
associ<strong>at</strong>ed with the effective map (16) can be rewritten<br />
ɛ ′ − ɛ = α(m)ɛ + β(m)ɛ µ(m) , (19)<br />
which, in the continuous time limit, yields<br />
dɛ<br />
dt = α(m)ɛ + β(m)ɛµ(m) . (20)<br />
Thus, for small ɛ, we obtain an exponential growth r<strong>at</strong>e<br />
while for large enough ɛ<br />
ɛt ∝ e α(m)t , (21)<br />
1<br />
−<br />
ɛt ∝ (tc − t) µ(m)−1 . (22)<br />
Of course, these behaviours become invalid for t too close to<br />
tc for which the reinjection mechanism oper<strong>at</strong>es and ensures<br />
th<strong>at</strong> the probability remains less than one.<br />
For example, for m = 60 with ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85, we can check on figure 6 th<strong>at</strong> µ(m) = 3,<br />
which yields for large ɛ:<br />
pt − 1<br />
2 ∝<br />
1<br />
√ . (23)<br />
tc − t<br />
The prediction (23) implies th<strong>at</strong> plotting (pt − 1/2) −2 as a<br />
function of t should be a straight line in this regime. This<br />
non-param<strong>et</strong>ric test is checked in figure 8 on five successive<br />
bubbles. This provi<strong>de</strong>s a confirm<strong>at</strong>ion of the effective powerlaw<br />
represent<strong>at</strong>ion (16) of the map. The fact th<strong>at</strong> it is the lowest<br />
estim<strong>at</strong>e µ ≈ 3 shown in table 1 which domin<strong>at</strong>es in figure 8<br />
results simply from the fact th<strong>at</strong> it is the longest transient<br />
corresponding to the regime where p is closest to the unstable<br />
fixed point 1/2. This is visualized in figure 8 by the horizontal<br />
dashed lines indic<strong>at</strong>ing the levels p−1/2 = 0.05, 0.01 and 0.2.<br />
This <strong>de</strong>monstr<strong>at</strong>es th<strong>at</strong> most of the visited values are close to<br />
the unstable fixed point, which d<strong>et</strong>ermines the effective value<br />
of the nonlinear exponent µ ≈ 3.<br />
With the price dynamics (8), the prediction (22) implies<br />
th<strong>at</strong> the r<strong>et</strong>urns rt should increase in an acceler<strong>at</strong>ing superexponential<br />
fashion <strong>at</strong> the end of a bubble, leading to a price<br />
trajectory<br />
πt = πc − C(tc − t) µ(m)−2<br />
µ(m)−1 , (24)<br />
where πc is the culmin<strong>at</strong>ing price of the bubble reached <strong>at</strong><br />
t = tc when µ(m) > 2, such th<strong>at</strong> the finite-time singularity<br />
in rt gives rise only to an infinite slope of the price trajectory.<br />
The behaviour (24) with an exponent 0 < µ(m)−2<br />
< 1 has been<br />
µ(m)−1<br />
documented in many bubbles (Sorn<strong>et</strong>te <strong>et</strong> al 1996, Johansen<br />
<strong>et</strong> al 1999, 2000, Johansen and Sorn<strong>et</strong>te 1999, 2000, Sorn<strong>et</strong>te<br />
and Johansen 2001, Sorn<strong>et</strong>te and An<strong>de</strong>rsen 2001, Sorn<strong>et</strong>te<br />
2001). The case m = 60 with ρhb = ρbh = 0.72 and<br />
272<br />
169<br />
1<br />
Figure 8. (pt −1/2) 2 versus t to qualify the finite-time singularity<br />
predicted by (23) for m = 60 with ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85. The points are obtained from the time series pt<br />
and the straight continuous lines are the best linear fits. The<br />
horizontal dashed lines indic<strong>at</strong>e the levels p − 1/2 = 0.05, 0.01 and<br />
0.2 to <strong>de</strong>monstr<strong>at</strong>e th<strong>at</strong> most of the visited values are close to the<br />
unstable fixed point, which d<strong>et</strong>ermines the effective value of the<br />
nonlinear exponent µ ≈ 3.<br />
ρhh = ρbb = 0.85 shown in figure 6 leads to µ(m)−2<br />
= 1/2,<br />
µ(m)−1<br />
which is in reasonable agreement with previously reported<br />
values.<br />
Interpr<strong>et</strong>ed within the present mo<strong>de</strong>l, the exponent µ(m)−2<br />
µ(m)−1<br />
of the price singularity gives an estim<strong>at</strong>ion of the ‘connectivity’<br />
number m through the <strong>de</strong>pen<strong>de</strong>nce of µ on m documented<br />
in figure 6. Such a rel<strong>at</strong>ionship has already been argued by<br />
Johansen <strong>et</strong> al (2000) <strong>at</strong> a phenomenological level using a<br />
mean-field equ<strong>at</strong>ion in which the exponent is directly rel<strong>at</strong>ed<br />
to the number of connections to a given agent.<br />
5. St<strong>at</strong>istical properties of price r<strong>et</strong>urns<br />
in the symm<strong>et</strong>ric case<br />
Using the price dynamics (8), the distribution of p − 1/2 is the<br />
same as the distribution of r<strong>et</strong>urns, which is the first st<strong>at</strong>istical<br />
property analysed in econom<strong>et</strong>ric work (Campbell <strong>et</strong> al 1997,<br />
Lo and MacKinlay 1999, Lux 1996, Pagan 1996, Plerou <strong>et</strong> al<br />
1999, Laherrère and Sorn<strong>et</strong>te 1998). Note th<strong>at</strong> the distribution<br />
of p − 1/2 is nothing but the invariant measure of the chaotic<br />
map p ′ (p) which can be shown to be continuous with respect to<br />
the Lebesgue measure (Eckmann and Ruelle 1985). Figure 9<br />
shows the cumul<strong>at</strong>ive distribution of rt ∝ pt −1/2. Notice the<br />
two breaks <strong>at</strong> |p − 1/2| =0.28, which are due to the existence<br />
of weakly unstable periodic orbits corresponding to a transient<br />
oscill<strong>at</strong>ion b<strong>et</strong>ween bullish and bearish st<strong>at</strong>es.<br />
Figure 10 plots in double-logarithmic scales the survival<br />
(i.e. complementary cumul<strong>at</strong>ive) distribution of rt ∝ pt − 1/2<br />
for m = 30, 60 and 100. For m = 60, we can observe an<br />
approxim<strong>at</strong>e power-law tail but the exponent is smaller than<br />
1 in contradiction to the empirical evi<strong>de</strong>nce which suggests<br />
a tail of the survival probability with exponents 3–5. In
170 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
Probability<br />
1.0<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
Effect of periodic orbits<br />
0.0<br />
–0.4 –0.3 –0.2 –0.1 0.0 0.1 0.2 0.3 0.4<br />
p–1/2<br />
Figure 9. Cumul<strong>at</strong>ive distribution for m = 60 polled agents and the<br />
param<strong>et</strong>ers ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85.<br />
the other cases, we cannot conclu<strong>de</strong> on the existence of a<br />
power-law regime, but it is obvious th<strong>at</strong> the tail behaviour of<br />
the distribution function <strong>de</strong>pends on the number m of polled<br />
agents.<br />
Figure 11 shows the behaviour of the autocorrel<strong>at</strong>ion<br />
function for m = 60 and 100, with the same values of the<br />
other param<strong>et</strong>ers ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85.<br />
For m = 100, the presence of the weakly unstable orbits is felt<br />
much stronger, which is reflected in (i) a very strong periodic<br />
component of the correl<strong>at</strong>ion function and (ii) its slow <strong>de</strong>cay.<br />
Even for m = 60, the correl<strong>at</strong>ion function does not <strong>de</strong>cay<br />
fast enough compared to the typical dur<strong>at</strong>ion of specul<strong>at</strong>ive<br />
bubbles to be in quantit<strong>at</strong>ive agreement with empirical d<strong>at</strong>a.<br />
This anomalously large correl<strong>at</strong>ion of the r<strong>et</strong>urns is obviously<br />
rel<strong>at</strong>ed to the d<strong>et</strong>erministic dynamics of the r<strong>et</strong>urns. We thus<br />
expect th<strong>at</strong> including stochastic noise due to a finite number N<br />
of agents (see below) and adding external noise due to ‘news’<br />
will whiten rt significantly.<br />
Figure 12 compares the correl<strong>at</strong>ion function for the r<strong>et</strong>urns<br />
time series rt ∝ pt − 1/2 and the vol<strong>at</strong>ility time series <strong>de</strong>fined<br />
as |rt|. The vol<strong>at</strong>ility is an important measure of risks and thus<br />
plays an important role in portfolio managements and option<br />
pricing and hedging. Note th<strong>at</strong> taking the absolute value of the<br />
r<strong>et</strong>urn removes the one source of irregularity stemming from<br />
the change of sign of rt ∝ pt − 1/2 to focus on the local<br />
amplitu<strong>de</strong>s. We observe in figure 12 a significantly longer<br />
correl<strong>at</strong>ion time for the vol<strong>at</strong>ility. Moreover, the correl<strong>at</strong>ion<br />
function of the vol<strong>at</strong>ility first <strong>de</strong>cays exponentially and then as a<br />
power law. This behaviour has previously been documented in<br />
many econom<strong>et</strong>ric works (Ding <strong>et</strong> al 1993, Ding and Granger<br />
1996, Müller <strong>et</strong> al 1997, Dacorogna <strong>et</strong> al 1998, Arneodo <strong>et</strong> al<br />
1998, Ballocchi <strong>et</strong> al 1999, Muzy <strong>et</strong> al 2001).<br />
6. Asymm<strong>et</strong>ric cases<br />
We have seen th<strong>at</strong> the symm<strong>et</strong>ric case ρhb = ρbh and ρhh = ρbb<br />
is plagued by the weakly unstable periodic orbits which put a<br />
Probability<br />
10 0<br />
10 –1<br />
10 –5<br />
m = 30<br />
m = 60<br />
m = 100<br />
10 –4 10 –3<br />
Survival distribution<br />
10 –2<br />
1/2–p<br />
10 –1<br />
Figure 10. Survival (i.e. complementary cumul<strong>at</strong>ive) distribution<br />
for m = 30, 60 and 100 polled ‘friends’ per agent and param<strong>et</strong>ers<br />
ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85. The total number N of<br />
agents is infinite.<br />
strong and unrealistic imprint on the st<strong>at</strong>istical properties of<br />
the r<strong>et</strong>urn time series. It is n<strong>at</strong>ural to argue th<strong>at</strong> breaking the<br />
symm<strong>et</strong>ry will <strong>de</strong>stroy the strength of these periodic orbits.<br />
From a behavioural point of view, it is also quite clear th<strong>at</strong><br />
the <strong>at</strong>titu<strong>de</strong> of an investor is not symm<strong>et</strong>ric. One can expect<br />
a priori a stronger bullish buy-and-hold in<strong>de</strong>x ρhb − 1/2 than<br />
bearish buy-and-hold in<strong>de</strong>x ρbh − 1/2: one is a priori more<br />
prone to hold a position in a bullish mark<strong>et</strong> than in a bearish one.<br />
Similarly, we expect a smaller bullish reversal in<strong>de</strong>x 1 − ρhh<br />
than bearish reversal in<strong>de</strong>x 1 − ρbb: specul<strong>at</strong>ive bubbles are<br />
rarely seen on downward trends as it is much more common<br />
th<strong>at</strong> increasing prices are favourably perceived and can be sustained<br />
much longer without reference to the fundamental price.<br />
Such an asymm<strong>et</strong>ry has been clearly <strong>de</strong>monstr<strong>at</strong>ed<br />
empirically in the difference b<strong>et</strong>ween the r<strong>at</strong>e of occurrence<br />
and size of extreme drawdowns compared to drawups in stock<br />
mark<strong>et</strong> time series (Johansen and Sorn<strong>et</strong>te 2001). Drawdowns<br />
(drawups) are <strong>de</strong>fined as the cumul<strong>at</strong>ive losses (gains) from<br />
the last local maximum (minimum) to the next local minimum<br />
(maximum). Drawdowns and drawups are very interesting<br />
because they offer a more n<strong>at</strong>ural measure of real mark<strong>et</strong> risks<br />
than the variance, the value-<strong>at</strong>-risk or other measures based<br />
on fixed time scale distributions of r<strong>et</strong>urns. For the major<br />
stock mark<strong>et</strong> indices, there are very large drawdowns which are<br />
‘outliers’ while drawups do not exhibit such a drastic change<br />
of regime. For major companies, drawups of amplitu<strong>de</strong> larger<br />
than 15% occur <strong>at</strong> a r<strong>at</strong>e about twice as large as the r<strong>at</strong>e of<br />
drawdowns, but with lower absolute amplitu<strong>de</strong>. An asymm<strong>et</strong>ry<br />
ρhh = ρbb could result from the mechanism of asymm<strong>et</strong>ric<br />
inform<strong>at</strong>ion, in which actors on one si<strong>de</strong> of the mark<strong>et</strong> have<br />
b<strong>et</strong>ter inform<strong>at</strong>ion than those on the other (see Riley (2001) for<br />
a survey of <strong>de</strong>velopments in the economics of inform<strong>at</strong>ion over<br />
the last 25 years and The Bank of Swe<strong>de</strong>n Prize in Economics<br />
Sciences in Memory of Alfred Nobel 2001, document entitled<br />
‘Mark<strong>et</strong>s with Asymm<strong>et</strong>ric Inform<strong>at</strong>ion’, 10 October 2001).<br />
10 0<br />
273
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
p t<br />
Autocorrel<strong>at</strong>ion<br />
1.0<br />
m = 60<br />
1.0<br />
m = 100<br />
0.8<br />
0.8<br />
0.6<br />
0.6<br />
0.4<br />
0.4<br />
0.2<br />
0.2<br />
0.0<br />
0.0<br />
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000<br />
t t<br />
1.0<br />
1.0<br />
0.5<br />
0.5<br />
0.0<br />
–0.5<br />
–1.0<br />
0 20 40 60 80 100 120 140 160 180<br />
200<br />
p t<br />
Autocorrel<strong>at</strong>ion<br />
0.0<br />
–0.5<br />
–1.0<br />
0 20 40 60 80 100 120 140 160 180 200<br />
Time lag Time lag<br />
Figure 11. The upper panels represent the time series pt for m = 60 (left) and m = 100 (right). The lower panels represent the<br />
corresponding autocorrel<strong>at</strong>ion function of rt ∝ p − 1/2 for m = 60 (left) and m = 100 (right) with the same param<strong>et</strong>ers ρhb = ρbh = 0.72<br />
and ρhh = ρbb = 0.85.<br />
1.2<br />
1.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
Exponential<br />
<strong>de</strong>cay<br />
Power-law<br />
<strong>de</strong>cay<br />
R<strong>et</strong>urn<br />
Vol<strong>at</strong>ility<br />
0 50 100 150 200 250 300<br />
Figure 12. Autocorrel<strong>at</strong>ion function of the r<strong>et</strong>urns and of the<br />
vol<strong>at</strong>ility for m = 60 polled agents and the param<strong>et</strong>ers<br />
ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85.<br />
Figure 13 compares the dynamics for the symm<strong>et</strong>ric<br />
system (upper panel (a)) and for the asymm<strong>et</strong>ric system<br />
(lower panel (b)). It is clear th<strong>at</strong>, as expected, the number<br />
of periodic orbits <strong>de</strong>creases significantly in the asymm<strong>et</strong>ric<br />
system. However, there are still an unrealistic number of<br />
neg<strong>at</strong>ive bubbles. It is not possible to increase the asymm<strong>et</strong>ry<br />
sufficiently strongly without exiting from the chaotic regime.<br />
This unrealistic fe<strong>at</strong>ure is thus an intrinsic property and<br />
limit<strong>at</strong>ion of the present mo<strong>de</strong>l. We shall indic<strong>at</strong>e in the<br />
conclusion possible extensions and remedies.<br />
Figure 14 compares the cumul<strong>at</strong>ive distributions of p−1/2<br />
for m = 60 for the symm<strong>et</strong>ric and asymm<strong>et</strong>ric cases. The<br />
strong effect of the weakly unstable periodic orbits observed<br />
in the periodic case has disappeared. In addition, the tail of<br />
274<br />
171<br />
(a)<br />
1.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000<br />
(b)<br />
t<br />
1.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000<br />
t<br />
p t<br />
p t<br />
Figure 13. Time evolution of pt over 100 00 time steps for m = 60<br />
polled agents in (a) a symm<strong>et</strong>ric case ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85 and (b) an asymm<strong>et</strong>ric case ρhb = 0.72,<br />
ρbh = 0.74, ρhh = 0.85 and ρbb = 0.87.<br />
the distribution <strong>de</strong>cays faster in the asymm<strong>et</strong>ric case, in b<strong>et</strong>ter<br />
(but still not very good) agreement with empirical d<strong>at</strong>a.<br />
Figure 15 shows the correl<strong>at</strong>ion function of the r<strong>et</strong>urns for<br />
a symm<strong>et</strong>ric and an asymm<strong>et</strong>ric case. In the asymm<strong>et</strong>ric case,<br />
there is no trace of oscill<strong>at</strong>ions but the <strong>de</strong>cay is slightly slower.<br />
7. Finite-size effects<br />
Until now, our analysis has focused on the limit of an infinite<br />
number N →∞of agents, in which each agent polls randomly<br />
m agents among N. In this limit, we have shown th<strong>at</strong>, for a large<br />
domain in the param<strong>et</strong>er space, the dynamics of the r<strong>et</strong>urns is<br />
chaotic with interesting and qualit<strong>at</strong>ively realistic properties.
172 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
Probability<br />
1.0<br />
Symm<strong>et</strong>ric dynamics<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
Asymm<strong>et</strong>ric dynamics<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
Effect of periodic<br />
orbits<br />
–0.5 –0.4 –0.3 –0.2 –0.1 0.0<br />
p–1/2<br />
0.1 0.2 0.3 0.4 0.5<br />
Figure 14. Cumul<strong>at</strong>ive distribution function of p − 1/2 for m = 60<br />
polled agents and param<strong>et</strong>ers ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85 (dashed line) and ρhb = 0.72, ρbh = 0.74,<br />
ρhh = 0.85 and ρbb = 0.87 (continuous line). Note th<strong>at</strong> the<br />
asymm<strong>et</strong>ric dynamics has the effect of shifting the cumul<strong>at</strong>ive<br />
distribution to the right.<br />
7.1. Finite-size effects in other mo<strong>de</strong>ls<br />
We now investig<strong>at</strong>e finite-size effects resulting from a finite<br />
number N of interacting agents trading on the stock mark<strong>et</strong>.<br />
This issue of the role of the number of agents has recently been<br />
investig<strong>at</strong>ed vigorously with surprising results.<br />
Hellthaler (1995) studied the N-<strong>de</strong>pen<strong>de</strong>nce of the<br />
dynamical properties of price time series of the Levy <strong>et</strong> al<br />
mo<strong>de</strong>l (1995, 2000). Egenter <strong>et</strong> al (1999) did the same for the<br />
Kim and Markowitz (1989) and the Lux and Marchesi (1999)<br />
mo<strong>de</strong>ls. They found th<strong>at</strong>, if this number N goes to infinity,<br />
nearly periodic oscill<strong>at</strong>ions occur and the st<strong>at</strong>istical properties<br />
of the price time series become compl<strong>et</strong>ely unrealistic. Stauffer<br />
(1999) reviewed this work and others such as the Levy <strong>et</strong> al<br />
(1995, 2000) mo<strong>de</strong>l: realistically looking price fluctu<strong>at</strong>ions<br />
are obtained for N ∝ 102 , but for N ∝ 106 the prices<br />
vary smoothly in a nearly periodic and thus unrealistic way.<br />
The mo<strong>de</strong>l proposed by Farmer (1998) suffers from the same<br />
problem: with a few hundred investors, most investors are<br />
fundamentalists during calm times, but bursts of high vol<strong>at</strong>ility<br />
coinci<strong>de</strong> with large fractions of noise tra<strong>de</strong>rs. When N<br />
becomes much larger, the fraction of noise tra<strong>de</strong>rs goes to zero<br />
in contradiction to reality. On a somewh<strong>at</strong> different issue,<br />
Huang and Solomon (2001) have studied finite-size effects<br />
in dynamical systems of price evolution with multiplic<strong>at</strong>ive<br />
noise. They find th<strong>at</strong> the exponent of the Par<strong>et</strong>o law<br />
obtained in stochastic multiplic<strong>at</strong>ive mark<strong>et</strong> mo<strong>de</strong>ls is crucially<br />
affected by a finite N and may cause, in the absence of<br />
an appropri<strong>at</strong>e social policy, extreme wealth inequality and<br />
mark<strong>et</strong> instability. Another mo<strong>de</strong>l (apart from ours) where<br />
the mark<strong>et</strong> may stay realistic even for N →∞seems to be<br />
the Cont and Bouchaud percol<strong>at</strong>ion mo<strong>de</strong>l (2001). However,<br />
this only occurs for an unrealistic tuning of the percol<strong>at</strong>ion<br />
concentr<strong>at</strong>ion to its critical value. Thus, in most cases, the<br />
limit N →∞leads to a behaviour of the simul<strong>at</strong>ed mark<strong>et</strong>s<br />
Correl<strong>at</strong>ion<br />
1.2<br />
1.0<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0.0<br />
–0.2<br />
Symm<strong>et</strong>ric dynamics<br />
Asymm<strong>et</strong>ric dynamics<br />
0 20 40 60 80 100 120 140 160 180 200<br />
Time lag<br />
Figure 15. Correl<strong>at</strong>ion function for m = 60 polled agents and<br />
param<strong>et</strong>ers ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85 (dashed line)<br />
and ρhb = 0.72, ρbh = 0.74, ρhh = 0.85 and ρbb = 0.87 (continuous<br />
line). Note th<strong>at</strong> the correl<strong>at</strong>ion function of r<strong>et</strong>urns goes to zero<br />
(within the noise level) <strong>at</strong> time lags larger than about 100. This<br />
unrealistic fe<strong>at</strong>ure of a long-range correl<strong>at</strong>ion in the r<strong>et</strong>urns makes it<br />
unnecessary to show the even longer-range correl<strong>at</strong>ion of the<br />
absolute values of the r<strong>et</strong>urns.<br />
which becomes quite smooth or periodic and thus predictable,<br />
in contrast to real mark<strong>et</strong>s. Our mo<strong>de</strong>l, which remains<br />
(d<strong>et</strong>erministically) chaotic, is thus a significant improvement<br />
upon this behaviour. We trace this improvement on the highly<br />
nonlinear behaviour resulting from the interplay b<strong>et</strong>ween<br />
the imit<strong>at</strong>ive and contrarian behaviour. It has thus been<br />
argued (Stauffer 1999) th<strong>at</strong>, if these previous mo<strong>de</strong>ls are good<br />
<strong>de</strong>scriptions of mark<strong>et</strong>s, then real mark<strong>et</strong>s with their strong<br />
random fluctu<strong>at</strong>ions are domin<strong>at</strong>ed by a r<strong>at</strong>her limited number<br />
of large players: this amounts to the assumption th<strong>at</strong> the<br />
hundred most important investors or investment companies<br />
have much more influence than the millions of less wealthy<br />
priv<strong>at</strong>e investors.<br />
There is another class of mo<strong>de</strong>ls, the minority games<br />
(Chall<strong>et</strong> and Zhang 1997), in which the dynamics remains<br />
complex even in the limit N →∞. It has been established<br />
th<strong>at</strong> the fluctu<strong>at</strong>ions of the sum of the aggreg<strong>at</strong>e <strong>de</strong>mand<br />
have an approxim<strong>at</strong>e scaling with similar sized fluctu<strong>at</strong>ions<br />
(vol<strong>at</strong>ility/standard <strong>de</strong>vi<strong>at</strong>ion) for any N and m for the scale<br />
scaled variable 2 m /N, where m is the memory length (Chall<strong>et</strong><br />
<strong>et</strong> al 2000). In a generaliz<strong>at</strong>ion, the so-called grand canonical<br />
version of the minority game (Jefferies <strong>et</strong> al 2001), where the<br />
agents have a confi<strong>de</strong>nce threshold th<strong>at</strong> prevents them from<br />
playing if their str<strong>at</strong>egies have not been successful over the<br />
last T turns, the dynamics can <strong>de</strong>pend more sensitively on N:<br />
as N becomes small, the dynamics can become quite different.<br />
For large N, the complexity remains.<br />
The difference b<strong>et</strong>ween the limit N →∞consi<strong>de</strong>red up<br />
to now in this paper and the case of finite N is th<strong>at</strong> pt is no<br />
longer the fraction of bullish agents. For finite N, pt must<br />
be interpr<strong>et</strong>ed as the probability for an agent to be bullish.<br />
Of course, in the limit of large N, the law of large numbers<br />
275
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
(a)<br />
1.0<br />
p t<br />
0.5<br />
0.0<br />
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000<br />
(b)<br />
t<br />
1.0<br />
p t<br />
p t<br />
0.5<br />
0.0<br />
(c)<br />
0.4<br />
0.2<br />
0<br />
–0.2<br />
–0.4<br />
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000<br />
t<br />
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000<br />
t<br />
Figure 16. Time evolution of pt over 10 000 time steps for m = 60<br />
polled agents with (a) N =∞, (b) N = m +1= 61 agents and<br />
param<strong>et</strong>ers ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85. Panel<br />
(c) represents the noise due to the finite size of the system and is<br />
obtained by subtracting the time series in panel (a) from the time<br />
series in panel (b).<br />
ensures th<strong>at</strong> the fraction of bullish agents becomes equal to the<br />
probability for an agent to be bullish. There are several ways<br />
to implement a finite-size effect. We here discuss only the two<br />
simplest ones.<br />
7.2. Finite external sampling of an infinite system<br />
Consi<strong>de</strong>r a system with an infinite number of agents for which<br />
the fraction pt of bullish agents is governed by the d<strong>et</strong>erministic<br />
dynamics (2). At each time step t, l<strong>et</strong> us sample a finite number<br />
N of them to d<strong>et</strong>ermine the fraction of bullish agents. We g<strong>et</strong><br />
a number n, which is in general close but not exactly equal to<br />
Npt due to st<strong>at</strong>istical fluctu<strong>at</strong>ions. The probability of finding n<br />
bullish agents among N agents is in<strong>de</strong>ed given by the binomial<br />
law<br />
<br />
N<br />
Pr(n) = p<br />
n<br />
n (1 − p) N−n . (25)<br />
This shows th<strong>at</strong> the observed proportion ˜p = n/N of bullish<br />
agents is asymptotically normal with mean p and standard<br />
<strong>de</strong>vi<strong>at</strong>ion 1/ √ p(1 − p)N :Pr( ˜p) ∝ N (p, 1/ √ p(1 − p)N).<br />
Iter<strong>at</strong>ing the sampling among N agents <strong>at</strong> each time step gives<br />
a noisy dynamics ˜pt shadowing the true d<strong>et</strong>erministic one.<br />
Figure 16 compares the dynamics of the d<strong>et</strong>erministic pt<br />
corresponding to N →∞(panel (a)) with ˜pt for a number<br />
N = m +1 = 61 of sampled agents among the infinite<br />
ensemble of them (panel (b)). Panel (c) is the ‘noise’ time<br />
series <strong>de</strong>fined as ˜pt − pt, i.e. by subtracting the time series<br />
of panel (a) from the time series of panel (b). The noise time<br />
series of panel (c) thus represents the st<strong>at</strong>istical fluctu<strong>at</strong>ions<br />
due to the finite sampling of agents’ opinions. Figure 16(b)<br />
shows the characteristic vol<strong>at</strong>ility clusters which is one of the<br />
most important stylized properties of empirical time series.<br />
276<br />
Correl<strong>at</strong>ion<br />
1.0<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
–0.1<br />
173<br />
0 5 10 15 20 25 30 35 40 45 50<br />
Time lag<br />
Figure 17. Correl<strong>at</strong>ion function for m = 60 polled agents with<br />
N =∞(thin curve), N = 600 (dashed curve) and N = 61<br />
(continuous curve) agents and param<strong>et</strong>ers ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85.<br />
For large N, we can write<br />
˜pt = pt +<br />
1<br />
√ pt(1 − pt)N Wt<br />
(26)<br />
where {Wt} are in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically distributed<br />
Gaussian variables with zero mean and unit variance.<br />
Therefore, the correl<strong>at</strong>ion function corrN(τ) <strong>at</strong> lag τ = 0is<br />
obtained from th<strong>at</strong> for N →∞by multiplic<strong>at</strong>ion by a constant<br />
factor:<br />
Nvar(p)<br />
corrN(τ) =<br />
E[1/{p(1 − p)}]+Nvar(p)<br />
× corr∞(τ) and τ = 0, (27)<br />
corr∞(τ) for large N, (28)<br />
where E[x] <strong>de</strong>notes the expect<strong>at</strong>ion of x with respect to the<br />
continuous invariant measure of the dynamical system (2).<br />
Note th<strong>at</strong> E[1/{p(1 − p)}] always exists for m
174 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
1.0<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
p¢ 0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0<br />
p<br />
Figure 18. R<strong>et</strong>urn map of the fraction of bullish agents for m = 60<br />
polled agents among N = 61 agents (points) and the d<strong>et</strong>erministic<br />
trajectory (continuous curve) corresponding to N =∞agents. The<br />
param<strong>et</strong>ers are ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85.<br />
agents in the mark<strong>et</strong> (this is realistic) and they are in contact<br />
with only m other agents th<strong>at</strong> they poll <strong>at</strong> each time period. Not<br />
knowing the true value of N but assuming it to be large, it is<br />
r<strong>at</strong>ional for them to <strong>de</strong>velop the best predictor of the dynamics<br />
by assuming the i<strong>de</strong>al case of an infinite number of agents with<br />
m polled agents and thus use the d<strong>et</strong>erministic dynamics (2) as<br />
their best predictor.<br />
At each time period t, each agent thus chooses randomly<br />
m agents th<strong>at</strong> she polls. She then counts the number of bullish<br />
and bearish agents among her polled sample of m agents.<br />
This number divi<strong>de</strong>d by m gives her an estim<strong>at</strong>ion ˆpt of<br />
the probability pt being bullish <strong>at</strong> time t. Introducing this<br />
estim<strong>at</strong>ion in the d<strong>et</strong>erministic equ<strong>at</strong>ion (2), the agent obtains<br />
a forecast ˆp ′ of the true probability p ′ being bullish <strong>at</strong> the next<br />
time step.<br />
Results of the simul<strong>at</strong>ions of this mo<strong>de</strong>l are shown in figure<br />
21. We observe a significantly stronger ‘noise’ compared<br />
to the previous section, which is expected since the noise is<br />
itself injected in the dynamical equ<strong>at</strong>ion <strong>at</strong> each time step. As<br />
a consequence, the correl<strong>at</strong>ion function of the r<strong>et</strong>urns and of<br />
the vol<strong>at</strong>ility <strong>de</strong>cay faster than their d<strong>et</strong>erministic counterpart.<br />
The correl<strong>at</strong>ion of the vol<strong>at</strong>ility still <strong>de</strong>cays about ten times<br />
slower than the correl<strong>at</strong>ion of the r<strong>et</strong>urns, but this clustering of<br />
vol<strong>at</strong>ility is not sufficiently strong compared to empirical facts.<br />
As in the <strong>de</strong>finition of the mo<strong>de</strong>l presented in section 2,<br />
we use here again a ‘mean-field’ approxim<strong>at</strong>ion, valid only if<br />
each agent selects randomly her m agents to w<strong>at</strong>ch, with a new<br />
selection compl<strong>et</strong>ely in<strong>de</strong>pen<strong>de</strong>nt of the previous one ma<strong>de</strong> <strong>at</strong><br />
every new time step. This unrealistic but simplifying fe<strong>at</strong>ure<br />
is the one we have followed in our simul<strong>at</strong>ions. If each agent<br />
has some neighbouring agents which she will w<strong>at</strong>ch again<br />
and again, then correl<strong>at</strong>ions will build up which are ignored<br />
compl<strong>et</strong>ely here; some correl<strong>at</strong>ions of this type are presumably<br />
more realistic. This is left for future works.<br />
Other more realistic mo<strong>de</strong>ls of a finite number of agents<br />
can be introduced. For instance, <strong>at</strong> time t, consi<strong>de</strong>r an agent<br />
1.0<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
p¢<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0.0<br />
0.0 0.1 0.2 0.3 0.4 0.5<br />
p<br />
0.6 0.7 0.8 0.9 1.0<br />
Figure 19. R<strong>et</strong>urn map of the fraction of bullish agents for m = 60<br />
polled agents among N = 600 agents (points) and the d<strong>et</strong>erministic<br />
trajectory (continuous curve) corresponding to N =∞agents. The<br />
param<strong>et</strong>ers are ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85.<br />
among the N. She chooses m other agents randomly and<br />
polls them. Each of them is either bullish or bearish as a<br />
result of <strong>de</strong>cisions taken during the previous time period. She<br />
then counts the number of bullish agents among the m, and<br />
then d<strong>et</strong>ermines her new <strong>at</strong>titu<strong>de</strong> using the rules (1). If she is<br />
polled <strong>at</strong> time t + 1 by another agent, her <strong>at</strong>titu<strong>de</strong> will be the<br />
one d<strong>et</strong>ermined from t to t + 1. In this way, we never refer<br />
to the d<strong>et</strong>erministic dynamics pt but only to its un<strong>de</strong>rlying<br />
rules. As a consequence, this d<strong>et</strong>erministic dynamics does<br />
not exert an <strong>at</strong>traction th<strong>at</strong> minimizes the effect of st<strong>at</strong>istical<br />
fluctu<strong>at</strong>ions due to finite sizes. This approach is similar to<br />
going from a Fokker–Planck equ<strong>at</strong>ion (equ<strong>at</strong>ion (2)) to a<br />
Langevin equ<strong>at</strong>ion with finite-size effects. This class of mo<strong>de</strong>ls<br />
will be investig<strong>at</strong>ed elsewhere.<br />
8. Conclusions<br />
The traditional concept of stock mark<strong>et</strong> dynamics envisions a<br />
stream of stochastic ‘news’ th<strong>at</strong> may move prices in random<br />
directions. This paper, in contrast, <strong>de</strong>monstr<strong>at</strong>es th<strong>at</strong> certain<br />
types of d<strong>et</strong>erministic behaviour—mimicry and contradictory<br />
behaviour alone—can already lead to chaotic prices.<br />
If the mark<strong>et</strong> prices are assumed to follow the pt<br />
behaviour, our <strong>de</strong>scription refers to the well-known evolution<br />
of the specul<strong>at</strong>ive bubbles. Such apparent regularities often<br />
occur in the stock mark<strong>et</strong> and form the basis of the so-called<br />
‘technical analysis’ whereby tra<strong>de</strong>rs <strong>at</strong>tempt to predict future<br />
price movements by extrapol<strong>at</strong>ing certain p<strong>at</strong>terns from recent<br />
historical prices. Our mo<strong>de</strong>l provi<strong>de</strong>s an explan<strong>at</strong>ion of birth,<br />
life and <strong>de</strong><strong>at</strong>h of the specul<strong>at</strong>ive bubbles in this context.<br />
While the traditional theory of r<strong>at</strong>ional anticip<strong>at</strong>ions<br />
exhibits and emphasizes self-reinforcing mechanisms, without<br />
either predicting their inception nor their collapse, the strength<br />
of our mo<strong>de</strong>l is to justify the occurrence of specul<strong>at</strong>ive<br />
bubbles. It allows for their collapse by taking into account<br />
277
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
p t<br />
p t<br />
p t<br />
0.5<br />
0.0<br />
–0.5<br />
0 1 2 3 4 5 6 7 8 9 10<br />
2.5<br />
2.0<br />
1.5<br />
1.0<br />
0.5<br />
0.0<br />
t<br />
0 1 2 3 4 5<br />
t<br />
6 7 8 9 10<br />
10 1<br />
10 0<br />
10 –1<br />
0 1 2 3 4 5<br />
t<br />
6 7 8 9 10<br />
Figure 20. Upper panel: r<strong>et</strong>urn trajectory ˜rt = γ ˜pt − 1/2 for<br />
m = 100, N = 100, ρhb = ρbh = 0.72 and ρhh = ρbb = 0.85 and<br />
γ = 0.01. Middle panel: price trajectory obtained by<br />
πt = πt−1 exp[˜rt]. Lower panel: same as the middle panel with πt<br />
shown on a logarithmic scale. Note the ‘fl<strong>at</strong> trough–sharp peak’<br />
structure of the log-price trajectory (Roehner and Sorn<strong>et</strong>te 1998).<br />
the combin<strong>at</strong>ion of mim<strong>et</strong>ic and antagonistic behaviour in the<br />
form<strong>at</strong>ion of expect<strong>at</strong>ions about prices.<br />
The specific fe<strong>at</strong>ure of the mo<strong>de</strong>l is to combine these two<br />
Keynesian aspects of specul<strong>at</strong>ion and enterprise and to <strong>de</strong>rive<br />
from them behavioural rules based on collective opinion: the<br />
agents can adopt an imit<strong>at</strong>ive and gregarious behaviour or,<br />
on the contrary, anticip<strong>at</strong>e a reversal of ten<strong>de</strong>ncy, thereby<br />
d<strong>et</strong>aching themselves from the current trend. It is this duality,<br />
the continuous coexistence of these two elements, which is <strong>at</strong><br />
the origin of the properties of our mo<strong>de</strong>l: chaotic behaviour<br />
and the gener<strong>at</strong>ion of bubbles.<br />
It is a common wisdom th<strong>at</strong> d<strong>et</strong>erministic chaos leads to<br />
fundamental limits of predictability because the tiny inevitable<br />
fluctu<strong>at</strong>ions in those chaotic systems quickly snowball in<br />
unpredictable ways. This has been investig<strong>at</strong>ed in rel<strong>at</strong>ion<br />
to, for instance, long-term we<strong>at</strong>her p<strong>at</strong>terns. However, in<br />
the context of our mo<strong>de</strong>ls, we have shown th<strong>at</strong> the chaotic<br />
dynamics of the r<strong>et</strong>urns alone cannot be the limiting factor<br />
for predictability, as it contains too many residual correl<strong>at</strong>ions.<br />
Endogenous fluctu<strong>at</strong>ions due to finite-size effects and external<br />
news (noise) seem to be nee<strong>de</strong>d as important factors leading to<br />
the observed randomness of stock mark<strong>et</strong> prices. The rel<strong>at</strong>ion<br />
b<strong>et</strong>ween these extrinsic factors and the intrinsic ones studied<br />
in this paper will be explored elsewhere.<br />
Remarks and acknowledgements<br />
This paper is an outgrowth and extension of unpublished work<br />
(1994) by three of us (AC, JPE, AM) which was in turn based<br />
on the PhD of Anne Corcos in 1993. We are gr<strong>at</strong>eful to<br />
J V An<strong>de</strong>rsen for useful discussions. This work was partially<br />
supported by the Fonds N<strong>at</strong>ional Suisse (JPE and AM) and<br />
278<br />
Figure 21. Evolution of the system over 10 000 time steps for<br />
m = 60 polled agents with (upper panel) N =∞, (second panel)<br />
N = m +1= 61 and param<strong>et</strong>ers ρhb = ρbh = 0.72 and<br />
ρhh = ρbb = 0.85. The lower panel represents the ‘noise’<br />
introduced by the finite size of the system and is obtained by<br />
subtracting the upper panel from the second panel.<br />
175<br />
by the James S Mc Donnell Found<strong>at</strong>ion 21st century scientist<br />
award/studying complex system (DS).<br />
Appendix<br />
We expand Fm(p) around the fixed point p = 1/2, so th<strong>at</strong>,<br />
using the symm<strong>et</strong>ry of Fm(p)<br />
Fm(p) = 1<br />
2 +F ′ m<br />
1<br />
(1/2)(p− 2<br />
(3)<br />
1<br />
)+F m (1/2)(p− 2 )3 +···. (29)<br />
First of all, it is obvious to show by recursion th<strong>at</strong><br />
F ′ m (1/2) = 1 − 2gm(1/2) − g ′ m (1/2)<br />
F<br />
(30)<br />
(2k+1)<br />
m (1/2) =−2(2k +1)g (2k)<br />
m (1/2)<br />
− g (2k+1)<br />
m (1/2) if k>0. (31)<br />
The problem thus amounts to calcul<strong>at</strong>ing the <strong>de</strong>riv<strong>at</strong>ives of gm.<br />
Some simple algebraic manipul<strong>at</strong>ions allow us to obtain<br />
g ′ m−1 <br />
<br />
m − 1<br />
m (p) = m<br />
p<br />
j<br />
j=0<br />
m1−j (1 − p) j<br />
<br />
j j +1<br />
× f − f<br />
m m<br />
m−1 <br />
<br />
m − 1<br />
=−m<br />
p<br />
j<br />
j=0<br />
(32)<br />
m1−j (1 − p) j 1fm(j), (33)<br />
where 1fm(·) is the first-or<strong>de</strong>r discr<strong>et</strong>e <strong>de</strong>riv<strong>at</strong>ive of f( ·<br />
m ),<br />
which yields<br />
<br />
1<br />
=−<br />
2<br />
m<br />
2m−1 m−1 <br />
<br />
m − 1<br />
1fm(j). (34)<br />
j<br />
g ′ m<br />
j=0
176 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
By recursion, it is easy to prove th<strong>at</strong><br />
g (k)<br />
m<br />
<br />
1<br />
2<br />
<br />
= (−1)k m!<br />
2m−k m−k <br />
<br />
m − k<br />
kfm(j) (35)<br />
k! j<br />
j=0<br />
and kfm(·) is the k th or<strong>de</strong>r discr<strong>et</strong>e <strong>de</strong>riv<strong>at</strong>ive of f( ·<br />
m ):<br />
Finally,<br />
kfm(j) =<br />
k<br />
i=0<br />
<br />
k<br />
(−1)<br />
i<br />
i f<br />
j + i<br />
m<br />
<br />
. (36)<br />
F (2k+1)<br />
m!<br />
m (1/2) =<br />
2m−2k−1 (2k)!<br />
m−2k−1<br />
1 <br />
<br />
m − 2k − 1<br />
×<br />
2k+1fm(j).<br />
2k +1<br />
j<br />
j=0<br />
m−2k <br />
<br />
m − 2k<br />
− (2k +1)<br />
2kfm(j) . (37)<br />
j<br />
References<br />
j=0<br />
Arneodo A, Muzy J F and Sorn<strong>et</strong>te D 1998 ‘Direct’ causal casca<strong>de</strong><br />
in the stock mark<strong>et</strong> Eur. Phys. J. B 2 277–82<br />
Arthur W B 1987 Self-reinforcing mechanisms in economics Center<br />
Econ. Policy Res. 111 1–20<br />
Azariadis C 1981 Self fulfilling prophecies J. Econ. Theory 25<br />
380–96<br />
Azariadis C and Guesnerie R 1982 Prophéties autoréalis<strong>at</strong>rices <strong>et</strong><br />
persistance <strong>de</strong>s <strong>théorie</strong>s, Rev. Économique 33 878–906<br />
Ballocchi G, Dacorogna M M and Gencay R 1999 Intraday<br />
st<strong>at</strong>istical properties of eurofutures by barbara piccin<strong>at</strong>o<br />
Deriv<strong>at</strong>ives Q. 6 28–44<br />
Benhabib J and Day R H 1981 R<strong>at</strong>ional choice and err<strong>at</strong>ic behaviour<br />
Rev. Econ. Studies July 153<br />
Bikhchandani S, Hirshleifer D and Welch I 1992 A theory of fads,<br />
fashion, custom and cultural changes as inform<strong>at</strong>ional casca<strong>de</strong>s<br />
J. Political Economy 100 992–1026<br />
Blanchard O and W<strong>at</strong>son M W 1982 Bubbles, r<strong>at</strong>ional expect<strong>at</strong>ions<br />
and financial mark<strong>et</strong>s ed P Wachtel Crises in the Economic and<br />
Financial Structure (Lexington Books) pp 295–315<br />
Bollerslev T 1987 A conditional h<strong>et</strong>eroskedastic time series mo<strong>de</strong>l<br />
for specul<strong>at</strong>ive prices and r<strong>at</strong>es of r<strong>et</strong>urn Rev. Econ. St<strong>at</strong>istics<br />
69 542–7<br />
Bollerslev T, Chou R Y, Jayaraman N and Kroner K F 1991 Les<br />
modèles ARCH en finance: un point sur la <strong>théorie</strong> <strong>et</strong> les<br />
résult<strong>at</strong>s empiriques Ann. d’Economie St<strong>at</strong>istiques 24 1–59<br />
Bouchaud J-P and Cont R 1998 A Langevin approach to stock<br />
mark<strong>et</strong> fluctu<strong>at</strong>ions and crashes Eur. Phys. J. B 6 543–50<br />
Brock W A 1988 Nonlinearity and complex dynamics in economics<br />
and finance The Economy as an Evolving Complex System ed<br />
P W An<strong>de</strong>rson, K J Arrow and D Pines (Reading, MA:<br />
Addison-Wesley)<br />
Brock W A and Dechert W D 1988 Theorems on distinguishing<br />
d<strong>et</strong>erministic from random systems Dynamic Econom<strong>et</strong>ric<br />
Mo<strong>de</strong>ling ed W A Barn<strong>et</strong>t, E R Berndt and H White<br />
(Cambridge: Cambridge University Press) ch 12 pp 247–65<br />
Brock W A, Dechert W D and Scheinkman J 1987 A test for<br />
in<strong>de</strong>pen<strong>de</strong>nce based on the correl<strong>at</strong>ion dimension Working<br />
paper University of Wisconsin <strong>at</strong> Madison, University of<br />
Houston and University of Chicago<br />
Brock W A, Hsieh D and LeBaron B 1991 Nonlinear Dynamics,<br />
Chaos and Instability: St<strong>at</strong>istical Theory and Economic<br />
Evi<strong>de</strong>nce (Cambridge: MAD Press)<br />
Campbell J Y, Lo A W and MacKinlay A C 1997 The Econom<strong>et</strong>rics<br />
of Financial Mark<strong>et</strong>s (Princ<strong>et</strong>on, NJ: Princ<strong>et</strong>on University<br />
Press)<br />
Chall<strong>et</strong> D, Marsili M and Zecchina R 2000 St<strong>at</strong>istical mechanics of<br />
systems with h<strong>et</strong>erogeneous agents: minority games Phys. Rev.<br />
L<strong>et</strong>t. 84 1824–7<br />
Chall<strong>et</strong> D, Marsili M and Zhang Y C 2000 Mo<strong>de</strong>ling mark<strong>et</strong><br />
mechanism with minority game Physica A 276 284–315<br />
Chall<strong>et</strong> D and Zhang Y C 1997 Emergence of cooper<strong>at</strong>ion and<br />
organiz<strong>at</strong>ion in an evolutionary game Physica A 246 407–18<br />
Coll<strong>et</strong> P and Eckmann J-P 1980 Iter<strong>at</strong>ed Maps on the Interval as<br />
Dynamical Systems (Boston, MA: Birkhäuser)<br />
Cont R and Bouchaud J-P 2000 Herd behaviour and aggreg<strong>at</strong>e<br />
fluctu<strong>at</strong>ions in financial mark<strong>et</strong>s Macroeconomic Dynamics 4<br />
170–96<br />
Corcos A 1993 Bruit <strong>et</strong> chaos sur les marchés financiers Thèse <strong>de</strong><br />
Doctor<strong>at</strong> Université Panthéon-Assas, Paris<br />
Dacorogna M M, Müller U A, Olsen R B and Pict<strong>et</strong> O V 1998<br />
Mo<strong>de</strong>lling short-term vol<strong>at</strong>ility with garch and harch mo<strong>de</strong>ls<br />
Nonlinear Mo<strong>de</strong>lling of High Frequency Financial Time Series<br />
ed C Dunis and B Zhou (New York: Wiley)<br />
Day R H 1982 Irregular growth cycles Am. Econ. Rev. 72 406–14<br />
Day R H 1983 The emergence of chaos from classical economic<br />
growth Q. J. Economics 48 201–13<br />
De Bondt WFMandThaler R H 1995 Financial <strong>de</strong>cision-making<br />
in mark<strong>et</strong>s and firms: a behavioural perspective Finance<br />
Handbooks in Oper<strong>at</strong>ions Research and Management Science<br />
vol 9 ed R A Jarrow, V Maksimovic, W T Ziemba (Amsterdam:<br />
Elsevier) pp 385–410<br />
De Grauwe P, Dewachter H and Embrechts M 1993 Foreign<br />
Exchange Mo<strong>de</strong>ls (Oxford: Blackwell)<br />
De Grauwe P and Vansanten K 1990 D<strong>et</strong>erministic chaos in the<br />
foreign exchange mark<strong>et</strong> Working Paper, CEPR K<strong>at</strong>holieke<br />
Universiteit Leuven, Belgium<br />
Ding Z and Granger CWJ1996 Mo<strong>de</strong>ling vol<strong>at</strong>ility persistence of<br />
specul<strong>at</strong>ive r<strong>et</strong>urns: a new approach J. Econom<strong>et</strong>rics 73<br />
185–215<br />
Ding Z, Granger CWJandEngle R 1993 A long memory property<br />
of stock r<strong>et</strong>urns and a new mo<strong>de</strong>l J. Empirical Finance 1<br />
83–106<br />
Eckmann J-P 1981 Roads to turbulence in dissip<strong>at</strong>ive dynamical<br />
systems Rev. Mod. Phys. 53 643–54 (P Cvitanović (ed) 1984<br />
Universality in Chaos (Bristol: Adam Hilger) (reprint))<br />
Eckmann J-P, Oliffson Kamphorst S, Ruelle D and Scheinkman J<br />
1988 Lyapunov exponents for stock r<strong>et</strong>urns The Economy as an<br />
Evolving Complex System ed P W An<strong>de</strong>rson, K J Arrow and<br />
D Pines (Reading, MA: Addison-Wesley)<br />
Eckmann J-P and Ruelle D 1985 Ergodic theory of chaos and<br />
strange <strong>at</strong>tractors Rev. Mod. Phys. 57 617–56<br />
Eckmann J-P and Ruelle D 1992 Fundamental limit<strong>at</strong>ions for<br />
estim<strong>at</strong>ing dimensions and lyapunov exponents in dynamical<br />
systems Physica D 56 185–7<br />
Egenter E, Lux T and Stauffer D 1999 Finite-size effects in Monte<br />
Carlo simul<strong>at</strong>ions of two stock mark<strong>et</strong> mo<strong>de</strong>ls Physica A 268<br />
250–6<br />
Engle R F 1982 Autoregressive conditional h<strong>et</strong>eroskedasticity with<br />
estim<strong>at</strong>es of the variance of UK infl<strong>at</strong>ion Econom<strong>et</strong>rica 50<br />
987–1008<br />
Fama E 1965 The behaviour of stock mark<strong>et</strong> prices J. Business 38<br />
34–105<br />
Farmer J D 1998 Mark<strong>et</strong> force, ecology and evolution Preprint<br />
adap-org/9812005<br />
Farmer J D and Joshi S 2002 The price dynamics of common<br />
str<strong>at</strong>egies J. Econ. Behav. Organ. (preprint) <strong>at</strong> press<br />
Feller W 1966 An Introduction to Probability Theory and its<br />
Applic<strong>at</strong>ions vol 2 (New York: Wiley)<br />
Flavin M A 1983 Excess vol<strong>at</strong>ility in the financial mark<strong>et</strong>s: a<br />
reassessment of the empirical evi<strong>de</strong>nce J. Political Economy 91<br />
929–56<br />
279
A Corcos <strong>et</strong> al Q UANTITATIVE F INANCE<br />
Galam S 1997 R<strong>at</strong>ional group <strong>de</strong>cision making: a random field Ising<br />
mo<strong>de</strong>l <strong>at</strong> T = 0 Physica A 238 66–80<br />
Galam S and Moscovici S 1991 Towards a theory of collective<br />
phenomena: consensus and <strong>at</strong>titu<strong>de</strong> changes in groups Eur. J.<br />
Social Psy. 21 49–74<br />
Goldberg J and von Nitzsch R 2001 (transl<strong>at</strong>ed Morris A)<br />
Behavioural Finance (Chichester: Wiley)<br />
Graham B and Dodd D L 1934 Security Analysis 1st edn (New<br />
York: McGraw-Hill)<br />
Grandmont J-P 1985 On endogenous comp<strong>et</strong>itive business cycles<br />
Econom<strong>et</strong>rica 53 995–1045<br />
Grandmont J-P 1987 Nonlinear Economic Dynamics (New York:<br />
Aca<strong>de</strong>mic)<br />
Grandmont J-P and Malgrange P 1986 Nonlinear economic<br />
dynamics: introduction J. Econ. Theory 40 3–11<br />
Grossman S J 1977 The existence of future mark<strong>et</strong>s, noisy r<strong>at</strong>ional<br />
expect<strong>at</strong>ions and inform<strong>at</strong>ional externalities Rev. Econ. Studies<br />
44 431–49<br />
Grossman S J 1981 An introduction to the theory of r<strong>at</strong>ional<br />
expect<strong>at</strong>ions un<strong>de</strong>r asymm<strong>et</strong>ric inform<strong>at</strong>ion Rev. Econ. Studies<br />
48 541–59<br />
Grossman S J and Shiller R J 1981 The d<strong>et</strong>erminants of the<br />
variability of stock mark<strong>et</strong> prices Am. Econ. Rev. 71 222–7<br />
Grossman S J and Stiglitz J 1980 The impossibility of<br />
inform<strong>at</strong>ionally efficient mark<strong>et</strong>s Am. Econ. Rev. 70 393–408<br />
Hellthaler T 1995 The influence of investor number on a<br />
microscopic mark<strong>et</strong> mo<strong>de</strong>l J. Mod. Phys. C 6 845–52<br />
Hsieh D 1989 Testing for nonlinearity in daily foreign exchange r<strong>at</strong>e<br />
changes J. Business 62 339–68<br />
Hsieh D 1991 Chaos and nonlinear dynamics: applic<strong>at</strong>ion to<br />
financial mark<strong>et</strong>s J. Finance 46 1839–77<br />
Hsieh D 1992 Implic<strong>at</strong>ions of nonlinear dynamics for financial risk<br />
management Workshop on Nonlinear Dynamics in Economics,<br />
July European University Institute<br />
Hsieh D and LeBaron B 1988 Finite sample properties of the BDS<br />
st<strong>at</strong>istics Working Paper University of Chicago and University<br />
of Wisconsin<br />
Huang Z-F and Solomon S 2001 Finite mark<strong>et</strong> size as a source of<br />
extreme wealth inequality and mark<strong>et</strong> instability Physica A 294<br />
503–13<br />
I<strong>de</strong> K and Sorn<strong>et</strong>te D 2001 Oscill<strong>at</strong>ory finite-time singularities in<br />
finance, popul<strong>at</strong>ion and rupture Preprint<br />
http://arXiv.org/abs/cond-m<strong>at</strong>/0106047<br />
Jefferies P, Hart M L, Hui P M and Johnson N F 2001 From mark<strong>et</strong><br />
games to real-world mark<strong>et</strong>s, Eur. Phys. J. B 20 493–501<br />
Jevons S 1871 Theory of Political Economy (Pelican Classics)<br />
Johansen A, Ledoit O and Sorn<strong>et</strong>te D 2000 Crashes as critical points<br />
Int. J. Theor. Appl. Finance 3 219–55<br />
Johansen A and Sorn<strong>et</strong>te D 1999 Critical crashes Risk 12 91–4<br />
Johansen A and Sorn<strong>et</strong>te D 2000 The Nasdaq crash of April 2000:<br />
y<strong>et</strong> another example of log-periodicity in a specul<strong>at</strong>ive bubble<br />
ending in a crash Eur. Phys. J. B 17 319–28<br />
Johansen A and Sorn<strong>et</strong>te D 2001 Large stock mark<strong>et</strong> price<br />
drawdowns are outliers J. Risk (Preprint<br />
http://arXiv.org/abs/cond-m<strong>at</strong>/0010050) <strong>at</strong> press<br />
Johansen A, Sorn<strong>et</strong>te D and Ledoit O 1999 Predicting financial<br />
crashes using discr<strong>et</strong>e scale invariance J. Risk 1 5–32<br />
Keynes J M 1936 The General Theory of Employment, Interest and<br />
Money (London: McMillan)<br />
Kim G W and Markowitz H M 1989 J. Portfolio Management 16 45<br />
Kreps D 1977 Fulfilled expect<strong>at</strong>ions equilibria J. Econ. Theory 14<br />
32–43<br />
Laherrère J and Sorn<strong>et</strong>te D 1998 Str<strong>et</strong>ched exponential distributions<br />
in n<strong>at</strong>ure and economy: ‘f<strong>at</strong> tails’ with characteristic scales<br />
Eur. Phys. J. B 2 525–39<br />
LeBaron B 1988 The changing structure of stock r<strong>et</strong>urns Working<br />
Paper University of Wisconsin<br />
Levy M, Levy H and Solomon S 1995 Microscopic simul<strong>at</strong>ion of<br />
the stock mark<strong>et</strong>—the effect of microscopic diversity J.<br />
280<br />
177<br />
Physique I 5 1087–107<br />
Levy M, Levy H and Solomon S 2000 The Microscopic Simul<strong>at</strong>ion<br />
of Financial Mark<strong>et</strong>s: from Investor Behaviour to Mark<strong>et</strong><br />
Phenomena (San Diego, CA: Aca<strong>de</strong>mic)<br />
Lo A W and MacKinlay A C 1999 A Non-Random Walk down Wall<br />
Stre<strong>et</strong> (Princ<strong>et</strong>on, NJ: Princ<strong>et</strong>on University Press)<br />
Lorenz E 1963 D<strong>et</strong>erministic nonperiodic flow J. Atmos. Sci. 20<br />
130–41<br />
Lux L 1996 The stable Par<strong>et</strong>ian hypothesis and the frequency of<br />
large r<strong>et</strong>urns: an examin<strong>at</strong>ion of major German stocks Appl.<br />
Financial Economics 6 463–75<br />
Lux T and Marchesi M 1999 Scaling and criticality in a stochastic<br />
multi-agent mo<strong>de</strong>l of a financial mark<strong>et</strong> N<strong>at</strong>ure 297 498–500<br />
Manneville P 1991 Structures Dissip<strong>at</strong>ives, Chaos <strong>et</strong> Turbulence<br />
(Saclay: CEA)<br />
Maug E and Naik N 1995 Herding and <strong>de</strong>leg<strong>at</strong>ed portfolio<br />
management: the impact of rel<strong>at</strong>ive performance evalu<strong>at</strong>ion on<br />
ass<strong>et</strong> alloc<strong>at</strong>ion Working Paper Duke University<br />
May R 1976 Simple m<strong>at</strong>hem<strong>at</strong>ical mo<strong>de</strong>ls with very complic<strong>at</strong>ed<br />
dynamics N<strong>at</strong>ure 261 459–67<br />
Müller U A, Dacorogna M M, Davé R, Olsen R B, Pict<strong>et</strong> OVand<br />
von Weizsäcker J E 1997 Vol<strong>at</strong>ilities of different time<br />
resolutions-analyzing the dynamics of mark<strong>et</strong> components J.<br />
Empirical Finance 4 213–40<br />
Muth J 1961 R<strong>at</strong>ional expect<strong>at</strong>ions and the theory of price<br />
movements Econom<strong>et</strong>rica 29 315–25<br />
Muzy J-F, Sorn<strong>et</strong>te D, Delour J and Arneodo A 2001 Multifractal<br />
r<strong>et</strong>urns and hierarchical portfolio theory Quantit<strong>at</strong>ive Finance<br />
1 131–s48<br />
Orléan A 1986 Mimétisme <strong>et</strong> anticip<strong>at</strong>ions r<strong>at</strong>ionnelles: perspectives<br />
keynésiennes Recherches Economiques <strong>de</strong> Louvain 52 45–66<br />
Orléan A 1989 Comportements mimétiques <strong>et</strong> diversité <strong>de</strong>s opinions<br />
sur les marchés financiers Théorie Economique <strong>et</strong> Crise <strong>de</strong>s<br />
Marchés Financiers ed H Bourguin<strong>at</strong> and P Artus (Paris:<br />
Economica) pp 45–65<br />
Orléan A 1990 Le rôle <strong>de</strong>s influences interpersonnelles dans la<br />
détermin<strong>at</strong>ion <strong>de</strong>s cours boursiers Rev. Economique 41 839–68<br />
Orléan A 1992 Contagion <strong>de</strong>s opinions <strong>et</strong> fonctionnement <strong>de</strong>s<br />
marchés financiers Rev. Economique 43 685–97<br />
Pagan A 1996 The econom<strong>et</strong>rics of financial mark<strong>et</strong>s J. Empirical<br />
Finance 3 15–102<br />
Pan<strong>de</strong>y R B and Stauffer D 2000 Search for log-periodicity<br />
oscill<strong>at</strong>ions in stock mark<strong>et</strong> simul<strong>at</strong>ions Int. Theor. Appl.<br />
Finance 3 479–82<br />
Plerou V, Gopikrishnan P, AmaralLAN,MeyerMandStanley H E<br />
1999 Scaling of distribution of price fluctu<strong>at</strong>ions of individual<br />
companies Phys. Rev. E 60 6519–29<br />
Plerou V, Gopikrishnan P, Gabaix X and Stanley H E 2001<br />
Quantifying stock price response to <strong>de</strong>mand fluctu<strong>at</strong>ions<br />
Preprint cond-m<strong>at</strong>/0106657<br />
Radner R 1972 Existence of equilibrium of plans, prices and price<br />
expect<strong>at</strong>ions in a sequence of mark<strong>et</strong>s Econom<strong>et</strong>rica 40<br />
289–304<br />
Radner R 1979 R<strong>at</strong>ional expect<strong>at</strong>ions equilibrium: generic existence<br />
and the inform<strong>at</strong>ion revealed by price Econom<strong>et</strong>rica 47 255–78<br />
Riley J 2001 Silver signals: twenty-five years of screening and<br />
signaling J. Econ. Liter<strong>at</strong>ure 39 432–78<br />
Roehner B M and Sorn<strong>et</strong>te D 1998 The sharp peak-fl<strong>at</strong> trough<br />
p<strong>at</strong>tern and critical specul<strong>at</strong>ion Eur. Phys. J. B 4 387–99<br />
Scharfstein D and Stein J 1990 Herd behaviour and investment Am.<br />
Econ. Rev. 80 465–79<br />
Scheinkman J A and LeBaron B 1989a Nonlinear dynamics and<br />
GNP d<strong>at</strong>a Economic Complexity: Chaos, Sunspots, Bubbles<br />
and Nonlinearity ed W A Barn<strong>et</strong>t, Geweke J and Shell K<br />
(Cambridge: Cambridge University Press)<br />
Scheinkman J A and LeBaron B 1989b Nonlinear dynamics and<br />
stock r<strong>et</strong>urns J. Business 62<br />
Shefrin H 2000 Beyond Greed and Fear: Un<strong>de</strong>rstanding<br />
Behavioural Finance and the Psychology of Investing (Boston,
178 6. Comportements mimétiques <strong>et</strong> antagonistes : bulles hyperboliques, krachs <strong>et</strong> chaos<br />
Q UANTITATIVE F INANCE Imit<strong>at</strong>ion and contrarian behaviour: hyperbolic bubbles, crashes and chaos<br />
MA: Harvard Business School Press)<br />
Shiller R 1981 Do stock prices move too much to be justified by<br />
subsequent changes in divi<strong>de</strong>nds? Am. Econ. Rev. 71 421–36<br />
Shiller R 1984 Stock prices and social dynamics Brookings Papers<br />
on Economic Activity pp 457–98<br />
Shiller R J 2000 Irr<strong>at</strong>ional Exuberance (Princ<strong>et</strong>on, NJ: Princ<strong>et</strong>on<br />
University Press)<br />
Shleifer A 2000 Inefficient Mark<strong>et</strong>s: An Introduction to Behavioural<br />
Finance (New York: Oxford University Press)<br />
Sorn<strong>et</strong>te D 2001 Critical Mark<strong>et</strong> Crashes (Princ<strong>et</strong>on, NJ: Princ<strong>et</strong>on<br />
University Press) <strong>at</strong> press<br />
Sorn<strong>et</strong>te D and An<strong>de</strong>rsen J V 2001 A nonlinear super-exponential<br />
r<strong>at</strong>ional mo<strong>de</strong>l of specul<strong>at</strong>ive financial bubbles Preprint<br />
webpage http://arXiv.org/abs/cond-m<strong>at</strong>/0104341<br />
Sorn<strong>et</strong>te D and I<strong>de</strong> K 2001 Theory of self-similar oscill<strong>at</strong>ory<br />
finite-time singularities in finance, popul<strong>at</strong>ion and rupture<br />
Preprint webpage http://arXiv.org/abs/cond-m<strong>at</strong>/0106054<br />
Sorn<strong>et</strong>te D and Johansen A 2001 Significance of log-periodic<br />
precursors to financial crashes Quantit<strong>at</strong>ive Finance 1 452–71<br />
Sorn<strong>et</strong>te D, Johansen A and Bouchaud J-P 1996 Stock mark<strong>et</strong><br />
crashes precursors and replicas J. Physique I 6 167–75<br />
Stauffer D 1999 Finite-size effects in lux-marchesi and other<br />
microscopic mark<strong>et</strong> mo<strong>de</strong>ls Genoa Economics Me<strong>et</strong>ing, June<br />
(electronic distribution only, ed M Marchesi<br />
http://dibe.unige.it/wehia)<br />
Stauffer D and Sorn<strong>et</strong>te D 1999 Self-organized percol<strong>at</strong>ion mo<strong>de</strong>l<br />
for stock mark<strong>et</strong> fluctu<strong>at</strong>ions Physica A 271 N3-4 496–506<br />
Stutzer M 1980 Chaotic dynamics and bifurc<strong>at</strong>ion in a macro-mo<strong>de</strong>l<br />
J. Econ. Dyn. Control<br />
Thaler R H (ed) 1993 Advances in Behavioural Finance (New York:<br />
Russell Sage Found<strong>at</strong>ion)<br />
Topol R 1991 Bubbles and vol<strong>at</strong>ility of stock prices: effect of<br />
mim<strong>et</strong>ic contagion Economic J. 101 786–800<br />
Trueman B 1994 Analyst forecasts and herding behaviour Rev.<br />
Financial Studies 7 97–124<br />
Van <strong>de</strong>r Ploeg F 1986 R<strong>at</strong>ional expect<strong>at</strong>ions, risk and chaos in<br />
financial mark<strong>et</strong>s Econ. J. 96 (suppl.)<br />
Welch I 1992 Sequential sales, learning, and casca<strong>de</strong>s J. Finance 47<br />
695–732<br />
see also webpage http://welch.som.yale.edu/casca<strong>de</strong>s for an<br />
annot<strong>at</strong>ed bibliography and resource reference on inform<strong>at</strong>ion<br />
casca<strong>de</strong>s<br />
West K D 1988 Bubbles, fads and stock price vol<strong>at</strong>ility tests: a<br />
partial evalu<strong>at</strong>ion J. Finance 43 639–55<br />
Zwiebel J 1995 Corpor<strong>at</strong>e conserv<strong>at</strong>ism and rel<strong>at</strong>ive compens<strong>at</strong>ion<br />
J. Political Economy 103 1–25<br />
281
Deuxième partie<br />
Etu<strong>de</strong> <strong>de</strong>s propriétés <strong>de</strong> dépendances entre<br />
actifs financiers<br />
179
Chapitre 7<br />
Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s<br />
copules<br />
Nous abordons maintenant l’étu<strong>de</strong> <strong>de</strong> la dépendance entres actifs financiers. Nous avons déjà évoqué aux<br />
chapitres 1 <strong>et</strong> 4 le problème <strong>de</strong> la dépendance temporelle pour un actif donné, <strong>et</strong> c’est pourquoi nous<br />
nous bornerons ici à présenter l’aspect que l’on peut qualifier - en référence au langage <strong>de</strong> la physique -<br />
<strong>de</strong> “sp<strong>at</strong>ial”, <strong>et</strong> qui consiste en l’étu<strong>de</strong> <strong>de</strong>s interactions entre différents actifs à un instant donné.<br />
Jusqu’à une pério<strong>de</strong> récente, le coefficient <strong>de</strong> corrél<strong>at</strong>ion était l’outil <strong>de</strong> prédilection utilisé pour quantifier<br />
l’intensité <strong>de</strong> la dépendance entre <strong>de</strong>ux actifs. Ceci est bien sûr à r<strong>at</strong>tacher au fait que l’on pensait<br />
que non seulement la distribution marginale <strong>de</strong>s ren<strong>de</strong>ments était gaussienne, mais aussi, par une extension<br />
somme toute très n<strong>at</strong>urelle, leur distribution multivariée. L’hypothèse gaussienne s’effondrant<br />
peu à peu (voir Richardson <strong>et</strong> Smith (1993) pour un test spécifique du caractère gaussien multivarié<br />
<strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments), la pertinence du coefficient corrél<strong>at</strong>ion comme mesure <strong>de</strong> dépendance<br />
était elle-même remise en cause, si bien qu’il semble désormais admis qu’une telle mesure est totalement<br />
insuffisante pour donner une idée juste <strong>de</strong>s propriétés <strong>de</strong> dépendance entre <strong>de</strong>ux actifs (Blum, Dias<br />
<strong>et</strong> Embrechts 2002, Embrechts, McNeil <strong>et</strong> Straumann 2002) <strong>et</strong> qu’il est nécessaire d’avoir recours à la<br />
distribution jointe <strong>de</strong>s probabilités <strong>de</strong> leurs ren<strong>de</strong>ments 1 pour obtenir une <strong>de</strong>scription aussi complète<br />
que possible <strong>de</strong> leur structure <strong>de</strong> dépendance 2 . En fait, cela est d’autant plus crucial que l’on souhaite<br />
s’intéresser aux événements rares (<strong>et</strong> donc extrêmes).<br />
Il convient toutefois <strong>de</strong> prendre conscience que la distribution bivariée <strong>de</strong>s ren<strong>de</strong>ments contient <strong>de</strong>ux<br />
types d’inform<strong>at</strong>ion : d’une part <strong>de</strong> l’inform<strong>at</strong>ion sur la distribution individuelle (monovariée) <strong>de</strong>s ren<strong>de</strong>ments<br />
<strong>de</strong> chacun <strong>de</strong>s actifs <strong>et</strong> d’autre part <strong>de</strong> l’inform<strong>at</strong>ion concernant la seule dépendance entre ces <strong>de</strong>ux<br />
actifs, indépendamment <strong>de</strong> leur comportement individuel. Il va donc être souhaitable <strong>de</strong> pouvoir séparer<br />
ces <strong>de</strong>ux sources d’inform<strong>at</strong>ion, afin <strong>de</strong> ne conserver que celle qui nous intéresse ici, à savoir l’inform<strong>at</strong>ion<br />
sur le comportement collectif (ou comportement joint) <strong>de</strong>s <strong>de</strong>ux actifs. Ceci est rendu possible grâce<br />
à l’emploi d’obj<strong>et</strong>s m<strong>at</strong>hém<strong>at</strong>iques dénommés copules, dont le rôle est précisément <strong>de</strong> décrire <strong>de</strong> manière<br />
complète <strong>et</strong> généralement unique la structure <strong>de</strong> dépendance présentée par les <strong>de</strong>ux actifs financiers, à<br />
l’exclusion <strong>de</strong> tout autre inform<strong>at</strong>ion sur leur comportement marginal.<br />
C<strong>et</strong>te approche <strong>de</strong> la dépendance sp<strong>at</strong>iale en terme <strong>de</strong> copule s’est abondamment développée ces <strong>de</strong>rnières<br />
1 Nous l’avons déjà fait remarquer dans la partie précé<strong>de</strong>nte, les ren<strong>de</strong>ments <strong>de</strong>s actifs financiers semblent être mieux adaptés<br />
que les prix eux-mêmes du fait du manque <strong>de</strong> st<strong>at</strong>ionnarité flagrant dont font preuve ces <strong>de</strong>rniers. Ainsi, dans toute la suite <strong>de</strong><br />
c<strong>et</strong>te partie la distribution bivariée <strong>de</strong>s ren<strong>de</strong>ments sera un <strong>de</strong>s obj<strong>et</strong>s fondamentaux <strong>de</strong> notre étu<strong>de</strong>.<br />
2 Pour <strong>de</strong>s raisons <strong>de</strong> simplicité <strong>et</strong> <strong>de</strong> clarté <strong>de</strong> l’exposé, nous limitons notre présent<strong>at</strong>ion au cas où seuls <strong>de</strong>ux actifs sont<br />
présents. Il va <strong>de</strong> soit que cela n’est en rien restrictif <strong>et</strong> que le cas général pourrait être traité <strong>de</strong> la même manière.<br />
181
182 7. Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s copules<br />
années <strong>et</strong> semble prom<strong>et</strong>teuse aussi bien d’un point <strong>de</strong> vue pr<strong>at</strong>ique que théorique. En eff<strong>et</strong>, sur le plan<br />
pr<strong>at</strong>ique, la dépendance entre les actifs est l’un <strong>de</strong>s piliers <strong>de</strong> la <strong>gestion</strong> <strong>de</strong>s risques, puisque c’est <strong>de</strong> c<strong>et</strong>te<br />
étu<strong>de</strong> <strong>de</strong> la dépendance que va découler la str<strong>at</strong>égie <strong>de</strong> <strong>gestion</strong> à m<strong>et</strong>tre en œuvre selon que l’on souhaite<br />
diversifier les risques en les agrégeant sous forme <strong>de</strong> <strong>portefeuille</strong>s, ou bien que l’on déci<strong>de</strong> <strong>de</strong> couvrir<br />
ces risques à l’ai<strong>de</strong> <strong>de</strong> produits dérivés, ou enfin que l’on choisisse <strong>de</strong> les titriser, c’est-à-dire les m<strong>et</strong>tre<br />
sur le marché, ce qui revient à les cé<strong>de</strong>r à une tierce partie. Or, l’apport <strong>de</strong>s copules est <strong>de</strong> perm<strong>et</strong>tre une<br />
meilleure compréhension <strong>et</strong> quantific<strong>at</strong>ion <strong>de</strong> l’eff<strong>et</strong> <strong>de</strong> l’interaction entre actifs par l’étu<strong>de</strong> <strong>de</strong> l’influence<br />
<strong>de</strong> diverses structures <strong>de</strong> dépendance entre les multiples sources <strong>de</strong> risques. Les exemples d’applic<strong>at</strong>ions<br />
sont nombreux, aussi bien en finance 3 que dans d’autres domaines tels que l’assurance 4 .<br />
Sur le plan théorique, l’étu<strong>de</strong> <strong>de</strong> la dépendance entre actifs est fondamentale car elle perm<strong>et</strong> <strong>de</strong> son<strong>de</strong>r<br />
la structure sous-jacente <strong>de</strong>s mécanismes <strong>de</strong> marché. En eff<strong>et</strong>, il est très raisonnable <strong>de</strong> penser que la<br />
dépendance entre actifs résulte, du moins en partie 5 , <strong>de</strong> l’interaction entre les agents agissant sur ces<br />
marchés. De part les positions qu’ils prennent, ceux-ci sont responsables <strong>de</strong> l’évolution individuelle<br />
(en fait, <strong>de</strong>s fluctu<strong>at</strong>ions) <strong>de</strong>s prix <strong>de</strong>s actifs, mais aussi <strong>de</strong> part les choix qu’ils opèrent - vendant ou<br />
ach<strong>et</strong>ant tel titre plutôt que tel autre - ils créent <strong>de</strong> la dépendance entre ces actifs. Donc, l’étu<strong>de</strong> <strong>de</strong> la<br />
dépendance entre actifs doit perm<strong>et</strong>tre <strong>de</strong> compléter la compréhension <strong>de</strong>s mécanismes en jeu sur les<br />
marchés financiers <strong>et</strong> ainsi <strong>de</strong> mieux saisir les interactions entre les agents. Elle doit aussi amener à<br />
cerner les paramètres macroscopiques auxquels les agents sont sensibles.<br />
Avant d’entrer dans le vif du suj<strong>et</strong> <strong>et</strong> <strong>de</strong> faire tout d’abord quelques rappels sur la notion <strong>de</strong> copule <strong>et</strong><br />
quelques unes <strong>de</strong> leurs propriétés fondamentales, puis <strong>de</strong> présenter certaines familles <strong>de</strong> copules dont<br />
nous aurons à nous servir par la suite, <strong>et</strong> enfin d’en venir aux résult<strong>at</strong>s empiriques que nous avons pu<br />
obtenir sur la dépendance entre actifs financiers, nous souhaitons préciser une hypothèse importante sur<br />
laquelle nous nous appuierons dans la suite. En eff<strong>et</strong>, nous avons affirmé qu’il était n<strong>at</strong>urel <strong>de</strong> découpler<br />
l’étu<strong>de</strong> <strong>de</strong> la dépendance sp<strong>at</strong>iale <strong>de</strong> celle <strong>de</strong> la dépendance temporelle. Or, pour <strong>de</strong>meurer dans un cadre<br />
complètement général, il convient <strong>de</strong> gar<strong>de</strong>r à l’esprit que la dépendance entre <strong>de</strong>ux actifs peut tout<br />
à fait évoluer au cours du temps, elle n’a aucune raison <strong>de</strong> rester constante (P<strong>at</strong>ton 2001, Rockinger <strong>et</strong><br />
Jon<strong>de</strong>au 2001), donc il conviendrait aussi d’étudier sa dynamique. Cela dit, dans un cadre aussi vaste, une<br />
étu<strong>de</strong> aussi bien empirique que théorique <strong>de</strong>vient extrêmement délic<strong>at</strong>e. C’est pourquoi nous avons choisi<br />
<strong>de</strong> faire une hypothèse simplific<strong>at</strong>rice en considérerant que la dépendance temporelle est entièrement<br />
capturée par l’évolution marginale <strong>de</strong> chaque actif <strong>et</strong> donc que leurs propriétés <strong>de</strong> dépendance sp<strong>at</strong>iale<br />
<strong>de</strong>meurent les mêmes à tout instant.<br />
7.1 Les copules<br />
Nous allons exposer brièvement dans ce paragraphe les propriétés principales <strong>de</strong>s copules. Pour une<br />
présent<strong>at</strong>ion complète du suj<strong>et</strong> le lecteur est invité à se référer aux ouvrages <strong>de</strong> Joe (1997) <strong>et</strong> Nelsen<br />
(1998) notamment <strong>et</strong> aux articles <strong>de</strong> revues <strong>de</strong> Frees <strong>et</strong> Val<strong>de</strong>z (1998) ou Bouyé, Durrleman, Nikeghbali,<br />
Riboul<strong>et</strong> <strong>et</strong> Roncalli (2000) pour une présent<strong>at</strong>ion plus orientée vers les applic<strong>at</strong>ions financières <strong>et</strong><br />
3<br />
Voir notamment Embrechts, Hoeing <strong>et</strong> Juri (2001) pour les conséquences sur le calcul <strong>de</strong> VaR <strong>et</strong> la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> ou<br />
encore Cherubini <strong>et</strong> Luciano (2000) <strong>et</strong> Coutant, Durrleman, Rapuch <strong>et</strong> Roncalli (2001) pour une applic<strong>at</strong>ion au calcul d’options<br />
<strong>et</strong> Frey <strong>et</strong> McNeil (2001), Frey, McNeil <strong>et</strong> Nyfeler (2001) pour ce qui est du risque <strong>de</strong> crédit.<br />
4<br />
Voir notamment Frees <strong>et</strong> Val<strong>de</strong>z (1998), Val<strong>de</strong>z (2001), ou Coss<strong>et</strong>te, Gaillard<strong>et</strong>z, Marceau <strong>et</strong> Rioux (2002) pour certaines<br />
applic<strong>at</strong>ions<br />
5<br />
Bien évi<strong>de</strong>mment, la dépendance observée entre différents actifs n’a pas pour seule <strong>et</strong> unique source l’action <strong>de</strong>s agents<br />
économiques sur les marchés financiers. Les facteurs macroéconomiques jouent aussi un rôle important en ce qui concerne<br />
c<strong>et</strong>te dépendance, qui est par exemple d’autant plus forte que les actifs considérés appartiennent au même secteur économique<br />
<strong>et</strong> sont donc <strong>de</strong> ce fait sensibles aux mêmes variables, aux mêmes évolutions <strong>de</strong> l’environnement économique.
7.2. Quelques familles <strong>de</strong> copules 183<br />
actuarielles.<br />
Commençons par rappeler les propriétés m<strong>at</strong>hém<strong>at</strong>iques que vérifient les copules :<br />
DÉFINITION 1 (COPULE)<br />
Une fonction C : [0, 1] × [0, 1] −→ [0, 1] est une copule si elle s<strong>at</strong>isfait les propriétés suivantes :<br />
– ∀u ∈ [0, 1], C(1, u) = C(u, 1) = u ,<br />
– ∀ui ∈ [0, 1], C(u1, u2) = 0 si au moins un <strong>de</strong>s ui est nul,<br />
– C est croissante dans le sens où le C-volume <strong>de</strong> chaque rectangle dont les somm<strong>et</strong>s se situent dans<br />
[0, 1] 2 est positif.<br />
Cela signifie simplement qu’une copule n’est rien d’autre qu’une distribution multivariée dont les distributions<br />
marginales sont uniformes. Leur principal intérêt provient en fait du théorème <strong>de</strong> représent<strong>at</strong>ion<br />
<strong>de</strong> Sklar (1959), qui stipule que toute distribution multivariée peut être exprimée comme une fonction <strong>de</strong><br />
ses marginales :<br />
THÉORÈME 1 (SKLAR (1959))<br />
Etant donné <strong>de</strong>ux variables alé<strong>at</strong>oires X <strong>et</strong> Y dont la fonction <strong>de</strong> distribution est notée F <strong>et</strong> dont les<br />
distributions marginales sont FX <strong>et</strong> FY , il existe une copule C : [0, 1] 2 −→ [0, 1] telle que :<br />
F (x, y) = C(FX(x), FY (y)) . (7.1)<br />
Dans le cas où les distributions marginales <strong>de</strong>s variables X <strong>et</strong> Y sont continues, c<strong>et</strong>te copule est unique<br />
<strong>et</strong> C(·, ·) est alors la copule du couple <strong>de</strong> variables alé<strong>at</strong>oires (X, Y ). Bien évi<strong>de</strong>mment, ce théorème<br />
s’étend au cas d’un nombre quelconque <strong>de</strong> variables alé<strong>at</strong>oires. Une <strong>de</strong> ces conséquences immédi<strong>at</strong>es est<br />
que la fonction F (FX −1 (x), FY −1 (y)) est une copule, <strong>et</strong> plus précisément la copule <strong>de</strong> (X, Y ). Donc,<br />
partant <strong>de</strong> n’importe quelle distribution jointe, il est aisé d’en dériver une copule. C’est une <strong>de</strong>s métho<strong>de</strong>s<br />
les plus employées pour construire <strong>de</strong>s copules.<br />
Enfin, cela perm<strong>et</strong> aisément <strong>de</strong> comprendre que la copule <strong>de</strong> <strong>de</strong>ux variables alé<strong>at</strong>oires est invariante par<br />
changement <strong>de</strong> variable strictement croissant, ce qui est démontré dans Lindskog (2000) notamment.<br />
Ce résult<strong>at</strong> est particulièrement important car il justifie que la copule est une mesure intrinsèque <strong>de</strong><br />
la dépendance entre variables alé<strong>at</strong>oires. En eff<strong>et</strong>, par changement <strong>de</strong> variable strictement croissant,<br />
l’ancienne <strong>et</strong> la nouvelle variable sont comonotones. Or, il est n<strong>at</strong>urel <strong>de</strong> requérir d’une mesure <strong>de</strong> la<br />
dépendance entre <strong>de</strong>ux variables alé<strong>at</strong>oires qu’elle soit indifférente à la substitution par une variable<br />
comonotone : si X <strong>et</strong> X ′ sont <strong>de</strong>ux variables comonotones, il est normal que la structure <strong>de</strong> dépendance<br />
entre (X, Y ) d’une part <strong>et</strong> (X ′ , Y ) d’autre part soit la même. C’est exactement ce que traduit ce théorème<br />
<strong>et</strong> dont rend compte la copule. Bien entendu, <strong>de</strong>s quantités telles que le coefficient <strong>de</strong> corrél<strong>at</strong>ion, qui<br />
sont fonction à la fois <strong>de</strong> la copule <strong>et</strong> <strong>de</strong>s marginales, ne sont pas invariantes par une telle transform<strong>at</strong>ion<br />
<strong>et</strong> ne constituent donc pas <strong>de</strong>s mesures <strong>de</strong> la seule dépendance.<br />
En résumé, les copules sont <strong>de</strong>s obj<strong>et</strong>s perm<strong>et</strong>tant <strong>de</strong> décrire <strong>de</strong> manière complète <strong>et</strong> généralement unique<br />
les propriétés <strong>de</strong> dépendance entre <strong>de</strong>ux variables alé<strong>at</strong>oires. En cela, elles autorisent l’étu<strong>de</strong> séparée <strong>de</strong>s<br />
distributions marginales <strong>de</strong> chaque actif financier <strong>et</strong> <strong>de</strong> leur dépendance, ce qui va être le point auquel<br />
nous allons nous intéresser maintenant. Mais avant cela, nous allons présenter quelques copules d’usage<br />
courant dont nous aurons à traiter dans la suite.<br />
7.2 Quelques familles <strong>de</strong> copules<br />
Comme nous l’avons indiqué après avoir rappelé le théorème <strong>de</strong> Sklar (1959), il est possible <strong>de</strong> dériver<br />
une copule <strong>de</strong> toute distribution multivariée. Leur nombre est donc considérable. Cependant quelques co-
184 7. Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s copules<br />
pules ou familles <strong>de</strong> copules occupent une place plus importante que d’autres <strong>et</strong> nous allons en présenter<br />
quelques unes.<br />
7.2.1 Les copules gaussiennes <strong>et</strong> copules <strong>de</strong> Stu<strong>de</strong>nt<br />
Comme leurs noms l’indiquent ces <strong>de</strong>ux classes <strong>de</strong> copules sont respectivement issues <strong>de</strong> la distribution<br />
gaussienne <strong>et</strong> <strong>de</strong>s distributions <strong>de</strong> Stu<strong>de</strong>nt. Et, <strong>de</strong> même que la distribution gaussienne est un cas limite <strong>de</strong><br />
distribution <strong>de</strong> Stu<strong>de</strong>nt, la copule gaussienne est aussi un cas limite <strong>de</strong> copule <strong>de</strong> Stu<strong>de</strong>nt. C’est pourquoi<br />
nous traitons ces copules, a priori différentes, dans le même paragraphe. Une autre raison vient <strong>de</strong> ce<br />
que l’une comme l’autre sont <strong>de</strong>s copules elliptiques, c’est-à-dire <strong>de</strong>s copules dérivées <strong>de</strong>s distributions<br />
elliptiques.<br />
L’intérêt <strong>de</strong> c<strong>et</strong>te classe <strong>de</strong> copules vient simplement du fait qu’elle ouvre la voie à la généralis<strong>at</strong>ion <strong>de</strong>s<br />
distributions gaussiennes <strong>et</strong> elliptiques à <strong>de</strong>s distributions dites “méta-gaussiennes” - dont l’introduction<br />
est originellement due à Krzysztofowicz <strong>et</strong> Kelly (1996) <strong>et</strong> dont l’intérêt pr<strong>at</strong>ique a été démontré par Karlen<br />
(1998) dans le domaine du traitement d’expériences <strong>de</strong> physique <strong>de</strong>s particules <strong>et</strong> par Sorn<strong>et</strong>te, Simon<strong>et</strong>ti<br />
<strong>et</strong> An<strong>de</strong>rsen (2000) pour ce qui est <strong>de</strong> la finance - <strong>et</strong> “méta-elliptiques” (Fang, Fang <strong>et</strong> Kotz 2002).<br />
Ces “méta-distributions” possè<strong>de</strong>nt la même structure <strong>de</strong> dépendance que les distributions gaussiennes<br />
ou elliptiques mais en différent par leurs marginales qui peuvent être quelconques. Or, on sait que les<br />
distributions elliptiques peuvent aisément être obtenues par <strong>de</strong>s modèles à vol<strong>at</strong>ilité stochastiques, ce qui<br />
fait <strong>de</strong>s copules elliptiques en général <strong>et</strong> <strong>de</strong>s copules gaussiennes <strong>et</strong> <strong>de</strong> Stu<strong>de</strong>nt en particulier, <strong>de</strong>s copules<br />
tout à fait pertinentes pour la modélis<strong>at</strong>ion <strong>de</strong> la dépendance entre actifs financiers.<br />
Utilisant le théorème <strong>de</strong> Sklar (1959), on obtient simplement leurs expressions, même s’il n’est pas<br />
possible d’en donner une forme fermée :<br />
DÉFINITION 2 (COPULE GAUSSIENNE)<br />
Soit Φ la distribution gaussienne standard <strong>et</strong> Φρ,N la distribution Gaussienne en dimension N, dont la<br />
m<strong>at</strong>rice <strong>de</strong> corrél<strong>at</strong>ion est ρ. La copule gaussienne <strong>de</strong> m<strong>at</strong>rice <strong>de</strong> corrél<strong>at</strong>ion ρ est alors :<br />
<strong>et</strong> sa <strong>de</strong>nsité est donnée par<br />
avec yk(u) = Φ −1 (uk).<br />
Cρ,N(u1, · · · , uN) = Φρ,N(Φ −1 (u1), · · · , Φ −1 (uN)) (7.2)<br />
cρ,N(u1, · · · , uN) =<br />
<br />
1<br />
√ exp −<br />
d<strong>et</strong> ρ 1<br />
2 yt (u) (ρ−1 <br />
− Id)y (u)<br />
DÉFINITION 3 (COPULE DE STUDENT)<br />
Soit Tν la distribution <strong>de</strong> Stu<strong>de</strong>nt avec ν <strong>de</strong>grés <strong>de</strong> liberté <strong>et</strong> Tν,ρ,N la distribution <strong>de</strong> Stu<strong>de</strong>nt avec ν<br />
<strong>de</strong>grés <strong>de</strong> liberté en dimension N, dont la m<strong>at</strong>rice <strong>de</strong> corrél<strong>at</strong>ion est ρ. La copule <strong>de</strong> Stu<strong>de</strong>nt avec ν<br />
<strong>de</strong>grés <strong>de</strong> liberté <strong>et</strong> <strong>de</strong> m<strong>at</strong>rice <strong>de</strong> corrél<strong>at</strong>ion ρ est alors :<br />
<strong>et</strong> sa <strong>de</strong>nsité est donnée par<br />
avec yk(u) = T −1<br />
ν (uk).<br />
(7.3)<br />
Cρ,ν,N (u1, · · · , uN) = Tν,ρ,N (T −1<br />
ν (u1), · · · , Tnu −1 (uN)) (7.4)<br />
cρ,ν,N (u1, · · · , uN) =<br />
1 Γ<br />
√<br />
d<strong>et</strong> ρ<br />
<br />
ν+N ν<br />
2 Γ<br />
<br />
ν+1 Γ 2<br />
N−1 2<br />
N <br />
N<br />
k=1<br />
<br />
1 + y2 k<br />
ν<br />
1 + yt ρy<br />
ν<br />
ν+1<br />
2<br />
ν+N<br />
2<br />
, (7.5)
7.2. Quelques familles <strong>de</strong> copules 185<br />
Par construction même, les copules <strong>de</strong> Stu<strong>de</strong>nt <strong>et</strong> la copule gaussienne sont très proches dans leur région<br />
centrale, <strong>et</strong> l’on observe que le domaine où ces <strong>de</strong>ux copules sont quasiment i<strong>de</strong>ntiques s’étend <strong>de</strong> plus<br />
en plus au fur <strong>et</strong> à mesure que le nombre <strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté <strong>de</strong> la copule <strong>de</strong> Stu<strong>de</strong>nt augmente. En<br />
conséquence, il est parfois délic<strong>at</strong> <strong>de</strong> les distinguer, même avec <strong>de</strong>s tests assez fins. Or, nous verrons plus<br />
loin en détail que la confusion entre ces <strong>de</strong>ux copules à <strong>de</strong>s conséquences importantes pour l’étu<strong>de</strong> <strong>de</strong>s<br />
risques extrêmes.<br />
Enfin, signalons que ces copules présentent en outre l’intérêt d’être aisément synthétisables, ce qui est<br />
très utile en pr<strong>at</strong>ique pour la simul<strong>at</strong>ion numérique ou l’étu<strong>de</strong> <strong>de</strong> scénarii. En eff<strong>et</strong>, il est facile <strong>de</strong> générer<br />
<strong>de</strong>s variables alé<strong>at</strong>oires gaussiennes ou <strong>de</strong> Stu<strong>de</strong>nt, ce qui, après un changement <strong>de</strong> variable (croissant),<br />
perm<strong>et</strong> d’obtenir les distributions marginales recherchées tout en conservant la copule inchangée.<br />
7.2.2 Les copules archimédiennes<br />
Les copules d’Archimè<strong>de</strong> revêtent un intérêt tout particulier dans la mesure où un très grand nombre <strong>de</strong><br />
copules appartiennent à c<strong>et</strong>te classe, qui <strong>de</strong> plus jouit d’un certain nombre <strong>de</strong> propriétés intéressantes. En<br />
outre, comme souligné par Frees <strong>et</strong> Val<strong>de</strong>z (1998), nombre <strong>de</strong> modèles - essentiellement issus du domaine<br />
<strong>de</strong> l’assurance - visant à rendre compte <strong>de</strong> la dépendance entre diverses sources <strong>de</strong> risques conduisent<br />
à <strong>de</strong>s copules archimédiennes. Une exception notable cependant est celle <strong>de</strong>s modèles à facteurs, qui<br />
jouent un rôle fondamental dans la <strong>de</strong>scription phénoménologique <strong>de</strong> l’interaction entre actifs financiers,<br />
<strong>et</strong> dont les copules archimédiennes ne suffisent à rendre compte.<br />
Avant d’aller plus loin, commençons par définir ce qu’est une copule archimédienne :<br />
DÉFINITION 4 (COPULE ARCHIMÉDIENNE)<br />
Soit ϕ une fonction continue, strictement décroissante <strong>de</strong> [0, 1] dans [0, ∞] <strong>et</strong> telle que ϕ(1) = 0. Soit<br />
ϕ [−1] le pseudo-inverse <strong>de</strong> ϕ :<br />
ϕ [−1] <br />
ϕ−1 (t), si 0 ≤ t ≤ ϕ(0) ,<br />
(t) =<br />
(7.6)<br />
0, si t ≥ ϕ(0) ,<br />
alors la fonction<br />
est une copule dite archimédienne <strong>de</strong> génér<strong>at</strong>eur ϕ.<br />
C(u, v) = ϕ [−1] (ϕ(u) + ϕ(v)) (7.7)<br />
On remarque donc que c<strong>et</strong>te classe <strong>de</strong> copules est particulièrement simple dans la mesure où toute la<br />
complexité <strong>de</strong> la structure <strong>de</strong> dépendance, décrite <strong>de</strong> manière générale par une fonction <strong>de</strong> <strong>de</strong>ux (ou N)<br />
variables, est totalement contenue dans l’expression du génér<strong>at</strong>eur ϕ, qui n’est qu’une fonction d’une<br />
seule variable. On passe ainsi d’un problème multidimensionnel - <strong>et</strong> donc généralement délic<strong>at</strong> - à un<br />
problème unidimensionnel, <strong>et</strong> ainsi beaucoup plus simple.<br />
Parmi les nombreux membres <strong>de</strong> c<strong>et</strong>te famille, citons par exemple la copule <strong>de</strong> Clayton :<br />
Cθ(u, v) = max u θ + v θ <br />
−1/θ<br />
− 1 , 0 , (7.8)<br />
mais aussi la copule <strong>de</strong> Gumbel, qui joue un rôle particulier pour la <strong>de</strong>scription <strong>de</strong> la dépendance dans la<br />
<strong>théorie</strong> <strong>de</strong>s valeurs extrêmes :<br />
<br />
Cθ(u, v) = exp − (− ln u) θ + (− ln v) θ 1/θ <br />
, (7.9)
186 7. Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s copules<br />
ou encore la copule <strong>de</strong> Frank :<br />
Cθ(u, v) = − 1<br />
θ ln<br />
<br />
1 + (e−θu − 1)(e−θv − 1)<br />
e−θ <br />
. (7.10)<br />
− 1<br />
Notons que les copules gaussiennes ou <strong>de</strong> Stu<strong>de</strong>nt ne font pas partie <strong>de</strong> c<strong>et</strong>te famille. En fait, on peut<br />
montrer que toutes les copules archimédiennes sont d’une part associ<strong>at</strong>ives 6 , ce que clairement les copules<br />
gaussiennes <strong>et</strong> <strong>de</strong> Stu<strong>de</strong>nt ne s<strong>at</strong>isfont pas, <strong>et</strong> d’autre part leurs valeurs sur la première bissectrice<br />
vérifient C(u, u) < u pour tout u ∈]0, 1[. Réciproquement, on peut démontrer que (Nelsen 1998) toute<br />
copule possédant ces <strong>de</strong>ux propriétés est archimédienne, ce qui donne une idée <strong>de</strong> la généralité <strong>de</strong> c<strong>et</strong>te<br />
famille.<br />
Enfin, précisons que Juri <strong>et</strong> Wüthrich (2002) ont établi, pour c<strong>et</strong>te classe <strong>de</strong> copules, un théorème limite<br />
équivalent au théorème <strong>de</strong> Gne<strong>de</strong>nko - Pikand - Balkema - <strong>de</strong> Haan (Embrechts <strong>et</strong> al. 1997, par exemple),<br />
montrant la convergence vers la copule <strong>de</strong> Clayton (7.8) <strong>de</strong> la copule <strong>de</strong> la distribution <strong>de</strong>s excé<strong>de</strong>nts <strong>de</strong><br />
plusieurs variables alé<strong>at</strong>oires au-<strong>de</strong>ssus d’un seuil, quand celui-ci tend vers l’infini. Ainsi, la copule <strong>de</strong><br />
Clayton joue un rôle similaire (en dimension N) à celui <strong>de</strong>s distributions <strong>de</strong> Par<strong>et</strong>o généralisées (un<br />
dimension un). Ceci est particulièrement intéressant lorsque l’on s’intéresse à la <strong>st<strong>at</strong>istique</strong> multivariée<br />
<strong>de</strong>s extrêmes.<br />
7.3 Tests empiriques<br />
Les tests empiriques visent à déterminer la n<strong>at</strong>ure <strong>de</strong> la copule perm<strong>et</strong>tant <strong>de</strong> décrire la structure <strong>de</strong><br />
dépendance entre <strong>de</strong>ux actifs <strong>et</strong> peuvent être soit paramétriques soit non paramétriques. C<strong>et</strong>te <strong>de</strong>rnière<br />
métho<strong>de</strong> est bien évi<strong>de</strong>mment beaucoup plus générale puisqu’elle ne requière pas le choix d’un modèle<br />
a priori, <strong>et</strong> n’est donc pas suj<strong>et</strong>te à l’erreur <strong>de</strong> spécific<strong>at</strong>ion. Cependant, si le modèle est correctement<br />
spécifié, l’estim<strong>at</strong>ion paramétrique est beaucoup plus précise. De plus, le p<strong>et</strong>it nombre <strong>de</strong> paramètres intervenant<br />
dans la <strong>de</strong>scription <strong>de</strong> la copule sélectionnée apparaît alors comme l’ensemble <strong>de</strong>s paramètres<br />
pertinents pour résumer les propriétés <strong>de</strong> dépendance entre actifs. Que l’on pense, par exemple, à la<br />
représent<strong>at</strong>ion gaussienne (ou plus généralement à toute représent<strong>at</strong>ion en terme <strong>de</strong> distribution elliptique)<br />
dont la dépendance est complètement capturée par le coefficient <strong>de</strong> corrél<strong>at</strong>ion linéaire. Ainsi, ces<br />
paramètres vont pouvoir jouer un rôle particulier <strong>et</strong> il est tentant <strong>de</strong> les interpréter comme <strong>de</strong>s variables<br />
macroscopiques (ou phénoménologiques) synthétisant l’ensemble <strong>de</strong>s interactions microscopiques entre<br />
les agents économiques dont résulte la dépendance observée. Or, l’i<strong>de</strong>ntific<strong>at</strong>ion <strong>de</strong>s “ bonnes variables”<br />
est la première étape <strong>de</strong> la construction <strong>de</strong> modèles dont on peut espérer qu’ils seront mieux à même <strong>de</strong><br />
rendre compte <strong>de</strong>s phénomènes observés. L’intérêt <strong>de</strong> l’estim<strong>at</strong>ion paramétrique est donc grand <strong>et</strong> c’est<br />
pour cela que nous nous sommes tout particulièrement focalisés sur c<strong>et</strong>te approche.<br />
7.3.1 Tests paramétriques<br />
Le paradigme gaussien a longtemps été en vigueur en finance. Certes, l’on a vu que les distributions<br />
marginales ne sauraient être modélisées par <strong>de</strong>s distributions gaussiennes dont les queues sont trop fines<br />
pour rendre compte <strong>de</strong>s gran<strong>de</strong>s dévi<strong>at</strong>ions observées sur les distributions <strong>de</strong>s ren<strong>de</strong>ments, mais a priori,<br />
la structure <strong>de</strong> dépendance entre <strong>de</strong>ux actifs étant fort peu connue, rien ne perm<strong>et</strong> <strong>de</strong> rej<strong>et</strong>er le fait que la<br />
copule gaussienne soit à même <strong>de</strong> décrire correctement c<strong>et</strong>te structure <strong>de</strong> dépendance. De plus, suivant<br />
une idée initialement développée par Karlen (1998) <strong>et</strong> Sorn<strong>et</strong>te, Simon<strong>et</strong>ti <strong>et</strong> An<strong>de</strong>rsen (2000), il apparaît<br />
6 Une copule est associ<strong>at</strong>ive si C(u, C(v, w)) = C(C(u, v), w).
7.3. Tests empiriques 187<br />
Y<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
Copule Gaussienne<br />
−5<br />
−5 −4 −3 −2 −1 0<br />
X<br />
1 2 3 4 5<br />
Y<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
Copule <strong>de</strong> Stu<strong>de</strong>nt<br />
−5<br />
−5 −4 −3 −2 −1 0<br />
X<br />
1 2 3 4 5<br />
FIG. 7.1 – Représent<strong>at</strong>ion <strong>de</strong>s réalis<strong>at</strong>ions <strong>de</strong> <strong>de</strong>ux variables alé<strong>at</strong>oires dont les marginales sont gaussiennes<br />
<strong>et</strong> la copule est soit gaussienne (figure <strong>de</strong> gauche) soit <strong>de</strong> Stu<strong>de</strong>nt avec trois <strong>de</strong>grés <strong>de</strong> liberté<br />
(figure <strong>de</strong> droite), le coefficient <strong>de</strong> corrél<strong>at</strong>ion étant le même : ρ = 0.8.<br />
que la copule gaussienne est obtenue <strong>de</strong> manière n<strong>at</strong>urelle à partir du principe <strong>de</strong> maximum d’entropie 7 .<br />
En outre, c<strong>et</strong>te copule est sans doute la copule la plus simple <strong>de</strong> la classe <strong>de</strong>s copules elliptiques, puisqu’elle<br />
est entièrement spécifiée par son seul coefficient <strong>de</strong> corrél<strong>at</strong>ion <strong>et</strong> ne dépend donc que d’un seul<br />
paramètre, contrairement à la copule <strong>de</strong> Stu<strong>de</strong>nt, par exemple, qui est aussi fonction d’un certain nombre<br />
<strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté. Ainsi, la copule gaussienne semble être un point <strong>de</strong> départ logique pour l’étu<strong>de</strong> <strong>de</strong> la<br />
dépendance entre actifs <strong>et</strong> nous avons donc choisi <strong>de</strong> tester sa capacité à rendre compte <strong>de</strong>s données (voir<br />
Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2001b), présenté au chapitre suivant). Les tests ont été conduits sur différents<br />
types d’actifs : <strong>de</strong>s actions, <strong>de</strong>s taux <strong>de</strong> change <strong>et</strong> <strong>de</strong>s m<strong>at</strong>ières premières (métaux). Pour ces <strong>de</strong>ux <strong>de</strong>rniers,<br />
la copule gaussienne ne semble pas une très bonne approxim<strong>at</strong>ion. En fait, pour les métaux, elle est<br />
quasi systém<strong>at</strong>iquement rej<strong>et</strong>ée, tandis que pour les monnaies la situ<strong>at</strong>ion est plus complexe. Il semble<br />
en eff<strong>et</strong> que pour les monnaies pour lesquelles n’existe pas <strong>de</strong> mécanisme <strong>de</strong> régul<strong>at</strong>ion <strong>de</strong> la parité l’hypothèse<br />
<strong>de</strong> copule gaussienne soit raisonnable alors que dans le cas contraire (cf. les monnaies <strong>de</strong>s pays<br />
membres du SME) c<strong>et</strong>te hypothèse est fortement rej<strong>et</strong>ée, ce qui, somme toute, est parfaitement cohérent.<br />
Pour ce qui est <strong>de</strong>s actions, la copule gaussienne semble fournir une bonne approxim<strong>at</strong>ion, y compris<br />
pour <strong>de</strong>s actions appartenant à un même secteur d’activité.<br />
D’autres tests paramétriques ont été menés par Klugman <strong>et</strong> Parsa (1999) sur <strong>de</strong>s données touchant aux<br />
domaines <strong>de</strong> l’assurance, où il semble que la famille <strong>de</strong>s copules <strong>de</strong> Frank soient à même <strong>de</strong> rendre<br />
compte <strong>de</strong> la dépendance pour ce genre <strong>de</strong> données. P<strong>at</strong>ton (2001) s’est quant à lui intéressé plus en<br />
détail à la dépendance entre taux <strong>de</strong> change, <strong>et</strong> conclut que la copule gaussienne n’est pas la mieux<br />
adaptée. Au contraire, la copule <strong>de</strong> Clayton fournit une meilleure <strong>de</strong>scription <strong>de</strong>s données.<br />
En fait, une <strong>de</strong>s limites <strong>de</strong>s tests que nous avons conduits sur les copules gaussiennes est <strong>de</strong> ne pas<br />
pouvoir distinguer une copule <strong>de</strong> Stu<strong>de</strong>nt d’une copule gaussienne dès lors que le nombre <strong>de</strong> <strong>de</strong>grés<br />
<strong>de</strong> liberté <strong>de</strong> celle-là <strong>de</strong>vient trop grand. Typiquement, lorsque le nombre <strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté dépasse<br />
quinze ou vingt, les tests que nous avons utilisés ne sont plus capables <strong>de</strong> distinguer entre ces <strong>de</strong>ux types<br />
<strong>de</strong> copules. Ceci n’est pas très grave tant que l’on ne s’occupe que <strong>de</strong>s événements “normaux” <strong>et</strong> pas <strong>de</strong>s<br />
événements extrêmes. En eff<strong>et</strong>, comme nous l’avons déjà signalé, les copules <strong>de</strong> Stu<strong>de</strong>nt <strong>et</strong> les copules<br />
gaussiennes sont très semblables dans leur cœur, <strong>et</strong> ce n’est que lorsque l’on s’en éloigne que l’on <strong>de</strong>vient<br />
7 Pour d’autres exemples <strong>de</strong> détermin<strong>at</strong>ion <strong>de</strong> distributions à l’ai<strong>de</strong> du principe <strong>de</strong> maximum d’entropie, voir notamment<br />
Rockinger <strong>et</strong> Jon<strong>de</strong>au (2002)
188 7. Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s copules<br />
sensible à la différence. Concrètement, la distinction essentielle entre ces <strong>de</strong>ux familles provient du fait<br />
que les copules gaussiennes produisent <strong>de</strong>s événements extrêmes indépendants tandis que les copules <strong>de</strong><br />
Stu<strong>de</strong>nt produisent <strong>de</strong>s événements extrêmes concomitants avec une probabilité non nulle, celle-ci étant<br />
d’autant plus importante que le nombre <strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté <strong>de</strong> la copule est faible <strong>et</strong> que la corrél<strong>at</strong>ion<br />
est importante. Ceci est clairement illustré par la figure 7.1. où l’on a représenté les réalis<strong>at</strong>ions <strong>de</strong> <strong>de</strong>ux<br />
variables alé<strong>at</strong>oires dont les marginales sont gaussiennes <strong>et</strong> la copule est soit gaussienne soit <strong>de</strong> Stu<strong>de</strong>nt<br />
avec trois <strong>de</strong>grés <strong>de</strong> liberté, le coefficient <strong>de</strong> corrél<strong>at</strong>ion étant le même : ρ = 0.8.<br />
Pour quantifier c<strong>et</strong>te propension <strong>de</strong>s extrêmes à se produire simultanément, il est commo<strong>de</strong> d’introduire<br />
le coefficient <strong>de</strong> dépendance <strong>de</strong> queue λ défini comme la probabilité que l’une <strong>de</strong>s variables, X par<br />
exemple, soit extrême conditionnée au fait que l’autre variable, Y , soit elle-même extrême. Ceci nous<br />
conduit à la définition m<strong>at</strong>hém<strong>at</strong>ique suivante :<br />
λ = lim<br />
u→1 Pr{X > FX −1 (u) | Y > FY −1 (u)} (7.11)<br />
= lim<br />
u→1<br />
1 − 2u + C(u, u)<br />
, (7.12)<br />
1 − u<br />
qui démontre que le coefficient <strong>de</strong> dépendance <strong>de</strong> queue n’est fonction que <strong>de</strong> la copule <strong>et</strong> non <strong>de</strong>s<br />
marginales.<br />
Dans le cas <strong>de</strong> la copule gaussienne, il est démontré que la dépendance <strong>de</strong> queue est nulle pour toute<br />
valeur du coefficient <strong>de</strong> corrél<strong>at</strong>ion (différent <strong>de</strong> un), tandis que pour les copules <strong>de</strong> Stu<strong>de</strong>nt, elle dépend<br />
à la fois <strong>de</strong> la corrél<strong>at</strong>ion <strong>et</strong> du nombre <strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté. Dans le premier cas, on dit que les extrêmes se<br />
produisent asymptotiquement indépendamment alors que dans le second cas ils sont asymptotiquement<br />
dépendants. C<strong>et</strong>te différence <strong>de</strong> comportement est illustrée sur la figure 7.1, où dans le cas <strong>de</strong> la copule<br />
<strong>de</strong> Stu<strong>de</strong>nt on remarque que les points se répartissent à l’intérieur <strong>de</strong> fuseaux <strong>de</strong> plus en plus étroits au fur<br />
<strong>et</strong> à mesure que l’on considère les réalis<strong>at</strong>ions <strong>de</strong> plus en plus extrêmes. On observe que ce phénomène<br />
se produit non seulement dans les quadrans inférieur-gauche <strong>et</strong> supérieur-droit, mais aussi dans les <strong>de</strong>ux<br />
autres quadrans, ce qui vient du fait que le coefficient <strong>de</strong> dépendance <strong>de</strong> queue n’est pas nul pour les<br />
coefficients <strong>de</strong> corrél<strong>at</strong>ion nég<strong>at</strong>ifs (voir figure 1 dans Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2001b) page 214).<br />
L’utilis<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue nous a permis d’estimer le risque potentiel qu’il y<br />
a à considérer une modélis<strong>at</strong>ion en terme <strong>de</strong> copule gaussienne, <strong>et</strong> il s’avère que pour <strong>de</strong>s corrél<strong>at</strong>ions<br />
élevées entre actifs il est dangereux d’utiliser c<strong>et</strong>te représent<strong>at</strong>ion tant que l’on n’a pas d’idée plus précise<br />
sur la dépendance <strong>de</strong> queue, quantité à laquelle ne nous donne pas accès les tests paramétriques. En eff<strong>et</strong>,<br />
choisir une forme paramétrique <strong>de</strong> la copule revient à fixer a priori l’existence ou non <strong>de</strong> dépendance <strong>de</strong><br />
queue <strong>et</strong> ne perm<strong>et</strong> donc pas <strong>de</strong> trancher c<strong>et</strong>te question. Donc, pour aller plus avant, une approche non<br />
paramétrique pourrait s’avérer utile.<br />
7.3.2 Tests non paramétriques<br />
Les tests non paramétriques ont, par rapport aux tests paramétriques, l’avantage d’être beaucoup plus<br />
généraux puisqu’ils ne conduisent pas à imposer a priori <strong>de</strong> modèle. Donc, on peut espérer qu’ils apporteront<br />
une réponse au problème auquel nous sommes désormais confrontés. Diverses métho<strong>de</strong>s d’estim<strong>at</strong>ion<br />
non paramétriques ont vu le jour. Citons notamment les approches basées sur l’estim<strong>at</strong>ion à<br />
l’ai<strong>de</strong> <strong>de</strong> familles <strong>de</strong> polynômes, tels les polynômes <strong>de</strong> Bernstein (Li, Mikusinski <strong>et</strong> Taylor 1998), <strong>et</strong> que<br />
Durrleman, Nikeghbali <strong>et</strong> Roncalli (2000) passent en revue. Citons encore les métho<strong>de</strong>s s’appuyant sur<br />
<strong>de</strong>s estim<strong>at</strong>eurs à noyau, telle celle développée par Scaill<strong>et</strong> (2000b).<br />
Toutes ces métho<strong>de</strong>s ont l’avantage <strong>de</strong> produire <strong>de</strong>s copules estimées qui sont lisses <strong>et</strong> dérivables en<br />
tout point, ce qui présente un grand nombre d’avantages lorsque l’on souhaite les réemployer à <strong>de</strong>s
7.4. Conclusion 189<br />
fins plus appliquées. Cela perm<strong>et</strong> notamment <strong>de</strong> réaliser <strong>de</strong>s étu<strong>de</strong>s <strong>de</strong> sensibilité, ou <strong>de</strong> simuler <strong>de</strong>s<br />
variables alé<strong>at</strong>oires ayant la même copule que celle qui a été estimée (Embrechts <strong>et</strong> al. 2002). Cependant,<br />
c<strong>et</strong> avantage est en même temps le principal défaut <strong>de</strong> ces métho<strong>de</strong>s. En eff<strong>et</strong>, on peut montrer très<br />
simplement (Durrleman <strong>et</strong> al. 2000) que toute copule dérivable dans le voisinage du point (1, 1) (ou du<br />
point (0, 0)) n’adm<strong>et</strong> pas <strong>de</strong> dépendance <strong>de</strong> queue. En eff<strong>et</strong>, il est nécessaire que la copule ne soit pas<br />
dérivable au voisinage <strong>de</strong> (1, 1) pour que λ soit non nul 8 .<br />
Ainsi, les métho<strong>de</strong>s d’estim<strong>at</strong>ion non paramétriques <strong>de</strong> la copule que nous venons <strong>de</strong> mentionner souffrent<br />
toutes du même défaut : négliger les événements extrêmes concomitants. Nous avons cherché<br />
d’autres métho<strong>de</strong>s d’estim<strong>at</strong>ion <strong>de</strong>s copules qui ne pâtissent pas <strong>de</strong> ce problème, mais nos recherches<br />
se sont révélées infructueuses. Aussi avons-nous choisi d’<strong>at</strong>taquer le problème <strong>de</strong> l’estim<strong>at</strong>ion <strong>de</strong> la<br />
dépendance <strong>de</strong> queue <strong>de</strong> manière directe. En eff<strong>et</strong>, cela nous est apparu être le seul moyen <strong>de</strong> déci<strong>de</strong>r<br />
si la copule gaussienne, qui est une bonne approxim<strong>at</strong>ion dans le cœur, <strong>de</strong>meurait s<strong>at</strong>isfaisante dans les<br />
extrêmes ou bien s’il valait mieux considérer les copules <strong>de</strong> Stu<strong>de</strong>nt, dont nous avons rappelé qu’elles<br />
étaient très proches <strong>de</strong> la copule gaussienne dans la région centrale mais qu’elles conduisaient à <strong>de</strong>s<br />
réalis<strong>at</strong>ions extrêmes pouvant se produirent simultanément.<br />
7.4 Conclusion<br />
Les copules fournissent le moyen le plus simple, le plus compl<strong>et</strong> <strong>et</strong> le plus n<strong>at</strong>urel <strong>de</strong> décrire <strong>et</strong> <strong>de</strong> mesurer<br />
la dépendance entre plusieurs variables alé<strong>at</strong>oires. Parmi la très gran<strong>de</strong> variété <strong>de</strong> copules, quelques<br />
familles présentent un intérêt particulier pour la modélis<strong>at</strong>ion <strong>de</strong>s interactions entre actifs financiers,<br />
comme par exemple la famille <strong>de</strong>s copules elliptiques, dont les copules gaussiennes <strong>et</strong> les copules <strong>de</strong><br />
Stu<strong>de</strong>nt sont les exemples les plus célèbres, mais aussi la famille <strong>de</strong>s copules archimédiennes. Le rôle<br />
prépondérant <strong>de</strong> ces <strong>de</strong>ux familles vient du fait qu’elles peuvent être reliées simplement à <strong>de</strong>s modèles<br />
phénoménologiques standards en sciences financières (ou actuarielles).<br />
Ceci nous a incité à tenter d’estimer la copule décrivant l’interaction entre actifs financiers. Il nous est<br />
apparu que la copule gaussienne pouvait être un bon candid<strong>at</strong>, ce que nos tests ont confirmé, du moins<br />
pour ce qui est <strong>de</strong>s actions. Cependant, une forte incertitu<strong>de</strong> <strong>de</strong>meure sur le fait <strong>de</strong> savoir si une copule<br />
<strong>de</strong> Stu<strong>de</strong>nt avec un nombre <strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté élevés ne serait pas mieux adaptée. Tout l’enjeu <strong>de</strong> c<strong>et</strong>te<br />
question est <strong>de</strong> savoir si les extrêmes ont tendance à se produire <strong>de</strong> manière simultanée pour les divers<br />
actifs ou indépendamment, point crucial pour la <strong>gestion</strong> <strong>de</strong>s risques <strong>et</strong> la str<strong>at</strong>égie à m<strong>et</strong>tre en œuvre pour<br />
les diversifier. Ceci n’ayant pas pu être résolu par une approche globale, nous allons <strong>de</strong>voir trouver une<br />
métho<strong>de</strong> perm<strong>et</strong>tant l’estim<strong>at</strong>ion directe <strong>de</strong> la dépendance <strong>de</strong> queue, ce qui sera l’obj<strong>et</strong> du chapitre 9.<br />
8 Notons au passage que c<strong>et</strong>te condition nécessaire n’est cependant pas suffisante comme le prouve l’exemple <strong>de</strong> la copule<br />
gaussienne qui ne présente pas <strong>de</strong> dépendance <strong>de</strong> queue, mais n’en est pas pour autant dérivable en (1, 1).
190 7. Etu<strong>de</strong> <strong>de</strong> la dépendance à l’ai<strong>de</strong> <strong>de</strong>s copules
Chapitre 8<br />
Tests <strong>de</strong> copule gaussienne<br />
Nous utilisons la propriété d’invariance <strong>de</strong>s copules par changement <strong>de</strong> variable strictement croissant<br />
pour tester l’hypothèse nulle selon laquelle la dépendance entre actifs financiers peut-être modélisée à<br />
l’ai<strong>de</strong> d’une copule gaussienne. Nous trouvons que certaines <strong>de</strong>vises ainsi que la plupart <strong>de</strong>s principales<br />
actions cotées sur le NYSE sont comp<strong>at</strong>ibles avec c<strong>et</strong>te hypothèse, tandis qu’elle est fortement rej<strong>et</strong>ée<br />
quand il s’agit <strong>de</strong> décrire la dépendance entre m<strong>at</strong>ières premières (métaux). Cependant, en dépit <strong>de</strong><br />
l’apparente qualific<strong>at</strong>ion <strong>de</strong> l’hypothèse <strong>de</strong> copule gaussienne pour la plupart <strong>de</strong>s actions <strong>et</strong> <strong>de</strong>vises, une<br />
copule non gaussienne, telle qu’une copule <strong>de</strong> Stu<strong>de</strong>nt, ne peut être rej<strong>et</strong>ée par nos tests si elle a un<br />
nombre <strong>de</strong> <strong>de</strong>grés <strong>de</strong> liberté suffisamment grand . En conséquence, il peut être dangereux d’accepter<br />
aveuglément l’hypothèse <strong>de</strong> copule gaussienne, tout particulièrement quand le coefficient <strong>de</strong> corrél<strong>at</strong>ion<br />
entre les paires d’actifs est important <strong>de</strong> sorte que l’existence d’une dépendance <strong>de</strong> queue, totalement<br />
ignorée par la copule gaussienne, peut conduire à négliger <strong>de</strong>s événements extrêmes se produisant <strong>de</strong><br />
manière concomitante.<br />
191
192 8. Tests <strong>de</strong> copule gaussienne
Testing the Gaussian Copula Hypothesis<br />
for Financial Ass<strong>et</strong>s Depen<strong>de</strong>nces ∗<br />
Y. Malevergne 1,2 and D. Sorn<strong>et</strong>te 1,3<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095, USA<br />
Corresponding author: D. Sorn<strong>et</strong>te<br />
Institute of Geophysics and Plan<strong>et</strong>ary Physics<br />
University of California, Los Angeles, California 90095, USA<br />
email: sorn<strong>et</strong>te@ess.ucla.edu tel: (310) 825 28 63 Fax: (310) 206 30 51<br />
Submitted to Quantit<strong>at</strong>ive Finance<br />
Abstract<br />
Using one of the key property of copulas th<strong>at</strong> they remain invariant un<strong>de</strong>r an arbitrary monotonous<br />
change of variable, we investig<strong>at</strong>e the null hypothesis th<strong>at</strong> the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween financial ass<strong>et</strong>s<br />
can be mo<strong>de</strong>led by the Gaussian copula. We find th<strong>at</strong> most pairs of currencies and pairs of major<br />
stocks are comp<strong>at</strong>ible with the Gaussian copula hypothesis, while this hypothesis can be rejected for<br />
the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween pairs of commodities (m<strong>et</strong>als). Notwithstanding the apparent qualific<strong>at</strong>ion<br />
of the Gaussian copula hypothesis for most of the currencies and the stocks, a non-Gaussian copula,<br />
such as the Stu<strong>de</strong>nt’s copula, cannot be rejected if it has sufficiently many “<strong>de</strong>grees of freedom”. As a<br />
consequence, it may be very dangerous to embrace blindly the Gaussian copula hypothesis, especially<br />
when the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the pair of ass<strong>et</strong> is too high as the tail <strong>de</strong>pen<strong>de</strong>nce neglected<br />
by the Gaussian copula can be as large as 0.6, i.e., three out five extreme events which occur in unison<br />
are missed.<br />
JEL Classific<strong>at</strong>ion: C12, C15, F31, G19<br />
Keywords: Copulas, Depen<strong>de</strong>nce Mo<strong>de</strong>lis<strong>at</strong>ion, Risk Management, Tail Depen<strong>de</strong>nce.<br />
∗ We acknowledge helpful discussions and exchanges with J. An<strong>de</strong>rsen, P. Embrechts, J.P. Laurent, F. Lindskog, V. Pisarenko<br />
and R. Valkanov. This work was partially supported by the James S. Mc Donnell Found<strong>at</strong>ion 21st century scientist<br />
award/studying complex system.<br />
1<br />
193
194 8. Tests <strong>de</strong> copule gaussienne<br />
1 Introduction<br />
The d<strong>et</strong>ermin<strong>at</strong>ion of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s un<strong>de</strong>rlies many financial activities, such as risk<br />
assessment and portfolio management, as well as option pricing and hedging. Following (Markovitz<br />
1959), the covariance and correl<strong>at</strong>ion m<strong>at</strong>rices have, for a long time, been consi<strong>de</strong>red as the main tools<br />
for quantifying the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s. But the dimension of risk captured by the correl<strong>at</strong>ion<br />
m<strong>at</strong>rices is only s<strong>at</strong>isfying for elliptic distributions and for mo<strong>de</strong>r<strong>at</strong>e risk amplitu<strong>de</strong>s (Sorn<strong>et</strong>te <strong>et</strong> al.<br />
2000a). In all other cases, this measure of risk is severely incompl<strong>et</strong>e and can lead to a very strong<br />
un<strong>de</strong>restim<strong>at</strong>ion of the real incurred risks (Embrechts <strong>et</strong> al. 1999).<br />
Although the unidimensional (marginal) distributions of ass<strong>et</strong> r<strong>et</strong>urns are reasonably constrained<br />
by empirical d<strong>at</strong>a and are more or less s<strong>at</strong>isfactorily <strong>de</strong>scribed by a power law with tail in<strong>de</strong>x ranging<br />
b<strong>et</strong>ween 2 and 4 (De Vries 1994, Lux 1996, Pagan 1996, Guillaume <strong>et</strong> al. 1997, Gopikrishnan <strong>et</strong> al. 1998)<br />
or by str<strong>et</strong>ched exponentials (Laherrère and Sorn<strong>et</strong>te 1998, Gouriéroux and Jasiak 1999, Sorn<strong>et</strong>te <strong>et</strong><br />
al. 2000a, Sorn<strong>et</strong>te <strong>et</strong> al. 2000b), no equivalent results have been obtained for multivari<strong>at</strong>e distributions<br />
of ass<strong>et</strong> r<strong>et</strong>urns. In<strong>de</strong>ed, a brute force d<strong>et</strong>ermin<strong>at</strong>ion of multivari<strong>at</strong>e distributions is unreliable due to the<br />
limited d<strong>at</strong>a s<strong>et</strong> (the curse of dimensionality), while the sole knowledge of marginals (one-point st<strong>at</strong>istics)<br />
of each ass<strong>et</strong> is not sufficient to obtain inform<strong>at</strong>ion on the multivari<strong>at</strong>e distribution of these ass<strong>et</strong>s which<br />
involves all the n-points st<strong>at</strong>istics.<br />
Some progress may be expected from the concept of copulas, recently proposed to be useful for financial<br />
applic<strong>at</strong>ions (Embrechts <strong>et</strong> al. 2001, Frees and Val<strong>de</strong>z 1998, Haas 1999, Klugman and Parsa 1999).<br />
This concept has the <strong>de</strong>sirable property of <strong>de</strong>coupling the study of the marginal distribution of each ass<strong>et</strong><br />
from the study of their collective behavior or <strong>de</strong>pen<strong>de</strong>nce. In<strong>de</strong>ed, the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s is<br />
entirely embed<strong>de</strong>d in the copula, so th<strong>at</strong> a copula allows for a simple <strong>de</strong>scription of the <strong>de</strong>pen<strong>de</strong>nce structure<br />
b<strong>et</strong>ween ass<strong>et</strong>s in<strong>de</strong>pen<strong>de</strong>ntly of the marginals. For instance, ass<strong>et</strong>s can have power law marginals<br />
and a Gaussian copula or altern<strong>at</strong>ively Gaussian marginals and a non-Gaussian copula, and any possible<br />
combin<strong>at</strong>ion thereof. Therefore, the d<strong>et</strong>ermin<strong>at</strong>ion of the multivari<strong>at</strong>e distribution of ass<strong>et</strong>s can be<br />
performed in two steps : (i) an in<strong>de</strong>pen<strong>de</strong>nt d<strong>et</strong>ermin<strong>at</strong>ion of the marginal distributions using standard<br />
techniques for distributions of a single variable ; (ii) a study of the n<strong>at</strong>ure of the copula characterizing<br />
compl<strong>et</strong>ely the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s. This exact separ<strong>at</strong>ion b<strong>et</strong>ween the marginal distributions<br />
and the <strong>de</strong>pen<strong>de</strong>nce is potentially very useful for risk management or option pricing and sensitivity analysis<br />
since it allows for testing several scenarios with different kind of <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween ass<strong>et</strong>s while<br />
the marginals can be s<strong>et</strong> to their well-calibr<strong>at</strong>ed empirical estim<strong>at</strong>es. Such an approach has been used by<br />
(Embrechts <strong>et</strong> al. 2001) to provi<strong>de</strong> various bounds for the Value-<strong>at</strong>-Risk of a portfolio ma<strong>de</strong> of <strong>de</strong>pend<br />
risks, and by (Rosenberg 1999) or (Cherubini and Luciano 2000) to price and to analyse the pricing<br />
sensitivity of binary digital options or options on the minimum of a bask<strong>et</strong> of ass<strong>et</strong>s.<br />
A fundamental limit<strong>at</strong>ion of the copula approach is th<strong>at</strong> there is in principle an infinite number of possible<br />
copulas (Genest and MacKay 1986, Genest 1987, Genest and Rivest 1993, Joe 1993, Nelsen 1998)<br />
and, up to now, no general empirical study has d<strong>et</strong>ermined the classes of copulas th<strong>at</strong> are acceptable for<br />
financial problems. In general, the choice of a given copula is gui<strong>de</strong>d both by the empirical evi<strong>de</strong>nces<br />
and the technical constraints, i.e., the number of param<strong>et</strong>ers necessary to <strong>de</strong>scribe the copula, the possibility<br />
to obtain efficient estim<strong>at</strong>ors of these param<strong>et</strong>ers and also the possiblity offered by the chosen<br />
param<strong>et</strong>eriz<strong>at</strong>ion to allow for tractable analytical calcul<strong>at</strong>ion. It is in<strong>de</strong>ed som<strong>et</strong>imes more advantageous<br />
to prefer a simplest copula to one th<strong>at</strong> fit b<strong>et</strong>ter the d<strong>at</strong>a, provi<strong>de</strong>d th<strong>at</strong> we can clearly quantify the effects<br />
of this substitution.<br />
In this vein, the first goal of the present article is to show th<strong>at</strong>, in most cases, the Gaussian copula<br />
2
can provi<strong>de</strong> an approxim<strong>at</strong>ion of the unknown true copula th<strong>at</strong> is sufficiently good so th<strong>at</strong> it cannot be<br />
rejected on a st<strong>at</strong>istical basis. Our second goal is to draw the consequences of the param<strong>et</strong>eriz<strong>at</strong>ion<br />
involved in the Gaussian copula in term of potential over/un<strong>de</strong>restim<strong>at</strong>ion of the risks, in particular for<br />
large and extreme events.<br />
The paper is organized as follows.<br />
In section 2, we first recall some important general <strong>de</strong>finitions and theorems about copulas th<strong>at</strong> will<br />
be useful in the sequel. We then introduce the concept of tail <strong>de</strong>pen<strong>de</strong>nce th<strong>at</strong> will allow us to quantify<br />
the probability th<strong>at</strong> two extreme events might occur simultaneously. We <strong>de</strong>fine and <strong>de</strong>scribe the two<br />
copulas th<strong>at</strong> will be <strong>at</strong> the core of our study : the Gaussian copula and the Stu<strong>de</strong>nt’s copula and compare<br />
their properties particularly in the tails.<br />
In section 3, we present our st<strong>at</strong>istical testing procedure which is applied to pairs of financial time<br />
series. First of all, we d<strong>et</strong>ermine a test st<strong>at</strong>istics which leads us to compare the empirical distribution of<br />
the d<strong>at</strong>a with a χ 2 -distribution using a bootstrap m<strong>et</strong>hod. We also test the sensitivity of our procedure<br />
by applying it to synth<strong>et</strong>ic multivari<strong>at</strong>e Stu<strong>de</strong>nt’s time series. This allows us to d<strong>et</strong>ermine the minimum<br />
st<strong>at</strong>istical test value nee<strong>de</strong>d to be able to distinguish b<strong>et</strong>ween a Gaussian and a Stu<strong>de</strong>nt’s copula, as a<br />
function of the number of <strong>de</strong>grees of freedom and of the correl<strong>at</strong>ion strength.<br />
Section 4 presents the empirical results obtained for the following ass<strong>et</strong>s which are combined pairwise<br />
in the test st<strong>at</strong>istics:<br />
• 6 currencies,<br />
• 6 m<strong>et</strong>als tra<strong>de</strong>d on the London M<strong>et</strong>al Exchange,<br />
• 22 stocks choosen among the largest companies quoted on the New York Stocks Exchange.<br />
We show th<strong>at</strong> the Gaussian copula hypothesis is very reasonnable for most stocks and currencies, while<br />
it is hardly comp<strong>at</strong>ible with the <strong>de</strong>scription of multivari<strong>at</strong>e behavior for m<strong>et</strong>als.<br />
Section 5 summarizes our results and conclu<strong>de</strong>s.<br />
2 Generalities about copulas<br />
2.1 Definitions and important results about copulas<br />
This section does not pr<strong>et</strong>end to provi<strong>de</strong> a rigorous m<strong>at</strong>hem<strong>at</strong>ical exposition of the concept of copula. We<br />
only recall a few basic <strong>de</strong>finitions and theorems th<strong>at</strong> will be useful in the following (for more inform<strong>at</strong>ion<br />
about the concept of copula, see for instance (Lindskog 1999, Nelsen 1998)).<br />
We first give the <strong>de</strong>finition of a copula of n random variables.<br />
DEFINITION 1 (COPULA)<br />
A function C : [0, 1] n −→ [0, 1] is a n-copula if it enjoys the following properties :<br />
• ∀u ∈ [0, 1], C(1, · · · , 1, u, 1 · · · , 1) = u ,<br />
• ∀ui ∈ [0, 1], C(u1, · · · , un) = 0 if <strong>at</strong> least one of the ui equals zero ,<br />
3<br />
195
196 8. Tests <strong>de</strong> copule gaussienne<br />
• C is groun<strong>de</strong>d and n-increasing, i.e., the C-volume of every boxes whose vertices lie in [0, 1] n is<br />
positive.<br />
It is clear from this <strong>de</strong>finition th<strong>at</strong> a copula is nothing but a multivari<strong>at</strong>e distribution with support<br />
in [0,1] n and with uniform marginals. The fact th<strong>at</strong> such copulas can be very useful for representing<br />
multivari<strong>at</strong>e distributions with arbitrary marginals is seen from the following result.<br />
THEOREM 1 (SKLAR’S THEOREM)<br />
Given an n-dimensional distribution function F with continuous marginal (cumul<strong>at</strong>ive) distributions<br />
F1, · · · , Fn, there exists a unique n-copula C : [0, 1] n −→ [0, 1] such th<strong>at</strong> :<br />
F (x1, · · · , xn) = C(F1(x1), · · · , Fn(xn)) . (1)<br />
This theorem provi<strong>de</strong>s both a param<strong>et</strong>eriz<strong>at</strong>ion of multivari<strong>at</strong>e distributions and a construction scheme<br />
for copulas. In<strong>de</strong>ed, given a multivari<strong>at</strong>e distribution F with marginals F1, · · · , Fn, the function<br />
C(u1, · · · , un) = F F −1<br />
1 (u1), · · · , F −1<br />
n (un) <br />
is autom<strong>at</strong>ically a n-copula. This copula is the copula of the multivari<strong>at</strong>e distribution F . We will use this<br />
m<strong>et</strong>hod in the sequel to <strong>de</strong>rive the expressions of standard copulas such as the Gaussian copula or the<br />
Stu<strong>de</strong>nt’s copula.<br />
A very powerful property of copulas is their invariance un<strong>de</strong>r arbitrary strictly increasing mapping<br />
of the random variables :<br />
THEOREM 2 (INVARIANCE THEOREM)<br />
Consi<strong>de</strong>r n continuous random variables X1, · · · , Xn with copula C. Then, if g1(X1), · · · , gn(Xn) are<br />
strictly increasing on the ranges of X1, · · · , Xn, the random variables Y1 = g1(X1), · · · , Yn = gn(Xn)<br />
have exactly the same copula C.<br />
It is this result th<strong>at</strong> shows us th<strong>at</strong> the full <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the n random variables is compl<strong>et</strong>ely<br />
captured by the copula, in<strong>de</strong>pen<strong>de</strong>ntly of the shape of the marginal distributions. This result is <strong>at</strong> the<br />
basis of our st<strong>at</strong>istical study presented in section 3.<br />
2.2 Depen<strong>de</strong>nce b<strong>et</strong>ween random variables<br />
The <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two time series is usually <strong>de</strong>scribed by their correl<strong>at</strong>ion coefficient. This measure<br />
is fully s<strong>at</strong>isfactory only for elliptic distributions (Embrechts <strong>et</strong> al. 1999), which are functions of a<br />
quadr<strong>at</strong>ic form of the random variables, when one is interested in mo<strong>de</strong>r<strong>at</strong>ely size events. However, an<br />
important issue for risk management concerns the d<strong>et</strong>ermin<strong>at</strong>ion of the <strong>de</strong>pen<strong>de</strong>nce of the distributions in<br />
the tails. Practically, the question is wh<strong>et</strong>her it is more probable th<strong>at</strong> large or extreme events occur simultaneously<br />
or on the contrary more or less in<strong>de</strong>pen<strong>de</strong>ntly. This is refered to as the presence or abscence<br />
of “tail <strong>de</strong>pen<strong>de</strong>nce”.<br />
The tail <strong>de</strong>pen<strong>de</strong>nce is also an interesting concept in studying the contagion of crises b<strong>et</strong>ween mark<strong>et</strong>s<br />
or countries. These questions have recently been addressed by (Ang and Cheng 2001, Longin and Solnik<br />
2001, Starica 1999) among several others. Large neg<strong>at</strong>ive moves in a country or mark<strong>et</strong> are often found<br />
to imply large neg<strong>at</strong>ive moves in others.<br />
Technically, we need to d<strong>et</strong>ermine the probability th<strong>at</strong> a random variable X is large, knowing th<strong>at</strong><br />
the random variable Y is large.<br />
4<br />
(2)
DEFINITION 2 (TAIL DEPENDENCE 1)<br />
L<strong>et</strong> X and Y be random variables with continuous marginals FX and FY . The (upper) tail <strong>de</strong>pen<strong>de</strong>nce<br />
coefficient of X and Y is, if it exists,<br />
lim<br />
u→1<br />
Pr{X > F −1<br />
X<br />
197<br />
−1<br />
(u)|Y > F (u)} = λ ∈ [0, 1] . (3)<br />
In words, given th<strong>at</strong> Y is very large (which occurs with probability 1 − u), the probability th<strong>at</strong> X is very<br />
large <strong>at</strong> the same probability level u <strong>de</strong>fines asymptotically the tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ.<br />
It turns out th<strong>at</strong> this tail <strong>de</strong>pen<strong>de</strong>nce is a pure copula property which is in<strong>de</strong>pen<strong>de</strong>nt of the marginals. L<strong>et</strong><br />
C be the copula of the variables X and Y , then<br />
THEOREM 3<br />
if the bivari<strong>at</strong>e copula C is such th<strong>at</strong><br />
lim<br />
u→1<br />
¯C(u, u)<br />
1 − u<br />
Y<br />
= λ (4)<br />
exists (where ¯ C(u, u) = 1 − 2u − C(u, u)), then C has an upper tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ.<br />
If λ > 0, the copula presents tail <strong>de</strong>pen<strong>de</strong>nce and large events tend to occur simultanously, with the<br />
probabilty λ. On the contrary, when λ = 0, the copula has no tail <strong>de</strong>pen<strong>de</strong>nce in this sense and large<br />
events appear to occur essentially in<strong>de</strong>pen<strong>de</strong>ntly. There is however a subtl<strong>et</strong>y in this <strong>de</strong>finition of tail<br />
<strong>de</strong>pen<strong>de</strong>nce. To make it clear, first consi<strong>de</strong>r the case where for large X and Y the distribution function<br />
F (x, y) factorizes such th<strong>at</strong><br />
lim<br />
x,y→∞<br />
F (x, y)<br />
= 1 . (5)<br />
FX(x)FY (y)<br />
This means th<strong>at</strong>, for X and Y sufficiently large, these two variables can be consi<strong>de</strong>red as in<strong>de</strong>pen<strong>de</strong>nt. It<br />
is then easy to show th<strong>at</strong><br />
lim<br />
u→1<br />
Pr{X > F −1<br />
X<br />
−1<br />
(u)|Y > F (u)} = lim<br />
Y<br />
u→1<br />
so th<strong>at</strong> in<strong>de</strong>pen<strong>de</strong>nt variables really have no tail <strong>de</strong>pen<strong>de</strong>nce, as one can expect.<br />
−1<br />
1 − FX(FX (u)) (6)<br />
= lim<br />
u→1 1 − u = 0, (7)<br />
Unfortun<strong>at</strong>ly, the converse does not holds : a value λ = 0 does not autom<strong>at</strong>ically imply true in<strong>de</strong>pen<strong>de</strong>nce,<br />
namely th<strong>at</strong> F (x, y) s<strong>at</strong>isfies equ<strong>at</strong>ion (5). In<strong>de</strong>ed, the tail in<strong>de</strong>pen<strong>de</strong>nce criterion λ = 0 may<br />
still be associ<strong>at</strong>ed with an absence of factoriz<strong>at</strong>ion of the multivari<strong>at</strong>e distribution for large X and Y .<br />
In a weaker sense, there may still be a <strong>de</strong>pen<strong>de</strong>nce in the tail even when λ = 0. Such behavior is for<br />
instance exhibited by the Gaussian copula, which has zero tail <strong>de</strong>pen<strong>de</strong>nce according to the <strong>de</strong>finition 2<br />
but nevertheless does not have a factorizable multivari<strong>at</strong>e distribution, since the non-diagonal term of the<br />
quadr<strong>at</strong>ic form in the exponential function does not become negligible in general as X and Y go to infinity.<br />
To summarize, the tail in<strong>de</strong>pen<strong>de</strong>nce, according to <strong>de</strong>finition 2, is not equivalent to the in<strong>de</strong>pen<strong>de</strong>nce<br />
in the tail as <strong>de</strong>fined in equ<strong>at</strong>ion (5).<br />
After this brief review of the main concepts un<strong>de</strong>rlying copulas, we now present two special families<br />
of copulas : the Gaussian copula and the Stu<strong>de</strong>nt’s copula.<br />
5
198 8. Tests <strong>de</strong> copule gaussienne<br />
2.3 The Gaussian copula<br />
The Gaussian copula is the copula <strong>de</strong>rived from the multivari<strong>at</strong>e Gaussian distribution. L<strong>et</strong> Φ <strong>de</strong>note<br />
the standard Normal (cumul<strong>at</strong>ive) distribution and Φρ,n the n-dimensional Gaussian distribution with<br />
correl<strong>at</strong>ion m<strong>at</strong>rix ρ. Then, the Gaussian n-copula with correl<strong>at</strong>ion m<strong>at</strong>rix ρ is<br />
whose <strong>de</strong>nsity<br />
reads<br />
−1<br />
Cρ(u1, · · · , un) = Φρ,n Φ (u1), · · · , Φ −1 (un) , (8)<br />
cρ(u1, · · · , un) =<br />
cρ(u1, · · · , un) = ∂Cρ(u1, · · · , un)<br />
∂u1 · · · ∂un<br />
<br />
1<br />
√ exp −<br />
d<strong>et</strong> ρ 1<br />
2 yt (u) (ρ−1 <br />
− Id)y (u)<br />
with yk(u) = Φ −1 (uk). Note th<strong>at</strong> theorem 1 and equ<strong>at</strong>ion (2) ensure th<strong>at</strong> Cρ(u1, · · · , un) in equ<strong>at</strong>ion (8)<br />
is a copula.<br />
As we said before, the Gaussian copula does not have a tail <strong>de</strong>pen<strong>de</strong>nce :<br />
lim<br />
u→1<br />
¯Cρ(u, u)<br />
1 − u<br />
(9)<br />
(10)<br />
= 0, ∀ρ ∈ (−1, 1). (11)<br />
This results is <strong>de</strong>rived for example in (Embrechts <strong>et</strong> al. 2001). But this does not mean th<strong>at</strong> the Gaussian<br />
copula goes to the in<strong>de</strong>pen<strong>de</strong>nt (or product) copula Π(u1, u2) = u1 · u2 when (u1, u2) goes to one.<br />
In<strong>de</strong>ed, consi<strong>de</strong>r a distribution F (x, y) with Gaussian copula :<br />
Its <strong>de</strong>nsity is<br />
where fX and fY are the <strong>de</strong>nsities of X and Y . Thus,<br />
F (x, y) = Cρ(FX(x), FY (y)). (12)<br />
f(x, y) = cρ(FX(x), FY (y)) · fX(x) · fY (y), (13)<br />
f(x, y)<br />
lim<br />
= lim<br />
(x,y)→∞ fX(x) · fY (y) (x,y)→∞ cρ(FX(x), FY (y)), (14)<br />
which should equal 1 if the variables X and Y were in<strong>de</strong>pen<strong>de</strong>nt in the tail. Reasoning in the quantile<br />
space, we s<strong>et</strong> x = F −1<br />
−1<br />
X (u) and y = FY (u), which yield<br />
f(x, y)<br />
lim<br />
= lim<br />
(x,y)→∞ fX(x) · fY (y) u→1 cρ(u, u). (15)<br />
Using equ<strong>at</strong>ion (10), it is now obvious to show th<strong>at</strong> cρ(u, u) goes to one when u goes to one, if and<br />
only if ρ = 0 which is equivalent to Cρ=0(u1, u2) = Π(u1, u2) for every (u1, u2). When ρ > 0, cρ(u, u)<br />
goes to infinity, while for ρ neg<strong>at</strong>ive, cρ(u, u) goes to zero as u → 1. Thus, the <strong>de</strong>pen<strong>de</strong>nce structure<br />
<strong>de</strong>scribed by the Gaussian copula is very different from the <strong>de</strong>pen<strong>de</strong>nce structure of the in<strong>de</strong>pen<strong>de</strong>nt<br />
copula, except for ρ = 0.<br />
The Gaussian copula is compl<strong>et</strong>ly d<strong>et</strong>ermined by the knowledge of the correl<strong>at</strong>ion m<strong>at</strong>rix ρ. The<br />
param<strong>et</strong>ers involved in the <strong>de</strong>scription of the Gaussian copula are very simple to estim<strong>at</strong>e, as we shall<br />
see in the following.<br />
6
In our tests presented below, we focus on pairs of ass<strong>et</strong>s, i.e., on Gaussian copulas involving only<br />
two random variables. Testing the Gaussian copula hypothesis for two random variables gives useful<br />
inform<strong>at</strong>ion for a larger number of <strong>de</strong>pen<strong>de</strong>nt variables constituting a large bask<strong>et</strong> or portfolio. In<strong>de</strong>ed,<br />
l<strong>et</strong> us assume th<strong>at</strong> each pair (a, b), (b, c) and (c, a) have a gaussian copula. Then, the tripl<strong>et</strong> (a, b, c) has<br />
also a Gaussian copula. This result generalizes to an arbitrary number of random variables.<br />
2.4 The Stu<strong>de</strong>nt’s copula<br />
The Stu<strong>de</strong>nt’s copula is <strong>de</strong>rived from the Stu<strong>de</strong>nt’s multivari<strong>at</strong>e distribution. Given a multivari<strong>at</strong>e Stu<strong>de</strong>nt’s<br />
distribution Tρ,ν with ν <strong>de</strong>grees of freedom and a correl<strong>at</strong>ion m<strong>at</strong>rix ρ<br />
1 Γ<br />
Tρ,ν(x) = √<br />
d<strong>et</strong> ρ<br />
<br />
ν+n<br />
2<br />
Γ <br />
ν<br />
2 (πν) N/2<br />
the corresponding Stu<strong>de</strong>nt’s copula reads :<br />
Cρ,ν(u1, · · · , un) = Tρ,ν<br />
x1<br />
−∞<br />
xN<br />
· · ·<br />
−∞<br />
<br />
dx<br />
1 + xt ρx<br />
ν<br />
ν+n<br />
2<br />
199<br />
, (16)<br />
t −1<br />
ν (u1), · · · , t −1<br />
ν (un) , (17)<br />
where tν is the univari<strong>at</strong>e Stu<strong>de</strong>nt’s distribution with ν <strong>de</strong>grees of freedom. The <strong>de</strong>nsity of the Stu<strong>de</strong>nt’s<br />
copula is thus<br />
where yk = t −1<br />
ν (uk).<br />
cρ,ν(u1, · · · , un) =<br />
1 Γ<br />
√<br />
d<strong>et</strong> ρ<br />
<br />
ν+n ν<br />
2 Γ<br />
<br />
ν+1 Γ 2<br />
n−1 2<br />
n<br />
<br />
n<br />
k=1<br />
<br />
1 + y2 k<br />
ν<br />
1 + yt ρy<br />
ν<br />
ν+1<br />
2<br />
ν+n<br />
2<br />
, (18)<br />
Since the Stu<strong>de</strong>nt’s distribution tends to the normal distribution when ν goes to infinity, the Stu<strong>de</strong>nt’s<br />
copula tends to the Gaussian copula as ν → +∞. In contrast to the Gaussian copula, the Stu<strong>de</strong>nt’s<br />
copula for ν finite presents a tail <strong>de</strong>pen<strong>de</strong>nce given by :<br />
λν(ρ) = lim<br />
u→1<br />
¯Cρ,ν(u, u)<br />
1 − u<br />
= 2¯tν+1<br />
√ √ <br />
ν + 1 1 − ρ<br />
√<br />
1 + ρ<br />
, (19)<br />
where ¯tν+1 is the complementary cumul<strong>at</strong>ive univari<strong>at</strong>e Stu<strong>de</strong>nt’s distribution with ν + 1 <strong>de</strong>grees of<br />
freedom (see (Embrechts <strong>et</strong> al. 2001) for the proof). Figure 1 shows the upper tail <strong>de</strong>pen<strong>de</strong>nce coefficient<br />
as a function of the correl<strong>at</strong>ion coefficient ρ for different values of the number ν of <strong>de</strong>grees of freedom.<br />
As expected from the fact th<strong>at</strong> the Stu<strong>de</strong>nt’s copula becomes i<strong>de</strong>ntical to the Gaussian copula for ν →<br />
+∞ for all ρ = 1, λν(ρ) exhibits a regular <strong>de</strong>cay to zero as ν increases. Moreover, for ν sufficiently large,<br />
the tail <strong>de</strong>pen<strong>de</strong>nce is significantly different from 0 only when the correl<strong>at</strong>ion coefficient is sufficiently<br />
close to 1. This suggests th<strong>at</strong>, for mo<strong>de</strong>r<strong>at</strong>e values of the correl<strong>at</strong>ion coefficient, a Stu<strong>de</strong>nt’s copula with<br />
a large number of <strong>de</strong>grees of freedom may be difficult to distinguish from the Gaussian copula from a<br />
st<strong>at</strong>istical point of view. This st<strong>at</strong>ement will be ma<strong>de</strong> quantit<strong>at</strong>ive in the following.<br />
Figure 2 presents the same inform<strong>at</strong>ion in a different way by showing the maximum value of the<br />
correl<strong>at</strong>ion coefficient ρ as a function of ν, below which the tail <strong>de</strong>pen<strong>de</strong>nce λν(ρ) of a Stu<strong>de</strong>nt’s copula<br />
is smaller than a given small value, here taken equal to 1%, 2.5%, 5% and 10%. The choice λν(ρ) = 5%<br />
for instance corresponds to 1 event in 20 for which the pair of variables are asymptotically coupled. At<br />
7
200 8. Tests <strong>de</strong> copule gaussienne<br />
the 95% probability level, values of λν(ρ) ≤ 5% are undistinguishable from 0, which means th<strong>at</strong> the<br />
Stu<strong>de</strong>nt’s copula can be approxim<strong>at</strong>ed by a Gaussian copula.<br />
The <strong>de</strong>scription of a Stu<strong>de</strong>nt’s copula relies on two param<strong>et</strong>ers : the correl<strong>at</strong>ion m<strong>at</strong>rix ρ, as in the<br />
Gaussian case, and in addition the number of <strong>de</strong>grees of freedom ν. The estim<strong>at</strong>ion of the param<strong>et</strong>er ν<br />
is r<strong>at</strong>her difficult and this has an important impact on the estim<strong>at</strong>ed value of the correl<strong>at</strong>ion m<strong>at</strong>rix. As a<br />
consequence, the Stu<strong>de</strong>nt’s copula is more difficult to calibr<strong>at</strong>e and use than the Gaussian copula.<br />
3 Testing the Gaussian copula hypothesis<br />
In view of the central role th<strong>at</strong> the Gaussian paradigm has played and still plays in particular in finance, it<br />
is n<strong>at</strong>ural to start with the simplest choice of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween different random variables, namely the<br />
Gaussian copula. It is also a n<strong>at</strong>ural first step as the Gaussian copula imposes itself in an approach which<br />
consists in (1) performing a nonlinear transform<strong>at</strong>ion on the random variables into Normal random variables<br />
(for the marginals) which is always possible and (2) invoking a maximum entropy principle (which<br />
amounts to add the least additional inform<strong>at</strong>ion in the Shannon sense) to construct the multivariable distribution<br />
of these Gaussianized random variables (Sorn<strong>et</strong>te <strong>et</strong> al. 2000a, Sorn<strong>et</strong>te <strong>et</strong> al. 2000b, An<strong>de</strong>rsen<br />
and Sorn<strong>et</strong>te 2001).<br />
In the sequel, we will <strong>de</strong>note by H0 the null hypothesis according to which the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
two (or more) random variables X and Y can be <strong>de</strong>scribed by the Gaussian copula.<br />
3.1 Test St<strong>at</strong>istics<br />
We now <strong>de</strong>rive the test st<strong>at</strong>istics which will allow us to reject or not our null hypothesis H0 and st<strong>at</strong>e the<br />
following proposition:<br />
PROPOSITION 1<br />
Assuming th<strong>at</strong> the N-dimensionnal random vector x = (x1, · · · , xN) with distribution function F and<br />
marginals Fi, s<strong>at</strong>isfies the null hypothesis H0, then, the variable<br />
where the m<strong>at</strong>rix ρ is<br />
z 2 =<br />
N<br />
j,i=1<br />
follows a χ 2 -distribution with N <strong>de</strong>grees of freedom.<br />
Φ −1 (Fi(xi)) (ρ −1 )ij Φ −1 (Fj(xj)), (20)<br />
ρij = Cov[Φ −1 (Fi(xi)), Φ −1 (Fj(xj))], (21)<br />
To proove the proposition above, first consi<strong>de</strong>r an N-dimensionnal random vector x = (x1, · · · , xN).<br />
L<strong>et</strong> us <strong>de</strong>note by F its distribution function and by Fi the marginal distribution of each xi. L<strong>et</strong> us now<br />
assume th<strong>at</strong> the distribution function F s<strong>at</strong>isfies H0, so th<strong>at</strong> F has a Gaussian copula with correl<strong>at</strong>ion<br />
m<strong>at</strong>rix ρ while the Fi’s can be any distribution function. According to theorem 1, the distribution F can<br />
be represented as :<br />
F (x1, · · · , xN) = Φρ,N(Φ −1 (F1(x1)), · · · , Φ −1 (FN(xN))) . (22)<br />
L<strong>et</strong> us now transform the xi’s into Normal random variables yi’s :<br />
yi = Φ −1 (Fi(xi)) . (23)<br />
8
Since the mapping Φ −1 (Fi(·)) is obviously increasing, theorem 2 allows us to conclu<strong>de</strong> th<strong>at</strong> the copula<br />
of the variables yi’s is i<strong>de</strong>ntical to the copula of the variables xi’s. Therefore, the variables yi’s have<br />
Normal marginal distributions and a Gaussian copula with correl<strong>at</strong>ion m<strong>at</strong>rix ρ. Thus, by <strong>de</strong>finition, the<br />
multivari<strong>at</strong>e distribution of the yi’s is the multivari<strong>at</strong>e Gaussian distribution with correl<strong>at</strong>ion m<strong>at</strong>rix ρ :<br />
201<br />
G(y) = Φρ,N(Φ −1 (F1(x1)), · · · , Φ −1 (FN(xN))) (24)<br />
= Φρ,N(y1, · · · , yN), (25)<br />
and y is a Gaussian random vector. From equ<strong>at</strong>ions (24-25), we obviously have<br />
Consi<strong>de</strong>r now the random variable<br />
ρij = Cov[Φ −1 (Fi(xi)), Φ −1 (Fj(xj))]. (26)<br />
z 2 = y t ρ −1 y =<br />
N<br />
yi (ρ −1 )ij yj , (27)<br />
i,j=1<br />
where · t <strong>de</strong>notes the transpose oper<strong>at</strong>or. This variable has already been consi<strong>de</strong>red in (Sorn<strong>et</strong>te <strong>et</strong> al.<br />
2000a) in preliminary st<strong>at</strong>istical tests of the transform<strong>at</strong>ion (23). It is well-known th<strong>at</strong> the variable z 2<br />
follows a χ 2 -distribution with N <strong>de</strong>grees of freedom. In<strong>de</strong>ed, since y is a Gaussian random vector with<br />
covariance m<strong>at</strong>rix 1 ρ, it follows th<strong>at</strong> the components of the vector<br />
˜y = ρ −1/2 y, (28)<br />
are in<strong>de</strong>pen<strong>de</strong>nt Normal random variables. Here, ρ −1/2 <strong>de</strong>notes the square root of the m<strong>at</strong>rix ρ −1 , which<br />
can be obtain by the Cholevsky <strong>de</strong>composition, for instance. Thus, the sum ˜y t ˜y = z 2 is the sum of the<br />
squares of N in<strong>de</strong>pen<strong>de</strong>nt Normal random variables, which follows a χ 2 -distribution with N <strong>de</strong>grees of<br />
freedom.<br />
3.2 Testing procedure<br />
The testing procedure used in the sequel is now <strong>de</strong>scribed. We consi<strong>de</strong>r two financial series (N = 2) of<br />
size T : {x1(1), · · · , x1(t), · · · , x1(T )} and {x2(1), · · · , x2(t), · · · , x2(T )}. We assume th<strong>at</strong> the vectors<br />
x(t) = (x1(t), x2(t)), t ∈ {1, · · · , T } are in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>nticaly distributed with distribution F ,<br />
which implies th<strong>at</strong> the variables x1(t) (respectively x2(t)), t ∈ {1, · · · , T }, are also in<strong>de</strong>pen<strong>de</strong>nt and<br />
i<strong>de</strong>ntically distributed, with distributions F1 (respectively F2).<br />
The cumul<strong>at</strong>ive distribution ˆ Fi of each variable xi, which is estim<strong>at</strong>ed empirically, is given by<br />
ˆFi(xi) = 1<br />
T<br />
T<br />
1 {xi
202 8. Tests <strong>de</strong> copule gaussienne<br />
The sample covariance m<strong>at</strong>rix ˆρ is estim<strong>at</strong>ed by the expression :<br />
which allows us to calcul<strong>at</strong>e the variable<br />
ˆz 2 (k) =<br />
ˆρ = 1<br />
T<br />
2<br />
i,j=1<br />
T<br />
ˆy(i) · ˆy(i) t<br />
i=1<br />
(31)<br />
ˆyi(k) ( ˆ<br />
ρ −1 )ij ˆyj(k) , (32)<br />
as <strong>de</strong>fined in (27) for k ∈ {1, · · · , T }, which should be asymptotically distributed according to a χ 2 -<br />
distribution if the Gaussian copula hypothesis is correct.<br />
The usual way for comparing an empirical with a theor<strong>et</strong>ical distribution is to measure the distance<br />
b<strong>et</strong>ween these two distributions and to perform the Kolmogorov test or the An<strong>de</strong>rson-Darling (An<strong>de</strong>rson<br />
and Darling 1952) test (for a b<strong>et</strong>ter accuracy in the tails of the distribution). The Kolmogorov distance<br />
is the maximum local distance along the quantile which most often occur in the bulk of the distribution,<br />
while the An<strong>de</strong>rson-Darling distance puts the emphasis on the tails of the two distributions by a suitable<br />
normaliz<strong>at</strong>ion. We propose to complement these two distances by two additional measures which are<br />
<strong>de</strong>fined as averages of the Kolmogorov distance and of the An<strong>de</strong>rson-Darling distance respectively:<br />
Kolmogorov : d1 = max |Fz2(z z<br />
2 ) − Fχ2(z 2 )|<br />
<br />
(33)<br />
average Kolmogorov : d2 = |Fz2(z 2 ) − Fz2(z 2 )| dFχ2(z 2 ) (34)<br />
An<strong>de</strong>rson − Darling : d3 = max<br />
z<br />
average An<strong>de</strong>rson − Darling : d4 =<br />
<br />
|F z 2(z 2 ) − F χ 2(z 2 )|<br />
<br />
F χ 2(z 2 )[1 − F χ 2(z 2 )]<br />
(35)<br />
|Fz2(z 2 ) − Fχ2(z2 )|<br />
<br />
Fχ2(z2 )[1 − Fχ2(z2 dFχ2(z )]<br />
2 ) (36)<br />
The Kolmogorov distance d1 and its average d2 are more sensitive to the <strong>de</strong>vi<strong>at</strong>ions occurring in the bulk<br />
of the distributions. In contrast, the An<strong>de</strong>rson-Darling distance d3 and its average d4 are more accur<strong>at</strong>e<br />
in the tails of the distributions. We present our st<strong>at</strong>istical tests for these four distances in or<strong>de</strong>r to be as<br />
compl<strong>et</strong>e as possible with respect to the different sensitivity of the tests.<br />
The distances d2 and d4 are not of common use in st<strong>at</strong>istics, so l<strong>et</strong> us justify our choice. One usually<br />
uses distances similar to d2 and d4 but which differ by the square instead of the modulus of F z 2(z 2 ) −<br />
F χ 2(z 2 ) and lead respectively to the ω-test and the Ω-test, whose st<strong>at</strong>itics are theor<strong>et</strong>ically known. The<br />
main advantage of the distances d2 and d4 with respect to the more usual distances ω and Ω is th<strong>at</strong> they<br />
are simply equal to the average of d1 and d3. This averaging is very interesting and provi<strong>de</strong>s important<br />
inform<strong>at</strong>ion. In<strong>de</strong>ed, the distances d1 and d3 are mainly controlled by the point th<strong>at</strong> maximizes the<br />
argument within the max(·) function. They are thus sensitive to the presence of an outlier. By averaging,<br />
d2 and d4 become less sensitive to outliers, since the weight of such points is only of or<strong>de</strong>r 1/T (where<br />
T is the size of the sample) while it equals one for d1 and d3. Of course, the distances ω and Ω also<br />
perform a smoothing since they are averaged quantities too. But they are the average of the square of<br />
d1 and d3 which leads to an un<strong>de</strong>sired overweighting of the largest events. In fact, this weight function<br />
is chosen as a convenient analytical form th<strong>at</strong> allows one to <strong>de</strong>rive explicitely the theor<strong>et</strong>ical asymptotic<br />
st<strong>at</strong>istics for the ω and Ω-tests. In contrast, using the modulus of F z 2(z 2 ) − F χ 2(z 2 ) instead of its<br />
square in the expression of d2 and d4, no theor<strong>et</strong>ical test st<strong>at</strong>istics can be <strong>de</strong>rived analytically. In other<br />
10
words, the presence of the square instead of the modulus of F z 2(z 2 ) − F χ 2(z 2 ) in the <strong>de</strong>finition of<br />
the distances ω and Ω is motiv<strong>at</strong>ed by m<strong>at</strong>hem<strong>at</strong>ical convenience r<strong>at</strong>her than by st<strong>at</strong>istical pertinence.<br />
In sum, the sole advantage of the standard distances ω and Ω with respect to the distances d2 and d4<br />
introduced here is the theor<strong>et</strong>ical knowledge of their distributions. However, this advantage disappears<br />
in our present case in which the covariance m<strong>at</strong>rix is not known a priori and needs to be estim<strong>at</strong>ed from<br />
the empirical d<strong>at</strong>a: in<strong>de</strong>ed, the exact knowledge of all the param<strong>et</strong>ers is necessary in the <strong>de</strong>riv<strong>at</strong>ion of<br />
the theor<strong>et</strong>ical st<strong>at</strong>istics of the ω and Ω-tests (as well as the Kolmogorov test). Therefore, we cannot<br />
directly use the results of these standard st<strong>at</strong>istical tests. As a remedy, we propose a bootstrap m<strong>et</strong>hod<br />
(Efron and Tibshirani 1986), whose accuracy is proved by (Chen and Lo 1997) to be <strong>at</strong> least as good<br />
as th<strong>at</strong> given by asymptotic m<strong>et</strong>hods used to <strong>de</strong>rive the theor<strong>et</strong>ical distributions. For the present work,<br />
we have d<strong>et</strong>ermined th<strong>at</strong> the gener<strong>at</strong>ion of 10,000 synth<strong>et</strong>ic time series was sufficient to obtain a good<br />
approxim<strong>at</strong>ion of the distribution of distances <strong>de</strong>scribed above. Since a bootstrap m<strong>et</strong>hod is nee<strong>de</strong>d to<br />
d<strong>et</strong>ermine the tests st<strong>at</strong>istics in every case, it is convenient to choose functional forms different from the<br />
usual ones in the ω and Ω-tests as they provi<strong>de</strong> an improvement with respect to st<strong>at</strong>istical reliability, as<br />
obtained with the d2 and d4 distances introduced here.<br />
To summarize, our test procedure is as follows.<br />
1. Given the original time series x(t), t ∈ {1, · · · , T }, we gener<strong>at</strong>e the Gaussian variables ˆy(t),<br />
t ∈ {1, · · · , T }.<br />
2. We then estim<strong>at</strong>e the covariance m<strong>at</strong>rix ˆρ of the Gaussian variables ˆy, which allows us to compute<br />
the variables ˆz 2 and then measure the distance of its estim<strong>at</strong>ed distribution to the χ 2 -distribution.<br />
3. Given this covariance m<strong>at</strong>rix ˆρ, we gener<strong>at</strong>e numerically a time series of T Gaussian random<br />
vectors with the same covariance m<strong>at</strong>rix ˆρ.<br />
4. For the time series of Gaussian vectors synth<strong>et</strong>ically gener<strong>at</strong>ed with covariance m<strong>at</strong>rix ˆρ, we estim<strong>at</strong>e<br />
its sample covariance m<strong>at</strong>rix ˜ρ.<br />
5. To each of the T vectors of the synth<strong>et</strong>ic Gaussian time series, we associ<strong>at</strong>e the corresponding<br />
realiz<strong>at</strong>ion of the random variable z 2 , called ˜z 2 (t).<br />
6. We can then construct the empirical distribution for the variable ˜z 2 and measure the distance<br />
b<strong>et</strong>ween this distribution and the χ 2 -distribution.<br />
7. Repe<strong>at</strong>ing 10,000 times the steps 3 to 6, we obtain an accur<strong>at</strong>e estim<strong>at</strong>e of the cumul<strong>at</strong>ive distribution<br />
of distances b<strong>et</strong>ween the distribution of the synth<strong>et</strong>ic Gaussian variables and the theor<strong>et</strong>ical<br />
χ 2 -distribution.<br />
8. Then, the distance obtained <strong>at</strong> step 2 for the true variables can be transformed into a significance<br />
level by reading the value of this synth<strong>et</strong>ically d<strong>et</strong>ermined distribution of distances b<strong>et</strong>ween the<br />
distribution of the synth<strong>et</strong>ic Gaussian variables and the theor<strong>et</strong>ical χ 2 -distribution as a function<br />
of the distance: this provi<strong>de</strong>s the probability to observe a distance smaller than the chosen or<br />
empirically d<strong>et</strong>ermined distance.<br />
3.3 Sensitivity of the m<strong>et</strong>hod<br />
Before presenting the st<strong>at</strong>istical tests, it is important to investig<strong>at</strong>e the sensitivity of our testing procedure.<br />
More precisely, can we distinguish for instance b<strong>et</strong>ween a Gaussian copula and a Stu<strong>de</strong>nt’s copula with<br />
11<br />
203
204 8. Tests <strong>de</strong> copule gaussienne<br />
a large number of <strong>de</strong>grees of freedom, for a given value of the correl<strong>at</strong>ion coefficient? Formaly, <strong>de</strong>noting<br />
by Hν the hypothesis according to which the true copula of the d<strong>at</strong>a is the Stu<strong>de</strong>nt’s copula with ν <strong>de</strong>grees<br />
of freedom, we want to d<strong>et</strong>ermine the minimum significance level allowing us to distinguish b<strong>et</strong>ween H0<br />
and Hν.<br />
3.3.1 Importance of the distinction b<strong>et</strong>ween Gaussian and Stu<strong>de</strong>nt’s copulas<br />
This question has important practical implic<strong>at</strong>ions because, as discussed in section 2.4, the Stu<strong>de</strong>nt’s<br />
copula presents a significant tail <strong>de</strong>pen<strong>de</strong>nce while the Gaussian copula has no asymptotic tail <strong>de</strong>pen<strong>de</strong>nce.<br />
Therefore, if our tests are unable to distinguish b<strong>et</strong>ween a Stu<strong>de</strong>nt’s and a Gaussian copula,<br />
we may be led to choose the l<strong>at</strong>er for the sake of simplicity and parsimony and, as a consequence, we<br />
may un<strong>de</strong>restim<strong>at</strong>e severely the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween extreme events if the correct <strong>de</strong>scription turns out<br />
to be the Stu<strong>de</strong>nt’s copula. This may have c<strong>at</strong>astrophic consequences in risk assessment and portfolio<br />
management.<br />
Figure 1 provi<strong>de</strong>s a quantific<strong>at</strong>ion of the dangers incurred by mistaking a Stu<strong>de</strong>nt’s copula for a<br />
Gaussian one. Consi<strong>de</strong>r the case of a Stu<strong>de</strong>nt’s copula with ν = 20 <strong>de</strong>grees of freedom with a correl<strong>at</strong>ion<br />
coefficient ρ lower than 0.3 ∼ 0.4 ; its tail <strong>de</strong>pen<strong>de</strong>nce λν(ρ) turns out to be less than 0.7%, i.e., the<br />
probability th<strong>at</strong> one variable becomes extreme knowing th<strong>at</strong> the other one is extreme is less than 0.7%.<br />
In this case. the Gaussian copula with zero probability of simultaneous extreme events is not a bad<br />
approxim<strong>at</strong>ion of the Stu<strong>de</strong>nt’s copula. In contrast, l<strong>et</strong> us take a correl<strong>at</strong>ion ρ larger than 0.7 − 0.8 for<br />
which the tail <strong>de</strong>pen<strong>de</strong>nce becomes larger than 10%, corresponding to a non-negligible probability of<br />
simultaneous extreme events. The effect of tail <strong>de</strong>pen<strong>de</strong>nce becomes of course much stronger as the<br />
number ν of <strong>de</strong>grees of freedom <strong>de</strong>creases.<br />
These examples stress the importance of knowing wh<strong>et</strong>her our testing procedure allows us to distinguish<br />
b<strong>et</strong>ween a Stu<strong>de</strong>nt’s copula with ν = 20 (or less) <strong>de</strong>grees of freedom and a given correl<strong>at</strong>ion<br />
coefficient ρ = 0.5, for instance, and a Gaussian copula with an appropri<strong>at</strong>e correl<strong>at</strong>ion coefficient ρ ′ .<br />
3.3.2 St<strong>at</strong>istical test on the distinction b<strong>et</strong>ween Gaussian and Stu<strong>de</strong>nt’s copulas<br />
To address this question, we have gener<strong>at</strong>ed 1,000 pairs of time series of size T = 1250, each pair of<br />
random variables following a Stu<strong>de</strong>nt’s bivari<strong>at</strong>e distribution with ν <strong>de</strong>grees of freedom and a correl<strong>at</strong>ion<br />
coefficient ρ b<strong>et</strong>ween the two simultaneous variables of the same pair, while the variables along the time<br />
axis are all in<strong>de</strong>pen<strong>de</strong>nt. We have then applied the previous testing procedure to each of the pairs of time<br />
series.<br />
Specifically, for each pair of time series, we construct the marginals distributions and transform the<br />
Stu<strong>de</strong>nt’s variables xi(k) into their Gaussian counterparts yi(k) via the transform<strong>at</strong>ion (23). For each<br />
pair (y1(k), y2(k)), k ∈ {1, · · · , T }, we estim<strong>at</strong>e its correl<strong>at</strong>ion m<strong>at</strong>rix, then construct the time series<br />
with T realiz<strong>at</strong>ions of the random variable z 2 (k) <strong>de</strong>fined in (27). The s<strong>et</strong> of T variables z 2 then allows<br />
us to construct the distribution of z 2 (with N = 2) and to compare it with the χ 2 -distribution with two<br />
<strong>de</strong>grees of freedom. We then measure the distances d1, d2, d3 and d4 <strong>de</strong>fined by (33-36) b<strong>et</strong>ween the<br />
distribution of z 2 and the χ 2 -distribution. Using the 1,000 pairs of such time series with the same ν<br />
and ρ, we then construct the distribution Di(di), i ∈ {1, 2, 3, 4} of each of these distances di. Using<br />
the previously d<strong>et</strong>ermined distribution of distances expected for the synth<strong>et</strong>ic Gaussian variables, we can<br />
transl<strong>at</strong>e each distance d obtained for the Stu<strong>de</strong>nt’s vectors into a corresponding Gaussian probability<br />
p: p is the probability th<strong>at</strong> pairs of Gaussian random variables with the correl<strong>at</strong>ion coefficient ρ have<br />
12
a distance equal to or larger than the distance obtained for the Stu<strong>de</strong>nt’s vector time series. A small<br />
p corresponds to a clear distinction b<strong>et</strong>ween Stu<strong>de</strong>nt’s and Gaussian vectors, as it is improbable th<strong>at</strong><br />
Gaussian vectors exhibit a distance larger than found for the Stu<strong>de</strong>nt’s vectors. The “distribution of<br />
probabilities” D(p) ≡ D(p(d)) then assesses how often this “improbable” event occurs among the s<strong>et</strong> of<br />
1,000 Stu<strong>de</strong>nt’s vectors, i.e., <strong>at</strong>tempts to quantify the rar<strong>et</strong>y of such large <strong>de</strong>vi<strong>at</strong>ions. In other words, the<br />
“distribution of probabilities” D(p) gives the number of Stu<strong>de</strong>nt’s vectors th<strong>at</strong> exhibit the value p for the<br />
probability th<strong>at</strong> Gaussian vectors can have a similar or larger distance. Then, fixing a confi<strong>de</strong>nce level<br />
D ∗ , this procedure allows us to reject or not the null hypothesis th<strong>at</strong> the empirical vector of r<strong>et</strong>urns is<br />
<strong>de</strong>scribed by a Gaussian copula: this will occur when the observed p gives a “distribution of probabilities”<br />
D(p) larger than D ∗ .<br />
The “distributions of probabilities” D(p) for each of the four distances di, i ∈ {1, 2, 3, 4} are shown<br />
in figure 3 for ν = 4 <strong>de</strong>grees of freedom and in figure 4 for ν = 20 <strong>de</strong>grees of freedom, for 5 different<br />
values of the correl<strong>at</strong>ion coefficient ρ = 0.1, 0.3, 0.5, 0.7 and 0.9. The very steep increase observed for<br />
almost all cases in figure 3 reflects the fact th<strong>at</strong> most of the 1,000 Stu<strong>de</strong>nt’s vectors with ν = 4 <strong>de</strong>grees of<br />
freedom have a small p, i.e., their copula is easily distinguishable from the Gaussian copula. The same<br />
cannot be st<strong>at</strong>ed for Stu<strong>de</strong>nt’s vectors with ν = 20 <strong>de</strong>grees of freedom. Note also th<strong>at</strong> the distances d1,<br />
d2 and d4 give essentially the same result while the An<strong>de</strong>rson-Darling distance d3 is more sensitive to ρ,<br />
especially for small ν.<br />
Fixing for instance the confi<strong>de</strong>nce level <strong>at</strong> D ∗ = 95%, we can read from each of these curves in<br />
figures 3 and 4 the minimum p 95%-value necessary to distinguish a Stu<strong>de</strong>nt’s copula with a given ν from<br />
a Gaussian copula. This p 95% is the abscissa corresponding to the ordin<strong>at</strong>e D(p 95%) = 0.95. These<br />
values p 95% are reported in table 1, for different values of the number ν of <strong>de</strong>grees of freedom ranging<br />
from ν = 3 to ν = 50 and correl<strong>at</strong>ion coefficients ρ = 0.1 to 0.9. The values of p 95%(ν, ρ) reported<br />
in table 1 are the maximum values th<strong>at</strong> the probability p should take in or<strong>de</strong>r to be able to reject the<br />
hypothesis th<strong>at</strong> a Stu<strong>de</strong>nt’s copula with ν <strong>de</strong>grees and correl<strong>at</strong>ion ρ can be mistaken with a Gaussian<br />
copula <strong>at</strong> the 95% confi<strong>de</strong>nce level.<br />
The results of the table 1 are <strong>de</strong>picted in figures 5-6 and represent the conventional “power/size”<br />
st<strong>at</strong>istics. The st<strong>at</strong>istical “power” is usually <strong>de</strong>fined as the rejection of null hypothesis when false. When<br />
the null hypothesis H0 and the altern<strong>at</strong>ive hypothesis Hν are i<strong>de</strong>ntical, the power should be equal to<br />
= 0.05, corresponding to the 95% confi<strong>de</strong>nce level. In our framework, this amounts to plot the abscissa<br />
as the inverse ν −1 of the number ν of <strong>de</strong>grees of freedom, which provi<strong>de</strong>s a n<strong>at</strong>ural “distance” b<strong>et</strong>ween<br />
the Gaussian copula hypothesis H0 and the Stu<strong>de</strong>nt’s copula hypothesis Hν. In the ordin<strong>at</strong>e, the “power”<br />
is represented by the minimum significance level (1−p 95%) necessary to distinguish b<strong>et</strong>ween H0 and Hν.<br />
The typical shape of these curves is a sigmoid, starting from a very small value for ν −1 → 0, increasing<br />
as ν −1 increases and going to 1 as ν −1 becomes large enough. This typical shape simply expresses the<br />
fact th<strong>at</strong> it is easy to separ<strong>at</strong>e a Gaussian copula from a Stu<strong>de</strong>nt’s copula with a small number of <strong>de</strong>grees<br />
of freedom, while it is difficult and even impossible for too large a number of <strong>de</strong>grees of freedom.<br />
The figure 5 shows us th<strong>at</strong> the distances d1, d2 and d3 are not sensitive to the value of the correl<strong>at</strong>ion<br />
coefficient ρ, while the discrimin<strong>at</strong>ing power of d3 increases with ρ. On figure 6, we note th<strong>at</strong> d2 and d4<br />
have the same discrimin<strong>at</strong>ing power for all ρ’s (which makes them somewh<strong>at</strong> redundant) and th<strong>at</strong> they<br />
are the most efficient to differenti<strong>at</strong>e Hν from H0 for small ρ. When ρ is about 0.5, d2, d3 and d4 (and<br />
maybe d1) are equivalent with respect to the differential power, while for large ρ, d3 becomes the most<br />
discrimin<strong>at</strong>ing one with high significance.<br />
This study of the test sensitivity involves a non-param<strong>et</strong>ric approach and the question may arise why<br />
it should be prefered to a direct param<strong>et</strong>ric test involving for instance the calibr<strong>at</strong>ion of the Stu<strong>de</strong>nt<br />
13<br />
205
206 8. Tests <strong>de</strong> copule gaussienne<br />
copula. First, a param<strong>et</strong>ric test of copulas would face the “curse of dimensionality”, i.e., the estim<strong>at</strong>ion<br />
of functions of several variables. With the limited d<strong>at</strong>a s<strong>et</strong> available, this does not seem a reasonable<br />
approach. Second, we have taken the Stu<strong>de</strong>nt copula as an example of an altern<strong>at</strong>ive to the Gaussian<br />
copula. However, our tests are in<strong>de</strong>pen<strong>de</strong>nt of this choice and aim mainly <strong>at</strong> testing the rejection of<br />
the Gaussian copula hypothesis. They are thus of a more general n<strong>at</strong>ure than would be a param<strong>et</strong>ric<br />
test which would be forced to choose one family of copulas with the problem of excluding others. The<br />
param<strong>et</strong>ric test would then be exposed to the criticism th<strong>at</strong> the rejection of a given choice might not be<br />
of a general n<strong>at</strong>ure.<br />
In the sequel, we will choose the level of 95% as the level of rejection, which leads us to neglect<br />
one extreme event out of twenty. This is not unreasonable in view of the other significant sources of<br />
errors resulting in particular from the empirical d<strong>et</strong>ermin<strong>at</strong>ion of the marginals and from the presence of<br />
outliers for instance.<br />
4 Empirical results<br />
We investig<strong>at</strong>e the following ass<strong>et</strong>s :<br />
• foreign exchange r<strong>at</strong>es,<br />
• m<strong>et</strong>als tra<strong>de</strong>d on the London M<strong>et</strong>al Exchange,<br />
• stocks tra<strong>de</strong>d on the New York Stocks Exchange.<br />
4.1 Currencies<br />
The sample we have consi<strong>de</strong>red is ma<strong>de</strong> of the daily r<strong>et</strong>urns for the spot foreign exchanges for 6 currencies<br />
2 : the Swiss Franc (CHF), the German Mark (DEM), the Japanese Yen (JPY), the Malaysian Ringgit<br />
(MYR), the Thai Baht (THA) and the Bristish Pound (UKP). All the exchange r<strong>at</strong>es are expressed against<br />
the US dollar. The time interval runs over ten years, from January 25, 1989 to December 31, 1998, so<br />
th<strong>at</strong> each sample contains 2500 d<strong>at</strong>a points.<br />
We apply our test procedure to the entire sample and to two sub-samples of 1250 d<strong>at</strong>a points so th<strong>at</strong><br />
the first one covers the time interval from January 25, 1989 to January 11, 1994 and the second one from<br />
January 12, 1994 to December 31, 1998. The results are presented in tables 2 to 4 and <strong>de</strong>picted in figures<br />
7 to 9.<br />
Tables 2-4 give, for the total time interval and for each of the two sub-intervals, the probability p(d) to<br />
obtain from the Gaussian hypothesis a <strong>de</strong>vi<strong>at</strong>ion b<strong>et</strong>ween the distribution of the z 2 and the χ 2 -distribution<br />
with two <strong>de</strong>grees of freedom larger than the observed one for each of the 15 pairs of currencies according<br />
to the distances d1-d4 <strong>de</strong>fined by (33)-(36).<br />
The figures 7-9 organize the inform<strong>at</strong>ion shown in the tables 2-4 by representing, for each distance<br />
d1 to d4, the number of currency pairs th<strong>at</strong> give a test-value p within a bin interval of width 0.05. A<br />
clustering close to the origin signals a significant rejection of the Gaussian copula hypothesis.<br />
At the 95% significance level, table 2 and figure 7 show th<strong>at</strong> only 40% (according to d1 and d3) but<br />
60% (according to d2 and d4) of the tested pairs of currencies are comp<strong>at</strong>ible with the Gaussian copula<br />
2 The d<strong>at</strong>a come from the historical d<strong>at</strong>abase of the Fe<strong>de</strong>ral Reserve Board.<br />
14
hypothesis over the entire time interval. During the first half-period from January 25, 1989 to Januray<br />
11, 1994 (table 3 and figure 8), 47% (according to d3) and up to about 75 % (according to d2 and d4)<br />
of the tested currency pairs are comp<strong>at</strong>ible with the assumption of Gaussian copula, while during the<br />
second sub-period from January 12, 1994 to December 31, 1998 (table 4 and figure 9), b<strong>et</strong>ween 66%<br />
(according to d1) and about 75% (according to d2, d3 and d4) of the currency pairs remain comp<strong>at</strong>ible<br />
with the Gaussian copula hypothesis. These results raise several comments both on a st<strong>at</strong>istical and an<br />
economic point of view.<br />
We first note th<strong>at</strong> the most significant rejection of the Gaussian copula hypothesis is obtained for<br />
the distance d3, which is in<strong>de</strong>ed the most sensitive to the events in the tail of the distributions. The test<br />
st<strong>at</strong>istics given by this distance can in<strong>de</strong>ed be very sensitive to the presence of a single large event in the<br />
sample, so much so th<strong>at</strong> the Gaussian copula hypothesis can be rejected only because of the presence<br />
of this single event (outlier). The difference b<strong>et</strong>ween the results given by d3 and d4 (the averaged d3)<br />
are very significant in this respect. Consi<strong>de</strong>r for instance the case of the German Mark and the Swiss<br />
Franc. During the time interval from January 12, 1994 to December 31, 1998, we check on table 4 th<strong>at</strong><br />
the non-rejection probability p(d) is very significant according to d1, d2 and d4 (p(d) ≥ 31%) while<br />
it is very low according to d3: p(d) = 0.05%, and should lead to the rejection of the Gaussian copula<br />
hypothesis. This suggests the presence of an outlier in the sample.<br />
To check this hypothesis, we show in the upper panel of figure 10 the function<br />
f3(t) = |Fz2(z2 (t) − Fχ2(χ2 (t))|<br />
<br />
Fχ2(χ2 )[1 − Fχ2(χ2 )]<br />
207<br />
, (37)<br />
used in the <strong>de</strong>finition of the An<strong>de</strong>rson-Darling distance d3 = maxz f3(z) (see <strong>de</strong>finition (35)), expressed<br />
in terms of time t r<strong>at</strong>her than z 2 . The function have been computed over the two time sub-intervals<br />
separ<strong>at</strong>ely.<br />
Apart from three extreme peaks occurring on June 20, 1989, August 19, 1991 and September 16,<br />
1992 during the first time sub-interval and one extreme peak on September 10, 1997 during the second<br />
time sub-interval, the st<strong>at</strong>istical fluctu<strong>at</strong>ions measured by f3(t) remain small and of the same or<strong>de</strong>r.<br />
Excluding the contribution of these outlier events to d3, the new st<strong>at</strong>istical significance <strong>de</strong>rived according<br />
to d3 becomes similar to th<strong>at</strong> obtained with d1, d2 and d4 on each sub-interval. From the upper pannel<br />
of figure 10, it is clear th<strong>at</strong> the An<strong>de</strong>rson-Darling distance d3 is equal to the height of the largest peak<br />
corresponding to the event on August 19, 1991 for the the first period and to the event on September 10,<br />
1997 for the second period. These events are <strong>de</strong>picted by a circled dot in the two lower panels of figure<br />
10, which represent the r<strong>et</strong>urn of the German Mark versus the r<strong>et</strong>urn of the Swiss Franc over the two<br />
consi<strong>de</strong>red time periods.<br />
The event on August 19, 1991 is associ<strong>at</strong>ed with the coup against Gorbachev in Moscow: the German<br />
mark (respectively the Swiss franc) lost 3.37% (respectively 0.74%) in daily annualized value against the<br />
US dollar. The 3.37% drop of the German Mark is the largest daily move of this currency against the<br />
US dollar over the whole first period. On September 10, 1997, the German Mark appreci<strong>at</strong>ed by 0.60%<br />
against the US dollar while the Swiss Franc lost 0.79% which represents a mo<strong>de</strong>r<strong>at</strong>e move for each<br />
currency, but a large joint move. This event is rel<strong>at</strong>ed to the contradictory announcements of the Swiss<br />
N<strong>at</strong>ional Bank about the mon<strong>et</strong>ary policy, which put an end to a rally of the Swiss Franc along with the<br />
German mark against the US dollar.<br />
Thus, neglecting the large moves associ<strong>at</strong>ed with major historical events or events associ<strong>at</strong>ed with<br />
unexpected incoming inform<strong>at</strong>ion, which cannot be taken into account by a st<strong>at</strong>istical study, we obtain,<br />
for d3, significance levels comp<strong>at</strong>ible with those obtained with the other distances. We can thus conclu<strong>de</strong><br />
15
208 8. Tests <strong>de</strong> copule gaussienne<br />
th<strong>at</strong>, according to the four distances, during the time interval from January 12, 1994 to December 31,<br />
1998 the Gaussian copula hypothesis cannot be rejected for the couple German Mark / Swiss Franc.<br />
However, the non-rejection of the Gaussian copula hypothesis does not always have minor consequences<br />
and may even lead to serious problem in stress scenarios. As shown in section 3.3, the nonrejection<br />
of the Gaussian copula hypothesis does not exclu<strong>de</strong>, <strong>at</strong> the 95% significance level, th<strong>at</strong> the<br />
<strong>de</strong>pen<strong>de</strong>nce of the currency pairs may be accounted for by a Stu<strong>de</strong>nt’s copula with a<strong>de</strong>qu<strong>at</strong>e values of ν<br />
and ρ. Still consi<strong>de</strong>ring the pair German Mark / Swiss Franc, we see in table 1 th<strong>at</strong>, according to d1, d2<br />
and d4, a Stu<strong>de</strong>nt’s copula with about five <strong>de</strong>grees of freedom allows to reach the test values given in table<br />
4. But, with the correl<strong>at</strong>ion coefficient ρ = 0.92 for the German Mark/Swiss Franc couple, the Gaussian<br />
copula assumption could lead to neglect a tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ5(0.92) = 63% according to the<br />
Stu<strong>de</strong>nt’s copula prediction. Such a large value of λ5(0.92) means th<strong>at</strong> when an extreme event occurs<br />
for the German Mark it also occurs for the Swiss Franc with a probabilty equals to 0.63. Therefore, a<br />
stress scenario based on a Gaussian copula assumption would fail to account for such coupled extreme<br />
events, which may represent as many as two third of all the extreme events, if it would turn out th<strong>at</strong> the<br />
true copula would be the Stu<strong>de</strong>nt’s copula with five <strong>de</strong>grees of freedom. In fact, with such a value of the<br />
correl<strong>at</strong>ion coefficient, the tail <strong>de</strong>pen<strong>de</strong>nce remains high even if the number of <strong>de</strong>grees of fredom reach<br />
twenty or more (see figure 1).<br />
The case of the Swiss Franc and the Malaysian Ringgit offers a striking difference. For instance,<br />
in the second half-period, the test st<strong>at</strong>istics p(d) are gre<strong>at</strong>er than 70% and even reach 91% while the<br />
correl<strong>at</strong>ion coefficient is only ρ = 0.16, so th<strong>at</strong> a Stu<strong>de</strong>nt’s copula with 7-10 <strong>de</strong>grees of freedom can be<br />
mistaken with the Gaussian copula (see table 1). Even in the most pessimistic situ<strong>at</strong>ion ν = 7, the choice<br />
of the Gaussian copula amounts to neglecting a tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ5(0.16) = 4% predicted by<br />
the Stu<strong>de</strong>nt’s copula. In this case, stress scenarios based on the Gaussian copula would predict uncoupled<br />
extreme events, which would be shown wrong only once out of twenty five times.<br />
These two examples show th<strong>at</strong>, more than the number of <strong>de</strong>grees of freedom of the Stu<strong>de</strong>nt’s copula<br />
necessary to <strong>de</strong>scribe the d<strong>at</strong>a, the key param<strong>et</strong>er is the correl<strong>at</strong>ion coefficient.<br />
¿From an economic point of view, the impact of regul<strong>at</strong>ory mechanisms b<strong>et</strong>ween currencies or mon<strong>et</strong>ary<br />
crisis can be well i<strong>de</strong>ntified by the rejection or absence of rejection of our null hypothesis. In<strong>de</strong>ed,<br />
consi<strong>de</strong>r the couple German Mark / British Pound. During the first half period, their correl<strong>at</strong>ion coefficient<br />
is very high (ρ = 0.82) and the Gaussian copula hypothesis is strongly rejected according to the<br />
four distances. On the contrary, during the second half period, the correl<strong>at</strong>ion coefficient significantly<br />
<strong>de</strong>creases (ρ = 0.56) and none of the four distances allows us to reject our null hypothesis. Such a<br />
non-st<strong>at</strong>ionarity can be easily explained. In<strong>de</strong>ed, on January 1, 1990, the British Pound entered the European<br />
Mon<strong>et</strong>ary System (EMS), so th<strong>at</strong> the exchange r<strong>at</strong>e b<strong>et</strong>ween the German Mark and the Bristish<br />
Pound was not allowed to fluctu<strong>at</strong>e beyond a margin of 2.25%. However, due to a strong specul<strong>at</strong>ive<br />
<strong>at</strong>tack, the British Pound was <strong>de</strong>valu<strong>at</strong>ed on September 1992 and had to leave the EMS. Thus, b<strong>et</strong>ween<br />
January 1990 and September 1992, the exchange r<strong>at</strong>e of the German Mark and the British Pound was<br />
confined within a narrow spread, incomp<strong>at</strong>ible with the Gaussian copula <strong>de</strong>scription. After 1992, the<br />
British Pound exchange r<strong>at</strong>e flo<strong>at</strong>ed with respect to German Mark, the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two currencies<br />
<strong>de</strong>creased, as shown by their correl<strong>at</strong>ion coefficient. In this regime, we can no more reject the<br />
Gaussian copula hypothesis.<br />
The impact of major crisis on the copula can be also clearly i<strong>de</strong>ntified. Such a case is exhibited by the<br />
couple Malaysian Ringgit/Thai Baht. In<strong>de</strong>ed, during the period from Januray 1989 to January 1994, these<br />
two currencies have only un<strong>de</strong>rgone mo<strong>de</strong>r<strong>at</strong>e and weakly correl<strong>at</strong>ed (ρ = 0.29) fluctu<strong>at</strong>ions, so th<strong>at</strong> our<br />
null hypothesis cannot be rejected <strong>at</strong> the 95% significance level. On the contrary, during the period from<br />
16
January 1994 to October 1998, the Gaussian copula hypothesis is strongly rejected. This rejection is<br />
obviously due to the persistent and <strong>de</strong>pen<strong>de</strong>nt (ρ = 0.44) shocks incured by the Asian financial and<br />
mon<strong>et</strong>ary mark<strong>et</strong>s during the seven months of the Asian Crisis from July 1997 to January 1998 (Baig and<br />
Goldfajn 1998, Kaminsky and Schlmukler 1999).<br />
These two cases show th<strong>at</strong> the Gaussian copula hypothesis can be consi<strong>de</strong>red reasonable for currencies<br />
in absence of regul<strong>at</strong>ory mechanisms and of strong and persistent crises. They also allows us to<br />
un<strong>de</strong>rstand why the results of the test over the entire sample are so much weaker than the results obtained<br />
for the two sub-intervals: the time series are strongly non-st<strong>at</strong>ionnary.<br />
4.2 Commodities: m<strong>et</strong>als<br />
We consi<strong>de</strong>r a s<strong>et</strong> of 6 m<strong>et</strong>als tra<strong>de</strong>d on the London M<strong>et</strong>al Exchange: aluminium, copper, lead, nickel,<br />
tin and zinc. Each sample contains 2270 d<strong>at</strong>a points and covers the time interval from January 4, 1989<br />
to December 30, 1997. The results are synth<strong>et</strong>ized in table 5 and in figure 11.<br />
Table 5 gives, for each of the 15 pairs of commodities, the probability p(d) to obtain from the Gaussian<br />
hypothesis a <strong>de</strong>vi<strong>at</strong>ion b<strong>et</strong>ween the distribution of the z 2 and the χ 2 -distribution with two <strong>de</strong>grees<br />
of freedom larger than the observed one for the commodity pair according to the distances d1-d4 <strong>de</strong>fined<br />
by (33)-(36).<br />
The figure 11 organizes the inform<strong>at</strong>ion shown in table 5 by representing, for each distance, the<br />
number of commodity pairs th<strong>at</strong> give a test-value p within a bin interval of width 0.05. A clustering<br />
close to the origin signals a significant rejection of the Gaussian copula hypothesis.<br />
According to the three distances d1, d2 and d4, <strong>at</strong> least two third and up to 93% of the s<strong>et</strong> of 15<br />
pairs of commodities are inconsistent with the Gaussian copula hypothesis. Surprisingly, according to<br />
the distance d3, <strong>at</strong> the 95% significance level, two third of the s<strong>et</strong> of 15 pairs of commodities remain<br />
comp<strong>at</strong>ible with the Gaussian copula hypothesis. This is the reverse to the previous situ<strong>at</strong>ion found for<br />
currencies. These test values lead to globally reject the Gaussian copula hypothesis.<br />
Moreover, the largest value obtained for the distance d3 is p = 65% for the pair copper-tin, which is<br />
significantly smaller than the 80% or 90% reached for some currencies over a similar time interval. Thus,<br />
even in the few cases where the Gaussian copula assumption is not rejected, the test values obtained are<br />
not really sufficient to distinguish b<strong>et</strong>ween the Gaussian copula and a Stu<strong>de</strong>nt’s copula with ν = 5 ∼ 6<br />
<strong>de</strong>grees of freedom. In such a case, with correl<strong>at</strong>ion coefficients ranging b<strong>et</strong>ween 0.31 and 0.46, the<br />
tail <strong>de</strong>pen<strong>de</strong>nce neglected by keeping the Gaussian copula is no less than 10% and can reach 15%. One<br />
extreme event out of seven or ten might occur simultaneously on both marginals, which would be missed<br />
by the Gaussian copula.<br />
To summarize, the Gaussian copula does not seem a reasonnable assumption for m<strong>et</strong>als, and it has<br />
not appeared necessary to test these d<strong>at</strong>a over smaller time interval.<br />
4.3 Stocks<br />
We now study the daily r<strong>et</strong>urns distibutions for 22 stocks among the largest compagnies quoted on the<br />
New York Stock Exchange 3 : Appl. M<strong>at</strong>erials (AMAT), AT&T (T), Citigroup (C), Coca Cola (KO),<br />
EMC, Exxon-Mobil (XOM), Ford (F), General Electric (GE), General Motors (GM), Hewl<strong>et</strong>t Packard<br />
3 The d<strong>at</strong>a come from the Center for Research in Security Prices (CRSP) d<strong>at</strong>abase.<br />
17<br />
209
210 8. Tests <strong>de</strong> copule gaussienne<br />
(HPW), IBM, Intel (INTC), MCI WorldCom (WCOM), Medtronic (MDT), Merck (MRK), Microsoft<br />
(MSFT), Pfizer (PFE), Procter&Gamble (PG), SBC Communic<strong>at</strong>ion (SBC), Sun Microsystem (SUNW),<br />
Texas Instruments (TXN), Wal Mart (WMT).<br />
Each sample contains 2500 d<strong>at</strong>a points and covers the time interval from February 8, 1991 to December<br />
29, 2000 and have been divi<strong>de</strong>d into two sub-samples of 1250 d<strong>at</strong>a points, so th<strong>at</strong> the first one<br />
covers the time interval from February 8, 1991 to January 18, 1996 and the second one from January<br />
19, 1996 to December 20, 2000. The results of fifteen randomly chosen pairs of ass<strong>et</strong>s are presented in<br />
tables 6 to 8 while the results obtain for the entire s<strong>et</strong> are represented in figures 12 to 14.<br />
At the 95% significance level, figure 12 shows th<strong>at</strong> 75% of the pairs of stocks are comp<strong>at</strong>ible with the<br />
Gaussian copula hypothesis. Figure 13 shows th<strong>at</strong> over the time interval from February 1991 to January<br />
1996, this percentage becomes larger than 99% for d1, d2 and d4 while it equals 94% according to d3. It<br />
is striking to note th<strong>at</strong>, during this period, according to d1, d2 and d4, more than a quarter of the stocks<br />
obtain a test-value p larger than 90%, so th<strong>at</strong> we can assert th<strong>at</strong> they are compl<strong>et</strong>ely inconsistent with the<br />
Stu<strong>de</strong>nt’s copula hypothesis for Stu<strong>de</strong>nt’s copulas with less than 10 <strong>de</strong>grees of freedom. Among this s<strong>et</strong><br />
of stocks, not a single one has a correl<strong>at</strong>ion coefficient larger than 0.4, so th<strong>at</strong> a scenario based on the<br />
Gaussian copula hypothesis leads to neglecting a tail <strong>de</strong>pen<strong>de</strong>nce of less than 5% as would be predicted<br />
by the Stu<strong>de</strong>nt’s copula with 10 <strong>de</strong>grees of freedom. In addition, about 80% of the pairs of stocks lead<br />
to a test-value p larger than 50% according to the distances d1, d2 and d4, so th<strong>at</strong> as much as 80% of<br />
the pairs of stocks are incomp<strong>at</strong>ible with a Stu<strong>de</strong>nt’s copula with a number of <strong>de</strong>grees of freedom less<br />
than or equal to 5. Thus, for correl<strong>at</strong>ion coefficients smaller than 0.3, the Gaussian copula hypothesis<br />
leads to neglecting a tail <strong>de</strong>pen<strong>de</strong>nce less than 10%. For correl<strong>at</strong>ion coefficients smaller than 0.1 which<br />
corresponds to 13% of the total number of pairs, the Gaussian copula hypothesis leads to neglecting a<br />
tail <strong>de</strong>pen<strong>de</strong>nce less than 5%.<br />
Figure 14 shows th<strong>at</strong>, over the time interval from January 1996 to December 2000, 92% of the pairs<br />
of stocks are comp<strong>at</strong>ible with the Gaussian copula hypothesis according to d1, d2 and d4 and more than<br />
79% according to d3. About a quarter of the pair of stocks have a test-value p larger than 50% according<br />
to the four measures and thus are inconsistent with a Stu<strong>de</strong>nt’s copula with less than five <strong>de</strong>grees of<br />
freedom.<br />
For compl<strong>et</strong>eness, we present in table 9 the results of the tests performed for five stocks belonging<br />
to the computer area : Hewl<strong>et</strong>t Packard, IBM, Intel, Microsoft and Sun Microsystem. We observe th<strong>at</strong>,<br />
during the first half period, all the pairs of stocks qualify the Gaussian copula Hypothesis <strong>at</strong> the 95%<br />
significance level. The results are r<strong>at</strong>her different for the second half period since about 40% of the pairs<br />
of stocks reject the Gaussian copula hypothesis according to d1, d2 and d3. This is probably due to the<br />
existence of a few shocks, notably associ<strong>at</strong>ed with the crash of the “new economy” in March-April 2000.<br />
On the whole, it appears however th<strong>at</strong> there is no system<strong>at</strong>ic rejection of the Gaussian copula hypothesis<br />
for stocks within the same industrial area, notwithstanding the fact th<strong>at</strong> one can expect stronger<br />
correl<strong>at</strong>ions b<strong>et</strong>ween such stocks than for currencies for instance.<br />
5 Conclusion<br />
We have studied the null hypothesis th<strong>at</strong> the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween pairs of financial ass<strong>et</strong>s can be mo<strong>de</strong>led<br />
by the Gaussian copula.<br />
Our test procedure is based on the following simple i<strong>de</strong>a. Assuming th<strong>at</strong> the copula of two ass<strong>et</strong>s<br />
18
X and Y is Gaussian, then the multivari<strong>at</strong>e distribution of (X, Y ) can be mapped into a Gaussian multivari<strong>at</strong>e<br />
distribution, by a transform<strong>at</strong>ion of each marginal into a normal distribution, which leaves the<br />
copula of X and Y unchanged. Testing the Gaussian copula hypothesis is therefore equivalent to the<br />
more standard problem of testing a two-dimensional multivari<strong>at</strong>e Gaussian distribution. We have used<br />
a bootstrap m<strong>et</strong>hod to d<strong>et</strong>ermine and calibr<strong>at</strong>e the test st<strong>at</strong>istics. Four different measures of distances<br />
b<strong>et</strong>ween distributions, more or less sensitive to the <strong>de</strong>parture in the bulk or in the tail of distributions,<br />
have been proposed to quantify the probability of rejection of our null hypothesis.<br />
Our tests have been performed over three types of ass<strong>et</strong>s: currencies, commodities (m<strong>et</strong>als) and<br />
stocks. In most cases, for currencies and stocks, the Gaussian copula hypothesis can not be rejected <strong>at</strong><br />
the 95% confi<strong>de</strong>nce level. For currencies, according to three of the four distances <strong>at</strong> least,<br />
• 40% of the pairs of currencies, over a 10 years time interval (due to non-st<strong>at</strong>ionnary d<strong>at</strong>a),<br />
• 67% of the pairs of currencies, over the first 5 years time interval,<br />
• 73% of the pairs of currencies, over the second 5 years time interval,<br />
are comp<strong>at</strong>ible with the Gaussian copula hypothesis. For stocks, we have shown th<strong>at</strong><br />
• 75% of the pairs of stocks, over a 10 years time interval,<br />
• 93% of the pairs of stocks, over the first 5 years time interval,<br />
• 92% of the pairs of stocks, over the second 5 years time interval,<br />
are comp<strong>at</strong>ible with the Gaussian copula hypothesis. In constrast, the Gaussian copula hypothesis cannot<br />
be consi<strong>de</strong>red as reasonable for m<strong>et</strong>als : b<strong>et</strong>ween 66% and 93% of the pairs of m<strong>et</strong>als reject the null<br />
hypothesis <strong>at</strong> the 95% confi<strong>de</strong>nce level.<br />
Notwithstanding the apparent qualific<strong>at</strong>ion of the Gaussian copula hypothesis for most of the currencies<br />
and the stocks we have analyzed, we must bear in mind the fact th<strong>at</strong> a non-Gaussian copula cannot<br />
be rejected. In particular, we have shown th<strong>at</strong> a Stu<strong>de</strong>nt’s copula can always be mistaken for a Gaussian<br />
copula if its number of <strong>de</strong>grees of freedom is sufficiently large. Then, <strong>de</strong>pending on the correl<strong>at</strong>ion coefficient,<br />
the Stu<strong>de</strong>nt’s copula can predict a non-negligible tail <strong>de</strong>pen<strong>de</strong>nce which is compl<strong>et</strong>ely missed<br />
by the Gaussian copula assumption. In other words, the Gaussian copula predicts no tail <strong>de</strong>pen<strong>de</strong>nces<br />
and therefore does not account for extreme events th<strong>at</strong> may occur simultaneously but nevertheless too<br />
rarely to modify the test st<strong>at</strong>istics. To quantify the probability for neglecting such events, we have investig<strong>at</strong>ed<br />
the situ<strong>at</strong>ions when one is unable to distinguish b<strong>et</strong>ween the Gaussian and Stu<strong>de</strong>nt’s copulas for<br />
a given number of <strong>de</strong>grees of freedom. Our study leads to the conclusion th<strong>at</strong> it may be very dangerous<br />
to embrace blindly the Gaussian copula hypothesis when the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the pair of<br />
ass<strong>et</strong> is too high as the tail <strong>de</strong>pen<strong>de</strong>nce neglected by the Gaussian copula can be as large as 0.6. In this<br />
respect, the case of the Swiss Franc and the German Mark is striking. The test values p obtained are very<br />
significant (about 33%), so th<strong>at</strong> we cannot mistake the Gaussian copula for a Stu<strong>de</strong>nt’s copula with less<br />
than 5-7 <strong>de</strong>grees of freedom. However, their correl<strong>at</strong>ion coefficient is so high (ρ = 0.9) th<strong>at</strong> a Stu<strong>de</strong>nt’s<br />
copula with, say ν = 30 <strong>de</strong>grees of freedom, still has a large tail <strong>de</strong>pen<strong>de</strong>nce.<br />
This remark shows th<strong>at</strong> it is highly <strong>de</strong>sirable to <strong>de</strong>velop tests th<strong>at</strong> are specific to the d<strong>et</strong>ection of a<br />
possible tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two time series. This task is very difficult but we hope to report useful<br />
progress in the near future. Another approach is to test for other non-Gaussian copulas, such as the<br />
Stu<strong>de</strong>nt’s copula.<br />
19<br />
211
212 8. Tests <strong>de</strong> copule gaussienne<br />
References<br />
An<strong>de</strong>rsen, J.V. and D. Sorn<strong>et</strong>te, 2001, Have your cake and e<strong>at</strong> it too: increasing r<strong>et</strong>urns while lowering<br />
large risks! Journal of Risk Finance 2, 70-82.<br />
An<strong>de</strong>rson T. W. and D.A. Darling, 1952, Asymptotic theory of certain “goodness of fit” criteria, Annals<br />
of M<strong>at</strong>hem<strong>at</strong>ical St<strong>at</strong>istics 23, 193-212.<br />
Ang, A. and J. Chen, 2001, Asymm<strong>et</strong>ric correl<strong>at</strong>ions of equity portfolio, Working Paper.<br />
Baig, T. and I. Goldfjn, 1998, Financial Mark<strong>et</strong>s Contagion in the Asian crisis, Mimeo, Intern<strong>at</strong>ional<br />
Mon<strong>et</strong>ary Fund.<br />
Chen, K. and S.-H. Lo, 1997, On a Mapping Approach to Investig<strong>at</strong>ing the Bootstrap Accuracy, Probability<br />
Theory and Rel<strong>at</strong>ed Fields 107, 197-217.<br />
Cherubini, U. and E. Luciano, 2000, Multivari<strong>at</strong>e option pricing with copulas, Working Paper.<br />
Efron, B., and R.Tibshirani, 1986, Bootstrap m<strong>et</strong>hod for standard errors, confi<strong>de</strong>nce intervals and other<br />
measures of st<strong>at</strong>istical accuracy, St<strong>at</strong>istical Science 1, 54-77.<br />
Embrechs P., A.J. McNeil and D. Straumann, 1999, Correl<strong>at</strong>ion : Pitfalls and Altern<strong>at</strong>ives. Risk, 69-71.<br />
2001. Embrechts, P., A. Hoeing, A. Juri, 2001, Using Copulae to bound the Value-<strong>at</strong>-Risk for functions<br />
of <strong>de</strong>pen<strong>de</strong>nt risk, ETH Zurick Working Paper.<br />
Embrechts P., A.J. McNeil and D. Straumann, 2001, Correl<strong>at</strong>ion and Depen<strong>de</strong>ncy in Risk Management :<br />
Properties and Pitfalls, in : Dempster, M., ed., Value <strong>at</strong> Risk and Beyond (Cambridge University<br />
Press).<br />
Frees, E. and E. Val<strong>de</strong>z, 1998, Un<strong>de</strong>rstanding Rel<strong>at</strong>ionships using copula,. North Americam Actuarial<br />
Journal 2, 1-25.<br />
Genest, C. and R. MacKay, 1986, The Joy of Copulas, American St<strong>at</strong>istician 40, 280-283.<br />
Genest, C., 1987, Frank’s Family of Bivari<strong>at</strong>e Distributions, Biom<strong>et</strong>rika 74, 549-555.<br />
Genest, C. and J.P. Rivest, 1993, St<strong>at</strong>istical Inference Procedures for Bivari<strong>at</strong>e Archime<strong>de</strong>an Copulas,<br />
Journal of American St<strong>at</strong>istical Associ<strong>at</strong>ion 88, 1034-1043.<br />
Gopikrishnan, P., M. Meyer, L.A.N. Amaral and H.E. Stanley, 1998, Inverse Cubic Law for the Distribution<br />
of Stock Price Vari<strong>at</strong>ion, European Physical Journal B 3, 139-140.<br />
Gouriéroux, C. and J. Jasiak, 1999, Trunc<strong>at</strong>ed Local Likelyhood and Non-paaram<strong>et</strong>ric tail analysis, DP<br />
99 CREST.<br />
Guillaume, D.M., M.M Dacorogna, R.R. Davé, J.A Muller, R.B. Olsen and O.V. Pict<strong>et</strong>, 1997, From<br />
the Bird Eye to the Microscope : A Survey of the New Stylized Facts of the Intra-Dayly Foreign<br />
Exchange Mark<strong>et</strong>s, Finance and Stochastics 1, 95-129.<br />
Haas, C.N., 1999, On Mo<strong>de</strong>ling Correl<strong>at</strong>ed Random Variables in Risk Assessment, Risk Analysis 19,<br />
1205-1214.<br />
Joe, H., 1993, Param<strong>et</strong>ric Families of Multivari<strong>at</strong>e Distributions with Given Marginals, Journal of Multivari<strong>at</strong>e<br />
Analysis 46, 262-282.<br />
20
Kaminsky, G.L. and S.L. Schmukler, 1999, Wh<strong>at</strong> Triggers Mark<strong>et</strong> Jitters? A Chronicle of the Asian<br />
Crisis, Journal of Intern<strong>at</strong>ional Money and Finance 18, 537-560.<br />
Klugman, S.A. and R. Parsa, 1999, Fitting Bivari<strong>at</strong>e Loss Distribution with Copulas, Insurance : M<strong>at</strong>hem<strong>at</strong>ics<br />
and Economics 24, 139-148.<br />
Laherrère, J. and D. Sorn<strong>et</strong>te, 1999, Str<strong>et</strong>ched exponential distributions in N<strong>at</strong>ure and Economy: “F<strong>at</strong><br />
tails” with characteristic scales, European Physical Journal B 2, 525-539.<br />
Longin F. and B. Solnik, Extreme Correl<strong>at</strong>ion of Intern<strong>at</strong>ional Equity Mark<strong>et</strong>s, The Journal of Finance<br />
LVI, 649-676 (2001).<br />
Lindskog, F., 1999, Mo<strong>de</strong>lling Depen<strong>de</strong>nce with Copulas,<br />
http : //www.risklab.ch/P apers.html#MT Lindskog<br />
Lux, L., 1996, The Stable Par<strong>et</strong>ian Hypothesis and the Frequency of Large R<strong>et</strong>urns : an Examin<strong>at</strong>ion of<br />
Major German Stocks, Applied Financial Economics 6, 463-475.<br />
Markovitz, H., 1959, Portfolio selection : Efficient diversific<strong>at</strong>ion of investments (John Wiley and Sons,<br />
New York).<br />
Nelsen, R.B., 1998, An Introduction to Copulas. Lectures Notes in st<strong>at</strong>istic 139 (Springer Verlag, New<br />
York).<br />
Pagan, A., 1996, The Econom<strong>et</strong>rics of Financial Mark<strong>et</strong>s, Journal of Empirical Finance 3, 15-102.<br />
Rosenberg, J.V., 1999, Semiparam<strong>et</strong>ric pricing of multivari<strong>at</strong>e contingent claims, NYU-Stern School of<br />
Business Working Paper.<br />
Sorn<strong>et</strong>te, D., P. Simon<strong>et</strong>ti and J. V. An<strong>de</strong>rsen, 2000, φ q -field theory for Portfolio optimiz<strong>at</strong>ion: “f<strong>at</strong> tails”<br />
and non-linear correl<strong>at</strong>ions, Physics Report 335, 19-92.<br />
Sorn<strong>et</strong>te, D., J. V. An<strong>de</strong>rsen and P. Simon<strong>et</strong>ti, 2000, Portfolio Theory for “F<strong>at</strong> Tails”, Intern<strong>at</strong>ional Journal<br />
of Theor<strong>et</strong>ical and Applied Finance 3 , 523-535.<br />
Starica, C., 1999, Multivari<strong>at</strong>e extremes for mo<strong>de</strong>ls with constant conditional correl<strong>at</strong>ions. Journal of<br />
Empirical Finance, 6, 515-553.<br />
Vries, C.G. <strong>de</strong>, 1994, Stylized Facts of Nominal Exchange R<strong>at</strong>e R<strong>et</strong>urns, in : Van <strong>de</strong>r Ploeg, F., ed., The<br />
Handbook of intern<strong>at</strong>ional Macroeconomics (Blackwell, Oxford) 348-389.<br />
21<br />
213
214 8. Tests <strong>de</strong> copule gaussienne<br />
λ ν (ρ)<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
ν=3<br />
ν=5<br />
ν=10<br />
ν=20<br />
ν=50<br />
ν=100<br />
0<br />
−1 −0.8 −0.6 −0.4 −0.2 0<br />
ρ<br />
0.2 0.4 0.6 0.8 1<br />
Figure 1: Upper tail <strong>de</strong>pen<strong>de</strong>nce coefficient λν(ρ) for the Stu<strong>de</strong>nt’s copula with ν <strong>de</strong>grees of freedom as<br />
a function of the correl<strong>at</strong>ion coefficient ρ, for different values of ν.<br />
22
ρ<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
−0.6<br />
−0.8<br />
λ ν (ρ) = 1%<br />
λ ν (ρ) = 2.5%<br />
λ ν (ρ) = 5%<br />
λ ν (ρ) = 10%<br />
−1<br />
0 10 20 30 40 50<br />
ν<br />
60 70 80 90 100<br />
Figure 2: Maximum value of the correl<strong>at</strong>ion coefficient ρ as a function of ν, below which the tail<br />
<strong>de</strong>pen<strong>de</strong>nce λν(ρ) of a Stu<strong>de</strong>nt’ copula is smaller than a given small value, here taken equal to<br />
λν(ρ) = 1%, 2.5%, 5% and 10%. The choice λν(ρ) = 5% for instance corresponds to 1 event in 20<br />
for which the pair of variables are asymptotically coupled. At the 1 − λν(ρ) probability level, values of<br />
λ ≤ λν(ρ) are undistinguishable from 0, which means th<strong>at</strong> the Stu<strong>de</strong>nt’s copula can be approxim<strong>at</strong>ed by<br />
a Gaussian copula.<br />
23<br />
215
216 8. Tests <strong>de</strong> copule gaussienne<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
ρ=0.1<br />
ρ=0.3<br />
0.2<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
ρ=0.1<br />
ρ=0.3<br />
0.2<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
d 1<br />
d 3<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
ρ=0.1<br />
ρ=0.3<br />
0.2<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
ρ=0.1<br />
ρ=0.3<br />
0.2<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
Figure 3: Cumul<strong>at</strong>ive “distribution of probabilities” D(p) ≡ D(p(d)) obtained as the fraction of Stu<strong>de</strong>nt’s<br />
pairs with ν = 4 <strong>de</strong>grees of freedom th<strong>at</strong> exhibit the value p for the probability th<strong>at</strong> Gaussian<br />
vectors can have a similar or larger distance. See the text for a d<strong>et</strong>ailled <strong>de</strong>scription of how D(p) is <strong>de</strong>fined<br />
and constructed. Each panel corresponds to one of the four distances di, i ∈ {1, 2, 3, 4}, <strong>de</strong>fined in<br />
the text by equ<strong>at</strong>ions (33-36). In each panel, we construct the cumul<strong>at</strong>ive “distribution of probabilities”<br />
D(p) for 5 different values of the correl<strong>at</strong>ion coefficient ρ = 0.1, 0.3, 0.5, 0.7 and 0.9 of the Stu<strong>de</strong>nt’s<br />
copula.<br />
24<br />
d 2<br />
d 4
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 1<br />
d 3<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 2<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 4<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
Figure 4: Same as figure 3 for Stu<strong>de</strong>nt’s distributions with ν = 20 <strong>de</strong>grees of freedom.<br />
25<br />
217
218 8. Tests <strong>de</strong> copule gaussienne<br />
ν = 3<br />
ν = 5<br />
ν = 8<br />
ν = 20<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.07 0.08 0.07 0.04 0.07<br />
d2 0.03 0.03 0.07 0.04 0.06<br />
d3 0.22 0.17 0.08 0.03 0.01<br />
d4 0.03 0.03 0.08 0.03 0.04<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.46 0.47 0.46 0.52 0.52<br />
d2 0.36 0.34 0.39 0.44 0.43<br />
d3 0.52 0.54 0.47 0.30 0.14<br />
d4 0.37 0.36 0.43 0.45 0.45<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.85 0.86 0.87 0.88 0.89<br />
d2 0.85 0.84 0.86 0.87 0.88<br />
d3 0.91 0.91 0.91 0.81 0.70<br />
d4 0.86 0.85 0.90 0.89 0.90<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.97 0.99 0.97 0.99 0.99<br />
d2 0.99 0.99 0.97 0.99 0.99<br />
d3 0.99 0.99 0.98 0.99 0.97<br />
d4 0.99 0.99 0.98 0.99 0.99<br />
ν = 4<br />
ν = 7<br />
ν = 10<br />
ν = 50<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.28 0.26 0.32 0.30 0.29<br />
d2 0.18 0.17 0.21 0.21 0.24<br />
d3 0.36 0.33 0.26 0.15 0.03<br />
d4 0.18 0.17 0.23 0.21 0.21<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.78 0.81 0.81 0.81 0.86<br />
d2 0.71 0.78 0.76 0.77 0.82<br />
d3 0.80 0.81 0.82 0.73 0.52<br />
d4 0.75 0.81 0.79 0.80 0.83<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.92 0.93 0.96 0.95 0.94<br />
d2 0.93 0.92 0.95 0.96 0.94<br />
d3 0.96 0.96 0.96 0.95 0.88<br />
d4 0.94 0.94 0.96 0.97 0.95<br />
ρ 0.1 0.3 0.5 0.7 0.9<br />
d1 0.99 0.99 0.99 0.99 0.99<br />
d2 0.99 0.99 0.99 0.99 0.99<br />
d3 0.99 0.99 0.99 0.99 0.99<br />
d4 0.99 0.99 0.99 0.99 0.99<br />
Table 1: The values p 95%(ν, ρ) shown in this table give the maximum values th<strong>at</strong> the probability p should<br />
take in or<strong>de</strong>r to be able to reject the hypothesis th<strong>at</strong> a Stu<strong>de</strong>nt’s copula with ν <strong>de</strong>grees and correl<strong>at</strong>ion<br />
ρ is undistinguishable from a Gaussian copula <strong>at</strong> the 95% confi<strong>de</strong>nce level. p 95% is the abscissa corresponding<br />
to the ordin<strong>at</strong>e D(p 95%) = 0.95 shown in figures 3 and 4. p is the probability th<strong>at</strong> pairs of<br />
Gaussian random variables with the correl<strong>at</strong>ion coefficient ρ have a distance (b<strong>et</strong>ween the distribution of<br />
z 2 and the theor<strong>et</strong>ical χ 2 distribution) equal to or larger than the corresponding distance obtained for the<br />
Stu<strong>de</strong>nt’s vector time series. A small p corresponds to a clear distinction b<strong>et</strong>ween Stu<strong>de</strong>nt’s and Gaussian<br />
vectors, as it is improbable th<strong>at</strong> Gaussian vectors exhibit a distance larger than found for the Stu<strong>de</strong>nt’s<br />
vectors. Different values of the number ν of <strong>de</strong>grees of freedom ranging from ν = 3 to ν = 50 and of<br />
the correl<strong>at</strong>ion coefficient ρ = 0.1 to 0.9 are shown. L<strong>et</strong> us take for instance the example with ν = 4 and<br />
ρ = 0.3. The table indic<strong>at</strong>es th<strong>at</strong> p should be less than about 0.3 (resp. 0.2) according to the distances d1<br />
and d3 (resp. d2 and d4) for being able to distinguish this Stu<strong>de</strong>nt’s copula from the Gaussian copula <strong>at</strong><br />
the 95% confi<strong>de</strong>nce level. This means th<strong>at</strong> less than 20−30% of Gaussian vectors should have a distance<br />
for their z 2 larger than the one found for the Stu<strong>de</strong>nt’s. See text for further explan<strong>at</strong>ions.<br />
26
1−p 95%<br />
1−p 95%<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 1<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 3<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3 0.4<br />
1−p 95%<br />
1−p 95%<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 2<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
d 4<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3<br />
Figure 5: Graph of the minimun significance level (1 − p 95%) necessary to distinguish the Gaussian<br />
copula hypothesis H0 from the hypothesis of a stu<strong>de</strong>nt copula with ν <strong>de</strong>grees of freedom, as a function<br />
of 1/ν, for a given distance di and various correl<strong>at</strong>ion coefficients ρ = 0.1, 0.3, 0.5, 0.7 and 0.9.<br />
27<br />
219
220 8. Tests <strong>de</strong> copule gaussienne<br />
1−p 95%<br />
1−p 95%<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
ρ =0.1 − 0.3<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3 0.4<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
ρ =0.7<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3 0.4<br />
1−p 95%<br />
1−p 95%<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
ρ =0.5<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3 0.4<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
ρ =0.9<br />
0<br />
0 0.1 0.2<br />
1/ν<br />
0.3 0.4<br />
Figure 6: Same as figure 5 but comparing different distances for the same correl<strong>at</strong>ion coefficient ρ.<br />
28
ˆρ d1 d2 d3 d4<br />
CHF DEM 0.92 1.01e-02 6.70e-03 0.00e+00 7.20e-03<br />
CHF JPY 0.53 3.44e-01 2.71e-01 2.32e-02 2.83e-01<br />
CHF MYR 0.23 7.27e-01 8.71e-01 5.77e-01 9.26e-01<br />
CHF THA 0.21 3.08e-02 9.47e-02 3.31e-02 9.52e-02<br />
CHF UKP 0.69 2.80e-03 1.80e-03 6.00e-04 1.30e-03<br />
DEM JPY 0.54 2.26e-02 1.33e-01 1.00e-01 1.51e-01<br />
DEM MYR 0.26 4.25e-01 6.77e-01 6.22e-01 7.35e-01<br />
DEM THA 0.24 6.53e-02 1.35e-01 3.26e-02 1.32e-01<br />
DEM UKP 0.72 1.70e-03 4.00e-04 0.00e+00 4.00e-04<br />
JPY MYR 0.31 2.45e-02 6.34e-02 2.26e-01 6.86e-02<br />
JPY THA 0.34 0.00e+00 0.00e+00 3.24e-02 0.00e+00<br />
JPY UKP 0.41 2.85e-02 3.72e-02 5.22e-02 3.09e-02<br />
MYR THA 0.40 0.00e+00 0.00e+00 2.22e-02 0.00e+00<br />
MYR UKP 0.21 6.94e-01 7.94e-01 6.23e-01 8.31e-01<br />
THA UKP 0.15 5.22e-01 6.23e-01 3.21e-02 7.05e-01<br />
Table 2: Each row gives the st<strong>at</strong>istics of our test for each of the 15 pairs of currencies over a 10 years<br />
time interval from January 25, 1989 to December 31, 1998. The column ˆρ gives the empirical correl<strong>at</strong>ion<br />
coefficient for each pair d<strong>et</strong>ermined as in section 3.1 and <strong>de</strong>fined in (31). The columns d1, d2, d3 and d4<br />
gives the probability to obtain, from the Gaussian hypothesis, a <strong>de</strong>vi<strong>at</strong>ion b<strong>et</strong>ween the distribution of the<br />
z 2 and the χ 2 -distribution with two <strong>de</strong>grees of freedom larger than the observed one for the currency pair<br />
according to the distances d1-d4 <strong>de</strong>fined by (33)-(36).<br />
29<br />
221
222 8. Tests <strong>de</strong> copule gaussienne<br />
ˆρ d1 d2 d3 d4<br />
CHF DEM 0.92 1.73e-02 1.33e-02 0.00e+00 1.31e-02<br />
CHF JPY 0.55 1.34e-01 1.49e-01 3.83e-01 1.41e-01<br />
CHF MYR 0.32 8.47e-01 7.00e-01 3.56e-01 7.40e-01<br />
CHF THA 0.17 4.40e-01 7.10e-01 3.53e-02 7.11e-01<br />
CHF UKP 0.79 3.10e-03 1.00e-03 0.00e+00 5.00e-04<br />
DEM JPY 0.56 2.46e-02 9.43e-02 1.63e-01 9.26e-02<br />
DEM MYR 0.35 9.32e-01 7.95e-01 3.51e-01 7.95e-01<br />
DEM THA 0.21 4.36e-01 8.77e-01 3.47e-02 8.74e-01<br />
DEM UKP 0.82 0.00e+00 0.00e+00 0.00e+00 0.00e+00<br />
JPY MYR 0.34 4.90e-01 5.49e-01 3.66e-01 5.94e-01<br />
JPY THA 0.27 3.89e-01 3.06e-01 3.37e-02 3.59e-01<br />
JPY UKP 0.53 9.00e-04 1.66e-02 6.72e-02 1.67e-02<br />
MYR THA 0.29 1.08e-01 8.71e-02 3.42e-02 9.30e-02<br />
MYR UKP 0.33 1.12e-01 2.86e-01 3.54e-01 3.45e-01<br />
THA UKP 0.21 4.34e-01 8.62e-01 3.13e-02 8.67e-01<br />
Table 3: Same as table 2 for currencies over a 5 years time interval from January 25, 1989 to Januay 11,<br />
1994.<br />
30
ˆρ d1 d2 d3 d4<br />
CHF DEM 0.92 3.15e-01 3.11e-01 5.00e-04 3.41e-01<br />
CHF JPY 0.52 5.84e-01 6.44e-01 1.98e-02 6.74e-01<br />
CHF MYR 0.16 7.11e-01 9.15e-01 8.83e-01 9.22e-01<br />
CHF THA 0.25 1.10e-02 3.87e-02 1.05e-01 3.34e-02<br />
CHF UKP 0.53 9.75e-02 1.03e-01 2.33e-01 9.29e-02<br />
DEM JPY 0.53 3.63e-01 5.40e-01 1.77e-02 6.54e-01<br />
DEM MYR 0.18 3.55e-01 5.00e-01 5.84e-01 5.67e-01<br />
DEM THA 0.28 1.28e-02 2.18e-02 1.08e-01 1.51e-02<br />
DEM UKP 0.56 1.15e-01 1.10e-01 3.02e-01 1.06e-01<br />
JPY MYR 0.29 7.63e-02 2.14e-01 6.67e-02 2.23e-01<br />
JPY THA 0.38 0.00e+00 2.00e-04 3.09e-02 2.00e-04<br />
JPY UKP 0.28 4.62e-01 2.30e-01 1.23e-01 2.07e-01<br />
MYR THA 0.44 5.00e-04 1.20e-03 5.34e-02 1.20e-03<br />
MYR UKP 0.11 5.94e-01 7.44e-01 6.95e-01 7.82e-01<br />
THA UKP 0.12 1.26e-02 7.66e-02 1.19e-01 6.51e-02<br />
Table 4: Same as table 2 for currencies over a 5 years time interval from January 12, 1994 to December<br />
31, 1998.<br />
31<br />
223
224 8. Tests <strong>de</strong> copule gaussienne<br />
9<br />
8<br />
7<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 7: For each distance d1-d4 <strong>de</strong>fined in equ<strong>at</strong>ions (33)-(36), this figure shows the number of currency<br />
pairs th<strong>at</strong> give a given p (shown on the abscissa) within a bin interval of width 0.05 for different currencies<br />
over a 10 years time interval from January 25, 1989 to December 31, 1998. p is the probability th<strong>at</strong><br />
pairs of Gaussian random variables with the same correl<strong>at</strong>ion coefficient ρ have a distance (b<strong>et</strong>ween the<br />
distribution of z 2 and the theor<strong>et</strong>ical chi 2 distribution) equal to or larger than the corresponding distance<br />
obtained for each currency pair. A clustering close to the origin signals a significant rejection of the<br />
Gaussian copula hypothesis.<br />
32<br />
d<br />
1<br />
d<br />
2<br />
d<br />
1<br />
d<br />
4
8<br />
7<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 8: Same as figure 7 for currencies over a 5 years time interval from January 25, 1989 to January<br />
11, 1994.<br />
33<br />
d<br />
1<br />
d<br />
2<br />
d<br />
1<br />
d<br />
4<br />
225
226 8. Tests <strong>de</strong> copule gaussienne<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 9: Same as figure 7 for currencies over a 5 years time interval from January 12, 1994 to December<br />
1998.<br />
34<br />
d<br />
1<br />
d<br />
2<br />
d<br />
1<br />
d<br />
4
ˆρ d1 d2 d3 d4<br />
aluminium copper 0.46 6.46e-02 4.48e-02 1.45e-02 4.00e-02<br />
aluminium lead 0.35 1.14e-01 5.01e-02 1.70e-01 4.59e-02<br />
aluminium nickel 0.36 3.30e-03 5.10e-03 3.41e-02 6.20e-03<br />
aluminium tin 0.34 1.34e-01 1.38e-01 1.25e-02 1.59e-01<br />
aluminium zinc 0.36 2.30e-03 2.20e-03 6.21e-02 2.30e-03<br />
copper lead 0.35 4.71e-02 1.74e-02 1.79e-01 1.34e-02<br />
copper nickel 0.38 4.91e-02 4.60e-02 1.48e-01 3.80e-02<br />
copper tin 0.32 1.94e-01 1.35e-01 6.53e-01 1.47e-01<br />
copper zinc 0.40 3.24e-02 2.05e-02 1.75e-01 1.94e-02<br />
lead nickel 0.32 6.71e-02 3.78e-02 2.74e-01 3.62e-02<br />
lead tin 0.33 7.86e-02 4.04e-02 4.91e-02 3 .31e-02<br />
lead zinc 0.42 2.00e-04 1.00e-04 4.59e-02 3.00e-04<br />
nickel tin 0.35 9.10e-03 9.20e-03 8.70e-02 7.60e-03<br />
nickel zinc 0.33 8.00e-04 3.40e-03 8.91e-02 3.50e-03<br />
tin zinc 0.31 5.30e-03 2.02e-02 1.03e-01 1.75e-02<br />
Table 5: Same as table 2 for m<strong>et</strong>als over a 9 years time interval from January 4, 1989 to December 30,<br />
1997.<br />
35<br />
227
228 8. Tests <strong>de</strong> copule gaussienne<br />
f 3<br />
German Mark<br />
4<br />
3<br />
2<br />
1<br />
0<br />
Jan 89 Jan 94 Dec 98<br />
4<br />
2<br />
0<br />
−2<br />
20/06/1989<br />
19/08/1991<br />
−4<br />
−4 −2 0<br />
Swiss Franc<br />
2 4<br />
16/09/1992<br />
German Mark<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
10/09/97<br />
−3<br />
−6 −4 −2 0 2 4<br />
Swiss Franc<br />
Figure 10: The upper panel represents the graph of the function f3(t) <strong>de</strong>fined in (37) used in the <strong>de</strong>finition<br />
of the distance d3 for the couple Swiss Franc/German Mark as a function of time t, over the time intervals<br />
from January 25, 1989 to January 11, 1994 and from January 12, 1994 to December 31, 1998. The two<br />
lower panels represent the sc<strong>at</strong>ter plot of the r<strong>et</strong>urn of the German Mark versus the r<strong>et</strong>urn of the Swiss<br />
Franc during the two previous time periods. The circled dot, in each figure, shows the pair of r<strong>et</strong>urns<br />
responsible for the largest <strong>de</strong>vi<strong>at</strong>ion of f3 during the consi<strong>de</strong>red time interval.<br />
36
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 11: Same as figure 7 for m<strong>et</strong>als over a 9 years time interval from January 4, 1989 to December<br />
30, 1997.<br />
37<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
229
230 8. Tests <strong>de</strong> copule gaussienne<br />
ˆρ d1 d2 d3 d4<br />
am<strong>at</strong> pfe 0.15 7.41e-02 1.12e-01 8.40e-03 1.14e-01<br />
c sunw 0.28 2.56e-01 4.87e-01 1.09e-01 5.39e-01<br />
f ge 0.33 2.52e-01 2.74e-01 1.15e-01 2.90e-01<br />
gm ibm 0.21 1.49e-01 3.85e-01 1.62e-01 4.18e-01<br />
hwp sbc 0.12 4.23e-01 1.69e-01 2.52e-01 1.72e-01<br />
intc mrk 0.17 2.48e-01 1.09e-01 6.46e-01 1.04e-01<br />
ko sunw 0.14 1.41e-01 1.01e-01 2.12e-01 9.35e-02<br />
mdt t 0.16 1.21e-01 2.81e-01 8.41e-02 2.98e-01<br />
mrk xom 0.19 1.54e-01 1.50e-01 1.12e-01 1.45e-01<br />
msft sunw 0.44 3.40e-02 1.85e-02 2.60e-03 1.74e-02<br />
pfe wmt 0.27 4.24e-02 4.12e-02 1.54e-01 3.74e-02<br />
t wcom 0.27 5.67e-02 8.02e-02 5.44e-02 9.07e-02<br />
txn wcom 0.28 4.79e-01 3.77e-01 1.52e-01 3.75e-01<br />
wmt xom 0.20 3.20e-03 0.00e+00 6.02e-02 0.00e+00<br />
Table 6: Same as table 2 for stocks over a 10 years time interval from February 8, 1991 to December 29,<br />
2000.<br />
38
ˆρ d1 d2 d3 d4<br />
am<strong>at</strong> pfe 0.10 5.83e-01 5.81e-01 1.18e-01 6.38e-01<br />
c sunw 0.23 4.66e-01 5.94e-01 4.34e-01 6.16e-01<br />
f ge 0.31 8.73e-01 7.87e-01 1.54e-01 8.48e-01<br />
gm ibm 0.21 6.00e-01 6.53e-01 1.03e-01 5.27e-01<br />
hwp sbc 0.11 8.73e-01 8.06e-01 2.84e-01 8.59e-01<br />
intc mrk 0.13 8.59e-01 8.21e-01 5.48e-02 8.65e-01<br />
ko sunw 0.20 3.53e-01 5.98e-01 4.51e-01 6.79e-01<br />
mdt t 0.14 9.09e-01 8.98e-01 1.68e-01 9.15e-01<br />
mrk xom 0.12 5.36e-01 6.21e-01 1.20e-01 6.18e-01<br />
msft sunw 0.40 2.68e-01 1.38e-01 1.60e-01 1.39e-01<br />
pfe wmt 0.23 2.94e-01 4.66e-01 1.41e-01 5.23e-01<br />
t wcom 0.19 7.92e-01 9.36e-01 4.95e-02 9.49e-01<br />
txn wcom 0.23 9.10e-01 9.83e-01 1.00e-01 9.93e-01<br />
wmt xom 0.22 7.16e-01 6.71e-01 7.35e-02 6.89e-01<br />
Table 7: Same as table 2 for stocks over a 5 years time interval from February 8, 1991 to January 18,<br />
1996.<br />
39<br />
231
232 8. Tests <strong>de</strong> copule gaussienne<br />
ˆρ d1 d2 d3 d4<br />
am<strong>at</strong> pfe 0.19 2.96e-01 3.39e-01 3.10e-02 3.95e-01<br />
c sunw 0.31 7.12e-01 6.58e-01 9.47e-01 7.08e-01<br />
f ge 0.34 3.80e-01 2.36e-01 3.22e-01 2.18e-01<br />
gm ibm 0.21 3.05e-02 1.79e-01 2.37e-01 2.19e-01<br />
hwp sbc 0.11 3.47e-01 6.13e-01 7.17e-01 6.40e-01<br />
intc mrk 0.20 1.31e-01 2.06e-01 5.57e-01 2.05e-01<br />
ko sunw 0.10 6.89e-01 3.44e-01 8.59e-01 3.52e-01<br />
mdt t 0.19 4.28e-01 6.11e-01 5.01e-01 5.79e-01<br />
mrk xom 0.23 3.57e-01 6.64e-01 1.13e-01 7.38e-01<br />
msft sunw 0.46 5.79e-02 7.60e-02 8.00e-04 8.07e-02<br />
pfe wmt 0.30 2.31e-01 2.12e-01 5.59e-01 1.98e-01<br />
t wcom 0.33 1.20e-01 1.37e-01 1.73e-01 1.40e-01<br />
txn wcom 0.31 5.63e-01 4.06e-01 4.64e-01 4.17e-01<br />
wmt xom 0.19 1.61e-01 5.38e-02 3.78e-02 4.94e-02<br />
Table 8: Same as table 2 for stocks over a 5 years time interval from January 19, 1996 to December 29,<br />
2000.<br />
40
70<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 12: Same as figure 7 for stocks over a 10 years time interval from February 8, 1991 to December<br />
29, 2000.<br />
41<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
233
234 8. Tests <strong>de</strong> copule gaussienne<br />
60<br />
50<br />
40<br />
30<br />
20<br />
10<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 13: Same as figure 7 for stocks over a 5 years time interval from February 8, 1991 to January 18,<br />
1996.<br />
42<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
20<br />
15<br />
10<br />
5<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Figure 14: Same as figure 7 for stocks over a 5 years time interval from January 19, 1996 to December<br />
30, 2000.<br />
43<br />
d<br />
1<br />
d<br />
2<br />
d<br />
3<br />
d<br />
4<br />
235
236 8. Tests <strong>de</strong> copule gaussienne<br />
Time interval from<br />
Frebruary 8, 1991<br />
to January 18, 1996<br />
Time interval from<br />
January 19, 1996 to<br />
December 29, 2000<br />
ˆρ d1 d2 d3 d4<br />
hwp ibm 0.34 3.36e-01 2.26e-01 3.33e-01 2.35e-01<br />
hwp intc 0.46 3.01e-01 4.73e-01 5.12e-01 5.21e-01<br />
hwp msft 0.41 7.63e-01 4.72e-01 3.23e-01 4.53e-01<br />
hwp sunw 0.40 2.96e-01 2.98e-01 7.66e-01 3.54e-01<br />
ibm intc 0.30 4.81e-01 3.54e-01 4.18e-02 3.34e-01<br />
ibm msft 0.24 3.93e-01 6.61e-01 5.88e-01 7.07e-01<br />
ibm sunw 0.29 9.65e-01 9.71e-01 3.46e-01 9.86e-01<br />
intc msft 0.47 2.59e-01 1.45e-01 4.50e-02 1.53e-01<br />
intc sunw 0.40 4.81e-01 3.86e-01 4.47e-02 3.95e-01<br />
msft sunw 0.40 2.68e-01 1.38e-01 1.66e-01 1.39e-01<br />
ˆρ d1 d2 d3 d4<br />
hwp ibm 0.46 2.02e-02 3.21e-02 9.60e-03 3.96e-02<br />
hwp intc 0.44 2.88e-02 4.89e-02 6.00e-04 5.80e-02<br />
hwp msft 0.37 5.23e-02 9.88e-02 3.36e-01 1.18e-01<br />
hwp sunw 0.45 5.66e-01 5.65e-01 1.08e-01 6.23e-01<br />
ibm intc 0.43 5.34e-02 3.31e-02 1.68e-02 2.44e-02<br />
ibm msft 0.39 1.00e-02 9.50e-03 2.28e-02 8.80e-03<br />
ibm sunw 0.46 2.35e-01 1.56e-01 3.38e-01 1.49e-01<br />
intc msft 0.57 3.18e-01 1.61e-01 1.15e-01 1.71e-01<br />
intc sunw 0.50 6.68e-02 3.55e-02 1.00e-04 4.37e-02<br />
msft sunw 0.46 5.79e-02 7.60e-02 8.00e-04 8.07e-02<br />
Table 9: Same as table 2 for stocks belonging to the inform<strong>at</strong>ic sector, over two time intervals of 5 years.<br />
44
Chapitre 9<br />
Mesure <strong>de</strong> la dépendance extrême entre<br />
<strong>de</strong>ux actifs financiers<br />
Les <strong>de</strong>ux chapitres précé<strong>de</strong>nts nous ont permis <strong>de</strong> comprendre l’importance <strong>de</strong> l’étu<strong>de</strong> <strong>de</strong> la dépendance<br />
extrême entre <strong>de</strong>ux actifs financiers. En eff<strong>et</strong>, nous avons montré qu’il était très difficile, par une estim<strong>at</strong>ion<br />
directe <strong>de</strong> la copule, <strong>de</strong> pouvoir obtenir une mesure fiable <strong>de</strong> ces propriétés <strong>de</strong> dépendances<br />
extrêmes. Or, celles-ci ont un impact très important sur la <strong>gestion</strong> <strong>de</strong>s risques, <strong>et</strong> plus particulièrement<br />
<strong>de</strong>s grands risques, puisque selon que les extrêmes auront tendance à se produire ensemble ou <strong>de</strong> manière<br />
indépendante, la constitution <strong>de</strong> <strong>portefeuille</strong>s agrégeant ces risques perm<strong>et</strong>tra, ou non, <strong>de</strong> les diversifier.<br />
Nous avons mentionné, au chapitre 7, le coefficient <strong>de</strong> dépendance <strong>de</strong> queue comme moyen <strong>de</strong> quantifier<br />
la propension <strong>de</strong>s extrêmes à se produire <strong>de</strong> manière concomitante. Cependant, bien d’autres mesures <strong>de</strong><br />
dépendances entre les événements <strong>de</strong> gran<strong>de</strong>s amplitu<strong>de</strong>s peuvent être imaginées comme par exemple <strong>de</strong>s<br />
coefficients <strong>de</strong> corrél<strong>at</strong>ion conditionnés, <strong>et</strong> il convient donc <strong>de</strong> se <strong>de</strong>man<strong>de</strong>r pourquoi choisir telle mesure<br />
plutôt que telle autre. Ce sera l’obj<strong>et</strong> <strong>de</strong> la première partie <strong>de</strong> ce chapitre, où nous comparons ces diverses<br />
métho<strong>de</strong>s <strong>de</strong> mesures <strong>de</strong>s dépendances extrêmes <strong>et</strong> montrons qu’elles conduisent à <strong>de</strong>s résult<strong>at</strong>s parfois<br />
surprenants <strong>et</strong> contradictoires. Ceci nous perm<strong>et</strong>tra <strong>de</strong> justifier notre choix concernant le coefficient <strong>de</strong><br />
dépendance <strong>de</strong> queue, dont nous passons également en revue les différents moyens d’estim<strong>at</strong>ion, ce qui<br />
est en fait assez délic<strong>at</strong> <strong>et</strong> ne peut être réalisé <strong>de</strong> manière (rel<strong>at</strong>ivement) précise que dans quelques cas<br />
particuliers.<br />
Une <strong>de</strong>s métho<strong>de</strong>s d’estim<strong>at</strong>ion à laquelle nous nous sommes particulièrement intéressés s’appuie sur<br />
les propriétés <strong>de</strong> la classe <strong>de</strong>s modèles à facteurs, dont on sait qu’elle joue un rôle prépondérant dans la<br />
modélis<strong>at</strong>ion <strong>de</strong>s ren<strong>de</strong>ments <strong>de</strong>s actifs financiers. C<strong>et</strong>te métho<strong>de</strong> est présentée en détail dans la secon<strong>de</strong><br />
partie <strong>de</strong> ce chapitre. Elle nous a permis d’estimer les coefficients <strong>de</strong> dépendance <strong>de</strong> queue d’un ensemble<br />
d’actions sous l’hypothèse que l’indice <strong>de</strong> marché peut être considéré comme le facteur commun<br />
principal expliquant la dépendance entre ces divers actifs.<br />
Enfin, dans la <strong>de</strong>rnière partie <strong>de</strong> ce chapitre, nous utilisons ces résult<strong>at</strong>s pour répondre à la question que<br />
nous nous posions à la fin du chapitre 7 sur la pertinence <strong>de</strong> l’approxim<strong>at</strong>ion par la copule gaussienne, <strong>et</strong><br />
plus généralement elliptique, <strong>de</strong> la structure <strong>de</strong> dépendance entre actifs.<br />
237
238 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
9.1 Les différentes mesures <strong>de</strong> dépendances extrêmes<br />
Nous allons étudier l’inform<strong>at</strong>ion rel<strong>at</strong>ive contenue dans diverses mesures <strong>de</strong> dépendance entre <strong>de</strong>ux<br />
variables alé<strong>at</strong>oires X <strong>et</strong> Y pour <strong>de</strong>s événements grands <strong>et</strong> extrêmes <strong>et</strong> différents modèles <strong>de</strong> séries<br />
financières. Les mesures considérées comprennent à la fois <strong>de</strong>s quantités conditionnées <strong>et</strong> non conditionnées<br />
telles que le coefficient <strong>de</strong> corrél<strong>at</strong>ion au-<strong>de</strong>là d’un seuil donné ou le coefficient <strong>de</strong> dépendance<br />
<strong>de</strong> queue. Nous présentons <strong>de</strong>s formules analytiques explicites ainsi que <strong>de</strong>s résult<strong>at</strong>s numériques <strong>et</strong> <strong>de</strong>s<br />
estim<strong>at</strong>ions empiriques <strong>de</strong> ces mesures <strong>de</strong> dépendance, ce qui nous perm<strong>et</strong> <strong>de</strong> prouver quantit<strong>at</strong>ivement<br />
que les dépendances conditionnées peuvent être très différentes <strong>de</strong>s dépendances non conditionnées. C<strong>et</strong><br />
eff<strong>et</strong> <strong>de</strong> conditionnement fournit un mécanisme simple <strong>et</strong> général pour expliquer les changements <strong>de</strong><br />
corrél<strong>at</strong>ions basés sur <strong>de</strong>s changements <strong>de</strong> vol<strong>at</strong>ilité ou <strong>de</strong> tendance. Ainsi, les outils basés sur <strong>de</strong>s quantités<br />
conditionnées doivent être utilisés avec précaution puisque le conditionnement à lui seul induit <strong>de</strong>s<br />
changements dans la structure <strong>de</strong> dépendance, qui n’ont rien à voir avec <strong>de</strong> réels changements <strong>de</strong> la<br />
dépendance non conditionnée. Pour cela, le coefficient <strong>de</strong> dépendance <strong>de</strong> queue nous semble <strong>de</strong>voir être<br />
préféré aux corrél<strong>at</strong>ions conditionnées. De plus, ces mesures <strong>de</strong> dépendances présentent <strong>de</strong>s comportements<br />
différents <strong>et</strong> parfois opposés, suggérant que les propriétés <strong>de</strong> dépendances extrêmes possè<strong>de</strong>nt un<br />
caractère multidimensionnel qui peut être quantifié <strong>de</strong> diverses manières conduisant à <strong>de</strong>s conclusions<br />
différentes.<br />
Nous appliquons nos résult<strong>at</strong>s théoriques au problème controversé <strong>de</strong> la contagion entre les marchés<br />
sud-américains durant les pério<strong>de</strong>s <strong>de</strong> turbulence associées aux crises financières mexicaine <strong>de</strong> 1994<br />
<strong>et</strong> argentine débutée en 2001 1 . Notre analyse <strong>de</strong>s différentes mesures <strong>de</strong> dépendance entre les marchés<br />
argentin, brésilien, chilien <strong>et</strong> mexicain montre que les eff<strong>et</strong>s <strong>de</strong> conditionnement n’expliquent pas totalement<br />
le comportement <strong>de</strong>s marchés sud-américains ce qui confirme l’existence d’une possible contagion.<br />
Nous trouvons que la crise mexicaine <strong>de</strong> 1994 s’est étendue aux marchés argentin <strong>et</strong> brésilien par<br />
un mécanisme <strong>de</strong> contagion alors que <strong>de</strong> simples co-mouvements suffisent à expliquer les fluctu<strong>at</strong>ions<br />
concomitantes du marché chilien. Au suj<strong>et</strong> <strong>de</strong> la récente crise argentine débutée en 2001, nous ne trouvons<br />
pas <strong>de</strong> preuve <strong>de</strong> contagion vers les autres marchés sud-américains (excepté peut-être le Brésil)<br />
mais i<strong>de</strong>ntifions <strong>de</strong>s co-mouvements signific<strong>at</strong>ifs.<br />
1 C<strong>et</strong>te étu<strong>de</strong> empirique semble déconnectée <strong>de</strong>s autres étu<strong>de</strong>s menées jusqu’ici, <strong>et</strong> l’on peut se <strong>de</strong>man<strong>de</strong>r pourquoi nous<br />
n’avons pas continué à étudier le même jeu d’actifs que précé<strong>de</strong>mment. Ceci est simplement dû à l’exigence du referee <strong>de</strong> la<br />
revue à laquelle le papier a été soumis.
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 239<br />
Investig<strong>at</strong>ing Extreme Depen<strong>de</strong>nces: Conditioning Effect Versus<br />
Contagion in L<strong>at</strong>in-American Crises ∗<br />
Y. Malevergne 1,2 and D. Sorn<strong>et</strong>te 1,3<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095, USA<br />
email: Yannick.Malevergne@unice.fr and sorn<strong>et</strong>te@unice.fr<br />
fax: (33) 4 92 07 67 54<br />
Abstract<br />
We investig<strong>at</strong>e the rel<strong>at</strong>ive inform<strong>at</strong>ion content of several measures of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random<br />
variables X and Y for large or extreme events in various mo<strong>de</strong>ls of financial time series. The consi<strong>de</strong>red<br />
measures involve both conditional and unconditional quantities such as the conditional correl<strong>at</strong>ion<br />
coefficient over a given threshold or the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce. We offer explicit analytical formulas<br />
as well as numerical and empirical estim<strong>at</strong>ions for these measures of <strong>de</strong>pen<strong>de</strong>nce, which allow us<br />
to provi<strong>de</strong> a quantit<strong>at</strong>ive proof th<strong>at</strong> conditional <strong>de</strong>pen<strong>de</strong>nces may be very different from the unconditional<br />
ones. This conditioning effect provi<strong>de</strong>s a straightforward and general mechanism for explaining changes<br />
of correl<strong>at</strong>ions based on changes of vol<strong>at</strong>ility or of trends. Thus, tools based upon conditional quantities<br />
should be used with caution since conditioning alone induces a change in the <strong>de</strong>pen<strong>de</strong>nce structure<br />
which has nothing to do with a genuine change of unconditional <strong>de</strong>pen<strong>de</strong>nce. In this respect, for its<br />
stability, the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce should be prefered to the conditional correl<strong>at</strong>ions. Moreover,<br />
the various measures of <strong>de</strong>pen<strong>de</strong>nce exhibit different and som<strong>et</strong>imes opposite behaviors, suggesting th<strong>at</strong><br />
extreme <strong>de</strong>pen<strong>de</strong>nce properties possess a multidimensional character th<strong>at</strong> can be quantified in various<br />
ways leading to different conclusions. We apply our theor<strong>et</strong>ical results to the controversial contagion<br />
problem across L<strong>at</strong>in American mark<strong>et</strong>s during the turmoil period associ<strong>at</strong>ed with the Mexican crisis in<br />
1994 and with the Argentina crisis th<strong>at</strong> started in 2001. Our analysis of several measures of <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween the Argentina, Brazilian, Chilean and Mexican mark<strong>et</strong>s shows th<strong>at</strong> the conditioning effect does<br />
not fully explain the behavior of the L<strong>at</strong>in American stock in<strong>de</strong>xes, confirming the existence of a possible<br />
contagion. We find th<strong>at</strong> the 1994 Mexican crisis has spread over to Argentina and Brazil through<br />
contagion mechanisms and to Chile only through co-movements. Concerning the recent Argentina crisis<br />
starting in 2001, we find no evi<strong>de</strong>nce of contagion to the other L<strong>at</strong>in American countries (except perhaps<br />
in the direction of Brazil) but i<strong>de</strong>ntify significant co-movements.<br />
∗ We acknowledge helpful discussions and exchanges with J.P. Laurent, F. Lindskog and V. Pisarenko. This work was partially<br />
supported by the James S. Mc Donnell Found<strong>at</strong>ion 21st century scientist award/studying complex system.<br />
1
240 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Introduction<br />
The October 19, 1987, stock-mark<strong>et</strong> crash stunned Wall Stre<strong>et</strong> professionals, hacked about $1 trillion off<br />
the value of all U.S. stocks, and elicited predictions of another Gre<strong>at</strong> Depression. On “Black Monday,” the<br />
Dow Jones industrial average plumm<strong>et</strong>ed 508 points, or 22.6 percent, to 1, 738.74. Contrary to common<br />
belief, the US was not the first to <strong>de</strong>cline sharply. Non-Japanese Asian mark<strong>et</strong>s began a severe <strong>de</strong>cline on<br />
October 19, 1987, their time, and this <strong>de</strong>cline was echoed first on a number of European mark<strong>et</strong>s, then in<br />
North American, and finally in Japan. However, most of the same mark<strong>et</strong>s had experienced significant but<br />
less severe <strong>de</strong>clines in the l<strong>at</strong>ter part of the previous week. With the exception of the US and Canada, other<br />
mark<strong>et</strong>s continued downward through the end of October, and some of these <strong>de</strong>clines were as large as the<br />
gre<strong>at</strong> crash on October 19.<br />
On December 19, 1994, the Mexican government, facing a solvency crisis, chose to <strong>de</strong>valu<strong>at</strong>e the peso and<br />
abandoned its exchange r<strong>at</strong>e parity. This <strong>de</strong>valu<strong>at</strong>ion plunged the country into a major financial crisis which<br />
quickly propag<strong>at</strong>ed to the rest of the L<strong>at</strong>in American countries.<br />
From July 1997 to December 1997, several East Asian mark<strong>et</strong>s crashed, starting with the Thai mark<strong>et</strong> on<br />
July 2, 1997 and ending with the Hong Kong mark<strong>et</strong> on October 17, 1997. After this regional event, the<br />
turmoil spread over to the American and European mark<strong>et</strong>s.<br />
The “slow” crash and in particular the turbulent behavior of the stock mark<strong>et</strong>s worldwi<strong>de</strong> starting midaugust<br />
1998 are wi<strong>de</strong>ly associ<strong>at</strong>ed with and even <strong>at</strong>tributed to the plunge of the Russian financial mark<strong>et</strong>s,<br />
the <strong>de</strong>valu<strong>at</strong>ion of its currency and the <strong>de</strong>fault of the government on its <strong>de</strong>bts oblig<strong>at</strong>ions.<br />
The Nasdaq Composite in<strong>de</strong>x dropped precipiteously with a low of 3227 on April 17, 2000, corresponding<br />
to a cumul<strong>at</strong>ive loss of 37% counted from its all-time high of 5133 reached on March 10, 2000. The drop<br />
was mostly driven by the so-called “New Economy” stocks which have risen nearly fourfold over 1998 and<br />
1999 compared to a gain of only 50% for the S&P500 in<strong>de</strong>x. And without technology, the benchmark would<br />
be fl<strong>at</strong>.<br />
All these events epitomize the observ<strong>at</strong>ion often reported by mark<strong>et</strong> professionals th<strong>at</strong>, “during major mark<strong>et</strong><br />
events, correl<strong>at</strong>ions change dram<strong>at</strong>ically” (Bookstaber 1997). The possible existence of changes of correl<strong>at</strong>ion,<br />
or more precisely of changes of <strong>de</strong>pen<strong>de</strong>nce, b<strong>et</strong>ween ass<strong>et</strong>s and b<strong>et</strong>ween mark<strong>et</strong>s in different mark<strong>et</strong><br />
phases has obvious implic<strong>at</strong>ions in risk assessment, portfolio management and in the way policy and regul<strong>at</strong>ion<br />
should be performed. Concerning portfolio management, (Ang and Bekaert 2001, Ang and Chen 2001)<br />
for instance, have stressed th<strong>at</strong> these questions rel<strong>at</strong>ed to st<strong>at</strong>e-varying-<strong>de</strong>pen<strong>de</strong>nce are important for practical<br />
applic<strong>at</strong>ions since in such a case the optimal portfolio will also become st<strong>at</strong>e-<strong>de</strong>pen<strong>de</strong>nt, and neglecting<br />
this point can lead to very inefficient ass<strong>et</strong> alloc<strong>at</strong>ions. In this spirit, the recent Argentine crisis in 2001 has<br />
triggered fears of a contagion to other L<strong>at</strong>in American mark<strong>et</strong>. Also, the Enron financial scandal <strong>at</strong> the end<br />
of 2001 seems to have opened a flux of similar bankrupcies in other “new economy” companies.<br />
From an aca<strong>de</strong>mic perspective, all these manifest<strong>at</strong>ions of propag<strong>at</strong>ing crisis have given birth to an intense<br />
activity concerning the notion of contagion (see (Claessens <strong>et</strong> al. 2001) for a review) which is <strong>de</strong>fined,<br />
according to the most commonly accepted <strong>de</strong>finition, as an increase in the correl<strong>at</strong>ion (or linkage) across<br />
mark<strong>et</strong>s during turmoil periods. In fact, as we shall see, there are two distinct classes of mechanisms for<br />
un<strong>de</strong>rstanding “changes of correl<strong>at</strong>ions”, not necessarily mutually exclusive.<br />
• It is possible th<strong>at</strong> there are genuine changes with time of the unconditional (with respect to amplitu<strong>de</strong>s)<br />
correl<strong>at</strong>ions and thus of the un<strong>de</strong>rlying structure of the dynamical processes, as observed by i<strong>de</strong>ntifying<br />
shifts in ARMA-ARCH/GARCH processes (Silvapulle and Granger 2001), in regime-switching<br />
mo<strong>de</strong>ls (Ang and Bekaert 2001, Ang and Chen 2001) or in contagion mo<strong>de</strong>ls<br />
2
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 241<br />
(Quintos 2001, Quintos <strong>et</strong> al. 2001). (Longin and Solnik 1995, Tsui and Yu 1999) and many others<br />
have shown th<strong>at</strong> the hypothesis of a constant conditional correl<strong>at</strong>ion for stock r<strong>et</strong>urns or intern<strong>at</strong>ional<br />
equity r<strong>et</strong>urns must be rejected. In fact, there is strong evi<strong>de</strong>nce th<strong>at</strong> the correl<strong>at</strong>ions are not only<br />
time <strong>de</strong>pen<strong>de</strong>nt but also st<strong>at</strong>e <strong>de</strong>pen<strong>de</strong>nt. In<strong>de</strong>ed, as shown by (King and Wadhwani 1990, Ramchand<br />
and Susmel 1998), the correl<strong>at</strong>ions increase in periods of large vol<strong>at</strong>ility. Moreover, (Longin<br />
and Solnik 2001) have proved th<strong>at</strong> the correl<strong>at</strong>ions across intern<strong>at</strong>ional equity mark<strong>et</strong>s are also trend<br />
<strong>de</strong>pen<strong>de</strong>nt.<br />
• In contrast, a second class of explan<strong>at</strong>ion is th<strong>at</strong> correl<strong>at</strong>ions b<strong>et</strong>ween two variables conditioned on<br />
signed exceedance (one-si<strong>de</strong>d) or on absolute value (vol<strong>at</strong>ility) exceedance of one or both variables<br />
may <strong>de</strong>vi<strong>at</strong>e significantly from the unconditional correl<strong>at</strong>ion (Boyer <strong>et</strong> al. 1997, Lor<strong>et</strong>an 2000, Lor<strong>et</strong>an<br />
and English 2000). In other words, with a fixed unconditional correl<strong>at</strong>ion ρ, the measured correl<strong>at</strong>ion<br />
conditioned of a given bullish trend, bearish trend, high or low mark<strong>et</strong> vol<strong>at</strong>ility, may in general differ<br />
from ρ and be a function of the specific mark<strong>et</strong> phase. According to this explan<strong>at</strong>ion, changes of<br />
correl<strong>at</strong>ion may be only a fallacious appearance th<strong>at</strong> stems from a change of vol<strong>at</strong>ility or a change of<br />
trend of the mark<strong>et</strong> and not from a real change of unconditional correl<strong>at</strong>ion or <strong>de</strong>pen<strong>de</strong>nce.<br />
The existence of the second class of explan<strong>at</strong>ion is appealing by its parsimony, as it posits th<strong>at</strong> observed<br />
“changes of correl<strong>at</strong>ion” may simply result from the way the measure of <strong>de</strong>pen<strong>de</strong>nce is performed. This<br />
approach has been followed by several authors but is often open to misinterpr<strong>et</strong><strong>at</strong>ion,<br />
as stressed by (Forbes and Rigobon 2002). In addition, it may also be misleading since it does not provi<strong>de</strong><br />
a sign<strong>at</strong>ure or procedure to i<strong>de</strong>ntify the existence of a genuine contagion phenomenon, if any. Therefore,<br />
in or<strong>de</strong>r to clarify the situ<strong>at</strong>ion and eventually <strong>de</strong>velop more a<strong>de</strong>qu<strong>at</strong>e tools for probing the <strong>de</strong>pen<strong>de</strong>nces<br />
b<strong>et</strong>ween ass<strong>et</strong>s and b<strong>et</strong>ween mark<strong>et</strong>s, it is highly <strong>de</strong>sirable to characterize the different possible ways with<br />
which higher or lower conditional <strong>de</strong>pen<strong>de</strong>nce can occur in mo<strong>de</strong>ls with constant unconditional <strong>de</strong>pen<strong>de</strong>nce.<br />
In or<strong>de</strong>r to make progress, it is necessary to first distinguish b<strong>et</strong>ween the different measures of <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween two variables for large or extreme events th<strong>at</strong> have been introduced in the liter<strong>at</strong>ure, because the<br />
conclusions th<strong>at</strong> one can draw about the variability of <strong>de</strong>pen<strong>de</strong>nce are sensitive to the choice of its measure.<br />
These measures inclu<strong>de</strong><br />
1. the correl<strong>at</strong>ion conditioned on signed exceedance of one or both variables (Boyer <strong>et</strong> al. 1997, Lor<strong>et</strong>an<br />
2000, Lor<strong>et</strong>an and English 2000, Cizeau <strong>et</strong> al. 2001), th<strong>at</strong> we call respectively ρ + v and ρu, where u<br />
and v <strong>de</strong>note the thresholds above which the exceedances are calcul<strong>at</strong>ed,<br />
2. the correl<strong>at</strong>ion conditioned on absolute value exceedance (or large vol<strong>at</strong>ility), above the threshold<br />
v, of one or both variables (Boyer <strong>et</strong> al. 1997, Lor<strong>et</strong>an 2000, Lor<strong>et</strong>an and English 2000, Cizeau <strong>et</strong><br />
al. 2001), th<strong>at</strong> we call ρ s v (for a condition of exceedance on one variable),<br />
3. the tail-<strong>de</strong>pen<strong>de</strong>nce param<strong>et</strong>er λ, which has a simple analytical expression when using copulas<br />
(Embrechts <strong>et</strong> al. 2001, Lindskog 1999) such as the Gumbel copula (Longin and Solnik 2001),<br />
and whose estim<strong>at</strong>ion provi<strong>de</strong>s useful inform<strong>at</strong>ion about the occurrence of extreme co-movements<br />
(Malevergne and Sorn<strong>et</strong>te 2001, Poon <strong>et</strong> al. 2001, Juri and Wüthrich 2002),<br />
4. the spectral measure associ<strong>at</strong>ed with the tail in<strong>de</strong>x (assumed to be the same of all ass<strong>et</strong>s) of extreme<br />
value multivari<strong>at</strong>e distributions (Davis <strong>et</strong> al. 1999, Starica 1999, Hauksson <strong>et</strong> al. 2001),<br />
5. tail indices of extremal correl<strong>at</strong>ions <strong>de</strong>fined as the upper or lower correl<strong>at</strong>ion of exceedances of or<strong>de</strong>red<br />
log-values (Quintos 2001),<br />
6. confi<strong>de</strong>nce weighted forecast correl<strong>at</strong>ions (Bhansali and Wise 2001) or algorithmic complexity measures<br />
(Mansilla 2001).<br />
3
242 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Our contribution to the liter<strong>at</strong>ure is both m<strong>et</strong>hological and empirical. On the m<strong>et</strong>hodological front, firstof-all,<br />
we review the existing tools available for probing the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween large or extreme events<br />
for several mo<strong>de</strong>ls of interest for financial time series; second, we provi<strong>de</strong> explicit analytical expressions<br />
for these measures of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two variables; third, this allows us to quantify the misleading<br />
intrepr<strong>et</strong><strong>at</strong>ions of certain conditional coefficients commomly used for exploring the evolution of the <strong>de</strong>pen<strong>de</strong>nce<br />
associ<strong>at</strong>ed with a change in the mark<strong>et</strong> conditions (an increase of the vol<strong>at</strong>ility, for instance). On the<br />
empirical front, we apply our theor<strong>et</strong>ical results to the controversial problem of the occurrence or not of a<br />
contagion phenomenon across L<strong>at</strong>in American mark<strong>et</strong>s during the turmoil period associ<strong>at</strong>ed with the Mexican<br />
crisis in 1994 or with the recent Argentina crisis. In this purpose, we use the novel insight <strong>de</strong>rived from<br />
our analysis on several measures of <strong>de</strong>pen<strong>de</strong>nce and apply them to the question of a possible evolution of<br />
the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the Argentina, Brazilian, Chilean and Mexican mark<strong>et</strong>s with respect to the mark<strong>et</strong><br />
conditions.<br />
The <strong>de</strong>pen<strong>de</strong>nce measures we study are the conditional correl<strong>at</strong>ion coefficients ρ + v , ρ s v, ρu, the conditional<br />
Spearman’s rho ρs(v) and the tail <strong>de</strong>pen<strong>de</strong>nce coefficients λ and ¯ λ, whose properties are investig<strong>at</strong>ed for<br />
several mo<strong>de</strong>ls among which are the bivari<strong>at</strong>e Gaussian distribution, the bivari<strong>at</strong>e Stu<strong>de</strong>nt’s distribution,<br />
and the one factor mo<strong>de</strong>l for various distributions of the factor. Initially, we hoped to show the existence<br />
of logical links b<strong>et</strong>ween some of these measures, such as a vanishing tail-<strong>de</strong>pen<strong>de</strong>nce param<strong>et</strong>er λ implies<br />
vanishing asymptotic conditional correl<strong>at</strong>ion coefficients. In fact, we will show th<strong>at</strong> this turns out to be<br />
wrong and one can construct simple examples for which all possible combin<strong>at</strong>ions occur. Therefore, each<br />
of these measures probe a different quality of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two variables for large or extreme<br />
events. In addition, even if the conditional correl<strong>at</strong>ion coefficients are asymptotically zero, they <strong>de</strong>cay in<br />
general extremely slowly, as inverse powers of the value of the threshold, and may thus remain significant<br />
for most pr<strong>at</strong>ical applic<strong>at</strong>ions. These results will allow us to assert th<strong>at</strong>, somewh<strong>at</strong> similarly to risks whose<br />
a<strong>de</strong>qu<strong>at</strong>e characteriz<strong>at</strong>ion requires an extension beyond the restricted one-dimensional measure in terms of<br />
the variance (vol<strong>at</strong>ility) to inclu<strong>de</strong> all higher or<strong>de</strong>r cumulants or more generally the knowledge of the full<br />
distribution (Sorn<strong>et</strong>te <strong>et</strong> al. 2000a, Sorn<strong>et</strong>te <strong>et</strong> al. 2000b, An<strong>de</strong>rsen and Sorn<strong>et</strong>te 2001), our results suggest<br />
th<strong>at</strong> tail-<strong>de</strong>pen<strong>de</strong>nce has also a multidimensional character.<br />
The paper is organized as follows.<br />
Section 1 <strong>de</strong>scribes three conditional correl<strong>at</strong>ion coefficients, namely the correl<strong>at</strong>ion ρ + v conditioned on<br />
signed exceedance of one variable, or on both variables (ρu) and the correl<strong>at</strong>ion ρ s v conditioned on absolute<br />
value exceedance (or large vol<strong>at</strong>ility) of one variable. (Boyer <strong>et</strong> al. 1997) have already provi<strong>de</strong>d the general<br />
expression of ρ + v and ρ s v for the Gaussian bivari<strong>at</strong>e mo<strong>de</strong>l, which are used to <strong>de</strong>rive their v <strong>de</strong>pen<strong>de</strong>nce for<br />
large v, and to show th<strong>at</strong>, for a given distribution, the conditional correl<strong>at</strong>ion coefficient changes even if the<br />
unconditional correl<strong>at</strong>ion is l<strong>et</strong> unchanged and the n<strong>at</strong>ure of this change <strong>de</strong>pends on the conditioning s<strong>et</strong>.<br />
We then provi<strong>de</strong> the general expression of ρ + v and ρ s v for the Stu<strong>de</strong>nt’s bivari<strong>at</strong>e mo<strong>de</strong>l with ν <strong>de</strong>grees of<br />
freedom and for the factor mo<strong>de</strong>l X = αY + ɛ, for which we give a general expression of the conditional<br />
correl<strong>at</strong>ion coefficient wh<strong>at</strong>ever the distributions of Y and ɛ may be. This leads us to conclu<strong>de</strong> by comparision<br />
with the Gaussian mo<strong>de</strong>l th<strong>at</strong>, for a fixed conditioning s<strong>et</strong>, the behavior of the conditional correl<strong>at</strong>ion<br />
change dram<strong>at</strong>ically from a distribution to another one. Conditioning now on both variables, we are able to<br />
provi<strong>de</strong> the asymptotic <strong>de</strong>pen<strong>de</strong>nce of ρu only for the bivari<strong>at</strong>e Gaussian mo<strong>de</strong>l and show th<strong>at</strong> it essentially<br />
behaves like ρ + v . We then apply these results to show th<strong>at</strong> we cannot entirely explain the behavior of the conditional<br />
correl<strong>at</strong>ion coefficient of the L<strong>at</strong>in American stock in<strong>de</strong>xes by the conditioning effect, suggesting<br />
the existence of a possible contagion.<br />
In section 2, to account for several <strong>de</strong>ficiencies of the correl<strong>at</strong>ion coefficient, we propose an altern<strong>at</strong>ive<br />
measure of <strong>de</strong>pen<strong>de</strong>nce, the conditional Spearman’s rho, which is rel<strong>at</strong>ed to the probability of concordance<br />
and discordance of several events drawn from the same probability distribution. This measure provi<strong>de</strong>s<br />
an important improvement with respect to the correl<strong>at</strong>ion coefficient since it only takes into account the<br />
4
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 243<br />
<strong>de</strong>pen<strong>de</strong>nce structure of the variable and is not sensitive to the marginal behavior of each variable. We<br />
perform numerical comput<strong>at</strong>ions to <strong>de</strong>rive the behavior of the conditional Spearman’s rho, <strong>de</strong>noted by ρs(v).<br />
This allows us to prove th<strong>at</strong> there is no direct rel<strong>at</strong>ion b<strong>et</strong>ween the Spearman’s rho conditioned on large<br />
values and the correl<strong>at</strong>ion coefficient conditioned on the same values. Therefore, each of these coefficients<br />
quantifies a different kind of extreme <strong>de</strong>pen<strong>de</strong>nce. Then, calibr<strong>at</strong>ing our mo<strong>de</strong>ls on the L<strong>at</strong>in American<br />
mark<strong>et</strong> d<strong>at</strong>a, we confirm th<strong>at</strong> the conditional effect cannot fully explain the observed <strong>de</strong>pen<strong>de</strong>nce and th<strong>at</strong><br />
contagion can therefore be invoked. This results are much clearer for the conditional Spearman’s rho than<br />
for the condition (linear) correl<strong>at</strong>ion coefficient, due to the much larger impact of large st<strong>at</strong>istical fluctu<strong>at</strong>ions<br />
in the l<strong>at</strong>er.<br />
Section 3 discusses the tail-<strong>de</strong>pen<strong>de</strong>nce param<strong>et</strong>ers λ and ¯ λ. We first recall their <strong>de</strong>finitions and their values<br />
for Gaussian and Stu<strong>de</strong>nt’s bivari<strong>at</strong>e distributions of X and Y , already known in the liter<strong>at</strong>ure. For the<br />
Gaussian factor mo<strong>de</strong>l, it is trivial to show th<strong>at</strong> λ = 0. A non-trivial result is obtained for the Stu<strong>de</strong>nt’s factor<br />
mo<strong>de</strong>l: λ is found non-zero and a function only of α and of the scale factor of ɛ. More generally, a theorem<br />
established in (Malevergne and Sorn<strong>et</strong>te 2002) allows one to calcul<strong>at</strong>e the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
for any distribution of the factor and shows th<strong>at</strong> λ vanishes for any rapidly varying factor. We then apply<br />
the (Poon <strong>et</strong> al. 2001)’s procedure to estim<strong>at</strong>e non-param<strong>et</strong>rically the tail <strong>de</strong>pen<strong>de</strong>nce coefficients. We find<br />
them significant and thus conclu<strong>de</strong> th<strong>at</strong> with or without contagion mechanism, extreme co-movements must<br />
n<strong>at</strong>urally occur on the various L<strong>at</strong>in American mark<strong>et</strong>s as soon as one of them enters a crisis.<br />
Section 4 provi<strong>de</strong>s a synthesis and comparison b<strong>et</strong>ween these different results. A first important message is<br />
th<strong>at</strong> there is no unique measure of extreme <strong>de</strong>pen<strong>de</strong>nce. Each of the coefficients of extreme <strong>de</strong>pen<strong>de</strong>nce th<strong>at</strong><br />
we have studied provi<strong>de</strong>s a specific quantific<strong>at</strong>ion th<strong>at</strong> is sensitive to a certain combin<strong>at</strong>ion of the marginals<br />
and of the copula of the two random variables. Similarly to risks whose a<strong>de</strong>qu<strong>at</strong>e characteriz<strong>at</strong>ion requires<br />
an extension beyond the restricted one-dimensional measure in terms of the variance (vol<strong>at</strong>ility) to inclu<strong>de</strong><br />
the knowledge of the full distribution, tail-<strong>de</strong>pen<strong>de</strong>nce has also a multidimensional character. A second<br />
important message is th<strong>at</strong> the increase of some of the conditional coefficients of extreme <strong>de</strong>pen<strong>de</strong>nce as<br />
one goes more in the tails does not necessarily signals a genuine increase of the unconditional correl<strong>at</strong>ion<br />
or <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two variables. Our calcul<strong>at</strong>ions firmly confirm th<strong>at</strong> this increase is a general<br />
and unvoidable result of the st<strong>at</strong>istical properties of many multivari<strong>at</strong>e mo<strong>de</strong>ls of <strong>de</strong>pen<strong>de</strong>nce. From the<br />
standpoint of the contagion across L<strong>at</strong>in American mark<strong>et</strong>s, our theor<strong>et</strong>ical and empirical results suggest an<br />
asymm<strong>et</strong>ric contagion phenomenon from Chile and Mexico onto Argentina and Brazil: large moves of the<br />
Chilean and Mexican mark<strong>et</strong>s tend to propag<strong>at</strong>e to Argentina and Brazil through contagion mechanisms, i.e.,<br />
with a change in the <strong>de</strong>pen<strong>de</strong>nce structure, while the converse does not hold. As a consequence, our study<br />
seems to prove th<strong>at</strong> the 1994 Mexican crisis have spread over to Argentina and Brazil through contagion<br />
mechanisms and to Chile only through co-movements. Concerning the recent Argentina crisis starting in<br />
2001, we find no evi<strong>de</strong>nce of contagion to the other L<strong>at</strong>in American countries (except perhaps in the direction<br />
of Brazil) but i<strong>de</strong>ntify only co-movements.<br />
1 Conditional correl<strong>at</strong>ion coefficient<br />
In this section, we discuss the properties of the correl<strong>at</strong>ion coefficient conditioned on one variable. We study<br />
the difference b<strong>et</strong>ween conditioning on the signed values or on absolute values of the variable (conditioning<br />
on the absolute value of the variable of interest is only meaningful when its distribution is symm<strong>et</strong>ric).<br />
This allows us to conclu<strong>de</strong> th<strong>at</strong> conditioning on signed values generally provi<strong>de</strong>s more inform<strong>at</strong>ion than<br />
conditioning on absolute values, and th<strong>at</strong>, as already un<strong>de</strong>rlined by (Boyer <strong>et</strong> al. 1997, for instance), the<br />
conditional correl<strong>at</strong>ion coefficient suffers from a bias which forbids its use as a measure of a change in the<br />
correl<strong>at</strong>ion b<strong>et</strong>ween two ass<strong>et</strong>s when the vol<strong>at</strong>ility increases, as seen in many papers about contagion. We<br />
5
244 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
then present an empirical illustr<strong>at</strong>ion of the evolution of the correl<strong>at</strong>ion b<strong>et</strong>ween several stock in<strong>de</strong>xes of<br />
L<strong>at</strong>in American mark<strong>et</strong>s.<br />
1.1 Definition<br />
We study the correl<strong>at</strong>ion coefficient ρA of two real random variables X and Y conditioned on Y ∈ A, where<br />
A is a subs<strong>et</strong> of R such th<strong>at</strong> Pr{Y ∈ A} > 0.<br />
By <strong>de</strong>finition, the conditional correl<strong>at</strong>ion coefficient ρA is given by<br />
ρA =<br />
Cov(X, Y | Y ∈ A)<br />
Var(X | Y ∈ A) · Var(Y | Y ∈ A) . (1)<br />
Applying this general expression of the conditional correl<strong>at</strong>ion coefficient, we will give closed formula for<br />
several standard distributions and mo<strong>de</strong>ls. This will allow us to investig<strong>at</strong>e the influence of the conditionning<br />
s<strong>et</strong> and the un<strong>de</strong>rlying mo<strong>de</strong>l on the behavior of ρA.<br />
1.2 Influence of the conditioning s<strong>et</strong><br />
L<strong>et</strong> the variables X and Y have a multivari<strong>at</strong>e Gaussian distribution with (unconditional) correl<strong>at</strong>ion coefficient<br />
ρ. The following result have been proved by (Boyer <strong>et</strong> al. 1997) :<br />
ρ<br />
<br />
. (2)<br />
ρA =<br />
ρ2 + (1 − ρ2 Var(Y )<br />
) Var(Y | Y ∈A)<br />
We can note th<strong>at</strong> ρ and ρA have the same sign, th<strong>at</strong> ρA = 0 if and only if ρ = 0 and th<strong>at</strong> ρA does not <strong>de</strong>pend<br />
directly on Var(X). Note also th<strong>at</strong> ρA can be either gre<strong>at</strong>er or smaller than ρ since Var(Y | Y ∈ A) can be<br />
either gre<strong>at</strong>er or smaller than Var(Y ). We will illustr<strong>at</strong>e this property in the two following examples, where<br />
we consi<strong>de</strong>r a conditioning on large positive (or neg<strong>at</strong>ive) r<strong>et</strong>urns and a conditioning on large vol<strong>at</strong>ility.<br />
The difference comes from the fact th<strong>at</strong> in the first case, we account for the trend while we neglect this<br />
inform<strong>at</strong>ion in the second case.<br />
Example 1: conditioning on large (positive) r<strong>et</strong>urns. L<strong>et</strong> us first consi<strong>de</strong>r the conditioning s<strong>et</strong> A =<br />
[v, +∞), with v ∈ R+. Thus ρA is the correl<strong>at</strong>ion coefficient conditioned on the r<strong>et</strong>urns Y larger than a<br />
given positive threshold v. It will be <strong>de</strong>noted by ρ + v in the sequel. Assuming for simplicity, but without loss<br />
of generality, th<strong>at</strong> Var(Y ) = 1, we can easily show (see appendix A.1.1 for an exact calcul<strong>at</strong>ion) th<strong>at</strong> for<br />
large v<br />
ρ + v ∼v→∞<br />
ρ 1<br />
· , (3)<br />
1 − ρ2 v<br />
which slowly goes to zero as v goes to infinity. Obviously, by symm<strong>et</strong>ry, the conditional correl<strong>at</strong>ion coefficient<br />
ρ − v , conditioned on Y smaller than v, obeys the same formula.<br />
Example 2: conditioning on large vol<strong>at</strong>ilities. L<strong>et</strong> now the conditioning s<strong>et</strong> be A = (−∞, −v] ∪<br />
[v, +∞), with v ∈ R+. Thus ρA is the correl<strong>at</strong>ion coefficient conditioned on |Y | larger than v, i.e., it<br />
is conditioned on large vol<strong>at</strong>ility of Y . Still assuming Var(Y ) = 1, we <strong>de</strong>note this correl<strong>at</strong>ion coefficient by<br />
ρ s v and, as shown in appendix A.1.2, we can conclu<strong>de</strong> th<strong>at</strong>, for large v,<br />
ρ s v ∼v→∞<br />
ρ<br />
<br />
ρ2 + 1−ρ2<br />
2+v2 , (4)<br />
6
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 245<br />
which goes to one as v goes to infinity as 1 − ρ s v ∼v→∞ 1−ρ2<br />
ρ 2<br />
These two simple examples clearly show th<strong>at</strong>, in the case of two Gaussian random variables, the two conditional<br />
correl<strong>at</strong>ion coefficients ρ + v and ρ s v exhibit opposite behavior since the conditional correl<strong>at</strong>ion coefficient<br />
ρ + v is a <strong>de</strong>creasing function of v which goes to zero as v → +∞ while the conditional correl<strong>at</strong>ion<br />
coefficient ρ s v is an increasing function of v and goes to one as v → ∞. These opposite behaviors are very<br />
general and do not <strong>de</strong>pend on the particular choice of the joint distribution of X and Y , namely the Gaussian<br />
distribution studied until now, as it will be seen in the sequel.<br />
This result un<strong>de</strong>rlines the importance of the choice of the conditioning s<strong>et</strong> with the following two cave<strong>at</strong>s.<br />
First, as already stressed by many authors, the conditional correl<strong>at</strong>ion ρ + v ou ρ s v change with the value of the<br />
threshod v even if the unconditional correl<strong>at</strong>ion ρ remains unchanged. Thus, the observ<strong>at</strong>ion of a change<br />
in the conditional correl<strong>at</strong>ion does not provi<strong>de</strong> a reliable sign<strong>at</strong>ure of a change in the true (unconditional)<br />
correl<strong>at</strong>ion. Second, the conditional correl<strong>at</strong>ions can exhibit really opposite behaviors <strong>de</strong>pending on the<br />
conditioning s<strong>et</strong>s. Specifically, we have seen th<strong>at</strong> accounting for a signed trend or only for its amplitu<strong>de</strong><br />
may yield a <strong>de</strong>crease or an increase of the conditional correl<strong>at</strong>ion with respect to the unconditional one, so<br />
th<strong>at</strong> these changes cannot be interpr<strong>et</strong>ed as a strengthening or a weakening of the correl<strong>at</strong>ions.<br />
v −2 .<br />
1.3 Influence of the un<strong>de</strong>rlying distribution for a given conditioning s<strong>et</strong><br />
For a fixed conditioning s<strong>et</strong> <strong>de</strong>fining a specific conditional correl<strong>at</strong>ion coefficient like ρ + v or ρ s v, the behavior<br />
of these coefficients can be dram<strong>at</strong>ically different from a pair of random variables to another one, <strong>de</strong>pending<br />
on their un<strong>de</strong>rlying joint distribution. As an example, l<strong>et</strong> the variables X and Y have a multivari<strong>at</strong>e Stu<strong>de</strong>nt’s<br />
distribution with ν <strong>de</strong>grees of freedom and an (unconditional) correl<strong>at</strong>ion coefficient ρ. According to the<br />
proposition st<strong>at</strong>ed in appendix B.1, we have the exact formula<br />
ρA =<br />
ρ<br />
<br />
ρ2 + E[E(X2 | Y )−ρ2Y 2 . (5)<br />
| Y ∈A]<br />
Var(Y | Y ∈A)<br />
Appendix B.1 gives the explicit formulas of E[E(X 2 | Y ) − ρ 2 Y 2 | Y ∈ A] and Var(Y | Y ∈ A).<br />
Expression (5) is the analog for a Stu<strong>de</strong>nt bivari<strong>at</strong>e distribution to (2) <strong>de</strong>rived above for the Gaussian bivari<strong>at</strong>e<br />
distribution. Again, ρ and ρA share the following properties: they have the same sign, ρA equals zero if and<br />
only if ρ equals zero and ρA can be either gre<strong>at</strong>er or smaller than ρ. We now apply this general formula (5)<br />
to the calculus of ρ + v and ρ s v and we find (see appendices B.3 and B.4) th<strong>at</strong>, conditioning on large r<strong>et</strong>urns,<br />
ρ + v −→<br />
<br />
while when conditionning on large vol<strong>at</strong>ility,<br />
ρ s v −→<br />
<br />
ρ 2 + (ν − 1)<br />
ρ 2 + 1<br />
(ν−1)<br />
ρ<br />
<br />
ν−2<br />
ν (1 − ρ2 )<br />
ρ<br />
ν−2<br />
ν (1 − ρ 2 )<br />
, (6)<br />
. (7)<br />
In both cases, ρ + v and ρ s v goes, <strong>at</strong> infinity, to non vanishing constant (exepted for ρ = 0). Moreover, for ν<br />
larger than νc 2.839, this constant is smaller than the unconditional correl<strong>at</strong>ion coefficient ρ, for all value<br />
of ρ, in the case of ρ + v , while for ρ s v it is always larger than ρ, wh<strong>at</strong>ever ν (larger than two) may be.<br />
These results show th<strong>at</strong>, conditioned on large r<strong>et</strong>urns, ρ + v is a <strong>de</strong>creasing function of the threshold v (<strong>at</strong><br />
least when ν ≥ 2.839), while, conditioned on large vol<strong>at</strong>ilities, ρ s v is an increasing function of v. Thus, for<br />
7
246 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Stu<strong>de</strong>nt’s random variables, ρ + v and ρ s v exhibit behaviors exactly opposite to those observed for Gaussian<br />
random variables, namely <strong>de</strong>creasing ρ + v and increasing ρ s v, as shown in the previous paragraph.<br />
To give another example, l<strong>et</strong> us now assume th<strong>at</strong> X and Y are two random variables following the equ<strong>at</strong>ion<br />
X = αY + ɛ , (8)<br />
where α is a non random real coefficient and ɛ an idiosyncr<strong>at</strong>ic noise in<strong>de</strong>pen<strong>de</strong>nt of Y , whose distribution<br />
admits a centered moment of second or<strong>de</strong>r σ 2 ɛ . L<strong>et</strong> us also <strong>de</strong>note by σ 2 y the second centered moment of the<br />
variable Y . This kind of rel<strong>at</strong>ion b<strong>et</strong>ween X and Y is the so-called one factor mo<strong>de</strong>l whose applic<strong>at</strong>ions<br />
in finance can be traced back to (Ross 1976). This one factor mo<strong>de</strong>l with in<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween Y and<br />
ɛ is of course naive for concr<strong>et</strong>e applic<strong>at</strong>ions, as it neglects the potential influence of other factors in the<br />
d<strong>et</strong>ermin<strong>at</strong>ion of X. However, it has been argued to be a useful mo<strong>de</strong>l in the context of the contagion,<br />
and several studies have been based upon it (see (Baig and Goldfajn 1998) or (Forbes and Rigobon 2002),<br />
for instance). Moreover for our purpose, it provi<strong>de</strong>s a simple illustr<strong>at</strong>ive mo<strong>de</strong>l with rich and somewh<strong>at</strong><br />
surprising results.<br />
Appendix C shows th<strong>at</strong> the conditional correl<strong>at</strong>ion coefficient of X and Y is<br />
ρ<br />
ρA = <br />
ρ2 + (1 − ρ2 )<br />
Var(y)<br />
, (9)<br />
where<br />
α · σy<br />
ρ = <br />
α2 · σ2 y + σ2 ɛ<br />
Var(y | y∈A)<br />
<strong>de</strong>notes the unconditional correl<strong>at</strong>ion coefficient of X and Y . Note th<strong>at</strong> the term σ 2 ɛ in the expression (10)<br />
of ρ is the only place where the influence of the idiosynchr<strong>at</strong>ic noise is felt.<br />
Expression (9) is the same as (2) for the bivari<strong>at</strong>e Gaussian situ<strong>at</strong>ion studied in 1.2. This is not surprising<br />
since, in the case where Y and ɛ have univari<strong>at</strong>e Gaussian distributions, the joint distribution of X and<br />
Y is a bivari<strong>at</strong>e Gaussian distribution. The new fact is th<strong>at</strong> this expression (9) remains true wh<strong>at</strong>ever the<br />
distribution of Y and ɛ, provi<strong>de</strong>d th<strong>at</strong> their second moments exist.<br />
We now present the asymptotic expression of ρA for Y with a Gaussian or a Stu<strong>de</strong>nt’s distribution. Note<br />
th<strong>at</strong> the expression of ρA is simple enough to allow for exact calcul<strong>at</strong>ions for a larger class of distributions,<br />
but for our purpose, these two simple case will be sufficient.<br />
Assuming th<strong>at</strong> Y has a Gaussian distribution, while the distribution of ɛ can be everything (provi<strong>de</strong>d th<strong>at</strong><br />
E[ɛ 2 ] < ∞) allows us to show th<strong>at</strong> the same results as those given by equ<strong>at</strong>ions (3) and (4) still hold, so th<strong>at</strong><br />
ρ + v goes to zero, while ρ s v goes to one.<br />
On the contrary, assuming th<strong>at</strong> Y has a Stu<strong>de</strong>nt’s distribution yields both for ρ + v and ρ s v:<br />
ρ +,s<br />
v<br />
∼<br />
1<br />
<br />
1 + K<br />
v 2<br />
(10)<br />
, (11)<br />
where K is a positive constant. ρ +,s<br />
v thus goes to 1 as v goes to infinity with 1 − ρ +,s<br />
v ∝ 1/v2 , which shows<br />
th<strong>at</strong> they can have similar behavior.<br />
1.4 Conditional correl<strong>at</strong>ion coefficient on both variables as an altern<strong>at</strong>ive?<br />
Since our intensive explor<strong>at</strong>ion of the behavior of the correl<strong>at</strong>ion coefficient conditionned on only one variable<br />
clearly indic<strong>at</strong>es th<strong>at</strong> it exhibits any kind of behavior, it is n<strong>at</strong>ural to look for the effect of a more<br />
8
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 247<br />
constraining conditioning. To this aim, we consi<strong>de</strong>r two random variables X and Y and <strong>de</strong>fine their conditional<br />
correl<strong>at</strong>ion coefficient ρA,B, conditioned upon X ∈ A and Y ∈ B, where A and B are two subs<strong>et</strong>s of<br />
R such th<strong>at</strong> Pr{X ∈ A, Y ∈ B} > 0, by<br />
ρA,B =<br />
Cov(X, Y | X ∈ A, Y ∈ B)<br />
Var(X | X ∈ A, Y ∈ B) · Var(Y | X ∈ A, Y ∈ B) . (12)<br />
In this case, it is much more difficult to obtain general results for any specified class of distributions compared<br />
to the previous case of conditioning on a single variable. We have only been able to give the asymptotic<br />
behavior for a Gaussian distribution in the situ<strong>at</strong>ion d<strong>et</strong>ailed below, using the expressions in (Johnson and<br />
Kotz 1972, p 113) or proposition A.1 of (Ang and Chen 2001).<br />
L<strong>et</strong> us assume th<strong>at</strong> the pair of random variables (X,Y) has a Normal distribution with unit unconditional variance<br />
and unconditional correl<strong>at</strong>ion coefficient ρ. The subs<strong>et</strong>s A and B are both choosen equal to [u, +∞),<br />
with u ∈ R+, so th<strong>at</strong> we focus on the correl<strong>at</strong>ion coefficient conditional on the r<strong>et</strong>urns of both X and Y<br />
larger than the threshold u. Denoting by ρu the correl<strong>at</strong>ion coefficient conditional on this particular choice<br />
for the subs<strong>et</strong>s A and B, we are able to show (see eppendix A.2) th<strong>at</strong>, for large u:<br />
ρu ∼u→∞ ρ<br />
1 + ρ<br />
1 − ρ<br />
1<br />
· , (13)<br />
u2 which goes to zero. This <strong>de</strong>cay is faster than in the case governed by (3) resulting from the conditioning on<br />
a single variable leading to ρ + v ∼v→+∞ 1/v, but, unfortun<strong>at</strong>ely, we do not observe a qualit<strong>at</strong>ive change.<br />
Thus, the correl<strong>at</strong>ion coefficient conditioned on both variables does not yield new significant inform<strong>at</strong>ion<br />
and does not provi<strong>de</strong> any special improvement with respect to the correl<strong>at</strong>ion coefficient conditioned on a<br />
single variable.<br />
1.5 Empirical evi<strong>de</strong>nce<br />
We consi<strong>de</strong>r four n<strong>at</strong>ional stock mark<strong>et</strong>s in L<strong>at</strong>in America, namely Argentina (MERVAL in<strong>de</strong>x), Brazil<br />
(IBOV in<strong>de</strong>x), Chile (IPSA in<strong>de</strong>x) and Mexico (MEXBOL in<strong>de</strong>x). We are particularly interested in the<br />
contagion effects which may have occurred across these mark<strong>et</strong>s. We will study this question for the mark<strong>et</strong><br />
in<strong>de</strong>xes expressed in US Dollar to emphasize the effect of the <strong>de</strong>valu<strong>at</strong>ions of local currencies and so to<br />
account for mon<strong>et</strong>ary crises. Doing so, we follow the same m<strong>et</strong>hodology as in most contagion papers (see<br />
(Forbes and Rigobon 2002), for instance). Our sample contains the daily (log) r<strong>et</strong>urns of each stock in<br />
local currency and US dollar during the time interval from January 15, 1992 to June 15, 2002 and thus<br />
encompasses both the Mexican crisis as well as the current Argentina crisis.<br />
Before applying the theor<strong>et</strong>ical results <strong>de</strong>rived above, we first need to check wh<strong>et</strong>her we are allowed to do<br />
so. Namely, we have to test wh<strong>et</strong>her the in<strong>de</strong>x r<strong>et</strong>urns distributions are not too f<strong>at</strong> tailed. In<strong>de</strong>ed, it its well<br />
known th<strong>at</strong> the correl<strong>at</strong>ion coefficient exists if and only if the tail of the distribution <strong>de</strong>cays faster than a<br />
power law with tail in<strong>de</strong>x α = 2, and its estim<strong>at</strong>or given by the Pearson’s coefficient is well behaved if <strong>at</strong><br />
least the fourth moment of the distribution is finite.<br />
Figure 1 represents the complementary distribution of the positive and neg<strong>at</strong>ive tails of the in<strong>de</strong>x r<strong>et</strong>urns in<br />
US dollar. We observe th<strong>at</strong> the positive tail clearly <strong>de</strong>cays faster than a power law with tail in<strong>de</strong>x µ = 2.<br />
In fact, Hill’s estim<strong>at</strong>or provi<strong>de</strong>s a value ranging b<strong>et</strong>ween 3 and 4 for the four in<strong>de</strong>xes. The case of the<br />
neg<strong>at</strong>ive tail is slightly different, particularly for the Brazilian in<strong>de</strong>x. In<strong>de</strong>ed, for the Argentina, the Chilean<br />
and the Mexican in<strong>de</strong>xes, the neg<strong>at</strong>ive tail behaves almost like the positive one, but for the Brazilian in<strong>de</strong>x,<br />
the neg<strong>at</strong>ive tail exponent is hardly larger than two, as confirmed by Hill’s estim<strong>at</strong>or. This means th<strong>at</strong>, in<br />
9
248 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
the Brazilian case, the estim<strong>at</strong>es of the correl<strong>at</strong>ion coefficient will be particularly noisy and thus of weak<br />
st<strong>at</strong>istical value.<br />
We have checked th<strong>at</strong> the f<strong>at</strong> tailness of the in<strong>de</strong>xes expressed in US dollar comes from the impact of the<br />
exchange r<strong>at</strong>es. Thus, an altern<strong>at</strong>ive should be to consi<strong>de</strong>r the in<strong>de</strong>xes in local currency, following (Longin<br />
and Solnik 1995) and (Longin and Solnik 2001)’s m<strong>et</strong>hodology, but it would lead to focus on the linkages<br />
b<strong>et</strong>ween mark<strong>et</strong>s only and to neglect the impact of the <strong>de</strong>valu<strong>at</strong>ions, which is precisely the main concern of<br />
many studies about the contagion in L<strong>at</strong>in America.<br />
Figures 2, 4 and 6 give the conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v<br />
(plain thick line) for the pairs Ar-<br />
gentina / Brazil, Brazil / Chile and Chile / Mexico while figures 3, 5 and 7 show the conditional correl<strong>at</strong>ion<br />
coefficient ρ s v for the same pairs. For each figure, the dashed thick line gives the theor<strong>et</strong>ical curve obtained<br />
un<strong>de</strong>r the bivari<strong>at</strong>e Gaussian assumption whose analytical expressions can be found in appendices A.1.1<br />
and A.1.2. The unconditional correl<strong>at</strong>ion coefficient of the Gaussian mo<strong>de</strong>l is s<strong>et</strong> to the empirically estim<strong>at</strong>ed<br />
unconditional correl<strong>at</strong>ion coefficent. The two dashed thin lines represent the interval within which<br />
we cannot reject, <strong>at</strong> the 95% confi<strong>de</strong>nce level, the hypothesis according to which the estim<strong>at</strong>ed conditional<br />
correl<strong>at</strong>ion coefficient is equal to the theor<strong>et</strong>ical one. This confi<strong>de</strong>nce interval has been estim<strong>at</strong>ed using the<br />
Fisher’s st<strong>at</strong>istics. Similarly, the thick dotted curve graphs the theor<strong>et</strong>ical curve obtained un<strong>de</strong>r the bivari<strong>at</strong>e<br />
Stu<strong>de</strong>nt’s assumption with ν = 3 <strong>de</strong>grees of freedom (whose expressions are given in appendices B.3<br />
and B.4) and the two thin dotted lines are its 95% confi<strong>de</strong>nce level. Here, the Fisher’s st<strong>at</strong>istics cannot be<br />
applied, since it requires <strong>at</strong> least th<strong>at</strong> the fourth moment of the distribution exists. In fact, (Meerschaert and<br />
Scheffler 2001) have shown th<strong>at</strong>, in such a case, the distribution of the sample correl<strong>at</strong>ion converges to a<br />
stable law with in<strong>de</strong>x 3/2, which justifies why the confi<strong>de</strong>nce interval for the Stu<strong>de</strong>nt’s mo<strong>de</strong>l with three<br />
<strong>de</strong>gres of freedom is much larger than the confi<strong>de</strong>nce interval for the Gaussian mo<strong>de</strong>l. In the present study,<br />
we have used a bootstrap m<strong>et</strong>hod to <strong>de</strong>rive this confi<strong>de</strong>nce interval since the scale factor intervening in the<br />
stable law is difficult to calcul<strong>at</strong>e.<br />
In figures 2, 4 and 6, we observe th<strong>at</strong> the change in the conditional correl<strong>at</strong>ion coefficients ρ +,−<br />
v<br />
are not sig-<br />
nificantly different, <strong>at</strong> the 95% confi<strong>de</strong>nce level, from those obtained with a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s mo<strong>de</strong>l with<br />
three <strong>de</strong>grees of freedom. In contrast, the Gaussian mo<strong>de</strong>l is often rejected. In fact, similar results hold (but<br />
are not <strong>de</strong>picted here) for the three others pairs Argentina / Chile, Argentina / Mexico and Brazil / Mexico.<br />
Thus, these observ<strong>at</strong>ions should lead us to conclu<strong>de</strong> th<strong>at</strong>, in these cases, no change in the correl<strong>at</strong>ions, and<br />
therefore no contagion mechanism, needs to be invoked to explain the d<strong>at</strong>a, since they are comp<strong>at</strong>ible with<br />
a Stu<strong>de</strong>nt’s mo<strong>de</strong>l with constant correl<strong>at</strong>ion.<br />
L<strong>et</strong> us now discuss the results obtained for the correl<strong>at</strong>ion coefficient conditioned on the vol<strong>at</strong>ility. Figures<br />
3 and 7 show th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficients conditioned on vol<strong>at</strong>ility remain consistent with the<br />
Stu<strong>de</strong>nt’s mo<strong>de</strong>l with three <strong>de</strong>gres of freedom, while they still reject the Gaussian mo<strong>de</strong>l. In contrast, figure<br />
5 shows th<strong>at</strong> the increase of the correl<strong>at</strong>ion cannot be explained by any of the Gaussian or Stu<strong>de</strong>nt mo<strong>de</strong>ls,<br />
when conditioning on the Mexican in<strong>de</strong>x vol<strong>at</strong>ilty. In<strong>de</strong>ed, when the Mexican in<strong>de</strong>x vol<strong>at</strong>ility becomes larger<br />
than 2.5 times its standard <strong>de</strong>vi<strong>at</strong>ion, none of our mo<strong>de</strong>ls can account for the increase of the correl<strong>at</strong>ion. The<br />
same discrepancy is observed for the pairs Argentina / Chile, Argentina / Mexico and Brazil / Mexico which<br />
have not been represented here. In each case, the Chilean or the Mexican mark<strong>et</strong> have an impact on the<br />
Argentina or the Brazilian mark<strong>et</strong> which cannot be accounted for by neither the Gaussian or Stu<strong>de</strong>nt mo<strong>de</strong>ls<br />
with constant correl<strong>at</strong>ion.<br />
To conclu<strong>de</strong> this empirical part, there is no significant increase in the real correl<strong>at</strong>ion b<strong>et</strong>ween Argentina<br />
and Brazil in the one hand and b<strong>et</strong>ween Chile and Mexico in the other hand, when the vol<strong>at</strong>ility or the<br />
r<strong>et</strong>urns exhibit large moves. But, in period of high vol<strong>at</strong>ility, the Chilean and Mexican mark<strong>et</strong> may have an<br />
genuine impact on the Argentina and Brazilian mark<strong>et</strong>s. Thus, a priori, this should lead us to conclu<strong>de</strong> on<br />
the existence of a contagion across these mark<strong>et</strong>s. However, this conclusion is based on the investig<strong>at</strong>ion of<br />
10
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 249<br />
only two theor<strong>et</strong>ical mo<strong>de</strong>ls and it would be a little bit too hasty to conclu<strong>de</strong> on the existence of contagion<br />
on the sole basis of these results. This is all the more so since our theor<strong>et</strong>ical mo<strong>de</strong>ls are all symm<strong>et</strong>ric in<br />
their positive and neg<strong>at</strong>ive tails, a crucial property nee<strong>de</strong>d for the <strong>de</strong>riv<strong>at</strong>ion of the expressions of ρ s v, while<br />
the empirical sample distributions are certainly not symm<strong>et</strong>ric, as shown in figure 1.<br />
1.6 Summary<br />
The previous sections have shown th<strong>at</strong> the conditional correl<strong>at</strong>ion coefficients can exhibit any behavior,<br />
<strong>de</strong>pending on their conditioning s<strong>et</strong> and the un<strong>de</strong>rlying distributions of r<strong>et</strong>urns. More precisely, we have<br />
shown th<strong>at</strong> the correl<strong>at</strong>ion coefficients, conditioned on large r<strong>et</strong>urns or vol<strong>at</strong>ility above a threshold v, can be<br />
either increasing or <strong>de</strong>creasing functions of the threshold, can go to any value b<strong>et</strong>ween zero and one when<br />
the threshold goes to infinity and can produce contradictory results in the sense th<strong>at</strong> accounting for a trend or<br />
not can lead to conclu<strong>de</strong> on an absence of linear correl<strong>at</strong>ion or on a perfect linear correl<strong>at</strong>ion. Moreover, due<br />
to the large st<strong>at</strong>istical fluctu<strong>at</strong>ions of the empirical estim<strong>at</strong>es, one should be very careful when concluding<br />
on an increase or <strong>de</strong>crease of the genuine correl<strong>at</strong>ions.<br />
Thus, from the general standpoint of the study of extreme <strong>de</strong>pen<strong>de</strong>nces, but more particularly for the specific<br />
problem of the contagion across countries, the use of conditional correl<strong>at</strong>ion does not seem very inform<strong>at</strong>ive<br />
and is som<strong>et</strong>imes misleading since it leads to spurious changes in the observed correl<strong>at</strong>ions: even when<br />
the unconditional correl<strong>at</strong>ion remains constant, conditional correl<strong>at</strong>ions yield artificial changes as we have<br />
forcefully shown. Since one of the most commonly accepted and used <strong>de</strong>finition of contagion is the d<strong>et</strong>ection<br />
of an increase of the conditional correl<strong>at</strong>ions during a period of turmoil, namely when the vol<strong>at</strong>ility increases,<br />
our results cast serious shadows on previous results. In this respect, the conclusions of (Calvo and Reinhart<br />
1996) about the occurrence of contagion across L<strong>at</strong>in American mark<strong>et</strong>s during the 1994 Mexican crisis but<br />
more generally also the results of (King and Wadhwani 1990), or (Lee and Kim 1993) on the effect of the<br />
October 1987 crash on the linkage of n<strong>at</strong>ional mark<strong>et</strong>s, must be consi<strong>de</strong>red with some caution. It is quite<br />
<strong>de</strong>sirable to find a more reliable tool for studying extreme <strong>de</strong>pen<strong>de</strong>nces.<br />
2 Conditional concordance measures<br />
The (conditional) correl<strong>at</strong>ion coefficients, which have just been investig<strong>at</strong>ed, suffer from several theor<strong>et</strong>ical<br />
as well as empirical <strong>de</strong>ficiencies. From the theor<strong>et</strong>ical point of view, it is only a measure of linear <strong>de</strong>pen<strong>de</strong>nce.<br />
Thus, as stressed by (Embrechts <strong>et</strong> al. 1999), it is fully s<strong>at</strong>isfying only for the <strong>de</strong>scription of the<br />
<strong>de</strong>pen<strong>de</strong>nce of variables with elliptical distributions. Moreover, the correl<strong>at</strong>ion coefficient aggreg<strong>at</strong>es the<br />
inform<strong>at</strong>ion contained both in the marginal and in the collective behavior. The correl<strong>at</strong>ion coefficient is<br />
not invariant un<strong>de</strong>r an increasing change of variable, a transform<strong>at</strong>ion which is known to l<strong>et</strong> unchanged the<br />
<strong>de</strong>pen<strong>de</strong>nce structure. From the empirical standpoint, we have seen th<strong>at</strong>, for some consi<strong>de</strong>red d<strong>at</strong>a, the correl<strong>at</strong>ion<br />
coefficient may not always exist, and even when it exits, it cannot always be accur<strong>at</strong>ly estim<strong>at</strong>ed, due<br />
to som<strong>et</strong>imes “wild” st<strong>at</strong>istical fluctu<strong>at</strong>ions. Thus, it is <strong>de</strong>sirable to find another measure of the <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween two ass<strong>et</strong>s or more generally b<strong>et</strong>ween two random variables, which, contrarily to the linear correl<strong>at</strong>ion<br />
coefficient, is always well-<strong>de</strong>fined and only <strong>de</strong>pends on the copula properties. This ensures th<strong>at</strong> this<br />
measure is not affected by a change in the marginal distributions (provi<strong>de</strong>d th<strong>at</strong> the mapping is increasing).<br />
It turns out th<strong>at</strong> this <strong>de</strong>sirable property is shared by all measures of concordance. Among these measures<br />
are the well-known Kendall’s tau, Spearman’s rho or Gini’s b<strong>et</strong>a (see (Nelsen 1998) for d<strong>et</strong>ails).<br />
However, these concordance measures are not well-adapted, as such, to the study of extreme <strong>de</strong>pen<strong>de</strong>nce,<br />
because they are functions of the whole distribution, including the mo<strong>de</strong>r<strong>at</strong>e and small r<strong>et</strong>urns. A simple i<strong>de</strong>a<br />
11
250 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
to investig<strong>at</strong>e the extreme concordance properties of two random variables is to calcul<strong>at</strong>e these quantities<br />
conditioned on values larger than a given threshold and l<strong>et</strong> this threshold go to infinity.<br />
In the sequel, we will only focus on the Spearman’s rho which can be easily estim<strong>at</strong>ed empirically. It offers a<br />
n<strong>at</strong>ural generaliz<strong>at</strong>ion of the (linear) correl<strong>at</strong>ion coefficient. In<strong>de</strong>ed, the correl<strong>at</strong>ion coefficient quantifies the<br />
<strong>de</strong>gree of linear <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random variables, while the Spearman’s rho quantifies the <strong>de</strong>gree<br />
of functional <strong>de</strong>pen<strong>de</strong>nce, wh<strong>at</strong>ever the functional <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two random variables may be.<br />
This represents a very interesting improvement. Perfect correl<strong>at</strong>ions (respectively anti-correl<strong>at</strong>ion) give a<br />
value 1 (respectively −1) both for the standard correl<strong>at</strong>ion coefficient and for the Spearman’s rho. Otherwise,<br />
there is no general rel<strong>at</strong>ion allowing us to <strong>de</strong>duce the Spearman’s rho from the correl<strong>at</strong>ion coefficient and<br />
vice-versa.<br />
2.1 Definition<br />
The Spearman’s rho, <strong>de</strong>noted ρs in the sequel, measures the difference b<strong>et</strong>ween the probability of concordance<br />
and the probability of discordance for the two pairs of random variables (X1, Y1) and (X2, Y3),<br />
where the pairs (X1, Y1), (X2, Y2) and (X3, Y3) are three in<strong>de</strong>pen<strong>de</strong>nt realiz<strong>at</strong>ions drawn from the same<br />
distribution:<br />
ρs = 3 (Pr[(X1 − X2)(Y1 − Y3) > 0] − Pr[(X1 − X2)(Y1 − Y3) < 0]) . (14)<br />
The Spearman’s rho can also be expressed with the copula C of the two variables X and Y (see (Nelsen<br />
1998), for instance):<br />
ρs = 12<br />
1 1<br />
0<br />
0<br />
C(u, v) du dv − 3, (15)<br />
which allows us to easily calcul<strong>at</strong>e ρs when the copula C is known in closed form.<br />
Denoting U = FX(X) and V = FY (V ), it is easy to show th<strong>at</strong> ρs is nothing but the (linear) correl<strong>at</strong>ion<br />
coefficient of the uniform random variables U and V :<br />
Cov(U, V )<br />
ρs = . (16)<br />
Var(U)Var(V )<br />
This justifies its name as a correl<strong>at</strong>ion coefficient of the rank, and shows th<strong>at</strong> it can easily be estim<strong>at</strong>ed.<br />
An <strong>at</strong>tractive fe<strong>at</strong>ure of the Spearman’s rho is to be in<strong>de</strong>pen<strong>de</strong>nt of the margins, as we can see in equ<strong>at</strong>ion<br />
(15). Thus, contrarily to the linear correl<strong>at</strong>ion coefficient, which aggreg<strong>at</strong>es the marginal properties<br />
of the variables with their collective behavior, the rank correl<strong>at</strong>ion coefficient takes into account only the<br />
<strong>de</strong>pen<strong>de</strong>nce structure of the variables.<br />
Using expression (16), we propose a n<strong>at</strong>ural <strong>de</strong>finition of the conditional rank correl<strong>at</strong>ion, conditioned on V<br />
larger than a given threshold ˜v:<br />
ρs(˜v) =<br />
whose expression in term of the copula C(·, ·) is given in appendix D.<br />
2.2 Example<br />
Cov(U, V | V ≥ ˜v)<br />
Var(U | V ≥ ˜v)Var(V | V ≥ ˜v) , (17)<br />
Contrarily to the conditional correl<strong>at</strong>ion coefficient, we have not been able to obtain analytical expressions<br />
for the conditional Spearman’s rho, <strong>at</strong> least for the distributions th<strong>at</strong> we have consi<strong>de</strong>red up to now. Obviously,<br />
for many families of copulas known in closed form, equ<strong>at</strong>ion (17) allows for an explicit calcul<strong>at</strong>ion<br />
12
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 251<br />
of ρs(v). However, most copulas of interest in finance have no simple closed form, so th<strong>at</strong> it is necessary to<br />
resort to numerical comput<strong>at</strong>ions.<br />
As an example, l<strong>et</strong> us consi<strong>de</strong>r the bivari<strong>at</strong>e Gaussian distribution (or copula) with unconditional correl<strong>at</strong>ion<br />
coefficent ρ. It is well-known th<strong>at</strong> its unconditional Spearman’s rho is given by<br />
ρs = 6 ρ<br />
· arcsin . (18)<br />
π 2<br />
The left panel of figure 8 shows the conditional Spearman’s rho ρs(v) <strong>de</strong>fined by (17) obtained from a<br />
numerical integr<strong>at</strong>ion. We observe the same bias as for the conditional correl<strong>at</strong>ion coefficient, namely the<br />
conditional rank correl<strong>at</strong>ion changes with v eventhough the unconditional correl<strong>at</strong>ion is fixed to a constant<br />
value. Non<strong>et</strong>heless, this conditional Spearman’s rho seems more sensitive than the conditional correl<strong>at</strong>ion<br />
coefficient since we can observe in the left panel of figure 8 th<strong>at</strong>, as v goes to one, the conditional Spearman’s<br />
rho ρs(v) does not go to zero for all values of ρ (<strong>at</strong> the precision of our bootstrap estim<strong>at</strong>es), as previously<br />
observed with the conditional correl<strong>at</strong>ion coefficient (see equ<strong>at</strong>ion (3)).<br />
The right panel of figure 8 <strong>de</strong>picts the conditional Spearman’s rho of the Stu<strong>de</strong>nt’s copula with three <strong>de</strong>gres<br />
of freedom. The results are the same concerning the bias, but this time ρs(v) goes to zero for all value<br />
of ρ when v goes to one. Thus, here again, many different behaviours can be observed <strong>de</strong>pending on the<br />
un<strong>de</strong>rlying copula of the random variables. Moreover, these two examples show th<strong>at</strong> the quantific<strong>at</strong>ion of<br />
extreme <strong>de</strong>pen<strong>de</strong>nce is a function of the tools used to quantify this <strong>de</strong>pen<strong>de</strong>nce. Here, the conditional Spearman’s<br />
ρ goes to a non-vanishing constant for the Gaussian mo<strong>de</strong>l, while the conditional (linear) correl<strong>at</strong>ion<br />
coefficient goes to zero, contrarily to the Stu<strong>de</strong>nt’s distribution for which the situ<strong>at</strong>ion is exactly the opposite.<br />
2.3 Empirical evi<strong>de</strong>nce<br />
Figures 9, 10 and 11 give the conditionnal Spearman’s rho respectively for the Argentina / Brazilian stock<br />
mark<strong>et</strong>s, the Brazilian / Chilean stock mark<strong>et</strong>s and the Chilean / Mexican stock mark<strong>et</strong>s. As previously, the<br />
plain thick line refers to the estim<strong>at</strong>ed correl<strong>at</strong>ion, while the dashed lines refer to the Gaussian copula and<br />
its 95% confi<strong>de</strong>nce levels and and dotted lines to the Stu<strong>de</strong>nt’s copula with three <strong>de</strong>grees of freedom and its<br />
95% confi<strong>de</strong>nce levels.<br />
We first observe th<strong>at</strong> contrarily to the cases of the conditional (linear) correl<strong>at</strong>ion coefficient exhibited in<br />
figures 2, 4 and 6, the empirical conditional Spearman’s ρ does not always comply with the Stu<strong>de</strong>nt’s mo<strong>de</strong>l<br />
(neither with the Gaussian one), and thus confirm the discrepancies observed in figures 3, 5 and 7. In all<br />
cases, for thresholds v larger than the quantile 0.5 corresponding to the positive r<strong>et</strong>urns, the Stu<strong>de</strong>nt’s mo<strong>de</strong>l<br />
with three <strong>de</strong>grees of freedom is always sufficient to explain the d<strong>at</strong>a. In contrast, for the neg<strong>at</strong>ive r<strong>et</strong>urns<br />
and thus thresholds v lower then the quantile 0.5, only the interaction b<strong>et</strong>ween the Chilean and the Mexican<br />
mark<strong>et</strong>s is well <strong>de</strong>scribed by the Stu<strong>de</strong>nt copula and does not nee<strong>de</strong>d any additional ingredient such as the<br />
contagion mechanism. For all other pairs, none of our mo<strong>de</strong>ls explain the d<strong>at</strong>a s<strong>at</strong>isfyingly. Therefore, for<br />
these cases and from the perspective of our mo<strong>de</strong>ls, the contagion hypothesis seems to be nee<strong>de</strong>d.<br />
There are however several cave<strong>at</strong>s. First, even though we have consi<strong>de</strong>red the most n<strong>at</strong>ural financial mo<strong>de</strong>ls,<br />
there may be other mo<strong>de</strong>ls, th<strong>at</strong> we have ignored, with constant <strong>de</strong>pen<strong>de</strong>nce structure which can account for<br />
the observed evolutions of the conditional Spearman’s ρ. If this is true, then the contagion hypothesis would<br />
not be nee<strong>de</strong>d. Second, the discrepancy b<strong>et</strong>ween the empirical conditional Spearman’s ρ and the prediction<br />
of the the Stu<strong>de</strong>nt’s mo<strong>de</strong>l does not occur in the tails the distribution, i.e for large and extreme movements,<br />
but in the bulk. Thus, during periods of turmoil, the Stu<strong>de</strong>nt’s mo<strong>de</strong>l with three <strong>de</strong>grees fo freedom seems to<br />
remain a good mo<strong>de</strong>l of co-movements. Third, the contagion effect is never necessary for upwards moves.<br />
In<strong>de</strong>ed, we observe the same asymm<strong>et</strong>ry or trend <strong>de</strong>pen<strong>de</strong>nce as found by (Longin and Solnik 2001) for five<br />
13
252 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
major equity mark<strong>et</strong>s. This was apparent in figures 2, 4 and 6 for ρ +,−<br />
v , and is strongly confirmed on the<br />
conditional Spearman’s ρ.<br />
Interestingly, there is also an asymm<strong>et</strong>ry or directivity in the mutual influence b<strong>et</strong>ween mark<strong>et</strong>s. For instance,<br />
the Chilean and Mexican mark<strong>et</strong>s have an influence on the Argentina and Brazilian mark<strong>et</strong>s, but the l<strong>at</strong>er do<br />
not have any impact on the Chile and Mexican mark<strong>et</strong>s. Chile and Mexico have no contagion effect on each<br />
other while Argentina and Brazil have.<br />
These empirical results on the conditional Spearman’s ρ are different from and often opposite to the conclu-<br />
sion <strong>de</strong>rived from the conditional correl<strong>at</strong>ion coefficients ρ +,−<br />
v . This puts in light the difficulty in obtaining<br />
reliable, unambiguous and sensitive estim<strong>at</strong>ions of conditional correl<strong>at</strong>ion measures. In particular, the Pearson’s<br />
coefficient usually employed to estim<strong>at</strong>e the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween two variables is known<br />
to be not very efficient when the variables are f<strong>at</strong>-tailed and when the estim<strong>at</strong>ion is performed on a small<br />
sample. In<strong>de</strong>ed, with small samples, the Pearson’s coefficient is very sensitive to the largest value, which<br />
can lead to an important bias in the estim<strong>at</strong>ion. Moreover, even with large sample sizes, (Meerschaert and<br />
Scheffler 2001) have shown th<strong>at</strong> the n<strong>at</strong>ure of convergence as the sample size T tends to infinity of the Pearson’s<br />
coefficient of two times series with tail in<strong>de</strong>x µ towards the theor<strong>et</strong>ical correl<strong>at</strong>ion is sensitive to the<br />
existence and strength of the theor<strong>et</strong>ical correl<strong>at</strong>ion. If there is no theor<strong>et</strong>ical correl<strong>at</strong>ion b<strong>et</strong>ween the two<br />
times series, the sample correl<strong>at</strong>ion tends to zero with Gaussian fluctu<strong>at</strong>ions. If the theor<strong>et</strong>ical correl<strong>at</strong>ion is<br />
non-zero, the difference b<strong>et</strong>ween the sample correl<strong>at</strong>ion and the theor<strong>et</strong>ical correl<strong>at</strong>ion times T 1−2/µ converges<br />
in distribution to a stable law with in<strong>de</strong>x µ/2. These large st<strong>at</strong>istical fluctu<strong>at</strong>ions are responsible for<br />
the lack of accuracy of the estim<strong>at</strong>ed conditional correl<strong>at</strong>ion coefficient encountered in the previous section.<br />
Thus, we think th<strong>at</strong> the conditional Spearman’s ρ provi<strong>de</strong>s a good altern<strong>at</strong>ive both from a theor<strong>et</strong>ical and an<br />
empirical viewpoint.<br />
3 Tail <strong>de</strong>pen<strong>de</strong>nce<br />
For the sake of compl<strong>et</strong>eness, and since it is directly rel<strong>at</strong>ed to the multivari<strong>at</strong>e extreme values theory, we<br />
study the so-called coefficient of tail <strong>de</strong>pen<strong>de</strong>nce λ. To our knowledge, its interest for financial applic<strong>at</strong>ions<br />
has been first un<strong>de</strong>rlined by (Embrechts <strong>et</strong> al. 2001).<br />
The coefficient of tail <strong>de</strong>pen<strong>de</strong>nce characterizes an important property of the extreme <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
X and Y , using the (original or unconditional) copula of X and Y . In constrast, the conditional spearman’s<br />
rho is <strong>de</strong>fined in terms of a conditional copula, and can be seen as the “unconditional Spearman’s rho” of the<br />
copula of X and Y conditioned on Y larger than the threshold v. This copula of X and Y conditioned on<br />
Y larger than the threshold v is not the true copula of X and Y because it is modified by the conditioning.<br />
In this sense, the tail <strong>de</strong>pen<strong>de</strong>nce param<strong>et</strong>er λ is a more n<strong>at</strong>ural property directly rel<strong>at</strong>ed to the copula of X<br />
and Y .<br />
To begin with, we recall the <strong>de</strong>finition of the coefficient λ as well as of ¯ λ (see below) which allows one to<br />
quantify the amount of <strong>de</strong>pen<strong>de</strong>nce in the tail. Then, we present several results concerning the coefficient λ<br />
of tail <strong>de</strong>pen<strong>de</strong>nce for various distributions and mo<strong>de</strong>ls, and finally, we discuss the problems encountered in<br />
the estim<strong>at</strong>ion of these quantities.<br />
14
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 253<br />
3.1 Definition<br />
The concept of tail <strong>de</strong>pen<strong>de</strong>nce is appealing by its simplicity. By <strong>de</strong>finition, the (upper) tail <strong>de</strong>pen<strong>de</strong>nce<br />
coefficient is:<br />
−1<br />
λ = lim<br />
(u)|Y > F (u)] , (19)<br />
Pr[X > F<br />
u→1 −1<br />
X<br />
and quantifies the probability to observe a large X, assuming th<strong>at</strong> Y is large itself. For a survey of the<br />
properties of the tail <strong>de</strong>pen<strong>de</strong>nce coefficient, the rea<strong>de</strong>r is refered to (Coles <strong>et</strong> al. 1999, Embrechts <strong>et</strong> al.<br />
2001, Lindskog 1999), for instance. In words, given th<strong>at</strong> Y is very large (which occurs with probability<br />
1 − u), the probability th<strong>at</strong> X is very large <strong>at</strong> the same probability level u <strong>de</strong>fines asymptotically the tail<br />
<strong>de</strong>pen<strong>de</strong>nce coefficient λ. As an example, consi<strong>de</strong>ring th<strong>at</strong> X and Y represent the vol<strong>at</strong>ility of two different<br />
n<strong>at</strong>ional mark<strong>et</strong>s, the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce λ gives the probabilty th<strong>at</strong> both mark<strong>et</strong>s exhibit tog<strong>et</strong>her<br />
very high vol<strong>at</strong>ility.<br />
One of the appeal of this <strong>de</strong>finition of tail <strong>de</strong>pen<strong>de</strong>nce is th<strong>at</strong> it is a pure copula property, i.e., it is in<strong>de</strong>pen<strong>de</strong>nt<br />
of the margins of X and Y . In<strong>de</strong>ed, l<strong>et</strong> C be the copula of the variables X and Y , then if the bivari<strong>at</strong>e copula<br />
C is such th<strong>at</strong><br />
Y<br />
1 − 2u + C(u, u)<br />
log C(u, u)<br />
lim<br />
= lim 2 − = λ (20)<br />
u→1 1 − u<br />
u→1 log u<br />
exists, then C has an upper tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ (see (Coles <strong>et</strong> al. 1999, Embrechts <strong>et</strong> al. 2001,<br />
Lindskog 1999)).<br />
If λ > 0, the copula presents tail <strong>de</strong>pen<strong>de</strong>nce and large events tend to occur simultaneously, with the<br />
probability λ. On the contrary, when λ = 0, the copula has no tail <strong>de</strong>pen<strong>de</strong>nce and the variables X and Y<br />
are said asymptotically in<strong>de</strong>pen<strong>de</strong>nt. There is however a subtl<strong>et</strong>y in this <strong>de</strong>finition (19) of tail <strong>de</strong>pen<strong>de</strong>nce.<br />
To make it clear, first consi<strong>de</strong>r the case where for large X and Y the cumul<strong>at</strong>ive distribution function H(x, y)<br />
factorizes such th<strong>at</strong><br />
F (x, y)<br />
lim<br />
= 1 , (21)<br />
x,y→∞ FX(x)FY (y)<br />
where FX(x) and FY (y) are the margins of X and Y respectively. This means th<strong>at</strong>, for X and Y sufficiently<br />
large, these two variables can be consi<strong>de</strong>red as in<strong>de</strong>pen<strong>de</strong>nt. It is then easy to show th<strong>at</strong><br />
lim<br />
u→1 Pr{X > FX −1 (u)|Y > FY −1 (u)} = lim 1 − FX(FX<br />
u→1 −1 (u)) (22)<br />
= lim<br />
u→1 1 − u = 0, (23)<br />
so th<strong>at</strong> in<strong>de</strong>pen<strong>de</strong>nt variables really have no tail <strong>de</strong>pen<strong>de</strong>nce λ = 0, as one can expect.<br />
However, the result λ = 0 does not imply th<strong>at</strong> the multivari<strong>at</strong>e distribution can be autom<strong>at</strong>ically factorized<br />
asymptotically, as shown by the Gaussian example. In<strong>de</strong>ed, the Gaussian multivari<strong>at</strong>e distribution does not<br />
have a factorizable multivari<strong>at</strong>e distribution, even asymptotically for extreme values, since the non-diagonal<br />
term of the quadr<strong>at</strong>ic form in the exponential function does not become negligible in general as X and Y go<br />
to infinity. Therefore, in a weaker sense, there may still be a <strong>de</strong>pen<strong>de</strong>nce in the tail even when λ = 0.<br />
To make this st<strong>at</strong>ement more precise, following (Coles <strong>et</strong> al. 1999), l<strong>et</strong> us introduce the coefficient<br />
¯λ = lim<br />
u→1<br />
2 log Pr{X > FX −1 (u)}<br />
log Pr{X > FX −1 (u), Y > FY −1 (u)}<br />
− 1 (24)<br />
2 log(1 − u)<br />
= lim<br />
− 1 . (25)<br />
u→1 log[1 − 2u + C(u, u)]<br />
It can be shown th<strong>at</strong> the coefficient ¯ λ = 1 if and only if the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce λ > 0, while ¯ λ<br />
takes values in [−1, 1) when λ = 0, allowing us to quantify the strength of the <strong>de</strong>pen<strong>de</strong>nce in the tail in such<br />
15
254 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
a case. In fact, it has been established th<strong>at</strong>, when ¯ λ > 0, the variables X and Y are simultaneously large more<br />
frequently than in<strong>de</strong>pen<strong>de</strong>nt variables, while simultaneous large <strong>de</strong>vi<strong>at</strong>ions of X and Y occur less frequenlty<br />
than un<strong>de</strong>r in<strong>de</strong>pen<strong>de</strong>nce when ¯ λ < 0 (the interested rea<strong>de</strong>r is refered to (Ledford and Tawn 1996, Ledford<br />
and Tawn 1998)).<br />
To summarize, in<strong>de</strong>pen<strong>de</strong>nce (factoriz<strong>at</strong>ion of the bivari<strong>at</strong>e distribution) implies no tail <strong>de</strong>pen<strong>de</strong>nce λ = 0.<br />
But λ = 0 is not sufficient to imply factoriz<strong>at</strong>ion and thus true in<strong>de</strong>pen<strong>de</strong>nce. It also requires, as a necessary<br />
condition, th<strong>at</strong> ¯ λ = 0.<br />
We will first recall the expression of the tail <strong>de</strong>pen<strong>de</strong>nce coefficient for usual distributions, and then calcul<strong>at</strong>e<br />
it in the case of a one-factor mo<strong>de</strong>l for different distributions of the factor.<br />
3.2 Tail <strong>de</strong>pen<strong>de</strong>nce for Gaussian distributions and Stu<strong>de</strong>nt’s distributions<br />
Assuming th<strong>at</strong> (X, Y ) are normally distributed with correl<strong>at</strong>ion coefficient ρ, (Embrechts <strong>et</strong> al. 2001) shows<br />
th<strong>at</strong> for all ρ ∈ [−1, 1), λ = 0, while (Heffernan 2000) gives ¯ λ = ρ, which expresses, as one can expect,<br />
th<strong>at</strong> extremes appear more likely tog<strong>et</strong>her for positively correl<strong>at</strong>ed variables.<br />
In constrast, if (X, Y ) have a Stu<strong>de</strong>nt’s distribution, (Embrechts <strong>et</strong> al. 2001) shows th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce<br />
coefficient is<br />
λ = 2 · ¯ <br />
√ν 1 − ρ<br />
Tν+1 + 1 ,<br />
1 + ρ<br />
(26)<br />
which is gre<strong>at</strong>er than zero for all ρ > −1, and thus ¯ λ = 1. This last example proves th<strong>at</strong> extremes appear<br />
more likely tog<strong>et</strong>her wh<strong>at</strong>ever the correl<strong>at</strong>ion coefficient may be, showing th<strong>at</strong>, in fact, there is no general<br />
rel<strong>at</strong>ionship b<strong>et</strong>ween the asymptotic <strong>de</strong>pen<strong>de</strong>nce and the linear correl<strong>at</strong>ion coefficient.<br />
The Gaussian and Stu<strong>de</strong>nt’s distributions are elliptical, for which the following general result is known:<br />
(Hult and Lindskog 2001) have shown th<strong>at</strong> ellipticaly distributed random variables presents tail <strong>de</strong>pen<strong>de</strong>nce<br />
if and only if they are regularly varing, i.e., behaves asymptotically like power laws with some exponent<br />
ν > 0. In such a case, for every regularly varying pair of random variables elliptically distributed, the<br />
coefficent of tail <strong>de</strong>pen<strong>de</strong>nce λ is given by expression (26). This result is very n<strong>at</strong>ural since the correl<strong>at</strong>ion<br />
coefficient is an invariant quantity within the class of elliptical distributions and since the coefficient of tail<br />
<strong>de</strong>pen<strong>de</strong>nce is only d<strong>et</strong>ermined by the asymptotic behavior of the distribution, so th<strong>at</strong> it does not m<strong>at</strong>ter th<strong>at</strong><br />
the distribution is a Stu<strong>de</strong>nt’s distribution with ν <strong>de</strong>grees of freedom or any other elliptical distribution as<br />
long as they have the same asymptotic behavior in the tail.<br />
3.3 Tail <strong>de</strong>pen<strong>de</strong>nce gener<strong>at</strong>ed by a factor mo<strong>de</strong>l<br />
Consi<strong>de</strong>r the one-factor mo<strong>de</strong>l<br />
X1 = α1 · Y + ɛ1, (27)<br />
X2 = α2 · Y + ɛ2, (28)<br />
where the ɛi’s are random variables in<strong>de</strong>pen<strong>de</strong>nt of Y and the αi’s non-random positive coefficients.<br />
(Malevergne and Sorn<strong>et</strong>te 2002) have shown th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce λ of X1 and X2 can be simply expressed<br />
as the minimum of the tail <strong>de</strong>pen<strong>de</strong>nce coefficients λ1 and λ2 b<strong>et</strong>ween the two random variables X1<br />
and Y and X2 and Y respectively:<br />
λ = min{λ1, λ2}. (29)<br />
16
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 255<br />
To un<strong>de</strong>rstand this result, note th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween X1 and X2 is cre<strong>at</strong>ed only through the<br />
common factor Y . It is thus n<strong>at</strong>ural th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween X1 and X2 is boun<strong>de</strong>d from above<br />
by the weakest tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the Xi’s and Y while <strong>de</strong>riving the equality requires more work<br />
(Malevergne and Sorn<strong>et</strong>te 2002). Thus, it it only necessary to focus our study on the tail <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween any Xi and Y . So, in or<strong>de</strong>r to simplify the not<strong>at</strong>ions, we neglect the subscripts 1 or 2, since<br />
they are irrelevant for the <strong>de</strong>pen<strong>de</strong>nce of X1 (or X2) and Y .<br />
A general result concerning the tail <strong>de</strong>pen<strong>de</strong>nce gener<strong>at</strong>ed by factor mo<strong>de</strong>ls for every kind of factor and<br />
noise distributions can be found in (Malevergne and Sorn<strong>et</strong>te 2002). It has been proved th<strong>at</strong> the coefficient<br />
of (upper) tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween X and Y is given by<br />
∞<br />
where, provi<strong>de</strong>d th<strong>at</strong> they exist,<br />
λ =<br />
max{1, l<br />
α}<br />
l = lim<br />
u→1<br />
dx f(x) , (30)<br />
FX −1 (u)<br />
FY −1 , (31)<br />
(u)<br />
t · PY (t · x)<br />
f(x) = lim<br />
. (32)<br />
t→∞ ¯FY (t)<br />
As a direct consequence, one can show th<strong>at</strong> any rapidly varying factor, which encompasses the Gaussian, the<br />
exponential or the gamma distributed factors for instance, leads to a vanishing coefficient of tail <strong>de</strong>pen<strong>de</strong>nce,<br />
wh<strong>at</strong>ever the distribution of the idiosyncr<strong>at</strong>ic noise may be. This resut is obvious when both the factor and<br />
the idiosyncr<strong>at</strong>ic are Gaussianly distributed, since then X and Y follow a bivari<strong>at</strong>e Gaussian distibution,<br />
whose tail <strong>de</strong>pen<strong>de</strong>nce has been said to be zero.<br />
On the contrary, regularly vaying factors, like the Stu<strong>de</strong>nt’s distributed factors, lead to a tail <strong>de</strong>pen<strong>de</strong>nce,<br />
provi<strong>de</strong>d th<strong>at</strong> the distribution of the idiosycr<strong>at</strong>ic noise does not become f<strong>at</strong>ter-tailed than the factor distribution.<br />
One can thus conclu<strong>de</strong> th<strong>at</strong>, in or<strong>de</strong>r to gener<strong>at</strong>e tail <strong>de</strong>pen<strong>de</strong>nce, the factor must have a sufficiently<br />
‘wild’ distribution. To present an explicit example, l<strong>et</strong> us assume now th<strong>at</strong> the factor Y and the idiosyncr<strong>at</strong>ic<br />
noise ɛ have centered Stu<strong>de</strong>nt’s distributions with the same number ν of <strong>de</strong>grees of freedom and scale factors<br />
respectively equal to 1 and σ. The choice of the scale factor equal to 1 for Y is not restrictive but only<br />
provi<strong>de</strong>s a convenient normaliz<strong>at</strong>ion for σ. Appendix E shows th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce coefficient is<br />
1<br />
λ =<br />
1 + <br />
σ ν . (33)<br />
α<br />
As is reasonable intuitively, the larger the typical scale σ of the fluctu<strong>at</strong>ion of ɛ and the weaker is the coupling<br />
coefficient α, the smaller is the tail <strong>de</strong>pen<strong>de</strong>nce.<br />
L<strong>et</strong> us recall th<strong>at</strong> the unconditional correl<strong>at</strong>ion coefficient ρ can be writen as ρ = (1+ σ2<br />
α 2 ) −1/2 , which allows<br />
us to rewrite the coefficient of upper tail <strong>de</strong>pen<strong>de</strong>nce as<br />
λ =<br />
ρν ρν + (1 − ρ2 . (34)<br />
) ν/2<br />
Surprinsingly, λ does not go to zero for all ρ’s as ν goes to infinity, as one would expect intuitively. In<strong>de</strong>ed,<br />
a n<strong>at</strong>ural reasoning would be th<strong>at</strong>, as ν goes to infinity, the Stu<strong>de</strong>nt’s distribution goes to the Gaussian<br />
distribution. Therefore, one could a priori expect to find again the result given in the previous section for<br />
the Gaussian factor mo<strong>de</strong>l. We note th<strong>at</strong> λ → 0 when ν → ∞ for all ρ’s smaller than 1/ √ 2. But, and<br />
here lies the surprise, λ → 1 for all ρ larger than 1/ √ 2 when ν → ∞. This counter-intuitive result is due<br />
17
256 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
to a non-uniform convergence which makes the or<strong>de</strong>r to two limits non-commut<strong>at</strong>ive: taking first the limit<br />
u → 1 and then ν → ∞ is different from taking first the limit ν → ∞ and then u → 1. In a sense, by taking<br />
first the limit u → 1, we always ensure somehow the power law regime even if ν is l<strong>at</strong>er taken to infinity.<br />
This is different from first “sitting” on the Gaussian limit ν → ∞. It then is a posteriori reasonable th<strong>at</strong> the<br />
absence of uniform convergence is ma<strong>de</strong> strongly apparent in its consequences when measuring a quantity<br />
probing the extreme tails of the distributions.<br />
As an illustr<strong>at</strong>ion, figure 12 represents the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce for the Stu<strong>de</strong>nt’s copula and Stu<strong>de</strong>nt’s<br />
factor mo<strong>de</strong>l as a function of ρ for various value of ν. It is interesting to note th<strong>at</strong> λ equals zero for all<br />
neg<strong>at</strong>ive ρ in the case of the factor mo<strong>de</strong>l, while λ remains non-zero for neg<strong>at</strong>ive values of the correl<strong>at</strong>ion<br />
coefficient for bivari<strong>at</strong>e Stu<strong>de</strong>nt’s variables.<br />
If Y and ɛ have different numbers νY and νɛ of <strong>de</strong>grees of freedom, two cases occur. For νY < νɛ, ɛ is<br />
negligible asymptotically and λ = 1. For νY > νɛ, X becomes asymptotically i<strong>de</strong>ntical to ɛ. Then, X and<br />
Y have the same tail-<strong>de</strong>pen<strong>de</strong>nce as ɛ and Y , which is 0 by construction.<br />
3.4 Estim<strong>at</strong>ion of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
It would seem th<strong>at</strong> the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce could provi<strong>de</strong> a useful measure of the extreme <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween two random variables which could then be useful for the analysis of contagion b<strong>et</strong>ween<br />
mark<strong>et</strong>s. In<strong>de</strong>ed, either the whole d<strong>at</strong>a s<strong>et</strong> does not exhibit tail <strong>de</strong>pen<strong>de</strong>nce, and a contagion mechanism<br />
seems necessary to explain the occurrence of concomitant large movements during turmoil periods, or it<br />
exhibits tail <strong>de</strong>pen<strong>de</strong>nce so th<strong>at</strong> the usual <strong>de</strong>pen<strong>de</strong>nce structure is such th<strong>at</strong> it is able to produce by itself<br />
concomitant extremes.<br />
Unfortun<strong>at</strong>ely, the empirical estim<strong>at</strong>ion of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is a strenuous task. In<strong>de</strong>ed, a<br />
direct estim<strong>at</strong>ion of the conditional probability Pr{X > FX −1 (u) | Y > FY −1 (u)}, which should tend<br />
to λ when u → 1 is impossible to put in practice due to the combin<strong>at</strong>ion of the curse of dimensionality<br />
and the drastic <strong>de</strong>crease of the number of realis<strong>at</strong>ions as u become close to one. A b<strong>et</strong>ter approach consists<br />
in using kernel estim<strong>at</strong>ors, which generally provi<strong>de</strong> smooth and accur<strong>at</strong>e estim<strong>at</strong>ors (Kulpa 1999, Li <strong>et</strong><br />
al. 1998, Scaill<strong>et</strong> 2000). However, these smooth estim<strong>at</strong>ors lead to differentiable estim<strong>at</strong>ed copulas which<br />
have autom<strong>at</strong>ically vanishing tail <strong>de</strong>pen<strong>de</strong>nce. In<strong>de</strong>ed, in or<strong>de</strong>r to obtain a non-vanishing coefficient of tail<br />
<strong>de</strong>pen<strong>de</strong>nce, it is necessary for the corresponding copula to be non-differentiable <strong>at</strong> the point (1, 1) (or <strong>at</strong><br />
(0, 0)). An altern<strong>at</strong>ive is then the fully param<strong>et</strong>ric approach. One can choose to mo<strong>de</strong>l <strong>de</strong>pen<strong>de</strong>nce via a<br />
specific copula, and thus to d<strong>et</strong>ermine the associ<strong>at</strong>ed tail <strong>de</strong>pen<strong>de</strong>nce (Longin and Solnik 2001, Malevergne<br />
and Sorn<strong>et</strong>te 2001, P<strong>at</strong>ton 2001). The problem with such a m<strong>et</strong>hod is th<strong>at</strong> the choice of the param<strong>et</strong>eriz<strong>at</strong>ion<br />
of the copula amounts to choose a priori wh<strong>et</strong>her or not the d<strong>at</strong>a presents tail <strong>de</strong>pen<strong>de</strong>nce.<br />
In fact, there exist three ways to estim<strong>at</strong>e the tail <strong>de</strong>pen<strong>de</strong>nce coefficient. The two first ones are specific to a<br />
class of copulas or of mo<strong>de</strong>ls, while the last one is very general, but obvioulsy less accur<strong>at</strong>e. The first m<strong>et</strong>hod<br />
is only reliable when it is known th<strong>at</strong> the un<strong>de</strong>rlying copula is Archimedian (see (Joe 1997) or (Nelsen 1998)<br />
for the <strong>de</strong>finition). In such a case, a limit theorem established by (Juri and Wüthrich 2002) allows to estim<strong>at</strong>e<br />
the tail <strong>de</strong>pen<strong>de</strong>nce. The problem is th<strong>at</strong> it is not obvious th<strong>at</strong> the Archimedian copulas provi<strong>de</strong> a good<br />
represent<strong>at</strong>ion of the <strong>de</strong>pen<strong>de</strong>nce structure for financial ass<strong>et</strong>s. For instance, the Achimedian copulas are<br />
generally inconsistent with a represent<strong>at</strong>ion of ass<strong>et</strong>s by factor mo<strong>de</strong>ls. In such case, a second m<strong>et</strong>hod<br />
provi<strong>de</strong>d by (Malevergne and Sorn<strong>et</strong>te 2002) offers good results allowing to estim<strong>at</strong>e the tail <strong>de</strong>pen<strong>de</strong>nce in<br />
a semi-param<strong>et</strong>ric way, which solely relies on the estim<strong>at</strong>ion of marginal distributions, a significantly easier<br />
task.<br />
When none of these situ<strong>at</strong>ions occur, or when the factors are too difficult to extract, a third and fully non-<br />
18
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 257<br />
param<strong>et</strong>ric m<strong>et</strong>hod exists, which is based upon the m<strong>at</strong>hem<strong>at</strong>ical results by (Ledford and Tawn 1996, Ledford<br />
and Tawn 1998) and (Coles <strong>et</strong> al. 1999) and has recently been applied by (Poon <strong>et</strong> al. 2001). The m<strong>et</strong>hod<br />
consists in tranforming the original random variables X and Y into Fréch<strong>et</strong> random variables <strong>de</strong>noted by S<br />
and T respectively. Then, consi<strong>de</strong>ring the variable Z = min{S, T }, its survival distribution is:<br />
Pr{Z > z} = d · z 1/η<br />
as z → ∞, (35)<br />
and ¯ λ = 2 · η − 1, with λ = 0 if ¯ λ < 1 or ¯ λ = 1 and λ = d. The param<strong>et</strong>ers η and d can be estim<strong>at</strong>ed<br />
by maximum likelihood, and <strong>de</strong>riving their asymptotic st<strong>at</strong>istics allows one to test wh<strong>et</strong>her the hypothesis<br />
¯λ = 1 can be rejected or not, and consequently, wh<strong>et</strong>her the d<strong>at</strong>a present tail <strong>de</strong>pen<strong>de</strong>nce or not.<br />
We have implemented this procedure and the estim<strong>at</strong>ed values of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce are given<br />
in table 1 both for the positive and the neg<strong>at</strong>ive tails. Our tests show th<strong>at</strong> we cannot reject the hypothesis of<br />
tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the four consi<strong>de</strong>red L<strong>at</strong>in American mark<strong>et</strong>s (Argentina, Brazil, Chile and Mexico).<br />
Notice th<strong>at</strong> the positive tail <strong>de</strong>pen<strong>de</strong>nce is almost always slightly smaller than the neg<strong>at</strong>ive one, which could<br />
be linked with the trend asymm<strong>et</strong>ry of (Longin and Solnik 2001), but it turns out th<strong>at</strong> these differences are<br />
not st<strong>at</strong>istically significant. These results indic<strong>at</strong>e th<strong>at</strong>, according to this analysis of the extreme <strong>de</strong>pen<strong>de</strong>nce<br />
coefficient, the propensity of extreme co-movements is almost the same for each pair of stock mark<strong>et</strong>s:<br />
even if the transmission mechanisms of a crisis are different from a country to another one, the propag<strong>at</strong>ion<br />
occur with the same probability overall. Thus, the subsequent risks are the same. In table 2, we also<br />
give the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce estim<strong>at</strong>ed un<strong>de</strong>r the Stu<strong>de</strong>nt’s copula (or in fact any copula <strong>de</strong>rived<br />
from an elliptical distribution) with three dregrees of freedom, given by expression (26). One can observe<br />
a remarkable agreement b<strong>et</strong>ween these values and the non-param<strong>et</strong>ric estim<strong>at</strong>es given in table 1. This<br />
is consistent with the results given by the conditional Spearman’s ρ, for which we have shown th<strong>at</strong> the<br />
Stu<strong>de</strong>nt’s copula seems to reasonably account for the extreme <strong>de</strong>pen<strong>de</strong>nce.<br />
4 Summary and Discussion<br />
Table 3 summarizes the asymptotic <strong>de</strong>pen<strong>de</strong>nces for large v and u of the signed conditional correl<strong>at</strong>ion<br />
coefficient ρ + v , the unsigned conditional correl<strong>at</strong>ion coefficient ρ s v and the correl<strong>at</strong>ion coefficient ρu conditioned<br />
on both variables for the bivari<strong>at</strong>e Gaussian, the Stu<strong>de</strong>nt’s mo<strong>de</strong>l, the Gaussian factor mo<strong>de</strong>l and<br />
the Stu<strong>de</strong>nt’s factor mo<strong>de</strong>l. Our results provi<strong>de</strong> a quantit<strong>at</strong>ive proof th<strong>at</strong> conditioning on exceedance leads<br />
to conditional correl<strong>at</strong>ion coefficients th<strong>at</strong> may be very different from the unconditional correl<strong>at</strong>ion. This<br />
provi<strong>de</strong>s a straightforward mechanism for fluctu<strong>at</strong>ions or changes of correl<strong>at</strong>ions, based on fluctu<strong>at</strong>ions of<br />
vol<strong>at</strong>ility or changes of trends. In other words, the many reported vari<strong>at</strong>ions of correl<strong>at</strong>ion structure might<br />
be in large part <strong>at</strong>tributed to changes in vol<strong>at</strong>ility (and st<strong>at</strong>isical uncertainty).<br />
We also suggest th<strong>at</strong> the distinct <strong>de</strong>pen<strong>de</strong>nces as a function of exceedance v and u of the conditional correl<strong>at</strong>ion<br />
coefficients may offer novel tools for characterizing the st<strong>at</strong>istical multivari<strong>at</strong>e distributions of extreme<br />
events. Since their direct characteriz<strong>at</strong>ion is in general restricted by the curse of dimensionality and the scarsity<br />
of d<strong>at</strong>a, the conditional correl<strong>at</strong>ion coefficients provi<strong>de</strong> reduced robust st<strong>at</strong>istics which can be estim<strong>at</strong>ed<br />
with reasonable accuracy and reliability. In this respect, our empirical results encourage us to assert th<strong>at</strong><br />
a Stu<strong>de</strong>nt’s copula, or more generally and elliptical copula, with a tail in<strong>de</strong>x of about three seems able to<br />
account for the main extreme <strong>de</strong>pen<strong>de</strong>nce properties investig<strong>at</strong>ed here.<br />
Table 4 gives the asymptotic values of ρ + v , ρ s v and ρu for v → +∞ and u → ∞ in or<strong>de</strong>r to compare them<br />
with the tail-<strong>de</strong>pen<strong>de</strong>nce λ.<br />
These two tables only scr<strong>at</strong>ch the surface of the rich s<strong>et</strong>s of measures of tail and extreme <strong>de</strong>pen<strong>de</strong>nces. We<br />
have shown th<strong>at</strong> compl<strong>et</strong>e in<strong>de</strong>pen<strong>de</strong>nce implies the absence of tail <strong>de</strong>pen<strong>de</strong>nce: λ = 0. But λ = 0 does not<br />
19
258 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
implies in<strong>de</strong>pen<strong>de</strong>nce, <strong>at</strong> least in the intermedi<strong>at</strong>e range, since it is only an asymptotic property. Conversely,<br />
a non-zero tail <strong>de</strong>pen<strong>de</strong>nce λ implies the absence of asymptotic in<strong>de</strong>pen<strong>de</strong>nce. Non<strong>et</strong>heless, it does not<br />
imply necessarily th<strong>at</strong> the conditional correl<strong>at</strong>ion coefficients ρ + v=∞ and ρ s v=∞ are non-zero, as one would<br />
have expected naively.<br />
Note th<strong>at</strong> the examples of Table 4 are such th<strong>at</strong> λ = 0 seems to go hand-in-hand with ρ + v→∞ = 0. However,<br />
the logical implic<strong>at</strong>ion (λ = 0) ⇒ (ρ + v→∞ = 0) does not hold in general. A counter example is offered by<br />
the Stu<strong>de</strong>nt’s factor mo<strong>de</strong>l in the case where νY > νɛ (the tail of the distribution of the idiosynchr<strong>at</strong>ic noise<br />
is f<strong>at</strong>ter than th<strong>at</strong> of the distribution of the factor). In this case, X and Y have the same tail-<strong>de</strong>pen<strong>de</strong>nce<br />
as ɛ and Y , which is 0 by construction. But, ρ + v=∞ and ρ s v=∞ are both 1 because a large Y almost always<br />
gives a large X and the simultaneous occurrence of a large Y and a large ɛ can be neglected. The reason for<br />
this absence of tail <strong>de</strong>pen<strong>de</strong>nce (in the sense of λ) coming tog<strong>et</strong>her with asymptotically strong conditional<br />
correl<strong>at</strong>ion coefficients stems from two facts:<br />
• first, the conditional correl<strong>at</strong>ion coefficients put much less weight on the extreme tails th<strong>at</strong> the tail<strong>de</strong>pen<strong>de</strong>nce<br />
param<strong>et</strong>er λ. In other words, ρ + v=∞ and ρ s v=∞ are sensitive to the marginals, i.e., there are<br />
d<strong>et</strong>ermined by the full bivari<strong>at</strong>e distribution, while, as we said, λ is a pure copula property in<strong>de</strong>pen<strong>de</strong>nt<br />
of the marginals. Since ρ + v=∞ and ρ s v=∞ are measures of tail <strong>de</strong>pen<strong>de</strong>nce weighted by the specific<br />
shapes of the marginals, it is n<strong>at</strong>ural th<strong>at</strong> they may behave differently.<br />
• Secondly, the tail <strong>de</strong>pen<strong>de</strong>nce λ probes the extreme <strong>de</strong>pen<strong>de</strong>nce property of the original copula of the<br />
random variables X and Y . On the contrary, when conditioning on Y , one changes the copula of X<br />
and Y , so th<strong>at</strong> the extreme <strong>de</strong>pen<strong>de</strong>nce properties investig<strong>at</strong>ed by the conditional correl<strong>at</strong>ions are not<br />
exactly those of the original copula. This last remark explains clearly why we observe wh<strong>at</strong> (Boyer <strong>et</strong><br />
al. 1997) call a “bias” in the conditional correl<strong>at</strong>ions. In<strong>de</strong>ed, changing the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two<br />
random variables obviously leads to changing their correl<strong>at</strong>ions.<br />
The consequences are of importance. In such a situ<strong>at</strong>ion, one measure (λ) would conclu<strong>de</strong> on asymptotic<br />
tail-in<strong>de</strong>pen<strong>de</strong>nce while the other measures ρ + v=∞ and ρ s v=∞ would conclu<strong>de</strong> the opposite. Thus, before<br />
concluding on a change in the <strong>de</strong>pen<strong>de</strong>nce structure with respect to a given param<strong>et</strong>er - the vol<strong>at</strong>ility or<br />
the trend, for instance - one should check th<strong>at</strong> this change does not result from the tool used to probe the<br />
<strong>de</strong>pen<strong>de</strong>nce. In this respect, our study allows us to shed new lights on recent controversial results about the<br />
occurrence or abscence of contagion during the L<strong>at</strong>in American crises. As in every previous works, we find<br />
no contagion evi<strong>de</strong>nce b<strong>et</strong>ween Chile and Mexico, but contrarily to (Forbes and Rigobon 2002), we think it<br />
is difficult to ignore the possibility of contagion towards Argentina and Brazil, and in this respect we agree<br />
with (Calvo and Reinhart 1996).<br />
In fact, we think th<strong>at</strong> most of the discrepancies b<strong>et</strong>ween these different studies stem from the fact th<strong>at</strong><br />
the conditional correl<strong>at</strong>ion coefficient does not provi<strong>de</strong> an accur<strong>at</strong>e tool for probing the potential changes<br />
of <strong>de</strong>pen<strong>de</strong>nce. In<strong>de</strong>ed, even when the bias have been accounted for, the f<strong>at</strong>-tailness of the distributions<br />
of r<strong>et</strong>urns are such th<strong>at</strong> the Pearson’s coefficient is subjected to very strong st<strong>at</strong>istical fluctu<strong>at</strong>ions which<br />
forbid an accur<strong>at</strong>e estim<strong>at</strong>ion of the correl<strong>at</strong>ion. Moreover, when studying the <strong>de</strong>pen<strong>de</strong>nce properties, it is<br />
interesting to free oneself from the marginal behavior of each random variable. This is why the conditional<br />
Spearman’s rho seems a good tool: it only <strong>de</strong>pends on the copula and is st<strong>at</strong>istically well-behaved.<br />
The conditional Spearman’s rho has allowed us to i<strong>de</strong>ntify a change in the <strong>de</strong>pen<strong>de</strong>nce structure during<br />
downward trends in L<strong>at</strong>in American mark<strong>et</strong>s, similar to th<strong>at</strong> found by (Longin and Solnik 2001) in their<br />
study of the contagion across five major equity mark<strong>et</strong>s. It has also enabled us to put in light the asymm<strong>et</strong>ry<br />
in the contagion effects: Mexico and Chile can be potential source sof contagion toward Argentina and<br />
Brazil, while reverse does not seem to hold. This phenomenom has been observed during the 1994 Mexican<br />
20
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 259<br />
crisis and appears to remain true in the recent Argentina crisis, for which only Brazil seems to exhibit the<br />
sign<strong>at</strong>ure of a possible contagion.<br />
We suggest th<strong>at</strong> a possible origin for the discovered asymm<strong>et</strong>ry may lie in the difference in the more mark<strong>et</strong>oriented<br />
countries versus more st<strong>at</strong>e intervention oriented economies, giving rise to either currency flo<strong>at</strong>ing<br />
regimes adapted to an important manufacturing sector which tend to <strong>de</strong>liver more comp<strong>et</strong>itive real exchange<br />
r<strong>at</strong>es (Chile and Mexico) or fixed r<strong>at</strong>e pegs (Argentina until the 2001 crisis and Brazil until the early 1999<br />
crisis) (Frie<strong>de</strong>n 1992, Frie<strong>de</strong>n <strong>et</strong> al. 2000a, Frie<strong>de</strong>n <strong>et</strong> al. 2000b). The asymm<strong>et</strong>ry of the contagion is comp<strong>at</strong>ible<br />
with the view th<strong>at</strong> fixed echange r<strong>at</strong>es tighten more strickly an economy and its stock mark<strong>et</strong> to external<br />
shocks (case of Argentina and Brazil) while a more flexible exchange r<strong>at</strong>e seems to provi<strong>de</strong> a cushion allowing<br />
a <strong>de</strong>coupling b<strong>et</strong>ween the stock mark<strong>et</strong> and external influences.<br />
Finally, the absence of contagion does not imply necessarily the absence of “contamin<strong>at</strong>ion.” In<strong>de</strong>ed, the<br />
study of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce has proven th<strong>at</strong> with or without contagion mechanisms - i.e.,<br />
increase in the linkage b<strong>et</strong>ween mark<strong>et</strong>s during crisis - the probability of extreme co-movements during the<br />
crisis - i.e., the contamin<strong>at</strong>ion - is almost the same for all pairs of mark<strong>et</strong>s. Thus, wh<strong>at</strong>ever the propag<strong>at</strong>ion<br />
mechanism may be - historically closed rel<strong>at</strong>ionship or irr<strong>at</strong>ional fear and herd behavior - the observed<br />
effects are the same: the propag<strong>at</strong>ion of the crisis. From the practical perspective of risk management or<br />
regul<strong>at</strong>ory policy, this last point is may be more important than the real knowledge of the occurrence or not<br />
of contagion.<br />
21
260 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Figure legends<br />
Figure 1: The upper (respectively lower) panel graphs the complementary distribution of the positive (respectively<br />
the minus neg<strong>at</strong>ive) r<strong>et</strong>urns in US dollar. The straight line represents the slope of a power law<br />
with tail exponent α = 2.<br />
Figure 2: In the upper panel, the thick plain curve <strong>de</strong>picts the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the Argentina<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns and the Mexican stock in<strong>de</strong>x daily r<strong>et</strong>urns conditional on the Mexican stock in<strong>de</strong>x<br />
daily r<strong>et</strong>urns larger than (smaller than) a given positive (neg<strong>at</strong>ive) value v (after normaliz<strong>at</strong>ion by the stan-<br />
dard <strong>de</strong>vi<strong>at</strong>ion). The thick dashed curve represents the theor<strong>et</strong>ical conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v<br />
calcul<strong>at</strong>ed for a bivari<strong>at</strong>e Gaussian mo<strong>de</strong>l, while the two thin dashed curves represent the area within which<br />
we cannot consi<strong>de</strong>r, <strong>at</strong> the 95% confi<strong>de</strong>nce level, th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficient is significantly<br />
different from its Gaussian theor<strong>et</strong>ical value. The dotted curves provi<strong>de</strong> the same inform<strong>at</strong>ion un<strong>de</strong>r the<br />
assumption of a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s mo<strong>de</strong>l with ν = 3 <strong>de</strong>grees of freedom.<br />
The lower panel gives the same kind of inform<strong>at</strong>ion for the correl<strong>at</strong>ion coefficient conditioned on the Argentina<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns larger than (smaller than) a given positive (neg<strong>at</strong>ive) value v.<br />
Figure 3: In the upper panel, the thick plain curve gives the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the Argentina<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns and the Mexican stock in<strong>de</strong>x daily r<strong>et</strong>urns conditioned on the Mexican stock in<strong>de</strong>x<br />
daily vol<strong>at</strong>ility larger than a give value v (after normaliz<strong>at</strong>ion by the standard <strong>de</strong>vi<strong>at</strong>ion). The thick dashed<br />
curve represents the theor<strong>et</strong>ical conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v calcul<strong>at</strong>ed for a bivari<strong>at</strong>e Gaussian<br />
mo<strong>de</strong>l, while the two thin dashed curves represent the area within which we cannot consi<strong>de</strong>r, <strong>at</strong> the 95%<br />
confi<strong>de</strong>nce level, th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficient is significantly different from its Gaussian theor<strong>et</strong>ical<br />
value. The dotted curves provi<strong>de</strong> the same inform<strong>at</strong>ion un<strong>de</strong>r the assumption of a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s<br />
mo<strong>de</strong>l with ν = 3 <strong>de</strong>grees of freedom.<br />
The lower panel gives the same kind of inform<strong>at</strong>ion for the correl<strong>at</strong>ion coefficient conditioned on the Argentina<br />
stock in<strong>de</strong>x daily vol<strong>at</strong>ility larger than a given value v.<br />
Figure 4: In the upper panel, the thick plain curve gives the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the Brazilian<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns and the Chilean stock in<strong>de</strong>x daily r<strong>et</strong>urns conditioned on the Chilean stock in<strong>de</strong>x<br />
daily r<strong>et</strong>urns larger than (smaller than) a given positive (neg<strong>at</strong>ive) value v (after normaliz<strong>at</strong>ion by the stan-<br />
dard <strong>de</strong>vi<strong>at</strong>ion). The thick dashed curve represents the theor<strong>et</strong>ical conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v<br />
calcul<strong>at</strong>ed for a bivari<strong>at</strong>e Gaussian mo<strong>de</strong>l, while the two thin dashed curves represent the area within which<br />
we cannot consi<strong>de</strong>r, <strong>at</strong> the 95% confi<strong>de</strong>nce level, th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficient is significantly<br />
different from its Gaussian theor<strong>et</strong>ical value. The dotted curves provi<strong>de</strong> the same inform<strong>at</strong>ion un<strong>de</strong>r the<br />
assumption of a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s mo<strong>de</strong>l with ν = 3 <strong>de</strong>grees of freedom.<br />
The lower panel gives the same kind of inform<strong>at</strong>ion for the correl<strong>at</strong>ion coefficient conditioned on the Brazilian<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns larger than (smaller than) a given positive (neg<strong>at</strong>ive) value v.<br />
Figure 5: In the upper panel, the thick plain curve shows the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the Brazilian<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns and the Chilean stock in<strong>de</strong>x daily r<strong>et</strong>urns conditional on the Chilean stock in<strong>de</strong>x<br />
daily vol<strong>at</strong>ility larger than a given value v (after normaliz<strong>at</strong>ion by the standard <strong>de</strong>vi<strong>at</strong>ion). The thick dashed<br />
curve represents the theor<strong>et</strong>ical conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v calcul<strong>at</strong>ed for a bivari<strong>at</strong>e Gaussian<br />
mo<strong>de</strong>l, while the two thin dashed curves represent the area within which we cannot consi<strong>de</strong>r, <strong>at</strong> the 95%<br />
confi<strong>de</strong>nce level, th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficient is significantly different from its Gaussian theor<strong>et</strong>ical<br />
value. The dotted curves provi<strong>de</strong> the same inform<strong>at</strong>ion un<strong>de</strong>r the assumption of a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s<br />
mo<strong>de</strong>l with ν = 3 <strong>de</strong>grees of freedom.<br />
The lower panel gives the same kind of inform<strong>at</strong>ion for the correl<strong>at</strong>ion coefficient conditioned on the Brazil-<br />
22
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 261<br />
ian stock in<strong>de</strong>x daily vol<strong>at</strong>ility larger than a given value v.<br />
Figure 6: In the upper panel, the thick plain curve shows the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the Chilean<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns and the Mexican stock in<strong>de</strong>x daily r<strong>et</strong>urns conditioned on the Mexican stock in<strong>de</strong>x<br />
daily r<strong>et</strong>urns larger than (smaller than) a given positive (neg<strong>at</strong>ive) value v (after normaliz<strong>at</strong>ion by the standard<br />
<strong>de</strong>vi<strong>at</strong>ion). The thick dashed curve represents the theor<strong>et</strong>ical conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v<br />
calcul<strong>at</strong>ed for a bivari<strong>at</strong>e Gaussian mo<strong>de</strong>l, while the two thin dashed curves represent the area within which<br />
we cannot consi<strong>de</strong>r, <strong>at</strong> the 95% confi<strong>de</strong>nce level, th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficient is significantly<br />
different from its Gaussian theor<strong>et</strong>ical value. The dotted curves provi<strong>de</strong> the same inform<strong>at</strong>ion un<strong>de</strong>r the<br />
assumption of a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s mo<strong>de</strong>l with ν = 3 <strong>de</strong>grees of freedom.<br />
The lower panel gives the same kind of inform<strong>at</strong>ion for the correl<strong>at</strong>ion coefficient conditioned on the Chilean<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns larger than (smaller than) a given positive (neg<strong>at</strong>ive) value v.<br />
Figure 7: In the upper panel, the thick plain curve <strong>de</strong>picts the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the Chilean<br />
stock in<strong>de</strong>x daily r<strong>et</strong>urns and the Mexican stock in<strong>de</strong>x daily r<strong>et</strong>urns conditioned on the Mexican stock in<strong>de</strong>x<br />
daily vol<strong>at</strong>ility larger than a give value v (after normaliz<strong>at</strong>ion by the standard <strong>de</strong>vi<strong>at</strong>ion). The thick dashed<br />
curve represents the theor<strong>et</strong>ical conditional correl<strong>at</strong>ion coefficient ρ +,−<br />
v calcul<strong>at</strong>ed for a bivari<strong>at</strong>e Gaussian<br />
mo<strong>de</strong>l, while the two thin dashed curves represent the area within which we cannot consi<strong>de</strong>r, <strong>at</strong> the 95%<br />
confi<strong>de</strong>nce level, th<strong>at</strong> the estim<strong>at</strong>ed correl<strong>at</strong>ion coefficient is significantly different from its Gaussian theor<strong>et</strong>ical<br />
value. The dotted curves provi<strong>de</strong> the same inform<strong>at</strong>ion un<strong>de</strong>r the assumption of a bivari<strong>at</strong>e Stu<strong>de</strong>nt’s<br />
mo<strong>de</strong>l with ν = 3 <strong>de</strong>grees of freedom.<br />
The lower panel gives the same kind of inform<strong>at</strong>ion for the correl<strong>at</strong>ion coefficient conditioned on the Chilean<br />
stock in<strong>de</strong>x daily vol<strong>at</strong>ility larger than a given value v.<br />
Figure 8: Conditional Spearman’s rho for a bivari<strong>at</strong>e Gaussian copula (left panel) and a Stu<strong>de</strong>nt’s copula<br />
with three <strong>de</strong>grees of freedom (right panel), with an unconditional linear correl<strong>at</strong>ion coefficient ρ =<br />
0.1, 0.3, 0.5, 0.7, 0.9.<br />
Figure 9: In the upper panel, the thick curve shows the Spearman’s rho b<strong>et</strong>ween the Argentina stock in<strong>de</strong>x<br />
daily r<strong>et</strong>urns and the Brazilian stock in<strong>de</strong>x daily r<strong>et</strong>urns. Above the quantile v = 0.5, the Spearman’s rho is<br />
conditioned on the Brazilian in<strong>de</strong>x daily r<strong>et</strong>urns whose quantiles are larger than v, while below the quantile<br />
v = 0.5 it is conditioned on the Brazilian in<strong>de</strong>x daily r<strong>et</strong>urns whose quantiles are smaller than v. As in the<br />
above figures for the correl<strong>at</strong>ion coefficients, the dashed lines refer to the prediction of the Gaussian copula<br />
and its 95% confi<strong>de</strong>nce levels and the dotted lines to the Stu<strong>de</strong>nt’s copula with three <strong>de</strong>grees of freedom<br />
and its 95% confi<strong>de</strong>nce levels. The lower panel gives the same kind of inform<strong>at</strong>ion for the Spearman’s rho<br />
conditioned on the realiz<strong>at</strong>ions of the Argentina in<strong>de</strong>x daily r<strong>et</strong>urns.<br />
Figure 10: In the upper panel, the thick curve shows the Spearman’s rho b<strong>et</strong>ween the Brazilian stock in<strong>de</strong>x<br />
daily r<strong>et</strong>urns and the chilean stock in<strong>de</strong>x daily r<strong>et</strong>urns. Above the quantile v = 0.5, the Spearman’s rho is<br />
conditioned on the Chilean in<strong>de</strong>x daily r<strong>et</strong>urns whose quantiles are larger than v, while below the quantile<br />
v = 0.5 it is conditioned on the Mexican in<strong>de</strong>x daily r<strong>et</strong>urns whose quantiles are smaller than v. The dashed<br />
lines refer to the prediction of the Gaussian copula and its 95% confi<strong>de</strong>nce levels and the dotted lines to the<br />
Stu<strong>de</strong>nt’s copula with three <strong>de</strong>grees of freedom and its 95% confi<strong>de</strong>nce levels. The lower panel gives the<br />
same kind of inform<strong>at</strong>ion for the Spearman’s rho conditioned on the realiz<strong>at</strong>ions of the Brazilian in<strong>de</strong>x daily<br />
r<strong>et</strong>urns.<br />
Figure 11: In the upper panel, the thick curve <strong>de</strong>picts the Spearman’s rho b<strong>et</strong>ween the Chilean stock in<strong>de</strong>x<br />
daily r<strong>et</strong>urns and the Mexican stock in<strong>de</strong>x daily r<strong>et</strong>urns. Above the quantile v = 0.5, the Spearman’s rho is<br />
conditioned on the Mexican in<strong>de</strong>x daily r<strong>et</strong>urns whose quantiles are larger than v, while below the quantile<br />
v = 0.5 it is conditioned on the Mexican in<strong>de</strong>x daily r<strong>et</strong>urns whose quantiles are smaller than v. The dashed<br />
lines refer to the prediction of the Gaussian copula and its 95% confi<strong>de</strong>nce levels and the dotted lines to the<br />
23
262 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Stu<strong>de</strong>nt’s copula with three <strong>de</strong>grees of freedom and its 95% confi<strong>de</strong>nce levels. The lower panel gives the<br />
same kind of inform<strong>at</strong>ion for the Spearman’s rho conditioned on the realiz<strong>at</strong>ions of the Chilean in<strong>de</strong>x daily<br />
r<strong>et</strong>urns.<br />
Figure 12: Coefficient of upper tail <strong>de</strong>pen<strong>de</strong>nce as a function of the correl<strong>at</strong>ion coefficient ρ for various<br />
values of the number of <strong>de</strong>gres of freedomn ν for the stu<strong>de</strong>nt’s Copula (left panel) and the Stu<strong>de</strong>nt’s factor<br />
mo<strong>de</strong>l (right panel).<br />
<br />
<br />
Figure 13: The graph of the function 1<br />
(1+ ɛu<br />
) x0 ν <br />
<br />
− 1<br />
(thick solid line), the string which gives an upper bound<br />
of the function within 1−x0<br />
ɛ , 0 (dashed line) and the tangent in 0 + which gives an upper bound of the<br />
function within 0, x0<br />
<br />
ɛ (dash dotted line).<br />
24
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 263<br />
A Conditional correl<strong>at</strong>ion coefficient for Gaussian variables<br />
L<strong>et</strong> us consi<strong>de</strong>r a pair of Normal random variables (X, Y ) ∼ N (0, Σ) where Σ is their covariance m<strong>at</strong>rix<br />
with unconditional correl<strong>at</strong>ion coefficient ρ. Without loss of generality, and for simplicity, we shall assume<br />
Σ with unit unconditional variances.<br />
A.1 Conditioning on one variable<br />
A.1.1 Conditioning on Y larger than v<br />
Given a conditioning s<strong>et</strong> A = [v, +∞), v ∈ R+, ρA = ρ + v is the correl<strong>at</strong>ion coefficient conditioned on Y<br />
larger than v:<br />
ρ + v =<br />
<br />
ρ 2 +<br />
ρ<br />
1−ρ 2<br />
Var(Y | Y >v)<br />
√ v<br />
πe 2 = v +<br />
v√2<br />
2 erfc<br />
1<br />
v<br />
. (A.1)<br />
We start with the calcul<strong>at</strong>ion of the first and the second moment of Y conditioned on Y larger than v:<br />
E(Y | Y > v) =<br />
√<br />
2<br />
<br />
2 1<br />
− + O<br />
v3 v5 <br />
, (A.2)<br />
E(Y 2 | Y > v) = 1 +<br />
√ 2v<br />
√ v<br />
πe 2 <br />
v√2<br />
2 erfc<br />
which allows us to obtain the variance of Y conditioned on Y larger than v:<br />
Var(Y | Y > v) = 1 +<br />
which, for large v, yields:<br />
√ 2v<br />
√ v<br />
πe 2 <br />
v√2<br />
2 erfc<br />
A.1.2 Conditioning on |Y | larger than v<br />
⎛<br />
− ⎝<br />
ρ + v ∼v→∞<br />
= v 2 + 2 − 2<br />
<br />
1<br />
+ O<br />
v2 v4 <br />
, (A.3)<br />
√ 2<br />
√ v<br />
πe 2 <br />
v√2<br />
2 erfc<br />
⎞<br />
⎠<br />
2<br />
= 1<br />
<br />
1<br />
+ O<br />
v2 v4 <br />
, (A.4)<br />
ρ 1<br />
· . (A.5)<br />
1 − ρ2 v<br />
Given a conditioning s<strong>et</strong> A = (−∞, −v] ∪ [v, +∞), v ∈ R+, ρA = ρ s v is the correl<strong>at</strong>ion coefficient<br />
conditioned on |Y | larger than v:<br />
ρ s v =<br />
<br />
ρ 2 +<br />
ρ<br />
1−ρ 2<br />
Var(Y | |Y |>v)<br />
. (A.6)<br />
The first and second moment of Y conditioned on |Y | larger than v can be easily calcul<strong>at</strong>ed:<br />
E(Y | |Y | > v) = 0, (A.7)<br />
E(Y 2 √<br />
2v<br />
| |Y | > v) = 1 +<br />
= v 2 + 2 − 2<br />
<br />
1<br />
+ O<br />
v2 v4 <br />
. (A.8)<br />
√ v<br />
πe 2 <br />
v√2<br />
2 erfc<br />
25
264 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Expression (A.8) is the same as (A.4) as it should. This gives the following conditional variance:<br />
Var(Y | |Y | > v) = 1 +<br />
√<br />
2v<br />
<br />
1<br />
v2 <br />
, (A.9)<br />
and finally yields, for large v,<br />
A.1.3 Intuitive meaning<br />
ρ s v ∼v→∞<br />
ρ<br />
<br />
ρ 2 + 1−ρ2<br />
2+v 2<br />
√ v<br />
πe 2 = v<br />
v√2<br />
2 erfc<br />
2 + 2 + O<br />
∼v→∞ 1 − 1<br />
2<br />
1 − ρ 2<br />
ρ 2<br />
1<br />
. (A.10)<br />
v2 L<strong>et</strong> us provi<strong>de</strong> an intuitive explan<strong>at</strong>ion (see also (Longin and Solnik 2001)). As seen from (A.1), ρ + v is<br />
controlled by the <strong>de</strong>pen<strong>de</strong>nce Var(Y | Y > v) ∝ 1/v 2 <strong>de</strong>rived in Appendix A.1.1. In contrast, as seen from<br />
(A.6), ρ s v is controlled by Var(Y | |Y | > v) ∝ v 2 given in Appendix A.1.2. The difference b<strong>et</strong>ween ρ + v and<br />
ρ s v can thus be traced back to th<strong>at</strong> b<strong>et</strong>ween Var(Y | Y > v) ∝ 1/v 2 and Var(Y | |Y | > v) ∝ v 2 for large v.<br />
This results from the following effect. For Y > v, one can picture the possible realiz<strong>at</strong>ions of Y as those of<br />
a random particle on the line, which is strongly <strong>at</strong>tracted to the origin by a spring (the Gaussian distribution<br />
th<strong>at</strong> prevents Y from performing significant fluctu<strong>at</strong>ions beyond a few standard <strong>de</strong>vi<strong>at</strong>ions) while being<br />
forced to be on the right to a wall <strong>at</strong> Y = v. It is clear th<strong>at</strong> the fluctu<strong>at</strong>ions of the position of this particle<br />
are very small as it is strongly glued to the unpen<strong>et</strong>rable wall by the restoring spring, hence the result<br />
Var(Y | Y > v) ∝ 1/v 2 . In constrast, for the condition |Y | > v, by the same argument, the fluctu<strong>at</strong>ions<br />
of the particle are hin<strong>de</strong>red to be very close to |Y | = v, i.e., very close to Y = +v or Y = −v. Thus, the<br />
fluctu<strong>at</strong>ions of Y typically flip from −v to +v and vice-versa. It is thus not surprising to find Var(Y | |Y | ><br />
v) ∝ v 2 .<br />
This argument makes intuitive the results Var(Y | Y > v) ∝ 1/v2 and Var(Y | |Y | > v) ∝ v2 for large<br />
v and thus the results for ρ + v and for ρs v if we use (A.1) and (A.6). We now <strong>at</strong>tempt to justify ρ + v ∼v→∞ 1<br />
v<br />
and 1 − ρs v ∼v→∞ 1/v2 directly by the following intuitive argument. Using the picture of particles, X<br />
and Y can be visualized as the positions of two particles which fluctu<strong>at</strong>e randomly. Their joint bivari<strong>at</strong>e<br />
Gaussian distribution with non-zero unconditional correl<strong>at</strong>ion amounts to the existence of a spring th<strong>at</strong> ties<br />
them tog<strong>et</strong>her. Their Gaussian marginals also exert a spring-like force <strong>at</strong>taching them to the origin. When<br />
Y > v, the X-particle is teared off b<strong>et</strong>ween two extremes, b<strong>et</strong>ween 0 and v. When the unconditional<br />
correl<strong>at</strong>ion ρ is less than 1, the spring <strong>at</strong>tracting to the origin is stronger than the spring <strong>at</strong>tracting to the<br />
wall <strong>at</strong> v. The particle X thus un<strong>de</strong>rgoes tiny fluctu<strong>at</strong>ions around the origin th<strong>at</strong> are rel<strong>at</strong>ively less and less<br />
<strong>at</strong>tracted by the Y -particle, hence the result ρ + v ∼v→∞ 1<br />
v → 0. In constrast, for |Y | > v, notwithstanding<br />
the still strong <strong>at</strong>traction of the X-particle to the origin, it can follow the sign of the Y -particle without<br />
paying too much cost in m<strong>at</strong>ching its amplitu<strong>de</strong> |v|. Rel<strong>at</strong>ively tiny fluctu<strong>at</strong>ion of the X-particle but of the<br />
same sign as Y ≈ ±v will result in a strong ρs v, thus justifying th<strong>at</strong> ρs v → 1 for v → +∞.<br />
A.2 Conditioning on both X and Y larger than u<br />
By <strong>de</strong>finition, the conditional correl<strong>at</strong>ion coefficient ρu, conditioned on both X and Y larger than u, is<br />
ρu =<br />
=<br />
Cov[X, Y | X > u, Y > u]<br />
Var[X | X > u, Y > u] Var[Y | X > u, Y > u] , (A.11)<br />
m11 − m10 · m01<br />
√<br />
m20 − m10 2√ , (A.12)<br />
m02 − m01<br />
2<br />
26
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 265<br />
where mij <strong>de</strong>notes E[X i · Y j | X > u, Y > u].<br />
Using the proposition A.1 of (Ang and Chen 2001) or the expressions in (Johnson and Kotz 1972, p 113),<br />
we can assert th<strong>at</strong><br />
m10 L(u, u; ρ) =<br />
<br />
1 − ρ<br />
(1 + ρ) φ(u) 1 − Φ<br />
1 + ρ u<br />
m20 L(u, u; ρ) =<br />
<br />
, (A.13)<br />
(1 + ρ 2 <br />
1 − ρ<br />
) u φ(u) 1 − Φ<br />
1 + ρ u<br />
<br />
+ ρ 1 − ρ2 <br />
2<br />
√ φ<br />
2π 1 + ρ u<br />
m11 L(u, u; ρ) =<br />
<br />
+ L(u, u; ρ),(A.14)<br />
<br />
1 − ρ<br />
2ρ u φ(u) 1 − Φ<br />
1 + ρ u<br />
<br />
1 − ρ2 2<br />
+ √ φ<br />
2π 1 + ρ u<br />
<br />
+ ρ L(u, u; ρ) , (A.15)<br />
where L(·, ·; ·) <strong>de</strong>notes the bivari<strong>at</strong>e Gaussian survival (or complementary cumul<strong>at</strong>ive) distribution:<br />
1<br />
L(h, k; ρ) =<br />
2π 1 − ρ2 ∞ ∞ <br />
dx dy exp −<br />
h k<br />
1 x<br />
2<br />
2 − 2ρxy + y2 1 − ρ2 <br />
, (A.16)<br />
φ(·) is the Gaussian <strong>de</strong>nsity:<br />
and Φ(·) is the cumul<strong>at</strong>ive Gaussian distribution:<br />
A.2.1 Asymptotic behavior of L(u, u; ρ)<br />
φ(x) = 1 x2<br />
− √ e 2 , (A.17)<br />
2π<br />
Φ(x) =<br />
x<br />
−∞<br />
We focus on the asymptotic behavior of<br />
1<br />
L(u, u; ρ) =<br />
2π 1 − ρ2 ∞ ∞<br />
dx<br />
u u<br />
du φ(u). (A.18)<br />
<br />
dy exp − 1<br />
2<br />
x 2 − 2ρxy + y 2<br />
1 − ρ 2<br />
for large u. Performing the change of variables x ′ = x − u and y ′ = y − u, we can write<br />
u2<br />
e− 1+ρ<br />
L(u, u; ρ) =<br />
2π 1 − ρ2 ∞<br />
dx<br />
0<br />
′<br />
∞<br />
0<br />
dy ′ <br />
exp −u x′ + y ′ <br />
1 + ρ<br />
<br />
exp − 1<br />
2<br />
x ′2 − 2ρx ′ y ′ + y ′2<br />
1 − ρ 2<br />
<br />
, (A.19)<br />
<br />
. (A.20)<br />
Using the fact th<strong>at</strong><br />
<br />
exp − 1 x<br />
2<br />
′2 − 2ρx ′ y ′ + y ′2<br />
1 − ρ2 <br />
= 1− x′2 − 2ρx ′ y ′ + y ′2<br />
2(1 − ρ2 +<br />
)<br />
(x′2 − 2ρx ′ y ′ + y ′2 ) 2<br />
8(1 − ρ2 ) 2 − (x′2 − 2ρx ′ y ′ + y ′2 ) 3<br />
48(1 − ρ2 ) 3 +· · · ,<br />
(A.21)<br />
and applying theorem 3.1.1 in (Jensen 1995, p 58) (Laplace’s m<strong>et</strong>hod), equ<strong>at</strong>ions (A.20) and (A.21) yield<br />
and<br />
(1 + ρ)2<br />
L(u, u; ρ) =<br />
2π 1 − ρ<br />
u2<br />
e− 1+ρ<br />
·<br />
2<br />
1/L(u, u; ρ) = 2π u2 1 − ρ 2<br />
(1 + ρ) 2<br />
u 2<br />
<br />
(2 − ρ)(1 + ρ)<br />
1 − ·<br />
1 − ρ<br />
1<br />
u2 + (2ρ2 − 6ρ + 7)(1 + ρ) 2<br />
(1 − ρ) 2<br />
−3 (12 − 13ρ + 8ρ2 − 2ρ 3 )(1 + ρ) 3<br />
(1 − ρ) 3<br />
· e u2<br />
<br />
(2 − ρ)(1 + ρ)<br />
1+ρ 1 + ·<br />
1 − ρ<br />
1<br />
+ (16 − 13ρ + 10ρ2 − 3ρ 3 )(1 + ρ) 3<br />
(1 − ρ) 3<br />
27<br />
· 1<br />
u4 · 1<br />
<br />
1<br />
+ O<br />
u6 u8 <br />
, (A.22)<br />
u2 − 3 − 2ρ + ρ2 )(1 + ρ) 2<br />
(1 − ρ) 2<br />
· 1<br />
u4 · 1<br />
<br />
1<br />
+ O<br />
u6 u8 <br />
. (A.23)
266 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
A.2.2 Asymptotic behavior of the first moment m10<br />
The first moment m10 = E[X | X > u, Y > u] is given by (A.13). For large u,<br />
<br />
1 − ρ<br />
1 − Φ<br />
1 + ρ u<br />
<br />
= 1<br />
2 erfc<br />
<br />
1 − ρ<br />
2(1 + ρ) u<br />
<br />
=<br />
<br />
1 + ρ<br />
1 − ρ<br />
<br />
1 + ρ<br />
−15<br />
1 − ρ<br />
1−ρ<br />
−<br />
e 2(1+ρ) u2<br />
√<br />
2π u<br />
3<br />
(A.24)<br />
<br />
2 1 + ρ 1 1 + ρ<br />
1 − · + 3 ·<br />
1 − ρ u2 1 − ρ<br />
1<br />
u4 · 1<br />
<br />
1<br />
+ O<br />
u6 u8 , (A.25)<br />
so th<strong>at</strong> multiplying by (1 + ρ) φ(u), we obtain<br />
u2 <br />
− <br />
(1 + ρ)2 e 1+ρ<br />
2<br />
1 + ρ 1 1 + ρ<br />
m10 L(u, u; ρ) = 1 − · + 3 ·<br />
1 − ρ2 2π u 1 − ρ u2 1 − ρ<br />
1<br />
3 1 + ρ<br />
− 15 ·<br />
u4 1 − ρ<br />
1<br />
<br />
1<br />
+ O<br />
u6 u8 .<br />
(A.26)<br />
Using the result given by equ<strong>at</strong>ion (A.22), we can conclu<strong>de</strong> th<strong>at</strong><br />
m10 = u + (1 + ρ) · 1<br />
u − (1 + ρ)2 (2 − ρ)<br />
·<br />
(1 − ρ)<br />
1<br />
u3 + (10 − 8ρ + 3ρ2 )(1 + ρ) 3<br />
(1 − ρ) 2<br />
In the sequel, we will also need the behavior of m10 2 :<br />
m10 2 = u 2 + 2 (1 + ρ) − (1 + ρ)2 (3 − ρ)<br />
(1 − ρ)<br />
A.2.3 Asymptotic behavior of the second moment m20<br />
· 1<br />
u2 + 2(8 − 5ρ + 2ρ2 )(1 + ρ) 3<br />
(1 − ρ) 2<br />
· 1<br />
<br />
1<br />
+ O<br />
u5 u7 <br />
. (A.27)<br />
· 1<br />
<br />
1<br />
+ O<br />
u4 u6 <br />
. (A.28)<br />
The second moment m20 = E[X2 | X > u, Y > u] is given by expression (A.14). The first term in the<br />
right hand si<strong>de</strong> of (A.14) yields<br />
(1 + ρ 2 <br />
1 − ρ<br />
) u φ(u) 1 − Φ<br />
1 + ρ u<br />
<br />
= (1 + ρ 2 u2 <br />
− <br />
1 + ρ e 1+ρ<br />
2<br />
1 + ρ 1 1 + ρ<br />
)<br />
1 − · + 3 ·<br />
1 − ρ 2π 1 − ρ u2 1 − ρ<br />
1<br />
u4 3 1 + ρ<br />
−15 ·<br />
1 − ρ<br />
1<br />
<br />
1<br />
+ O<br />
u6 u8 (A.29) ,<br />
while the second term gives<br />
ρ 1 − ρ 2<br />
√ 2π<br />
<br />
2<br />
φ<br />
1 + ρ u<br />
<br />
= ρ 1 − ρ<br />
2 e− u2<br />
1+ρ<br />
2π<br />
. (A.30)<br />
Putting these two expressions tog<strong>et</strong>her and factorizing the term (1 + ρ)/(1 + ρ 2 ) allows us to obtain<br />
m20 L(u, u; ρ) =<br />
(1 + ρ)2<br />
1 − ρ 2<br />
u2<br />
−<br />
e 1+ρ<br />
2π<br />
<br />
1 + ρ2 1<br />
1 − ·<br />
1 − ρ<br />
−15 (1 + ρ2 )(1 + ρ) 2<br />
(1 − ρ) 3<br />
28<br />
u2 + 3(1 + ρ2 )(1 + ρ)<br />
(1 − ρ) 2<br />
· 1<br />
+ O<br />
u6 1<br />
u 8<br />
· 1<br />
u 4<br />
<br />
+ L(u, u; ρ) , (A.31)
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 267<br />
which finally yields<br />
m20 = u 2 (1 + ρ)2 1<br />
+ 2 (1 + ρ) − 2 ·<br />
1 − ρ u2 + 2(5 + 4ρ + ρ3 )(1 + ρ) 2<br />
(1 − ρ) 2<br />
<br />
1 1<br />
+ O<br />
u4 u6 <br />
. (A.32)<br />
A.2.4 Asymptotic behavior of the cross moment m11<br />
The cross moment m11 = E[X · Y | X > u, Y > u] is given by expression (A.15). The first and second<br />
terms in the right hand si<strong>de</strong> of (A.15) respectively give<br />
2ρ u φ(u)[1 − Φ(u)] = 2ρ<br />
1 − ρ 2<br />
√ 2π<br />
1 + ρ<br />
1 − ρ<br />
which, after factoriz<strong>at</strong>ion by (1 + ρ)/ρ, yields<br />
and finally<br />
m11 L(u, u; ρ) =<br />
(1 + ρ)2<br />
1 − ρ 2<br />
u2<br />
−<br />
e 1+ρ<br />
2π<br />
<br />
2 1 + ρ 1 1 + ρ<br />
1 − · + 3 ·<br />
1 − ρ u2 1 − ρ<br />
1<br />
u4 3 1 + ρ<br />
−15 ·<br />
1 − ρ<br />
1<br />
<br />
1<br />
+ O<br />
u6 u8 , (A.33)<br />
<br />
2<br />
φ<br />
1 + ρ u<br />
<br />
= 1 − ρ<br />
−30<br />
m11 = u 2 + 2 (1 + ρ) − (1 + ρ)2 (3 − ρ)<br />
(1 − ρ)<br />
u2<br />
−<br />
e 1+ρ<br />
2π<br />
<br />
ρ<br />
1 − 2<br />
1 − ρ<br />
ρ(1 + ρ)2 1<br />
·<br />
(1 − ρ) 3 u<br />
A.2.5 Asymptotic behavior of the correl<strong>at</strong>ion coefficient<br />
6 + O<br />
2 e− u2<br />
1+ρ<br />
2π<br />
· 1<br />
u2 + (16 − 9ρ + 3ρ2 )(1 + ρ) 3<br />
(1 − ρ) 2<br />
, (A.34)<br />
1 ρ(1 + ρ) 1<br />
· + 6 ·<br />
u2 (1 − ρ) 2 u4 <br />
1<br />
u8 <br />
+ ρ L(u, u; ρ), (A.35)<br />
· 1<br />
<br />
1<br />
+ O<br />
u4 u6 <br />
. (A.36)<br />
The conditional correl<strong>at</strong>ion coefficient conditioned on both X and Y larger than u is <strong>de</strong>fined by (A.12).<br />
Using the symm<strong>et</strong>ry b<strong>et</strong>ween X and Y , we have m10 = m01 and m20 = m02, which allows us to rewrite<br />
(A.12) as follows<br />
ρu = m11 − m10 2<br />
. (A.37)<br />
m20 − m10<br />
2<br />
Putting tog<strong>et</strong>her the previous results, we have<br />
which proves th<strong>at</strong><br />
m20 − m10 2 =<br />
m11 − m10 2 = ρ<br />
ρu = ρ<br />
(1 + ρ)2<br />
u 2<br />
(1 + ρ)3<br />
1 − ρ<br />
1 + ρ<br />
1 − ρ<br />
− 2 (4 − ρ + 3ρ2 + 3ρ3 )(1 + ρ) 2<br />
·<br />
1 − ρ<br />
1<br />
<br />
1<br />
+ O<br />
u4 u6 <br />
, (A.38)<br />
<br />
1 1<br />
· + O<br />
u4 u6 <br />
, (A.39)<br />
<br />
1 1<br />
· + O<br />
u2 u4 <br />
29<br />
and ρ ∈ [−1, 1). (A.40)
268 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
B Conditional correl<strong>at</strong>ion coefficient for Stu<strong>de</strong>nt’s variables<br />
B.1 Proposition<br />
L<strong>et</strong> us consi<strong>de</strong>r a pair of Stu<strong>de</strong>nt’s random variables (X, Y ) with ν <strong>de</strong>grees of freedom and unconditional<br />
correl<strong>at</strong>ion coefficient ρ. L<strong>et</strong> A be a subs<strong>et</strong> of R such th<strong>at</strong> Pr{Y ∈ A} > 0. The correl<strong>at</strong>ion coefficient of<br />
(X, Y ), conditioned on Y ∈ A <strong>de</strong>fined by<br />
can be expressed as<br />
with<br />
and<br />
⎡<br />
ρA =<br />
Cov(X, Y | Y ∈ A)<br />
Var(X | Y ∈ A) Var(Y | Y ∈ A) . (B.41)<br />
ρA =<br />
⎢ν<br />
− 1<br />
Var(Y | Y ∈ A) = ν ⎣<br />
ν − 2 ·<br />
ρ<br />
<br />
ρ 2 + E[E(x2 | Y )−ρ 2 Y 2 | Y ∈A]<br />
Var(Y | Y ∈A)<br />
<br />
ν Pr ν−2 Y ∈ A | ν − 2<br />
Pr{Y ∈ A | ν}<br />
E[E(X 2 | Y ) − ρ 2 Y 2 | Y ∈ A] = (1 − ρ 2 )<br />
B.2 Proof of the proposition<br />
ν<br />
ν − 2 ·<br />
<br />
ν Pr ν−2<br />
, (B.42)<br />
⎤<br />
<br />
2 ⎥ dy y · ty(y)<br />
y∈A<br />
− 1⎦<br />
−<br />
, (B.43)<br />
Pr{Y ∈ A | ν}<br />
<br />
Y ∈ A | ν − 2<br />
Pr{Y ∈ A | ν}<br />
. (B.44)<br />
L<strong>et</strong> the variables X and Y have a multivari<strong>at</strong>e Stu<strong>de</strong>nt’s distribution with ν <strong>de</strong>grees of freedom and a<br />
correl<strong>at</strong>ion coefficient ρ :<br />
PXY (x, y) =<br />
=<br />
Γ <br />
ν+2<br />
2<br />
νπ Γ <br />
ν+1<br />
2 1 − ρ2 <br />
ν + 1<br />
ν + y2 1/2 <br />
1 + x2 − 2ρxy + y 2<br />
ν (1 − ρ 2 )<br />
1<br />
<br />
1 − ρ2 tν(y)<br />
ν + 1<br />
· tν+1<br />
ν + y2 where tν(·) <strong>de</strong>notes the univari<strong>at</strong>e Stu<strong>de</strong>nt’s <strong>de</strong>nsity with ν <strong>de</strong>grees of freedom<br />
tν(x) =<br />
Γ <br />
ν+1<br />
2<br />
Γ <br />
ν<br />
2 (νπ) 1/2 ·<br />
1<br />
<br />
1 + x2<br />
Cν<br />
ν+1 = <br />
2<br />
ν 1 + x2<br />
ν<br />
L<strong>et</strong> us evalu<strong>at</strong>e Cov(X, Y | Y ∈ A):<br />
− ν+2<br />
2<br />
, (B.45)<br />
<br />
1/2<br />
x − ρy<br />
, (B.46)<br />
1 − ρ2 ν+1<br />
2<br />
. (B.47)<br />
Cov(X, Y | Y ∈ A) = E(X · Y | Y ∈ A) − E(X | Y ∈ A) · E(Y | Y ∈ A), (B.48)<br />
= E(E(X | Y ) · Y | Y ∈ A) − E(E(X | Y ) | Y ∈ A) · E(Y | Y ∈ A).(B.49)<br />
As it can be seen in equ<strong>at</strong>ion (B.46), E(X | Y ) = ρY , which gives<br />
Cov(X, Y | Y ∈ A) = ρ · E(Y 2 | Y ∈ A) − ρ · E(Y | Y ∈ A) 2 , (B.50)<br />
= ρ · Var(Y | Y ∈ A). (B.51)<br />
30
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 269<br />
Thus, we have<br />
<br />
ρA = ρ<br />
Var(Y | Y ∈ A)<br />
.<br />
Var(X | Y ∈ A)<br />
(B.52)<br />
Using the same m<strong>et</strong>hod as for the calcul<strong>at</strong>ion of Cov(X, Y | Y ∈ A), we find<br />
which yields<br />
as asserted in (B.42).<br />
Var(X | Y ∈ A) = E[E(X 2 | Y ) | Y ∈ A)] − E[E(X | Y ) | Y ∈ A)] 2 , (B.53)<br />
= E[E(X 2 | Y ) | Y ∈ A)] − ρ 2 · E[Y | Y ∈ A] 2 , (B.54)<br />
= E[E(X 2 | Y ) − ρ 2 Y 2 | Y ∈ A)] − ρ 2 · Var[Y | Y ∈ A], (B.55)<br />
ρA =<br />
<br />
ρ<br />
ρ 2 + E[E(x2 | Y )−ρ 2 Y 2 | Y ∈A]<br />
Var(Y | Y ∈A)<br />
(B.56)<br />
, (B.57)<br />
To go one step further, we have to evalu<strong>at</strong>e the three terms E(Y | Y ∈ A), E(Y 2 | Y ∈ A), and<br />
E[E(X 2 | Y ) | Y ∈ A].<br />
The first one is trivial to calcul<strong>at</strong>e :<br />
<br />
y∈A dy y · ty(y)<br />
E(Y | Y ∈ A) =<br />
. (B.58)<br />
Pr{Y ∈ A | ν}<br />
The second one gives<br />
so th<strong>at</strong><br />
<br />
y∈A dy y2 · ty(y)<br />
E(Y 2 | Y ∈ A) =<br />
, (B.59)<br />
Pr{Y ∈ A | ν}<br />
⎡<br />
⎢ν<br />
− 1<br />
= ν ⎣<br />
ν − 2 ·<br />
⎤<br />
ν Pr ν−2 Y ∈ A | ν − 2<br />
⎥<br />
− 1⎦<br />
, (B.60)<br />
Pr{Y ∈ A | ν}<br />
⎡<br />
⎢ν<br />
− 1<br />
Var(Y | Y ∈ A) = ν ⎣<br />
ν − 2 ·<br />
<br />
ν Pr ν−2 Y ∈ A | ν − 2<br />
Pr{Y ∈ A | ν}<br />
⎤<br />
<br />
2 ⎥ y∈A dy y · ty(y)<br />
− 1⎦<br />
−<br />
. (B.61)<br />
Pr{Y ∈ A | ν}<br />
To calcul<strong>at</strong>e the third term, we first need to evalu<strong>at</strong>e E(X2 | Y ). Using equ<strong>at</strong>ion (B.46) and the results given<br />
in (Abramovitz and Stegun 1972), we find<br />
which yields<br />
E(X 2 | Y ) =<br />
= ν + y2<br />
<br />
dx<br />
<br />
ν + 1<br />
ν + y2 1/2 x2 ν + 1<br />
· tν+1<br />
1 − ρ2 ν + y2 <br />
1/2<br />
x − ρy<br />
, (B.62)<br />
1 − ρ2 ν − 1 (1 − ρ2 ) − ρ 2 y 2 , (B.63)<br />
E[E(X 2 | Y ) − ρ 2 Y 2 | Y ∈ A] = ν<br />
ν − 1 (1 − ρ2 1 − ρ2<br />
) +<br />
ν − 1 E[Y 2 | Y ∈ A] , (B.64)<br />
31
270 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
and applying the result given in eq<strong>at</strong>ion (B.60), we finally obtain<br />
E[E(X 2 | Y ) − ρ 2 Y 2 | Y ∈ A] = (1 − ρ 2 )<br />
which conclu<strong>de</strong>s the proof.<br />
B.3 Conditioning on Y larger than v<br />
The conditioning s<strong>et</strong> is A = [v, +∞), thus<br />
Pr{Y ∈ A | ν} = ¯ Tν(v) = ν ν−1<br />
2<br />
y∈A<br />
Cν<br />
ν<br />
ν − 2 ·<br />
Pr<br />
+ O<br />
vν <br />
ν<br />
Pr Y ∈ A | ν − p =<br />
ν − p ¯ <br />
ν − p<br />
Tν−p v =<br />
ν<br />
ν ν−p<br />
2<br />
<br />
<br />
ν<br />
dy y · ty(y) =<br />
ν − 2 tν−2<br />
<br />
ν − 2<br />
v =<br />
ν<br />
ν ν<br />
<br />
ν<br />
ν−2 Y ∈ A | ν − 2<br />
Pr{Y ∈ A | ν}<br />
, (B.65)<br />
<br />
v −(ν+2)<br />
, (B.66)<br />
(ν − p) 1<br />
2<br />
<br />
Cν−p<br />
+ O v<br />
vν−p −(ν−p+2)<br />
, (B.67)<br />
2<br />
√ ν − 2<br />
<br />
Cν−2<br />
+ O<br />
vν−1 v −(ν−3)<br />
(B.68) ,<br />
where tν(·) and ¯ Tν(·) <strong>de</strong>note respectively the <strong>de</strong>nsity and the Stu<strong>de</strong>nt’s survival distribution with ν <strong>de</strong>grees<br />
of freedom and Cν is <strong>de</strong>fined in (B.47).<br />
Using equ<strong>at</strong>ion (B.42), one can thus give the exact expression of ρ + v . Since it is very cumbersomme, we will<br />
not write it explicitely. We will only give the asymptotic expression of ρ + v . In this respect, we can show th<strong>at</strong><br />
Thus, for large v,<br />
Var(Y | Y ∈ A) =<br />
E[E(X 2 | Y ) − ρ 2 Y 2 | Y ∈ A] =<br />
B.4 Conditioning on |Y | larger than v<br />
ν<br />
(ν − 2)(ν − 1) 2 v2 + O(1) (B.69)<br />
<br />
ν 1 − ρ<br />
ν − 2<br />
2<br />
ν − 1 v2 + O(1) . (B.70)<br />
ρ + ρ<br />
v −→ <br />
ρ2 <br />
ν−2<br />
+ (ν − 1) ν (1 − ρ2 . (B.71)<br />
)<br />
The conditioning s<strong>et</strong> is now A = (−∞, −v]∪[v, +∞), with v ∈ R+. Thus, the right hand si<strong>de</strong>s of equ<strong>at</strong>ions<br />
(B.66) and (B.67) have to be multiplied by two while<br />
<br />
dy y · ty(y) = 0, (B.72)<br />
y∈A<br />
for symm<strong>et</strong>ry reasons. So the equ<strong>at</strong>ion (B.70) still holds while<br />
Thus, for large v,<br />
Var(Y | Y ∈ A) =<br />
ρ s v −→<br />
<br />
ν<br />
(ν − 2) v2 + O(1) . (B.73)<br />
ρ<br />
ρ2 + 1<br />
<br />
ν−2<br />
(ν−1) ν (1 − ρ2 )<br />
32<br />
. (B.74)
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 271<br />
B.5 Conditioning on Y > v versus on |Y | > v<br />
The results (B.71) and (B.74) are valid for ν > 2, as one can expect since the second moment has to<br />
exist for the correl<strong>at</strong>ion coefficient to be <strong>de</strong>fined. We remark th<strong>at</strong> here, contrarily to the Gaussian case, the<br />
conditioning s<strong>et</strong> is not really important. In<strong>de</strong>ed with both conditioning s<strong>et</strong>, ρ + v and ρ s v goes to a constant<br />
different from zero and one, when v goes to infinity. This striking difference can be explained by the large<br />
fluctu<strong>at</strong>ions allowed by the Stu<strong>de</strong>nt’s distribution, and can be rel<strong>at</strong>ed to the fact th<strong>at</strong> the coefficient of tail<br />
<strong>de</strong>pen<strong>de</strong>nce for this distribution does not vanish even though the variables are anti-correl<strong>at</strong>ed (see section<br />
3.2 below).<br />
Contrarily to the Gaussian distribution which binds the fluctu<strong>at</strong>ions of the variables near the origin, the Stu<strong>de</strong>nt’s<br />
distribution allows for ‘wild’ fluctu<strong>at</strong>ions. These properties are thus responsible for the result th<strong>at</strong>,<br />
contrarily to the Gaussian case for which the conditional correl<strong>at</strong>ion coefficient goes to zero when conditioned<br />
on large signed values and goes to one when conditioned on large unsigned values, the conditional<br />
correl<strong>at</strong>ion coefficient for Stu<strong>de</strong>nt’s variables have a similar behavior in both cases. Intuitively, the large<br />
fluctu<strong>at</strong>ions of X for large v domin<strong>at</strong>e and control the asymptotic <strong>de</strong>pen<strong>de</strong>nce.<br />
33
272 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
C Proof of equ<strong>at</strong>ion (9)<br />
We assume th<strong>at</strong> X and Y are rel<strong>at</strong>ed by the equ<strong>at</strong>ion<br />
X = αY + ɛ , (C.75)<br />
where α is a non random real coefficient and ɛ an idiosyncr<strong>at</strong>ic noise in<strong>de</strong>pen<strong>de</strong>nt of Y , whose distribution<br />
is assumed to admit a moment of second or<strong>de</strong>r σ 2 ɛ . L<strong>et</strong> us also <strong>de</strong>note by σ 2 y the second moment of the<br />
variable Y .<br />
We have<br />
Cov(X, Y | Y ∈ A) = Cov(αY + ɛ, Y | Y ∈ A), (C.76)<br />
since Y and ɛ are in<strong>de</strong>pen<strong>de</strong>nt. We have also<br />
= αVar(Y | Y ∈ A) + Cov(ɛ, Y | Y ∈ A), (C.77)<br />
= αVar(Y | Y ∈ A), (C.78)<br />
Var(X | Y ∈ A) = = α 2 Var(Y | Y ∈ A) + 2 Cov(ɛ, Y | Y ∈ A) + Var(ɛ | Y ∈ A), (C.79)<br />
= α 2 Var(Y | Y ∈ A) + σ 2 ɛ , (C.80)<br />
where, again, we have used the in<strong>de</strong>pen<strong>de</strong>nce of Y and ɛ. This allows us to write<br />
Since<br />
we finally obtain<br />
which conclu<strong>de</strong> the proof.<br />
ρA =<br />
=<br />
αVar(Y | Y ∈ A)<br />
, (C.81)<br />
Var(Y | Y ∈ A)(α2Var(Y | Y ∈ A) + σ2 ɛ )<br />
sgn(α)<br />
<br />
1 + σ2 ɛ<br />
α2 1 · Var(Y | Y ∈A)<br />
ρA =<br />
ρ =<br />
sgn(α)<br />
<br />
1 + σ2 ɛ<br />
α 2 · 1<br />
Var(Y )<br />
ρ<br />
<br />
ρ2 + (1 − ρ2 )<br />
34<br />
. (C.82)<br />
Var(y)<br />
Var(y | y∈A)<br />
, (C.83)<br />
, (C.84)
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 273<br />
D Conditional Spearman’s rho<br />
The conditional Spearman’s rho has been <strong>de</strong>fined by<br />
We have<br />
ρs(˜v) =<br />
Cov(U, V | V ≥ ˜v)<br />
Var(U | V ≥ ˜v)Var(V | V ≥ ˜v) , (D.85)<br />
1 1<br />
<br />
˜v 0 · dC(u, v)<br />
1 1 1<br />
E[· | V ≥ ˜v] =<br />
=<br />
dC(u, v) 1 − ˜v ˜v 0<br />
1<br />
˜v<br />
1<br />
0<br />
· dC(u, v) , (D.86)<br />
thus, performing a simple integr<strong>at</strong>ion by parts, we obtain<br />
E[U | V ≥ ˜v] = 1 + 1<br />
1<br />
du C(u, ˜v) −<br />
1 − ˜v 0<br />
1<br />
<br />
,<br />
2<br />
(D.87)<br />
E[V | V ≥ ˜v]<br />
E[U<br />
=<br />
1 + ˜v<br />
,<br />
2<br />
(D.88)<br />
2 | V ≥ ˜v] = 1 + 2<br />
1<br />
du u C(u, ˜v) −<br />
1 − ˜v<br />
1<br />
<br />
,<br />
3<br />
(D.89)<br />
which yields<br />
so th<strong>at</strong><br />
0<br />
E[V 2 | V ≥ ˜v] = ˜v2 + ˜v + 1<br />
, (D.90)<br />
3<br />
1 1<br />
1<br />
1 + ˜v 1<br />
E[U · V | V ≥ ˜v] = + dv du C(u, v) + ˜v du C(u, ˜v) −<br />
2 1 − ˜v<br />
1<br />
<br />
, (D.91)<br />
2<br />
Cov(U, V | V ≥ ˜v) =<br />
ρs(˜v) = <br />
Var(U | V ≥ ˜v) =<br />
Var(V | V ≥ ˜v) =<br />
1 − 4˜v + 24 (1 − ˜v) 1<br />
˜v<br />
0<br />
1 1<br />
1<br />
dv du C(u, v) −<br />
1 − ˜v ˜v 0<br />
1<br />
1<br />
du C(u, ˜v) −<br />
2 0<br />
1<br />
, (D.92)<br />
4<br />
1<br />
1 − 4˜v 2<br />
2˜v − 1<br />
+ du u C(u, ˜v) +<br />
12 (1 − ˜v) 2 1 − ˜v 0<br />
(1 − ˜v) 2<br />
1<br />
du C(u, ˜v)<br />
0<br />
1<br />
−<br />
(1 − ˜v) 2<br />
1<br />
2<br />
du C(u, ˜v) , (D.93)<br />
(1 − ˜v)2<br />
12<br />
0<br />
, (D.94)<br />
<br />
12 1<br />
1−˜v ˜v dv 1<br />
0 du C(u, v) − 6 1<br />
0<br />
0 du u C(u, ˜v) + 12 (2˜v − 1) 1<br />
0<br />
0<br />
du C(u, ˜v) − 3<br />
du C(u, ˜v) − 12<br />
E Tail <strong>de</strong>pen<strong>de</strong>nce gener<strong>at</strong>ed by the Stu<strong>de</strong>nt’s factor mo<strong>de</strong>l<br />
We consi<strong>de</strong>r two random variables X and Y , rel<strong>at</strong>ed by the rel<strong>at</strong>ion<br />
<br />
1<br />
2<br />
0 du C(u, ˜v)<br />
(D.95)<br />
X = αY + ɛ, (E.96)<br />
35
274 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
where ɛ is a random variable in<strong>de</strong>pen<strong>de</strong>nt of Y and α a non random positive coefficient. Assume th<strong>at</strong> Y and<br />
ɛ have a Stu<strong>de</strong>nt’s distribution with <strong>de</strong>nsity:<br />
Cν<br />
PY (y) = <br />
1 + y2<br />
ν<br />
Pɛ(ɛ) =<br />
Cν<br />
<br />
σ 1 + ɛ2<br />
ν σ2 ν+1<br />
2<br />
ν+1<br />
2<br />
, (E.97)<br />
. (E.98)<br />
We first give a general expression for the probability for X to be larger than F −1<br />
X (u) knowing th<strong>at</strong> Y is<br />
(u) :<br />
larger than F −1<br />
Y<br />
LEMMA 1<br />
The probability th<strong>at</strong> X is larger than F −1<br />
X<br />
with<br />
Pr[X > F −1<br />
−1<br />
X (u)|Y > FY (u)] = ¯ Fɛ(η) + α<br />
∞<br />
1 − u F −1<br />
Y (u)<br />
−1<br />
(u) knowing th<strong>at</strong> Y is larger than F (u) is given by :<br />
η = F −1<br />
X<br />
Y<br />
dy ¯ FY (y) · Pɛ[αF −1<br />
Y (u) + η − αy] , (E.99)<br />
−1<br />
(u) − αF (u). (E.100)<br />
The proof of this lemma relies on a simple integr<strong>at</strong>ion by part and a change of variable, which are d<strong>et</strong>ailed<br />
in appendix E.1.<br />
Introducing the not<strong>at</strong>ion<br />
we can show th<strong>at</strong><br />
η = α 1 +<br />
Y<br />
˜Yu = F −1<br />
Y (u) , (E.101)<br />
<br />
σ<br />
<br />
ν1/ν − 1 ˜Yu + O(<br />
α<br />
˜ Y −1<br />
u ), (E.102)<br />
which allows us to conclu<strong>de</strong> th<strong>at</strong> η goes to infinity as u goes to 1 (see appendix E.2 for the <strong>de</strong>riv<strong>at</strong>ion of this<br />
result). Thus, ¯ Fɛ(η) goes to zero as u goes to 1 and<br />
∞ α<br />
λ = lim dy<br />
u→1 1 − u<br />
¯ FY (y) · Pɛ(α ˜ Yu + η − αy) . (E.103)<br />
Now, using the following result :<br />
LEMMA 2<br />
Assuming ν > 0 and x0 > 1,<br />
1<br />
lim<br />
ɛ→0 ɛ<br />
∞<br />
1<br />
˜Yu<br />
dx 1<br />
x ν<br />
<br />
Cν<br />
1 + x−x0<br />
ɛ<br />
2 ν+1<br />
2<br />
whose proof is given in appendix E.3, it is straigthforward to show th<strong>at</strong><br />
The final steps of this calcul<strong>at</strong>ion are given in appendix E.4.<br />
= 1<br />
xν , (E.104)<br />
0<br />
1<br />
λ =<br />
1 + <br />
σ ν . (E.105)<br />
α<br />
36
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 275<br />
E.1 Proof of Lemma 1<br />
By <strong>de</strong>finition,<br />
Pr[X > F −1<br />
X<br />
−1<br />
(u), Y > F (u)] =<br />
L<strong>et</strong> us perform an integr<strong>at</strong>ion by part :<br />
Defining η = F −1<br />
X<br />
Pr[X > F −1<br />
−1<br />
X (u), Y > F<br />
we obtain the result given in (E.99)<br />
Y<br />
=<br />
∞<br />
F −1<br />
X (u)<br />
∞<br />
F −1<br />
Y (u)<br />
dx<br />
∞<br />
F −1<br />
Y (u)<br />
dy PY (y) · Pɛ(x − αy) (E.106)<br />
dy PY (y) · ¯ Fɛ[F −1<br />
X (u) − αy]. (E.107)<br />
Y (u)] = − ¯ FY (y) · ¯ Fɛ(F −1<br />
X (u) − αy) ∞<br />
+ α<br />
∞<br />
F −1<br />
Y (u)<br />
F −1<br />
Y<br />
(u) +<br />
dy ¯ FY (y) · Pɛ(F −1<br />
X (u) − αy) (E.108)<br />
= (1 − u) ¯ Fɛ(F −1<br />
−1<br />
(u) − αF (u)) +<br />
+ α<br />
∞<br />
F −1<br />
Y (u)<br />
−1<br />
(u) − αF (u) (see equ<strong>at</strong>ion(E.100)), and dividing each term by<br />
E.2 Deriv<strong>at</strong>ion of equ<strong>at</strong>ion (E.102)<br />
Y<br />
X<br />
Y<br />
dy ¯ FY (y) · Pɛ(F −1<br />
X (u) − αy) (E.109)<br />
Pr[Y > F −1<br />
Y (u)] = 1 − u, (E.110)<br />
The factor Y and the idiosyncr<strong>at</strong>ic noise ɛ have Stu<strong>de</strong>nt’s distributions with ν <strong>de</strong>grees of freedom given by<br />
(E.97) and (E.98) respectively. It follows th<strong>at</strong> the survival distributions of Y and ɛ are :<br />
and<br />
¯FY (y) =<br />
ν ν−1<br />
2 Cν<br />
y ν + O(y −(ν+2) ), (E.111)<br />
¯Fɛ(ɛ) = σν ν ν−1<br />
2 Cν<br />
ɛ ν + O(ɛ −(ν+2) ), (E.112)<br />
(E.113)<br />
¯FX(x) = (αν + σ ν ) ν ν−1<br />
2 Cν<br />
x ν + O(x −(ν+2) ). (E.114)<br />
Using the not<strong>at</strong>ion (E.101), equ<strong>at</strong>ion (E.100) can be rewritten as<br />
whose solution for large ˜ Yu (or equivalently as u goes to 1) is<br />
<br />
σ<br />
<br />
ν1/ν η = α 1 + − 1<br />
α<br />
¯FX(η + α ˜ Yu) = ¯ FY ( ˜ Yu) = 1 − u, (E.115)<br />
˜Yu + O( ˜ Y −1<br />
u ). (E.116)<br />
To obain this equ<strong>at</strong>ion, we have used the asymptotic expressions of ¯ FX and ¯ FY given in (E.114) and (E.111).<br />
37
276 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
E.3 Proof of lemma 2<br />
We want to prove th<strong>at</strong>, assuming ν > 0 and x0 > 1,<br />
The change of variable<br />
gives<br />
1<br />
ɛ<br />
∞<br />
1<br />
dx 1<br />
x ν<br />
1<br />
lim<br />
ɛ→0 ɛ<br />
<br />
1 + 1<br />
ν<br />
∞<br />
1<br />
Cν<br />
x−x0<br />
ɛ<br />
Consi<strong>de</strong>r the second integral. We have<br />
which allows us to write<br />
so th<strong>at</strong><br />
<br />
<br />
<br />
<br />
<br />
∞<br />
x 0<br />
ɛ<br />
du<br />
1<br />
dx 1<br />
x ν<br />
2 ν+1<br />
2<br />
1<br />
<br />
1 + 1<br />
ν<br />
u =<br />
=<br />
(1 + u2 ) ν+1<br />
2<br />
(1 + ɛu<br />
x0 )ν<br />
Cν<br />
(1 + u2 ) ν+1<br />
2<br />
The next step of the proof is to show th<strong>at</strong><br />
L<strong>et</strong> us calcul<strong>at</strong>e<br />
<br />
<br />
x0 ɛ<br />
du<br />
<br />
1−x 0<br />
ɛ<br />
1<br />
(1 + ɛu<br />
x0 )ν<br />
x 0<br />
ɛ<br />
1−x 0<br />
ɛ<br />
Cν<br />
du<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
1<br />
(1 + ɛu<br />
x0 )ν<br />
<br />
<br />
<br />
− 1<br />
=<br />
x − x0<br />
ɛ<br />
Cν<br />
x−x0<br />
ɛ<br />
∞<br />
1−x 0<br />
ɛ<br />
= 1<br />
xν ∞<br />
1−x0 0 ɛ<br />
= 1<br />
x ν 0<br />
2 ν+1<br />
2<br />
= 1<br />
xν . (E.117)<br />
0<br />
, (E.118)<br />
x 0<br />
ɛ<br />
1−x0 ɛ<br />
+ 1<br />
xν ∞<br />
x0 0 ɛ<br />
1<br />
du<br />
(ɛu + x0) ν<br />
du<br />
du<br />
du<br />
1<br />
(1 + ɛu<br />
x0 )ν<br />
1<br />
(1 + ɛu<br />
x0 )ν<br />
1<br />
(1 + ɛu<br />
x0 )ν<br />
Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
+<br />
(E.119)<br />
(E.120)<br />
. (E.121)<br />
u ≥ x0<br />
, (E.122)<br />
ɛ<br />
<br />
<br />
<br />
<br />
<br />
≤<br />
Cν<br />
ν ν+1<br />
2 ɛ ν+1<br />
x ν+1<br />
0<br />
≤ ν ν+1<br />
2 ɛ ν+1<br />
x ν+1<br />
= ν ν+1<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
<br />
<br />
<br />
<br />
<br />
x0 ɛ<br />
0<br />
2 ɛ ν<br />
x ν 0<br />
, (E.123)<br />
∞<br />
1<br />
∞<br />
x 0<br />
ɛ<br />
du<br />
Cν<br />
(1 + ɛu<br />
x0 )ν<br />
dv Cν<br />
(1 + v) ν<br />
(E.124)<br />
(E.125)<br />
= O(ɛ ν ). (E.126)<br />
1−x 0<br />
ɛ<br />
38<br />
du<br />
−→ 1 as ɛ −→ 0. (E.127)<br />
1<br />
(1 + ɛu<br />
x0 )ν<br />
Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
−
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 277<br />
∞<br />
Cν<br />
− du<br />
−∞ (1 + u2<br />
<br />
<br />
<br />
ν+1 <br />
(E.128)<br />
ν ) 2 <br />
<br />
<br />
x <br />
0<br />
ɛ 1<br />
= du<br />
1−x0 (1 +<br />
ɛ<br />
ɛu<br />
<br />
Cν<br />
− 1<br />
)ν<br />
x0 (1 + u2<br />
ν+1 −<br />
ν ) 2<br />
1−x0 ɛ Cν<br />
− du<br />
−∞ (1 + u2<br />
∞<br />
Cν<br />
ν+1 − du<br />
x<br />
ν ) 2<br />
0<br />
ɛ (1 + u2<br />
<br />
<br />
<br />
ν+1 <br />
ν ) 2 (E.129)<br />
<br />
<br />
x <br />
0<br />
ɛ 1<br />
≤ du<br />
1−x0 (1 +<br />
ɛ<br />
ɛu<br />
<br />
Cν<br />
− 1<br />
)ν<br />
x0 (1 + u2<br />
<br />
<br />
<br />
ν+1 <br />
ν ) 2 +<br />
<br />
<br />
1−x0 ɛ Cν<br />
+ du<br />
−∞ (1 + u2<br />
<br />
<br />
<br />
ν+1 <br />
ν ) 2 +<br />
<br />
<br />
∞<br />
<br />
Cν<br />
du<br />
x0 ɛ (1 + u2<br />
<br />
<br />
<br />
ν+1 <br />
ν ) 2 . (E.130)<br />
The second and third integrals obviously behave like O(ɛν ) when ɛ goes to zero since we have assumed<br />
x0<br />
→ −∞ and ɛ → ∞ when ɛ → 0. For the first integral, we have<br />
<br />
<br />
x0 ɛ<br />
<br />
<br />
<br />
1<br />
du<br />
(1 + ɛu<br />
<br />
− 1<br />
)ν<br />
x0<br />
Cν<br />
<br />
<br />
<br />
<br />
≤<br />
x0 ɛ<br />
<br />
<br />
1<br />
du <br />
(1<br />
+ ɛu<br />
<br />
<br />
<br />
− 1<br />
)ν <br />
x0<br />
Cν<br />
. (E.131)<br />
x0 > 1 wh<strong>at</strong> ensures th<strong>at</strong> 1−x0<br />
ɛ<br />
1−x 0<br />
ɛ<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
1−x 0<br />
ɛ<br />
The function <br />
<br />
1<br />
(1 + ɛu<br />
<br />
<br />
<br />
− 1<br />
)ν <br />
x0<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
(E.132)<br />
vanishes <strong>at</strong> u = 0, is convex for u ∈ [ 1−x0<br />
ɛ , 0] and concave for u ∈ [0, x0<br />
ɛ ] (see also figure 13), so th<strong>at</strong> there<br />
are two constants A, B > 0 such th<strong>at</strong><br />
<br />
<br />
1<br />
<br />
(1<br />
+ ɛu<br />
<br />
<br />
<br />
− 1<br />
)ν x0 ≤ −xν <br />
<br />
1<br />
<br />
(1<br />
+<br />
<br />
0 − 1<br />
1 − x0<br />
ɛ · u = −A · ɛ · u, ∀u ∈ , 0<br />
x0 − 1 ɛ<br />
(E.133)<br />
ɛu<br />
<br />
<br />
<br />
− 1<br />
)ν x0<br />
≤<br />
νɛ<br />
<br />
u = B · ɛ · u, ∀u ∈ 0,<br />
x0<br />
x0<br />
<br />
.<br />
ɛ<br />
(E.134)<br />
We can thus conclu<strong>de</strong> th<strong>at</strong><br />
<br />
<br />
x <br />
0<br />
ɛ 1<br />
du<br />
(1 + ɛu<br />
<br />
− 1<br />
)ν<br />
x0<br />
1−x 0<br />
ɛ<br />
Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
<br />
<br />
<br />
<br />
<br />
0<br />
≤ −A · ɛ<br />
1−x0 ɛ<br />
x0 ɛ<br />
+ B · ɛ du<br />
0<br />
du<br />
u · Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
u · Cν<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
(E.135)<br />
= O(ɛ α ), (E.136)<br />
with α = min{ν, 1}. In<strong>de</strong>ed, the two integrals can be perfomed exactly, which shows th<strong>at</strong> they behave as<br />
O(1) if ν > 1 and as O(ɛν−1 ) otherwise. Thus, we finally obtain<br />
<br />
<br />
x0 ɛ 1<br />
du<br />
(1 + ɛu<br />
x0 )ν<br />
<br />
<br />
Cν <br />
− 1<br />
= O(ɛα ). (E.137)<br />
1−x 0<br />
ɛ<br />
(1 + u2<br />
ν+1<br />
ν ) 2<br />
39
278 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Putting tog<strong>et</strong>her equ<strong>at</strong>ions (E.126) and (E.137) we obtain<br />
<br />
<br />
∞<br />
1<br />
dx<br />
ɛ<br />
1<br />
1<br />
xν Cν<br />
which conclu<strong>de</strong>s the proof.<br />
<br />
1 + 1<br />
ν<br />
E.4 Deriv<strong>at</strong>ion of equ<strong>at</strong>ion (E.105)<br />
From equ<strong>at</strong>ion (E.111), we can <strong>de</strong>duce<br />
¯FY (y) =<br />
Using equ<strong>at</strong>ions (E.98) and (E.102), we obtain<br />
where<br />
x−x0<br />
ɛ<br />
2 ν+1<br />
2<br />
Pɛ(α ˜ Yu + η − αy) = Pɛ(γ ˜ Yu − αy) ·<br />
− 1<br />
xν <br />
<br />
<br />
<br />
= O(ɛ<br />
0 <br />
<br />
min{ν,1} ) , (E.138)<br />
ν−1<br />
ν 2 Cν<br />
yν −2<br />
1 + O(y ) . (E.139)<br />
<br />
γ = α 1 +<br />
Putting tog<strong>et</strong>her these results yields for the leading or<strong>de</strong>r<br />
∞<br />
dy ¯ FY (y) · Pɛ(α ˜ Yu + η − αy) =<br />
˜Yu<br />
= ν ν−1<br />
<br />
1 + O( ˜ Y −2<br />
<br />
u ) , (E.140)<br />
<br />
σ<br />
ν1/ν . (E.141)<br />
α<br />
∞<br />
dy<br />
˜Yu<br />
2 Cν<br />
ν<br />
α ˜ Yu<br />
ν−1<br />
ν 2 Cν<br />
yν ·<br />
∞<br />
1<br />
<br />
σ<br />
dx 1<br />
·<br />
xν where the change of variable x = y<br />
has been performed in the last equ<strong>at</strong>ion.<br />
˜Yu<br />
Cν<br />
1 + (γ ˜ Yu−αy) 2<br />
ν σ 2<br />
<br />
1 + 1<br />
ν<br />
Cν α ˜ Yu<br />
σ<br />
x− γ<br />
α<br />
σ<br />
α ˜ Yu<br />
ν+1<br />
2<br />
2 ν+1<br />
2<br />
, (E.142)<br />
We now apply lemma 2 with x0 = γ<br />
σ<br />
α > 1 and ɛ =<br />
α ˜ which goes to zero as u goes to 1. This gives<br />
Yu<br />
which shows th<strong>at</strong><br />
thus<br />
which finally yields<br />
∞<br />
dy ¯ FY (y) · Pɛ(α ˜ Yu + η − αy) ∼u→1<br />
˜Yu<br />
Pr[X > F −1<br />
−1<br />
X (u), Y > FY (u)] ∼u→1 F −1<br />
Y ( ˜ Yu)<br />
Pr[X > F −1<br />
X<br />
−1<br />
(u)|Y > F (u)] ∼u→1<br />
λ =<br />
Y<br />
1<br />
1 + σ<br />
α<br />
40<br />
ν ν−1<br />
2 Cν<br />
α ˜ Y ν u<br />
ν α<br />
γ<br />
ν α<br />
= (1 − u)<br />
γ<br />
, (E.143)<br />
, (E.144)<br />
ν α<br />
, (E.145)<br />
γ<br />
ν α<br />
, (E.146)<br />
γ<br />
ν . (E.147)
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 279<br />
References<br />
Abramovitz, E. and I.A. Stegun, 1972, Handbook of M<strong>at</strong>hem<strong>at</strong>ical functions (Dover Public<strong>at</strong>ions, New<br />
York).<br />
An<strong>de</strong>rsen, J.V. and D. Sorn<strong>et</strong>te, 2001, Have your cake and e<strong>at</strong> it too: increasing r<strong>et</strong>urns while lowering large<br />
risks! Journal of Risk Finance 2, 70-82.<br />
Ang, A. and G. Bekaert, 2001, Intern<strong>at</strong>ional ass<strong>et</strong> alloc<strong>at</strong>ion with regime shifts, forthcoming Review of<br />
Financial Studies. .<br />
Ang, A. and J. Chen, 2001, Asymm<strong>et</strong>ric correl<strong>at</strong>ions of equity portfolios, forthcoming Journal of Financial<br />
Economics.<br />
Baig, T. and I. Golgfajn, 1998, Financial mark<strong>et</strong> contagion in the Asian crisis, IMF Working Paper.<br />
Bhansali, V. and M.B. Wise, 2001, Forecasting portfolio risk in normal and stressed mark<strong>et</strong>s, working paper<br />
(preprint <strong>at</strong> http://xxx.lanl.gov/abs/nlin.AO/0108022)<br />
Bookstaber, R., 1997, Global risk management: are we missing the point? Journal of Portfolio Management,<br />
23, 102-107.<br />
Boyer, B.H., M.S Gibson and M. Laur<strong>et</strong>an, 1997, Pitfalls in tests for changes in correl<strong>at</strong>ions, Intern<strong>at</strong>ional<br />
Finance Discussion Paper 597, Board of the Governors of the Fe<strong>de</strong>ral Reserve System.<br />
Davis, R.A., T. Mikosch and B. Basrak, 1999, Sample ACF of multivari<strong>at</strong>e stochastic recurrence equ<strong>at</strong>ions<br />
with applic<strong>at</strong>ion to GARCH, Working paper.<br />
Calvo, S. and C.M. Reinhart, 1995, Capital flows to L<strong>at</strong>in Americ: Is there evi<strong>de</strong>nce of contagion effects?,<br />
in G.A. Calvo, M. Goldstein and E. Haochreiter,eds: Priv<strong>at</strong>e Captial Flows to Emerging Mark<strong>et</strong> After<br />
the Mexican Crisis (Institue for Interno<strong>at</strong>ional Economics, Washington DC)<br />
Cizeau, P., M. Potters and J.P. Bouchaud, 2001, Correl<strong>at</strong>ion structure of extreme stock r<strong>et</strong>urns, Quantit<strong>at</strong>ive<br />
Finance 1, 217-222.<br />
Claessen, S., R.W. Dornbush and Y.C. Park, 2001, Contagion: Why crises spread and how this can be<br />
stopped, in S. Cleassens and K.J. Forbes eds, Intern<strong>at</strong>ional Financial Contagion (Kluwer Aca<strong>de</strong>mic<br />
Press).<br />
Coles, S., J. Heffernan and J. Tawn, 1999, Depen<strong>de</strong>nce measures for extreme value analyses, Extremes 2,<br />
339-365.<br />
Embrechs, P., A.J. McNeil and D. Straumann, 1999, Correl<strong>at</strong>ion : Pitfalls and Altern<strong>at</strong>ives. Risk, 69-71.<br />
Embrechts, P., A.J. McNeil and D. Straumann, 2001, Correl<strong>at</strong>ion and Depen<strong>de</strong>ncy in Risk Management :<br />
Properties and Pitfalls, in : Dempster, M., ed., Value <strong>at</strong> Risk and Beyond (Cambridge University<br />
Press).<br />
Forbes, K.J. and R. Rigobon, 2002, No contagion, only inter<strong>de</strong>pen<strong>de</strong>nce: measuring stock mark<strong>et</strong> comovements,<br />
forthcoming Journal of Finance.<br />
Frees, E. and E. Val<strong>de</strong>z, 1998, Un<strong>de</strong>rstanding Rel<strong>at</strong>ionships using copulas, North Americam Actuarial Journal<br />
2, 1-25.<br />
41
280 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Frie<strong>de</strong>n, J.A., Debt, Development, and Democracy: Mo<strong>de</strong>rn Political Economy and L<strong>at</strong>in America, 1965-<br />
1985, Princ<strong>et</strong>on University Press, 1992.<br />
J.A. Frie<strong>de</strong>n, P. Ghezzi and E. Stein, Politics and Exchange R<strong>at</strong>es: A Cross-Country Approach to L<strong>at</strong>in<br />
America, Harvard University working paper, October 2000<br />
J.A. Frie<strong>de</strong>n and E. Stein, The Political Economy of Exchange R<strong>at</strong>e Policy in L<strong>at</strong>in America: An Analytical<br />
Overview, Harvard University working paper, October 2000<br />
Hartman, P., S. Stra<strong>et</strong>mans and C.G. <strong>de</strong> Vries, 2001, Ass<strong>et</strong> mark<strong>et</strong> linkages in crisis periods, European<br />
Central Bank, Working paper n o 71.<br />
Hauksson, H.A., M.M. Dacorogna, T. Domenig, U.A. Müller and G. Samorodnitsky, 2001, Multivari<strong>at</strong>e<br />
Extremes, Aggreg<strong>at</strong>ion and Risk Estim<strong>at</strong>ion, Quantit<strong>at</strong>ive Finance 1, 79-95.<br />
Herffernan J.E., 2000, A directory of tail <strong>de</strong>pen<strong>de</strong>nce, Extremes 3, 279-290.<br />
Hult, H. and F. Lindskog, 2001, Multivari<strong>at</strong>e extremes, aggreg<strong>at</strong>ion and <strong>de</strong>pen<strong>de</strong>nce in elliptical distributions,<br />
Risklab working paper.<br />
Jensen, J.L., 1995, Saddlepoint Approxim<strong>at</strong>ions (Oxford University Press).<br />
Joe, H., 1997, Multivari<strong>at</strong>e mo<strong>de</strong>ls and <strong>de</strong>pen<strong>de</strong>nce concepts (Chapman & Hall, London)<br />
Johnson, N.L. and S. Kotz, 1972, Distributions in st<strong>at</strong>istics: Continuous multivari<strong>at</strong>e distributions (John<br />
Willey and Sons).<br />
Juri, A. and M.V. Wüthrich, 2002, Copula convergence theorem for tail events, working paper, Risklab.<br />
King, M. and S. Wadhwani, 1990, Transmission of vol<strong>at</strong>ility b<strong>et</strong>ween stock mark<strong>et</strong>s, The Review of Financial<br />
Studies 3, 5-330.<br />
Kulpa, T., 1999, On approxim<strong>at</strong>ions of copulas, intern<strong>at</strong>ional Journal of M<strong>at</strong>hem<strong>at</strong>ics and M<strong>at</strong>hem<strong>at</strong>ical<br />
sciences 22, 259-269.<br />
Ledford, A.W. and J.A. Tawn,1996 St<strong>at</strong>istics for near in<strong>de</strong>pen<strong>de</strong>nce in multivari<strong>at</strong>e extrem values,<br />
Biom<strong>et</strong>rika 83, 169-187.<br />
Ledford, A.W. and J.A. Tawn, 1998, Concomitant tail behavior for extremes, Adv. Appl. Prob. 30, 197-215.<br />
Lee, S.B. and K.J. Kim, 1993, Does the October 1987 crash strengthen the co-movements among n<strong>at</strong>ional<br />
stock mark<strong>et</strong>s?, Review of Financial Economics 3, 89-102.<br />
Li, X., P. Mikusincki and M.D. Taylor, 1998, Strong approxim<strong>at</strong>ion of copulas, Journal of M<strong>at</strong>hem<strong>et</strong>ical<br />
Analisys and Applic<strong>at</strong>ions 225, 608-623.<br />
Lindskog, F., Mo<strong>de</strong>lling Depen<strong>de</strong>nce with Copulas, Risklab working paper.<br />
Longin F. and B. Solnik, 1995, Is the correl<strong>at</strong>ion in intern<strong>at</strong>ional equity r<strong>et</strong>urns constant: 1960-1990? Journal<br />
of Intern<strong>at</strong>ional Money and Finance 14, 3-26.<br />
Longin F. and B. Solnik, 2001, Extreme Correl<strong>at</strong>ion of Intern<strong>at</strong>ional Equity Mark<strong>et</strong>s, The Journal of Finance<br />
LVI, 649-676.<br />
Lor<strong>et</strong>an, M., 2000, Evalu<strong>at</strong>ing changes in correl<strong>at</strong>ions during periods of high mark<strong>et</strong> vol<strong>at</strong>ility, Global Investor<br />
135, 65-68.<br />
42
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 281<br />
Lor<strong>et</strong>an, M. and W.B. English, 2000, Working paper 000-658, Board of Governors of the Fe<strong>de</strong>ral Reserve<br />
System<br />
Malevergne, Y. and D.Sorn<strong>et</strong>te, 2001, Testing the Gaussian copula hypothesis for financial ass<strong>et</strong>s <strong>de</strong>pen<strong>de</strong>nce,<br />
Working paper.<br />
Malevergne, Y. and D.Sorn<strong>et</strong>te, 2002, Tail <strong>de</strong>pen<strong>de</strong>nce for factor mo<strong>de</strong>ls, Working paper.<br />
Mansilla, R., 2001, Algorithmic complexity of real financial mark<strong>et</strong>s, Physica A 301, 483-492.<br />
Meerschaert, M.M. and H.P. Scheffler, 2001, Sample cross-correl<strong>at</strong>ions for moving averages with regularly<br />
varying tails, Journal of Time Series Analysis 22, 481-492.<br />
Nelsen, R.B., 1998, An Introduction to Copulas. Lectures Notes in st<strong>at</strong>istic 139 (Springer Verlag, New<br />
York).<br />
P<strong>at</strong>ton, J.A., 2001, Estim<strong>at</strong>ion of copula mo<strong>de</strong>ls for time series of possibly different lengths, U of California,<br />
Econ. Disc. Paper No. 2001-17.<br />
Poon, S.H., M. Rockinger and J. Tawn, 2001, New extreme-value <strong>de</strong>pen<strong>de</strong>nce measures and finance applic<strong>at</strong>ions,<br />
working paper.<br />
Quintos, C.E., 2001, Estim<strong>at</strong>ing tail <strong>de</strong>pen<strong>de</strong>nce and testing for contagion using tail indices, working paper.<br />
Quintos, C.E., Z.H. Fan and P.C.B. Phillips, 2001, Structural change tests in tail behaviour and the Asian<br />
crisis, Review of Economic Studies 68, 633-663.<br />
Ramchand, L. and R. Susmel, 1998, Vol<strong>at</strong>ility and cross correl<strong>at</strong>ion across major stock mark<strong>et</strong>s, Journal of<br />
Empirical Finance 5, 397-416.<br />
Ross, S., 1976, The arbitrage theory of capital ass<strong>et</strong> pricing, Journal of Economic Theory 17, 254-286.<br />
Scaill<strong>et</strong>, O. 2000, Nonparam<strong>et</strong>ric estim<strong>at</strong>ion of copulas for time series, Working paper.<br />
Sharpe, W., 1964, Capital ass<strong>et</strong>s prices: a theory of mark<strong>et</strong> equilibrium un<strong>de</strong>r conditions of risk, Journal of<br />
Finance, 19, 425-442.<br />
Silvapulle, P. and C.W.J. Granger, 2001, Large r<strong>et</strong>urns, conditional correl<strong>at</strong>ion and portfolio diversific<strong>at</strong>ion:<br />
a value-<strong>at</strong>-risk approach, Quantit<strong>at</strong>ive Finance 1, 542-551.<br />
Sorn<strong>et</strong>te, D. P. Simon<strong>et</strong>ti and J. V. An<strong>de</strong>rsen, 2000, φ q -field theory for Portfolio optimiz<strong>at</strong>ion: “f<strong>at</strong> tails” and<br />
non-linear correl<strong>at</strong>ions, Physics Report 335, 19-92.<br />
Sorn<strong>et</strong>te, D., J.V. An<strong>de</strong>rsen and P. Simon<strong>et</strong>ti, 2000, Portfolio Theory for “F<strong>at</strong> Tails”, Intern<strong>at</strong>ional Journal<br />
of Theor<strong>et</strong>ical and Applied Finance 3, 523-535.<br />
Starica, C., 1999, Multivari<strong>at</strong>e extremes for mo<strong>de</strong>ls with constant conditional correl<strong>at</strong>ions, Journal of Empirical<br />
Finance 6, 515-553.<br />
Tsui, A.K. and Q. Yu, 1999, Constant conditional correl<strong>at</strong>ion in a bivari<strong>at</strong>e GARCH mo<strong>de</strong>l: evi<strong>de</strong>nce from<br />
the stock mark<strong>et</strong>s of China, M<strong>at</strong>hem<strong>at</strong>ics and Computers in Simul<strong>at</strong>ion 48, 503-509.<br />
43
282 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Neg<strong>at</strong>ive Tail<br />
Argentina Brazil Chile Mexico<br />
Argentina - 0.28 (0.04) 0.25 (0.04) 0.25 (0.05)<br />
Brazil - 0.19 (0.03) 0.25 (0.05)<br />
Chile - 0.24 (0.07)<br />
Mexico -<br />
Positive Tail<br />
Argentina Brazil Chile Mexico<br />
Argentina - 0.21 (0.06) 0.20 (0.04) 0.22 (0.04)<br />
Brazil - 0.28 (0.04) 0.19 (0.04)<br />
Chile - 0.19 (0.03)<br />
Mexico -<br />
Table 1: Coefficients of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the four L<strong>at</strong>in American mark<strong>et</strong>s. The figure within<br />
parenthesis gives the standard <strong>de</strong>vi<strong>at</strong>ion of the estim<strong>at</strong>ed value <strong>de</strong>rived un<strong>de</strong>r the assumption of asymptotic<br />
normality of the estim<strong>at</strong>ors. Only the coefficients above the diagonal are indic<strong>at</strong>ed since they are symm<strong>et</strong>ric.<br />
44
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 283<br />
Stu<strong>de</strong>nt Hypothesis nu=3<br />
Argentina Brazil Chile Mexico<br />
Argentina - 0.24 0.25 0.27<br />
Brazil - 0.24 0.27<br />
Chile - 0.28<br />
Mexico -<br />
Table 2: Coefficients of tail <strong>de</strong>pen<strong>de</strong>nce <strong>de</strong>rived un<strong>de</strong>r the assumption of a Stu<strong>de</strong>nt’s copula with three<br />
<strong>de</strong>grees of freedom<br />
45
284 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Bivari<strong>at</strong>e Gaussian<br />
Bivari<strong>at</strong>e Stu<strong>de</strong>nt’s<br />
√ ρ 1 ·<br />
1−ρ2 v<br />
ρ<br />
ρ<br />
2 +(ν−1)ν−2<br />
(1−ρ ν<br />
2 )<br />
ρ + v ρ s v ρu<br />
(3) 1 − 1<br />
2<br />
(6)<br />
1−ρ 2<br />
ρ 2<br />
ρ<br />
ρ<br />
2 + 1<br />
(1−ρ (ν−1)ν−2<br />
ν<br />
2 )<br />
1<br />
v2 (4) ρ 1+ρ 1<br />
1−ρ · u2 (13)<br />
(7) -<br />
Gaussian Factor Mo<strong>de</strong>l same as (3) same as (4) same as (13)<br />
Stu<strong>de</strong>nt’s Factor Mo<strong>de</strong>l 1 − K<br />
v 2 (11) 1 − K<br />
v 2 (11) -<br />
Table 3: Large v and u <strong>de</strong>pen<strong>de</strong>nce of the conditional correl<strong>at</strong>ions ρ + v (signed condition), ρ s v (unsigned<br />
condition) and ρu (on both variables) for the different mo<strong>de</strong>ls studied in the present paper, <strong>de</strong>scribed in the<br />
first column. The numbers in parentheses give the equ<strong>at</strong>ion numbers from which the formulas are <strong>de</strong>rived.<br />
The factor mo<strong>de</strong>l is <strong>de</strong>fined by (8), i.e., X = αY + ɛ. ρ is the unconditional correl<strong>at</strong>ion coefficient.<br />
46
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 285<br />
ρ + v=∞ ρ s v=∞ ρu=∞ λ ¯ λ<br />
Bivari<strong>at</strong>e Gaussian 0 1 0 0 ρ<br />
Bivari<strong>at</strong>e Stu<strong>de</strong>nt’s see Table 3 see Table 3 - 2 · ¯ Tν+1<br />
√ν + 1<br />
1−ρ<br />
1+ρ<br />
Gaussian Factor Mo<strong>de</strong>l 0 1 0 0 ρ<br />
Stu<strong>de</strong>nt’s Factor Mo<strong>de</strong>l 1 1 -<br />
ρ ν<br />
ρ ν +(1−ρ 2 ) ν/2<br />
Table 4: Asymptotic values of ρ + v , ρs v and ρu for v → +∞ and u → ∞ and comparison with the tail<strong>de</strong>pen<strong>de</strong>nce<br />
λ and ¯ λ for the four mo<strong>de</strong>ls indic<strong>at</strong>ed in the first column. The factor mo<strong>de</strong>l is <strong>de</strong>fined by (8),<br />
i.e., X = αY + ɛ. ρ is the unconditional correl<strong>at</strong>ion coefficient. For the Stu<strong>de</strong>nt’s factor mo<strong>de</strong>l, Y and ɛ<br />
have centered Stu<strong>de</strong>nt’s distributions with the same number ν of <strong>de</strong>grees of freedom and their scale factors<br />
are respectively equal to 1 and σ, so th<strong>at</strong> ρ = (1 + σ2<br />
α2 ) −1/2 . For the Bivari<strong>at</strong>e Stu<strong>de</strong>nt’s distribution, we<br />
refer to Table 1 for the constant values of ρ + v=∞ and ρs v=∞.<br />
47<br />
<br />
1<br />
1
286 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −3<br />
10 −4<br />
10 0<br />
10 −1<br />
10 −2<br />
10 −3<br />
10 −4<br />
10 −4<br />
10 −4<br />
Argentina<br />
Brazil<br />
Chile<br />
Mexico<br />
Argentina<br />
Brazil<br />
Chile<br />
Mexico<br />
10 −3<br />
10 −3<br />
Positive Tail<br />
10 −2<br />
Neg<strong>at</strong>ive Tail<br />
10 −2<br />
Figure 1:<br />
48<br />
μ=2<br />
10 −1<br />
μ=2<br />
10 −1<br />
10 0<br />
10 0
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 287<br />
+,−<br />
ρ ,y=Brazil<br />
v<br />
+,−<br />
ρ ,y=Argentina<br />
v<br />
1<br />
0.5<br />
0<br />
−0.5<br />
Argentina−Brazil<br />
−1<br />
−5 −4 −3 −2 −1<br />
v<br />
0 1 2 3<br />
1<br />
0.5<br />
0<br />
−0.5<br />
−1<br />
−4 −3 −2 −1 0<br />
v<br />
1 2 3 4<br />
Figure 2:<br />
49
288 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
s<br />
ρ ,y=Brazil<br />
v<br />
s<br />
ρ ,y=Argentina<br />
v<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Argentina−Brazil<br />
−0.2<br />
0 0.5 1 1.5 2 2.5<br />
v<br />
3 3.5 4 4.5 5<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
0 0.5 1 1.5 2<br />
v<br />
2.5 3 3.5 4<br />
Figure 3:<br />
50
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 289<br />
+,−<br />
ρ ,y=Chile<br />
v<br />
+,−<br />
ρ ,y=Brazil<br />
v<br />
1<br />
0.5<br />
0<br />
−0.5<br />
Brazil−Chile<br />
−1<br />
−4 −3 −2 −1 0<br />
v<br />
1 2 3 4<br />
1<br />
0.5<br />
0<br />
−0.5<br />
−1<br />
−4 −3 −2 −1 0 1 2 3<br />
v<br />
Figure 4:<br />
51
290 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
s<br />
ρ ,y=Chile<br />
v<br />
s<br />
ρ ,y=Brazil<br />
v<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Brazil−Chile<br />
−0.2<br />
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5<br />
v<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
0 1 2 3<br />
v<br />
4 5 6<br />
Figure 5:<br />
52
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 291<br />
+,−<br />
ρ ,y=Mexico<br />
v<br />
+,−<br />
ρ ,y=Chile<br />
v<br />
1<br />
0.5<br />
0<br />
−0.5<br />
Chile−Mexico<br />
−1<br />
−4 −3 −2 −1 0<br />
v<br />
1 2 3 4<br />
1<br />
0.5<br />
0<br />
−0.5<br />
−1<br />
−4 −3 −2 −1 0<br />
v<br />
1 2 3 4<br />
Figure 6:<br />
53
292 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
s<br />
ρ ,y=Mexico<br />
v<br />
s<br />
ρ ,y=Chile<br />
v<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
Chile−Mexico<br />
−0.2<br />
0 1 2 3<br />
v<br />
4 5 6<br />
1.2<br />
1<br />
0.8<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5<br />
v<br />
Figure 7:<br />
54
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 293<br />
0.9<br />
0.9<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
0.8<br />
ρ=0.1<br />
ρ=0.3<br />
ρ=0.5<br />
ρ=0.7<br />
ρ=0.9<br />
0.8<br />
0.7<br />
0.7<br />
0.6<br />
0.6<br />
0.5<br />
0.5<br />
ρ s (v)<br />
ρ s (v)<br />
0.4<br />
0.4<br />
55<br />
0.3<br />
0.3<br />
0.2<br />
0.2<br />
0.1<br />
0.1<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
v<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
v<br />
Figure 8:
294 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
+,−<br />
ρ ,y=Brazil<br />
v<br />
+,−<br />
ρ ,y=Argentina<br />
v<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
Argentina−Brazil<br />
0 0.1 0.2 0.3 0.4 0.5<br />
v<br />
0.6 0.7 0.8 0.9 1<br />
−0.4<br />
0 0.1 0.2 0.3 0.4 0.5<br />
v<br />
0.6 0.7 0.8 0.9 1<br />
Figure 9:<br />
56
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 295<br />
+,−<br />
ρ ,y=Chile<br />
v<br />
+,−<br />
ρ ,y=Brazil<br />
v<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
0.2<br />
0<br />
−0.2<br />
−0.4<br />
−0.6<br />
Brazil−Chile<br />
0 0.1 0.2 0.3 0.4 0.5<br />
v<br />
0.6 0.7 0.8 0.9 1<br />
−0.8<br />
0 0.1 0.2 0.3 0.4 0.5<br />
v<br />
0.6 0.7 0.8 0.9 1<br />
Figure 10:<br />
57
296 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
+,−<br />
ρ ,y=Mexico<br />
v<br />
+,−<br />
ρ ,y=Chile<br />
v<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
−0.1<br />
Chile−Mexico<br />
−0.2<br />
0 0.1 0.2 0.3 0.4 0.5<br />
v<br />
0.6 0.7 0.8 0.9 1<br />
0.6<br />
0.4<br />
0.2<br />
0<br />
−0.2<br />
0 0.1 0.2 0.3 0.4 0.5<br />
v<br />
0.6 0.7 0.8 0.9 1<br />
Figure 11:<br />
58
9.1. Les différentes mesures <strong>de</strong> dépendances extrêmes 297<br />
Stu<strong>de</strong>nt’s Factor Mo<strong>de</strong>l<br />
Stu<strong>de</strong>nt’s Copula<br />
1<br />
1<br />
ν=3<br />
ν=5<br />
ν=10<br />
ν=20<br />
ν=50<br />
ν=100<br />
0.9<br />
0.9<br />
0.8<br />
ν=3<br />
ν=5<br />
ν=10<br />
ν=20<br />
ν=50<br />
ν=100<br />
0.8<br />
0.7<br />
0.7<br />
0.6<br />
0.6<br />
0.5<br />
λ<br />
0.5<br />
λ<br />
0.4<br />
0.4<br />
59<br />
0.3<br />
0.3<br />
0.2<br />
0.2<br />
0.1<br />
0.1<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
0<br />
0<br />
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1<br />
ρ<br />
ρ<br />
Figure 12:
298 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
0<br />
(1−x )/ε<br />
0<br />
−<br />
ν<br />
x −1<br />
0 ⋅ ε ⋅ u<br />
x −1<br />
0<br />
0<br />
u<br />
Figure 13:<br />
60<br />
ν ⋅ ε<br />
⋅ u<br />
x<br />
0<br />
x 0 /ε
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 299<br />
9.2 Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue<br />
Utilisant le cadre <strong>de</strong>s modèles à facteurs, nous étudions les co-mouvements extrêmes entre <strong>de</strong>ux actifs<br />
financiers ou entre un actif <strong>et</strong> le marché. A c<strong>et</strong>te fin, nous établissons l’expression générale du coefficient<br />
<strong>de</strong> dépendance <strong>de</strong> queue entre le marché <strong>et</strong> un actif (c’est-à-dire, la probabilité qu’un actif subisse une<br />
perte extrême, sachant que le marché a lui-même subi une perte extrême) <strong>et</strong> entre <strong>de</strong>ux actifs comme<br />
une fonction <strong>de</strong>s paramètres du modèle à facteur <strong>et</strong> <strong>de</strong>s paramètres <strong>de</strong> queue <strong>de</strong>s distributions du facteur<br />
<strong>et</strong> du bruit idiosynchr<strong>at</strong>ique. Notre formule est valable pour <strong>de</strong>s distributions marginales quelconques<br />
<strong>et</strong> ne requièrt aucune paramètris<strong>at</strong>ion <strong>de</strong> la distribution jointe du marché <strong>et</strong> <strong>de</strong>s actifs. La détermin<strong>at</strong>ion<br />
<strong>de</strong> ce paramètre extrême, qui n’est pas accessible par inférence <strong>st<strong>at</strong>istique</strong> directe, est rendue possible<br />
par la mesure <strong>de</strong> paramètres dont l’estim<strong>at</strong>ion implique une quantité signific<strong>at</strong>ive <strong>de</strong> données. Nos tests<br />
empiriques démontrent un bon accord entre le coefficient <strong>de</strong> dépendance <strong>de</strong> queue calibré <strong>et</strong> les gran<strong>de</strong>s<br />
pertes réalisées entre 1962 <strong>et</strong> 2000. Néanmoins, un biais systém<strong>at</strong>ique est détecté, suggérant l’existence<br />
d’un “ outlier” lors du krach d’octobre 1987 <strong>et</strong> pouvant inciter à penser que le modèle à un facteur<br />
(CAPM) que nous avons considéré ne suffit pas totalement à rendre compte <strong>de</strong>s propriétés extrêmes dans<br />
certaines phases critiques <strong>de</strong> marché.
300 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
How to account for extreme co-movements b<strong>et</strong>ween individual<br />
stocks and the mark<strong>et</strong> ∗<br />
Y. Malevergne 1,2 and D. Sorn<strong>et</strong>te 1,3<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x, France<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095, USA<br />
email: Yannick.Malevergne@unice.fr and sorn<strong>et</strong>te@unice.fr<br />
fax: (33) 4 92 07 67 54<br />
August 8, 2002<br />
Abstract<br />
Using the framework of factor mo<strong>de</strong>ls, we study the extreme co-movements b<strong>et</strong>ween two<br />
stocks and b<strong>et</strong>ween a stock and the mark<strong>et</strong>. In this goal, we establish the general expression of<br />
the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the mark<strong>et</strong> and a stock (th<strong>at</strong> is, the probability th<strong>at</strong><br />
the stock incurs a large loss, assuming th<strong>at</strong> the mark<strong>et</strong> has also un<strong>de</strong>rgone a large loss) and<br />
b<strong>et</strong>ween two stocks as a function of the param<strong>et</strong>ers of the un<strong>de</strong>rlying factor mo<strong>de</strong>l and of the<br />
tail param<strong>et</strong>ers of the distributions of the factor and of the idiosyncr<strong>at</strong>ic noise of each stock.<br />
Our formula holds for arbitrary marginal distributions and in addition does not require any<br />
param<strong>et</strong>eriz<strong>at</strong>ion of the multivari<strong>at</strong>e distributions of the mark<strong>et</strong> and stocks. The d<strong>et</strong>ermin<strong>at</strong>ion<br />
of the extreme param<strong>et</strong>er, which is not accessible by a direct st<strong>at</strong>istical inference, is ma<strong>de</strong><br />
possible by the measurement of param<strong>et</strong>ers whose estim<strong>at</strong>ion involves a significant part of the<br />
d<strong>at</strong>a with sufficient st<strong>at</strong>istics. Our empirical tests find a good agreement b<strong>et</strong>ween the calibr<strong>at</strong>ion<br />
of the tail <strong>de</strong>pen<strong>de</strong>nce coefficient and the realized large losses over the period from 1962 to 2000.<br />
Nevertheless, a bias is d<strong>et</strong>ected which suggests the presence of an outlier in the form of the crash<br />
of October 1987.<br />
∗ We acknowledge helpful discussions and exchanges with C.W.G. Granger, J.P. Laurent, V. Pisarenko, R. Valkanov<br />
and D. Zaj<strong>de</strong>nweber. This work was partially supported by the James S. Mc Donnell Found<strong>at</strong>ion 21st century scientist<br />
award/studying complex system.<br />
1
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 301<br />
Introduction<br />
The concept of extreme or “tail <strong>de</strong>pen<strong>de</strong>nce” probes the reaction of a variable to the realiz<strong>at</strong>ion<br />
of another variable when this realiz<strong>at</strong>ion is of extreme amplitu<strong>de</strong> and very low probability. The<br />
<strong>de</strong>pen<strong>de</strong>nce, and especially the extreme <strong>de</strong>pen<strong>de</strong>nce, b<strong>et</strong>ween two ass<strong>et</strong>s or b<strong>et</strong>ween an ass<strong>et</strong> and<br />
any other exogeneous economic variable is an issue of major importance both for practioners and<br />
for aca<strong>de</strong>mics. The d<strong>et</strong>ermin<strong>at</strong>ion of extreme <strong>de</strong>pen<strong>de</strong>nces is crucial for financial and for insurance<br />
institutions involved in risk management. It is also fundamental for the establishment of a r<strong>at</strong>ional<br />
investment policy striving for the best diversific<strong>at</strong>ion of the various sources of risk. In all these<br />
situ<strong>at</strong>ions, the objective is to prevent or <strong>at</strong> least minimize the simultaneous occurrence of large<br />
losses across the different positions held in the portfolio.<br />
From an aca<strong>de</strong>mic perspective, taking into account the extreme <strong>de</strong>pen<strong>de</strong>nce properties provi<strong>de</strong><br />
useful yardsticks and important constraints for the construction of mo<strong>de</strong>ls, which should not un<strong>de</strong>restim<strong>at</strong>e<br />
or overestim<strong>at</strong>e risks. From the point of view of univari<strong>at</strong>e st<strong>at</strong>istics, extreme values<br />
theory provi<strong>de</strong>s the m<strong>at</strong>hem<strong>at</strong>ical framework for the classific<strong>at</strong>ion and quantific<strong>at</strong>ion of very large<br />
risks. This has been ma<strong>de</strong> possible by the existence of a “universal” behavior summarized by<br />
the Gne<strong>de</strong>nko-Pickands-Balkema-<strong>de</strong> Haan theorem which gives a n<strong>at</strong>ural limit law for peak-overthreshold<br />
values in the form of the Generalized Par<strong>et</strong>o Distribution (see Embrechts, Kluppelberg,<br />
and Mikosh (1997, pp 152-168)). Moreover, most of these univari<strong>at</strong>e extreme values results are robust<br />
with respect to the time <strong>de</strong>pen<strong>de</strong>nces observed in financial time series (see <strong>de</strong> Haan, Resnick,<br />
Rootzen and <strong>de</strong> Vries (1989) or Starica (1999) for instance). In contrast, no such result is y<strong>et</strong> available<br />
in the multivari<strong>at</strong>e case. In such absence of theor<strong>et</strong>ical gui<strong>de</strong>lines, the altern<strong>at</strong>ive is therefore<br />
to impose some <strong>de</strong>pen<strong>de</strong>nce structure in a r<strong>at</strong>her ad hoc and arbitrary way. This was the stance<br />
taken for instance in Longin and Solnik (2001) in their study of the phenomenon of contagion across<br />
intern<strong>at</strong>ional equity mark<strong>et</strong>s.<br />
This approach, where the <strong>de</strong>pen<strong>de</strong>nce structure is not d<strong>et</strong>ermined from empirical facts or from<br />
an economic mo<strong>de</strong>l, is not fully s<strong>at</strong>isfying. As a remedy, we propose a new approach, which does<br />
not directly rely on multivari<strong>at</strong>e extreme values theory, but r<strong>at</strong>her <strong>de</strong>rives the extreme <strong>de</strong>pen<strong>de</strong>nce<br />
structure from the characteristics of a financial mo<strong>de</strong>l of ass<strong>et</strong>s. Specifically, we use the general<br />
class of factor mo<strong>de</strong>ls, which is probably one of the most vers<strong>at</strong>ile and relevant one, and whose<br />
introduction in finance can be traced back <strong>at</strong> least to Ross (1976). The factor mo<strong>de</strong>ls are now wi<strong>de</strong>ly<br />
used in many branches of finance, including stock r<strong>et</strong>urn mo<strong>de</strong>ls, interest r<strong>at</strong>e mo<strong>de</strong>ls (Vasicek<br />
(1977), Brennan and Schwarz (1978), Cox, Ingersoll, and Ross (1985)), credit risks mo<strong>de</strong>ls (Carey<br />
(1998), Gorby (2000), Lucas, Klaassen, Spreij and Stra<strong>et</strong>mans (2001)), and so on, and are found<br />
<strong>at</strong> the core of many theories and equilibrium mo<strong>de</strong>ls.<br />
Here, we shall first focus on the characteriz<strong>at</strong>ion of the extreme <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween stock r<strong>et</strong>urns<br />
and the mark<strong>et</strong> r<strong>et</strong>urn and then on the extreme <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two stocks which share a<br />
common explaining factor. The role of the mark<strong>et</strong> r<strong>et</strong>urn as a factor explaining the evolution of<br />
individual stock r<strong>et</strong>urns is supported both by theor<strong>et</strong>ical mo<strong>de</strong>ls such as the Capital Ass<strong>et</strong> Pricing<br />
Mo<strong>de</strong>l (CAPM) (Sharpe (1964), Lintner (1965), Mossin (1966)) or the Arbitrage Pricing Theory<br />
(APT) (Ross (1976)) and by empirical studies (Fama and B<strong>et</strong>h (1973), Kan<strong>de</strong>l and Staumbaugh<br />
(1987) among many others). It has even been shown in Roll (1988) th<strong>at</strong> in certain dram<strong>at</strong>ic<br />
circumstances, such as the October 1987 stock-mark<strong>et</strong> crash, the (global) mark<strong>et</strong> was the sole<br />
relevant factor nee<strong>de</strong>d to explain the stock mark<strong>et</strong> movements and the propag<strong>at</strong>ion of the crash<br />
across countries. Thus, the choice of factor mo<strong>de</strong>ls is a very n<strong>at</strong>ural starting point for studying<br />
extreme <strong>de</strong>pen<strong>de</strong>nces from a general point of view. The main gain is th<strong>at</strong>, without imposing any<br />
a priori ad hoc <strong>de</strong>pen<strong>de</strong>nces other than the <strong>de</strong>finition of the factor mo<strong>de</strong>l, we shall be able to<br />
2
302 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
<strong>de</strong>rive the general properties of extreme <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween an ass<strong>et</strong> and one of its factor and to<br />
empirically d<strong>et</strong>ermine these properties by a simple estim<strong>at</strong>ion of the factor mo<strong>de</strong>l param<strong>et</strong>ers.<br />
Our results are directly relevant to a portfolio manager using any of the factor mo<strong>de</strong>ls such as the<br />
CAPM or the APT to estim<strong>at</strong>e the impact on her extreme risks upon the addition or removal of an<br />
ass<strong>et</strong> in her portfolio. In this framework, our results st<strong>at</strong>ed for single ass<strong>et</strong>s can easily be exten<strong>de</strong>d<br />
to an entire portfolio, and some examples will be given. This problem is acute in particular in<br />
funds of funds. From a more global perspective, our analysis of the tail <strong>de</strong>pen<strong>de</strong>nce of two ass<strong>et</strong>s is<br />
the correct s<strong>et</strong>ting for analyzing the str<strong>at</strong>egic ass<strong>et</strong> alloc<strong>at</strong>ion facing a portofolio manager striving<br />
to diversify b<strong>et</strong>ween a portfolio of stocks and a portfolio of bonds or b<strong>et</strong>ween portfolios constituted<br />
of domestic and of intern<strong>at</strong>ional ass<strong>et</strong>s.<br />
Our main addition to the liter<strong>at</strong>ure is to provi<strong>de</strong> a compl<strong>et</strong>ely general analytical formula for the<br />
extreme <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween any two ass<strong>et</strong>s, which holds for any distribution of r<strong>et</strong>urns of these<br />
two ass<strong>et</strong>s and of their common factor and which thus embodies their intrinsic <strong>de</strong>pen<strong>de</strong>nce. Our<br />
second innov<strong>at</strong>ion is to provi<strong>de</strong> a novel and robust m<strong>et</strong>hod for estim<strong>at</strong>ing empirically the extreme<br />
<strong>de</strong>pen<strong>de</strong>nce which we test on twenty majors stocks of the NYSE. Comparing with historical comovements<br />
in the last forty years, we check th<strong>at</strong> our prediction is valid<strong>at</strong>ed out-of-sample and thus<br />
provi<strong>de</strong> an ex-ante m<strong>et</strong>hod to quantify futur stressful periods, so th<strong>at</strong> our results can be directly<br />
used to construct a portfolio aiming <strong>at</strong> minimizing the impact of extreme events. We are also able<br />
to d<strong>et</strong>ect an anomalous co-monoticity associ<strong>at</strong>ed with the October 1987 crash.<br />
The plan of our present<strong>at</strong>ion is as follows. The first section <strong>de</strong>fines the concepts nee<strong>de</strong>d for the<br />
characteriz<strong>at</strong>ion and quantific<strong>at</strong>ion of extreme <strong>de</strong>pen<strong>de</strong>nces. In particular, we recall the <strong>de</strong>finition<br />
of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce, which captures in a single number the properties of extreme<br />
<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random variables: the tail <strong>de</strong>pen<strong>de</strong>nce is <strong>de</strong>fined as the probability for<br />
a given random variable to be large assuming th<strong>at</strong> another random variable is large, <strong>at</strong> the same<br />
probability level. We shall also need some basic notions on <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween random variables<br />
using the m<strong>at</strong>hem<strong>at</strong>ical concept of copulas. In or<strong>de</strong>r to provi<strong>de</strong> some perspective on the following<br />
results, this section also contains the expression of some classical exemples of tail <strong>de</strong>pen<strong>de</strong>nce<br />
coefficients for specific multivari<strong>at</strong>e distributions.<br />
The second section st<strong>at</strong>es our main result in the form of a general theorem allowing the calcul<strong>at</strong>ion<br />
of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce for any factor mo<strong>de</strong>l with arbitrary distribution functions of<br />
the factors and of the idiosyncr<strong>at</strong>ic noise. We find th<strong>at</strong> the factor must have sufficiently “wild”<br />
fluctu<strong>at</strong>ions (to be ma<strong>de</strong> precise below) in or<strong>de</strong>r for the tail <strong>de</strong>pen<strong>de</strong>nce not to vanish. For normal<br />
distributions of the factor, the tail <strong>de</strong>pen<strong>de</strong>nce is i<strong>de</strong>ntically zero, while for regularly varying distributions<br />
(power laws), the tail <strong>de</strong>pen<strong>de</strong>nce is in general non-zero. We also show th<strong>at</strong> the most<br />
interesting coefficients of tail <strong>de</strong>pen<strong>de</strong>nce are those b<strong>et</strong>ween each individual stock and their common<br />
factor, since the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween any pair of ass<strong>et</strong>s is shown to be nothing but the<br />
minimum of the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween each ass<strong>et</strong> and their common factor.<br />
The third section is <strong>de</strong>voted to the empirical estim<strong>at</strong>ion of the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
individual stock r<strong>et</strong>urns and the mark<strong>et</strong> r<strong>et</strong>urn. The tests are performed for daily stock r<strong>et</strong>urns.<br />
The estim<strong>at</strong>ed coefficients of tail <strong>de</strong>pen<strong>de</strong>nce are found in good agreement with the fraction of<br />
historically realized extreme events th<strong>at</strong> occur simultaneously with any of the ten largest losses of<br />
the mark<strong>et</strong> factor (these ten largest losses were not used to calibr<strong>at</strong>e the tail <strong>de</strong>pen<strong>de</strong>nce coefficient).<br />
We also find some evi<strong>de</strong>nce for comonotonicity in the crash of Oct. 1987, suggesting th<strong>at</strong> this event<br />
is an “outlier,” providing additional support to a previous analysis of large and extreme drawdowns.<br />
We summarize our results and conclu<strong>de</strong> in the fourth section.<br />
3
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 303<br />
1 Intrinsic measure of casual and of extreme <strong>de</strong>pen<strong>de</strong>nces<br />
This section provi<strong>de</strong>s a brief informal summary of the m<strong>at</strong>hem<strong>at</strong>ical concepts used in this paper to<br />
characterize the normal and extreme <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween ass<strong>et</strong> r<strong>et</strong>urns.<br />
1.1 How to characterize uniquely the full <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random<br />
variables?<br />
The answer to this question is provi<strong>de</strong>d by the m<strong>at</strong>hem<strong>at</strong>ical notion of “copulas,” initially introduced<br />
by Sklar (1959) 1 , which allows one to study the <strong>de</strong>pen<strong>de</strong>nce of random variables in<strong>de</strong>pen<strong>de</strong>ntly<br />
of the behavior of their marginal distributions. Our present<strong>at</strong>ion focuses on two variables<br />
but is easily exten<strong>de</strong>d to the case of N random variables, wh<strong>at</strong>ever N may be. Sklar’s Theorem<br />
st<strong>at</strong>es th<strong>at</strong>, given the joint distribution function F (·, ·) of two random variables X and Y with<br />
marginal distribution FX(·) and FY (·) respectively, there exists a function C(·, ·) with range in<br />
[0, 1] × [0, 1] such th<strong>at</strong><br />
F (x, y) = C(FX(x), FY (y)) , (1)<br />
for all (x, y). This function C is the copula of the two random variables X and Y , and is unique if<br />
the random variables have continous marginal distributions. Moreover, the following result shows<br />
th<strong>at</strong> copulas are intrinsic measures of <strong>de</strong>pen<strong>de</strong>nce. If g1(X), g2(Y ) are strictly increasing on the<br />
ranges of X, Y , the random variables ˜ X = g1(X), ˜ Y = g2(Y ) have exactly the same copula C<br />
(see Lindskog (2000)). The copula is thus invariant un<strong>de</strong>r strictly increasing transform<strong>at</strong>ion of the<br />
variables. This provi<strong>de</strong>s a powerful way of studying scale-invariant measures of associ<strong>at</strong>ions. It is<br />
also a n<strong>at</strong>ural starting point for construction of multivari<strong>at</strong>e distributions.<br />
1.2 Tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random variables<br />
A standard measure of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random variables is provi<strong>de</strong>d by the correl<strong>at</strong>ion<br />
coefficient. However, it suffers from <strong>at</strong> least three <strong>de</strong>ficiencies. First, as stressed by Embrechts,<br />
McNeil, and Straumann (1999), the correl<strong>at</strong>ion coefficient is an a<strong>de</strong>qu<strong>at</strong>e measure of <strong>de</strong>pen<strong>de</strong>nce<br />
only for elliptical distributions and for events of mo<strong>de</strong>r<strong>at</strong>e sizes. Second, the correl<strong>at</strong>ion coefficient<br />
measures only the <strong>de</strong>gree of linear <strong>de</strong>pen<strong>de</strong>nce and does not account of any other nonlinear functional<br />
<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the random variables. Third, it agreg<strong>at</strong>es both the marginal behavior<br />
of each random variable and their <strong>de</strong>pen<strong>de</strong>nce. For instance, a simple change in the marginals<br />
implies in general a change in the correl<strong>at</strong>ion coefficient, while the copula and, therefore the <strong>de</strong>pen<strong>de</strong>nce,<br />
remains unchanged. M<strong>at</strong>hem<strong>at</strong>ically speaking, the correl<strong>at</strong>ion coefficient is said to lack<br />
the property of invariance un<strong>de</strong>r increasing changes of variables.<br />
Since the copula is the unique and intrinsic measure of <strong>de</strong>pen<strong>de</strong>nce, it is <strong>de</strong>sirable to <strong>de</strong>fine measures<br />
of <strong>de</strong>pen<strong>de</strong>nces which <strong>de</strong>pend only on the copula. Such measures have in fact been known for a long<br />
time. Examples are provi<strong>de</strong>d by the concordance measures, among which the most famous are the<br />
Kendall’s tau and the Spearman’s rho (see Nelsen (1998) for a d<strong>et</strong>ailed exposition). In particular,<br />
the Spearman’s rho quantifies the <strong>de</strong>gres of functional <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two random variables: it<br />
equals one (minus one) when and only when the first variable is an increasing (<strong>de</strong>creasing) function<br />
of the second variable. However, as for the correl<strong>at</strong>ion coefficient, these concordance measures do<br />
1 The rea<strong>de</strong>r is refered to Joe (1997), Frees and Val<strong>de</strong>z (1998) or Nelsen (1998) for a d<strong>et</strong>ailed survey of the notion<br />
of copulas and a m<strong>at</strong>hem<strong>at</strong>ically rigorous <strong>de</strong>scription of their properties.<br />
4
304 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
not provi<strong>de</strong> a useful measure of the <strong>de</strong>pen<strong>de</strong>nce for extreme events, since they are constructed over<br />
the whole distributions.<br />
Another n<strong>at</strong>ural i<strong>de</strong>a, wi<strong>de</strong>ly used in the contagion liter<strong>at</strong>ure, is to work with the conditional correl<strong>at</strong>ion<br />
coefficient, conditioned only on the largest events. But, as stressed by Boyer, Gibson,<br />
and Laur<strong>et</strong>an (1997), such conditional correl<strong>at</strong>ion coefficient suffers from a bias: even for a constant<br />
unconditional correl<strong>at</strong>ion coefficient, the conditional correl<strong>at</strong>ion coefficient changes with the<br />
conditioning s<strong>et</strong>. Therefore, changes in the conditional correl<strong>at</strong>ion do not provi<strong>de</strong> a characteristic<br />
sign<strong>at</strong>ure of a change in the true correl<strong>at</strong>ions. The conditional concordance measures suffer from<br />
the same problem.<br />
In view of these <strong>de</strong>ficiencies, it is n<strong>at</strong>ural to come back to a fundamental <strong>de</strong>finition of <strong>de</strong>pen<strong>de</strong>nce<br />
through the use of probabilities. We thus study the conditional probability th<strong>at</strong> the first variable is<br />
large conditioned on the second variable being large too: ¯ F (x|y) = Pr{X > x|Y > y}, when x and<br />
y goes to infinity. Since the convergence of ¯ F (x|y) may <strong>de</strong>pend on the manner with which x and y<br />
go to infinity (the convergence is not uniform), we need to specify the p<strong>at</strong>h taken by the variables<br />
to reach the infinity. Recalling th<strong>at</strong> it would be preferable to have a measure which is in<strong>de</strong>pen<strong>de</strong>nt<br />
of the marginal distributions of X and Y , it is n<strong>at</strong>ural to reason in the quantile space. This leads<br />
to choose x = FX −1 (u) and y = FY −1 (u) and replace the conditions x, y → ∞ by u → 1. Doing so,<br />
we <strong>de</strong>fine the so-called coefficient of upper tail <strong>de</strong>pen<strong>de</strong>nce (see Coles, Heffernan, and Tawn (1999),<br />
Lindskog (2000), or Embrechts, McNeil, and Straumann (2001)):<br />
λ+ = lim<br />
u→1 − Pr{X > FX −1 (u) | Y > FY −1 (u)} . (2)<br />
As required, this measure of <strong>de</strong>pen<strong>de</strong>nce is in<strong>de</strong>pen<strong>de</strong>nt of the marginals, since it can be expressed<br />
in term of the copula of X and Y as<br />
λ+ = lim<br />
u→1− 1 − 2u + C(u, u)<br />
1 − u<br />
. (3)<br />
This represent<strong>at</strong>ion shows th<strong>at</strong> λ+ is symm<strong>et</strong>ric in X and Y , as it should for a reasonable measure<br />
of <strong>de</strong>pen<strong>de</strong>nce.<br />
In a similar way, we <strong>de</strong>fine the coefficient of lower tail <strong>de</strong>pen<strong>de</strong>nce as the probability th<strong>at</strong> X incurs<br />
a large loss assuming th<strong>at</strong> Y incurs a large loss <strong>at</strong> the same probability level<br />
λ− = lim<br />
u→0 + Pr{X < FX −1 (u) | Y < FY −1 (u)} = lim<br />
u→0 +<br />
C(u, u)<br />
u<br />
. (4)<br />
This last expression has a simple interpr<strong>et</strong><strong>at</strong>ion in term of Value-<strong>at</strong>-Risk. In<strong>de</strong>ed, the quantiles<br />
F −1<br />
−1<br />
X (u) and FY (u) are nothing but the Values-<strong>at</strong>-Risk of ass<strong>et</strong>s (or portfolios) X and Y <strong>at</strong> the<br />
confi<strong>de</strong>nce level 1 − u. Thus, the coefficient λ− simply provi<strong>de</strong>s the probability th<strong>at</strong> X exceeds the<br />
VaR <strong>at</strong> level 1 − u, assuming th<strong>at</strong> Y has excee<strong>de</strong>d the VaR <strong>at</strong> the same level confi<strong>de</strong>nce level 1 − u,<br />
when this level goes to one. As a consequence, the probability th<strong>at</strong> both X and Y exceed their VaR<br />
<strong>at</strong> the level 1 − u is asymptotically given by λ− · (1 − u) as u → 0. As an example, consi<strong>de</strong>r a daily<br />
VaR calcul<strong>at</strong>ed <strong>at</strong> the 99% confi<strong>de</strong>nce level. Then, the probability th<strong>at</strong> both X and Y un<strong>de</strong>rgo a<br />
loss larger than their VaR <strong>at</strong> the 99% level is approxim<strong>at</strong>ely given by λ−/100. Thus, when λ− is<br />
about 0.1, the typical recurrence time b<strong>et</strong>ween such concomitant large losses is about four years,<br />
while for λ− ≈ 0.5 it is less than ten months.<br />
The values of the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce are known explicitely for a large number of different<br />
copulas. For instance, the Gaussian copula, which is the copula <strong>de</strong>rived from <strong>de</strong> Gaussian multivari<strong>at</strong>e<br />
distribution, has a zero coefficient of tail <strong>de</strong>pen<strong>de</strong>nce. In contrast, the Gumbel’s copula used<br />
5
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 305<br />
by Longin and Solnik (2001) in the study of the contagion b<strong>et</strong>ween intern<strong>at</strong>ional equity mark<strong>et</strong>s,<br />
which is <strong>de</strong>fined by<br />
<br />
Cθ(u, v) = exp − (− ln u) θ 1 <br />
+ (− ln v)<br />
θ θ<br />
, θ ∈ [0, 1], (5)<br />
has an upper tail coefficient λ+ = 2 − 2 θ . For all θ’s smaller than one, λ+ is positive and the<br />
Gumbel’s copula is said to present tail <strong>de</strong>pen<strong>de</strong>nce, while for θ = 1, the Gumbel copula is said to<br />
be asymptotically in<strong>de</strong>pen<strong>de</strong>nt. One should however use this terminology with a grain of salt as<br />
“tail in<strong>de</strong>pen<strong>de</strong>nce” (quantified by λ+ = 0 or λ− = 0) does not imply necessarily th<strong>at</strong> large events<br />
occur in<strong>de</strong>pen<strong>de</strong>ntly (see Coles, Heffernan, and Tawn (1999) for a precise discussion of this point).<br />
2 Tail <strong>de</strong>pen<strong>de</strong>nce of factor mo<strong>de</strong>ls<br />
2.1 Tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween an ass<strong>et</strong> and one of its explaining factors<br />
Now we st<strong>at</strong>e the first part of our main theor<strong>et</strong>ical result. L<strong>et</strong> us consi<strong>de</strong>r two random variables<br />
X and Y of cumul<strong>at</strong>ive distribution functions FX(X) and FY (Y ), where X represents the r<strong>et</strong>urn<br />
of a single stock and Y is the mark<strong>et</strong> r<strong>et</strong>urn for instance. L<strong>et</strong> us also introduce an idiosyncr<strong>at</strong>ic<br />
noise ε, which is assumed in<strong>de</strong>pen<strong>de</strong>nt of the mark<strong>et</strong> r<strong>et</strong>urn Y . The factor mo<strong>de</strong>l is <strong>de</strong>fined by<br />
the following rel<strong>at</strong>ionship b<strong>et</strong>ween the individual stock r<strong>et</strong>urn X, the mark<strong>et</strong> r<strong>et</strong>urn Y and the<br />
idiosyncr<strong>at</strong>ic noise ε:<br />
X = β · Y + ε . (6)<br />
β is the usual coefficient introduced by the Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>l Sharpe (1964). L<strong>et</strong> us stress<br />
th<strong>at</strong> ε may embody other factors Y ′ , Y ′′ , . . ., as long as they remain in<strong>de</strong>pen<strong>de</strong>nt of Y . Un<strong>de</strong>r such<br />
conditions and a few other technical assumptions d<strong>et</strong>ailed in the theorem established in appendix<br />
A.1, the coefficient of (upper) tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween X and Y <strong>de</strong>fined in (2) is obtained as<br />
∞<br />
λ+ =<br />
max1, l β dx f(x) , (7)<br />
where l <strong>de</strong>notes the limit, when u → 1, of the r<strong>at</strong>io of the quantiles of X and Y ,<br />
l = lim<br />
u→1<br />
FX −1 (u)<br />
FY −1 (u)<br />
and f(x) is the limit, when t → +∞, of t · PY (tx)/ ¯ FY (t):<br />
f(x) = lim<br />
t→+∞ t PY (tx)<br />
¯FY (t)<br />
, (8)<br />
. (9)<br />
PY is the distribution <strong>de</strong>nsity of Y and ¯ FY = 1 − FY is the complementary cumul<strong>at</strong>ive distribution<br />
function of Y . A similar expression obviously holds, mut<strong>at</strong>is mutandis, for the coefficient of lower<br />
tail <strong>de</strong>pen<strong>de</strong>nce.<br />
The measure of tail <strong>de</strong>pen<strong>de</strong>nce given by equ<strong>at</strong>ion (7) <strong>de</strong>pends on two limits <strong>de</strong>fined in (8) and<br />
(9) and thus seems likely difficult to estim<strong>at</strong>e. As it turns out, we will show th<strong>at</strong> this is not the<br />
case in the empirical section below. In<strong>de</strong>ed, the first limit (8) is nothing but a r<strong>at</strong>io of quantiles,<br />
while the second limit (9) can be easily calcul<strong>at</strong>ed for almost all distributions of the factor. For<br />
6
306 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
instance, l<strong>et</strong> us consi<strong>de</strong>r the Par<strong>et</strong>o distribution ¯ FY (y) = 1/(y/y0) µ <strong>de</strong>fined for y ≥ y0, whose<br />
<strong>de</strong>nsity is equal to PY (y) = (µ/y0)/(y/y0) 1+µ ; the limit (9) gives f(x) = µ/x 1+µ . In contrast,<br />
for the Poisson law ¯ FY (y) = e −ry <strong>de</strong>fined for y ≥ 0 with <strong>de</strong>nsity PY (y) = re −ry , the limit (9)<br />
gives f(x) = limt→∞ r t e −rt(x−1) = 0 for x > 1. Thus an estim<strong>at</strong>ion of the tail of the factor<br />
distribution is sufficient to infer the limit function f(x). Moreover, equ<strong>at</strong>ion (7) has a r<strong>at</strong>her simple<br />
interpr<strong>et</strong><strong>at</strong>ion since it shows th<strong>at</strong> a non-vanishing coefficient of tail <strong>de</strong>pen<strong>de</strong>nce results from the<br />
combin<strong>at</strong>ion of two phenomena. First, the limit function f(x), which only <strong>de</strong>pends on the behavior<br />
of the factor distribution, must be non-zero. Second, the constant l must remain finite to ensure<br />
th<strong>at</strong> the integral in (7) does not vanish. Thus, the value of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is<br />
controlled by f(x) solely function of the factor and a second variable l quantifying the comp<strong>et</strong>ition<br />
of the tails of the distribution of the factor Y and of the idiosyncr<strong>at</strong>ic noise ε.<br />
The fundamental result (7) should be of vivid interest to financial economists because it provi<strong>de</strong>s a<br />
general, rigorous and simple m<strong>et</strong>hod for estim<strong>at</strong>ing one of the key variable embodying the occurrence<br />
of and the risks associ<strong>at</strong>ed with extremes in joint distributions. From a theor<strong>et</strong>ical view point, it<br />
also anchors the <strong>de</strong>riv<strong>at</strong>ion and quantific<strong>at</strong>ion of a key variable on extremes in the general class of<br />
financial factor mo<strong>de</strong>ls, thus extending their use and relevance also to this r<strong>at</strong>her novel domain of<br />
extreme <strong>de</strong>pen<strong>de</strong>nce, extreme risks and extreme losses.<br />
Up to now, we have assumed th<strong>at</strong> the factor Y and the idiosyncr<strong>at</strong>ic noise ε were in<strong>de</strong>pen<strong>de</strong>nt. In<br />
fact, it is important to stress for the sake of generality th<strong>at</strong> the result (7) holds even when they are<br />
<strong>de</strong>pen<strong>de</strong>nt, provi<strong>de</strong>d th<strong>at</strong> this <strong>de</strong>pen<strong>de</strong>nce is not too strong, as explained and ma<strong>de</strong> specific <strong>at</strong> the<br />
end of Appendix A.1.<br />
We now <strong>de</strong>rive two direct consequences of this result (7) (see corollary 1 and 2 in appendix B),<br />
concerning rapidly varying and regularly varying factors 2 , which clearly illustr<strong>at</strong>e the role the<br />
factor itself and the impact of the tra<strong>de</strong> off b<strong>et</strong>ween the factor and the idiosyncr<strong>at</strong>ic noise.<br />
2.2 Absence of tail <strong>de</strong>pen<strong>de</strong>nce for rapidly varying factors<br />
L<strong>et</strong> us assume th<strong>at</strong> the factor Y and the idiosyncr<strong>at</strong>ic noise ε are normally distributed (the second<br />
assumption is ma<strong>de</strong> for simplicity and will be relaxed below). As a consequence, the joint distribution<br />
of (X, Y ) is the bivari<strong>at</strong>e Gaussian distribution. Refering to the results st<strong>at</strong>ed in section 1.1.2,<br />
we conclu<strong>de</strong> th<strong>at</strong> the copula of (X, Y ) is the Gaussian copula whose coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
is zero. In fact, it is easy to show th<strong>at</strong> λ = 0 for any non-<strong>de</strong>gener<strong>at</strong>ed distribution of ε.<br />
More generally, l<strong>et</strong> us assume th<strong>at</strong> the distribution of the factor Y is rapidly varying, which <strong>de</strong>scribes<br />
the Gaussian, exponential and any distribution <strong>de</strong>caying faster than any power-law. Then, the<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is i<strong>de</strong>ntically zero. This result holds for any arbitrary distribution of<br />
the idiosyncr<strong>at</strong>ic noise (see corollary 1 in appendix B). It also holds for mixtures of normals or other<br />
distributions f<strong>at</strong>ter than Gaussians, some of which are thought to be reasonable approxim<strong>at</strong>ions to<br />
empirical stock r<strong>et</strong>urn distributions.<br />
These st<strong>at</strong>ements are somewh<strong>at</strong> counter-intuitive since one could expect a priori th<strong>at</strong> the coefficient<br />
of tail <strong>de</strong>pen<strong>de</strong>nce does not vanish as soon as the tail of the distribution of factor r<strong>et</strong>urns is f<strong>at</strong>ter<br />
than the tail the distribution noise r<strong>et</strong>urns. Said differently, when the standard <strong>de</strong>vi<strong>at</strong>ion of the<br />
idiosynchr<strong>at</strong>ic noise ε is small (but not zero), then the idiosynchr<strong>at</strong>ic noise component is small and<br />
X and Y are practically i<strong>de</strong>ntical, and it seems strange th<strong>at</strong> their tail <strong>de</strong>pen<strong>de</strong>nce can be equal<br />
2 see Bigham, Goldie, and Teugel (1987) or Embrechts, Kluppelberg, and Mikosh (1997) for a survey of the<br />
properties of rapidly and regularly varying functions<br />
7
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 307<br />
to zero. This non-intuitive result stems from the fact th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce is quantifying not<br />
just a <strong>de</strong>pen<strong>de</strong>nce but a specific <strong>de</strong>pen<strong>de</strong>nce for extreme co-movements. Thus, in or<strong>de</strong>r to g<strong>et</strong><br />
a non-vanishing tail-<strong>de</strong>pen<strong>de</strong>nce, the fluctu<strong>at</strong>ions of the factor must be ‘wild’ enough, which is<br />
not realized with rapidly varying distributions, irrespective of the rel<strong>at</strong>ive values of the standard<br />
<strong>de</strong>vi<strong>at</strong>ions of the factor and the idosyncr<strong>at</strong>ic noise.<br />
2.3 Coefficient of tail <strong>de</strong>pen<strong>de</strong>nce for regularly varying factors<br />
2.3.1 Example of the factor mo<strong>de</strong>l with Stu<strong>de</strong>nt distribution<br />
In or<strong>de</strong>r to account for the power-law tail behavior observed for the distributions of ass<strong>et</strong>s r<strong>et</strong>urns<br />
it is logical to consi<strong>de</strong>r th<strong>at</strong> the factor and the indiosyncr<strong>at</strong>ic noise also have power-law tailed<br />
distributions. As an illustr<strong>at</strong>ion, we will assume th<strong>at</strong> Y and ε are distributed according to a<br />
Stu<strong>de</strong>nt’s distribution with the same number of <strong>de</strong>grees of freedom ν (and thus same tail exponent<br />
ν). L<strong>et</strong> us <strong>de</strong>note by σ the scale factor of the distribution of ε while the scale factor of the<br />
distribution of Y is chosen equal to one 3 . Applying the theorem previously established, we find<br />
th<strong>at</strong> f(x) = ν/x ν+1 and l = β<br />
<br />
1 +<br />
σ<br />
β<br />
1<br />
λ± =<br />
1 +<br />
ν 1/ν<br />
, so th<strong>at</strong> the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is<br />
σ<br />
β<br />
ν , and β > 0. (10)<br />
As expected, the tail <strong>de</strong>pen<strong>de</strong>nce increases as β increases and as σ <strong>de</strong>creases. Since the idiosyncr<strong>at</strong>ic<br />
vol<strong>at</strong>ility of the ass<strong>et</strong> increases when the scale factor σ increases, this results simply means th<strong>at</strong><br />
the tail <strong>de</strong>pen<strong>de</strong>nce <strong>de</strong>creases when the idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility of a stock increases rel<strong>at</strong>ive to the<br />
mark<strong>et</strong> vol<strong>at</strong>ility. The <strong>de</strong>pen<strong>de</strong>nce with respect to ν is less intuitive. In particular, l<strong>et</strong> ν go to<br />
infinity. Then, λ → 0 if σ > β and λ → 1 for σ < β. This is surprising as one could argue th<strong>at</strong>, as<br />
ν → ∞, the Stu<strong>de</strong>nt distribution tends to the Gaussian law. As a consequence, one would expect<br />
the same coefficient of <strong>de</strong>pen<strong>de</strong>nce λ± = 0 as for rapidly varying functions. The reason for the<br />
non-certain convergence of λ± to zero as ν → ∞ is rooted in a subtle non-commut<strong>at</strong>ivity (and<br />
non-uniform convergence) of the two limits ν → ∞ and u → 1. In<strong>de</strong>ed, when taking first the limit<br />
u → 1, the result λ → 1 for β > σ indic<strong>at</strong>es th<strong>at</strong> a sufficiently strong factor coefficient β always<br />
ensures the validity of the power law regime, wh<strong>at</strong>ever the value of ν. Correl<strong>at</strong>ively, in this regime<br />
β > σ, λ± is an increasing function of ν.<br />
The result (10) is of interest for financial economics purpose because it provi<strong>de</strong>s a simple param<strong>et</strong>ric<br />
illustr<strong>at</strong>ion and interpr<strong>et</strong><strong>at</strong>ion of how the risk of large co-movements is affected by the three key<br />
param<strong>et</strong>ers entering in the <strong>de</strong>finition of the factor mo<strong>de</strong>l. It allows one to weight how the ingredients<br />
of the factor mo<strong>de</strong>l impact on the large risks captured by λ± and thus links the financial basis<br />
un<strong>de</strong>rlying the factor mo<strong>de</strong>l to the extreme multivari<strong>at</strong>e risks.<br />
2.3.2 General result<br />
We now provi<strong>de</strong> the general result valid for any regularly varying distribution. L<strong>et</strong> the factor<br />
Y follows a regularly varying distribution with tail in<strong>de</strong>x α: in other words, the complementary<br />
cumul<strong>at</strong>ive distribution of Y is such th<strong>at</strong> ¯ FY (y) = L(y) · y −α , where L(y) is a slowly varying<br />
3 Such a choice is always possible via a rescaling of the coefficient β.<br />
8
308 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
function, i.e:<br />
Corollary 2 in appendix B.2 shows th<strong>at</strong><br />
L(ty)<br />
lim = 1, ∀y > 0. (11)<br />
t→∞ L(t)<br />
λ =<br />
1<br />
<br />
max<br />
1, l<br />
β<br />
α , (12)<br />
where l <strong>de</strong>notes the limit, when u → 1, of the r<strong>at</strong>io FX −1 (u)/FY −1 (u). In the case of particular<br />
interest when the distribution of ε is also regularly varying with tail in<strong>de</strong>x α and if, in addition, we<br />
have ¯ FY (y) ∼ Cy ·y −α and ¯ Fε(ε) ∼ Cε ·ε −α , for large y and ε, then the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
is a simple function of the r<strong>at</strong>io Cε/Cy of the scale factors:<br />
λ =<br />
1<br />
1 + β −α · Cε<br />
Cy<br />
. (13)<br />
When the tail in<strong>de</strong>xes αY and αε of the distribution of the factor and the residue are different,<br />
then λ = 0 for αY < αε and λ = 1 for αY > αε.<br />
The results (12) and (13) are very important both for a financial and and economic perseptive<br />
because they express in the most general and straightforward way the risk of extreme co-movements<br />
quantified by the tail <strong>de</strong>pen<strong>de</strong>nce param<strong>et</strong>er λ within the important class of factor mo<strong>de</strong>ls. Th<strong>at</strong> λ<br />
increases with the β factor is intuitively clear. Less obvious is the <strong>de</strong>pen<strong>de</strong>nce of λ on the structure<br />
of the marginal distributions of the factor and of the idiosynchr<strong>at</strong>ic noise, which is found to be<br />
captured uniquely in terms of the r<strong>at</strong>io of their scale factors Cε and Cy. The scale factors Cε and<br />
Cy tog<strong>et</strong>her with the factor β thus replace the variance and covariance in their role as the sole<br />
quantifiers of the extreme risks occurring in co-movements.<br />
Until now, we have only consi<strong>de</strong>red a single ass<strong>et</strong> X. L<strong>et</strong> us now consi<strong>de</strong>r a portfolio of ass<strong>et</strong>s Xi,<br />
each of the ass<strong>et</strong>s following exactly the one factor mo<strong>de</strong>l (6)<br />
Xi = βi · Y + εi, (14)<br />
with in<strong>de</strong>pen<strong>de</strong>nt noises εi, whose scale factors are Cεi . The portfolio X = wiXi, with weights<br />
wi, also follows the factor mo<strong>de</strong>l with a param<strong>et</strong>er β = wiβi and noise ε, whose scale factor is<br />
Cε = |wi| α · 4<br />
Cεi . Thus, equ<strong>at</strong>ion (13) shows th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the portfolio and<br />
the factor is<br />
<br />
|wi|<br />
λ = 1 +<br />
α · Cεi<br />
( wiβi) α −1 . (15)<br />
· CY<br />
When unlimited short sells are allowed, one can follow a “mark<strong>et</strong> neutral” str<strong>at</strong>egy yielding β = 0<br />
and thus λ = 0. But in the more realistic case where only limited short sells are authorized, one<br />
cannot reach β = 0, and the best portfolio, which is the less “correl<strong>at</strong>ed” with the large mark<strong>et</strong><br />
moves, has to minimze the tail <strong>de</strong>pen<strong>de</strong>nce (15).<br />
This simple example clearly shows th<strong>at</strong> it is very different to minimize the extreme co-movements,<br />
according to (15), and to minimize the (linear) correl<strong>at</strong>ion ρ b<strong>et</strong>ween the portfolio and the mark<strong>et</strong><br />
factor given by<br />
ρ =<br />
<br />
1 +<br />
w 2 i · V ar(εi)<br />
( wiβi) 2 V ar(Y )<br />
−1/2<br />
. (16)<br />
4 In the more realistic case where the εi’s are not in<strong>de</strong>pen<strong>de</strong>nt but still embody one or several common factors<br />
Y ′ , Y ′′ , · · ·, the resulting scale factor Cε can be calcul<strong>at</strong>ed with the m<strong>et</strong>hod <strong>de</strong>scribed in Bouchaud, Sorn<strong>et</strong>te, Walter<br />
and Aguilar (1998)<br />
9
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 309<br />
Thus, since the minimum of ρ may be very different from the minimum of λ, minimizing ρ almost<br />
surely leads to accept a level of extreme risks which is not optimal.<br />
2.4 Tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two ass<strong>et</strong>s rel<strong>at</strong>ed by a factor mo<strong>de</strong>l<br />
We now present the second part of our theor<strong>et</strong>ical result. L<strong>et</strong> X1 and X2 be two random variables<br />
(two ass<strong>et</strong>s) of cumul<strong>at</strong>ive distributions functions F1, F2 with a common factor Y . L<strong>et</strong> ε1 and ε2<br />
be the idiosyncr<strong>at</strong>ic noises associ<strong>at</strong>ed with these two ass<strong>et</strong>s X1 and X2. We allow the idiosyncr<strong>at</strong>ic<br />
noises to be <strong>de</strong>pen<strong>de</strong>nt random variables, as occurs for instance if they embody the effect of other<br />
factors Y ′ , Y ′′ , ... which are in<strong>de</strong>pen<strong>de</strong>nt of Y . Our essential assumption is th<strong>at</strong> the distribution of<br />
the factor Y must have a tail not thinner than the tail of the distributions of the other factors Y ′ ,<br />
Y ′′ , ... This hypothesis is crucial in or<strong>de</strong>r to d<strong>et</strong>ect the existence of tail-<strong>de</strong>pen<strong>de</strong>nce. This means<br />
th<strong>at</strong>, for purposes of characterizing tail <strong>de</strong>pen<strong>de</strong>ncies in factor mo<strong>de</strong>ls, our mo<strong>de</strong>l can always be<br />
re-st<strong>at</strong>ed as a single factor mo<strong>de</strong>l where the single factor is the factor with the thickest tail. This<br />
makes our results quite general. Then, the mo<strong>de</strong>l can be written as<br />
X1 = β1 · Y + ε1 , (17)<br />
X2 = β2 · Y + ε2 . (18)<br />
We prove in appendix A.2, th<strong>at</strong> the coefficient of (upper) tail <strong>de</strong>pen<strong>de</strong>nce λ+ = limu→1 Pr{X1 ><br />
F1 −1 (u) | X2 > F2 −1 (u)}, b<strong>et</strong>ween the ass<strong>et</strong>s X1 and X2, is given by the expression<br />
λ+ =<br />
∞<br />
maxl 1<br />
β1 , l 2<br />
β2 dx f(x) , (19)<br />
which is very similar to th<strong>at</strong> found for the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween an ass<strong>et</strong> and one of its explaining<br />
factor (see equ<strong>at</strong>ion (7)). As previously, l1,2 <strong>de</strong>notes the limit, when u → 1, of the r<strong>at</strong>io<br />
F1,2 −1 (u)/FY −1 (u), and f(x) is the limit, when t → +∞, of t · PY (tx)/ ¯ FY (t).<br />
The result (19) can be cast in a different illumin<strong>at</strong>ing way. L<strong>et</strong> λ(X1, Y ) (resp. λ(X2, Y )) <strong>de</strong>note<br />
the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong> X1 (resp. X2) and their common factor Y . L<strong>et</strong><br />
λ(X1, X2) <strong>de</strong>note the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two ass<strong>et</strong>s. Equ<strong>at</strong>ion (19) allows us to assert<br />
th<strong>at</strong><br />
λ(X1, X2) = min{λ(X1, Y ), λ(X2, Y )}. (20)<br />
The tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two ass<strong>et</strong>s X1 and X2 is nothing but the smallest tail <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween each ass<strong>et</strong> and the common factor. Therefore, a <strong>de</strong>crease of the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
the ass<strong>et</strong>s and the mark<strong>et</strong> will also lead autom<strong>at</strong>ically to a <strong>de</strong>crease of the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
the two ass<strong>et</strong>s. This result also shows th<strong>at</strong> it is sufficient to study the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the<br />
ass<strong>et</strong>s and their common factor to obtain the tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween any pair of ass<strong>et</strong>s.<br />
The result (20) is also useful in the context of portfolio analysis. Not only does it provi<strong>de</strong> a tool<br />
for assessing the probability of large losses of a portfolio composed of ass<strong>et</strong>s driven by a common<br />
factor, it also allows us to <strong>de</strong>fine novel str<strong>at</strong>egies of portfolio optimiz<strong>at</strong>ion based on the selection<br />
and weighting of stocks chosen so as to balance to risks associ<strong>at</strong>ed with extreme co-movements.<br />
Such an approach has been tested in (Malevergne and Sorn<strong>et</strong>te 2002) with encouraging results.<br />
3 Empirical study<br />
We now apply our theor<strong>et</strong>ical results to the daily r<strong>et</strong>urns of a s<strong>et</strong> of stocks tra<strong>de</strong>d on the New York<br />
Stock Exchange. In or<strong>de</strong>r to estim<strong>at</strong>e the param<strong>et</strong>ers of the factor mo<strong>de</strong>l (6), the Standard and<br />
10
310 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Poor’s 500 in<strong>de</strong>x is chosen to represent the common “mark<strong>et</strong> factor.” It has been prefered over the<br />
Dow Jones Industrial Averages In<strong>de</strong>x for instance, because, although less diversified, it represents<br />
about 80% of the total mark<strong>et</strong> capitaliz<strong>at</strong>ion.<br />
We <strong>de</strong>scribe the s<strong>et</strong> of selected stocks in the next sub-section. Next, we estim<strong>at</strong>e the param<strong>et</strong>er<br />
β in (6) and check the in<strong>de</strong>pen<strong>de</strong>nce of the mark<strong>et</strong> r<strong>et</strong>urns and the residues. Then, applying the<br />
commonly used hypothesis according to which the tail of the distribution of ass<strong>et</strong>s r<strong>et</strong>urn is a power<br />
law or <strong>at</strong> leastregularly varying (see Longin (1996), Lux (1996), Pagan (1996), or Gopikrishnan,<br />
Meyer, Amaral, and Stanley (1998)), we estim<strong>at</strong>e the tail in<strong>de</strong>x and the scale factor of these<br />
distributions, which allows us to calcul<strong>at</strong>e the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween each ass<strong>et</strong><br />
r<strong>et</strong>urn and the mark<strong>et</strong> r<strong>et</strong>urn. Finally, we perform an analysis of the historical d<strong>at</strong>a to check<br />
the comp<strong>at</strong>ibility of our prediction on the fraction of realized large losses of the ass<strong>et</strong>s th<strong>at</strong> occur<br />
simultaneously with the large losses of the mark<strong>et</strong>.<br />
The results of our analysis are reported below in terms of the r<strong>et</strong>urns r<strong>at</strong>her than in terms of the<br />
excess r<strong>et</strong>urns above the risk free interest r<strong>at</strong>e, in apparent contradiction with the prescription of the<br />
CAPM. However, for daily r<strong>et</strong>urns, the difference b<strong>et</strong>ween r<strong>et</strong>urns and excess r<strong>et</strong>urns is negligible.<br />
In<strong>de</strong>ed, we checked th<strong>at</strong> neglecting the difference b<strong>et</strong>ween the r<strong>et</strong>urns and the excess r<strong>et</strong>urns does<br />
not affect our results by re-running all the study <strong>de</strong>scribed below in terms of the excess r<strong>et</strong>urns and<br />
found th<strong>at</strong> the tail <strong>de</strong>pen<strong>de</strong>nce did not change by more than 0.1%.<br />
3.1 Description of the d<strong>at</strong>a<br />
We study a s<strong>et</strong> of twenty ass<strong>et</strong>s tra<strong>de</strong>d on the New York Stock Exchange. The criteria presiding<br />
over the selection of the ass<strong>et</strong>s (see column 1 of table 1) are th<strong>at</strong> (1) they are among the stocks<br />
with the largest capitaliz<strong>at</strong>ions, but (2) each of them should have a weight smaller than 1% in the<br />
Standard and Poor’s 500 in<strong>de</strong>x, so th<strong>at</strong> the <strong>de</strong>pen<strong>de</strong>nce studied here does not stem trivially from<br />
their overlap with the mark<strong>et</strong> factor (taken as the Standard and Poor’s 500 in<strong>de</strong>x).<br />
The time interval we have consi<strong>de</strong>red ranges from July 03, 1962 to December 29, 2000, corresponding<br />
to 9694 d<strong>at</strong>a points, and represents the largest s<strong>et</strong> of daily d<strong>at</strong>a available from the Center for<br />
Research in Security Prices (CRSP). This large time interval is important to l<strong>et</strong> us collect as many<br />
large fluctu<strong>at</strong>ions of the r<strong>et</strong>urns as is possible in or<strong>de</strong>r to sample the extreme tail <strong>de</strong>pen<strong>de</strong>nce.<br />
Moreover, in or<strong>de</strong>r to allow for a non-st<strong>at</strong>ionarity over the four <strong>de</strong>ca<strong>de</strong>s of the study, to check the<br />
stability of our results and to test the st<strong>at</strong>ionnarity of the tail <strong>de</strong>pen<strong>de</strong>nce over the time, we split<br />
this s<strong>et</strong> into two subs<strong>et</strong>s. The first one ranges from July 1962 to December 1979, a period with<br />
few very large r<strong>et</strong>urn amplitu<strong>de</strong>s, while the second one ranges from January 1980 to December<br />
2000, a period which witnessed several very large price changes (see table 1 which shows the good<br />
stability of the standard <strong>de</strong>vi<strong>at</strong>ion b<strong>et</strong>ween the two sub-periods while the higher cumulants such<br />
as the excess kurtosis often increased dram<strong>at</strong>ically in the second sub-period for most ass<strong>et</strong>s). The<br />
table 1 presents the main st<strong>at</strong>istical properties of our s<strong>et</strong> of stocks during the three time intervals.<br />
All ass<strong>et</strong>s exhibit an excess kurtosis significantly different from zero over the three time interval,<br />
which is inconsistent with the assumption of Gaussianly distributed r<strong>et</strong>urns. While the standard<br />
<strong>de</strong>vi<strong>at</strong>ions remain stable over time, the excess kurtosis increases significantly from the first to the<br />
second period. This is in resonance with the financial community’s belief th<strong>at</strong> stock price vol<strong>at</strong>ility<br />
has increased over time, a still controversial result (Jones and J.W.Wilson (1989), Campbell, L<strong>et</strong>tau,<br />
Malkiel, and Xu (2001) or Xu and Malkiel (2002)).<br />
11
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 311<br />
3.2 Calibr<strong>at</strong>ion of the factor mo<strong>de</strong>l<br />
The d<strong>et</strong>ermin<strong>at</strong>ion of the param<strong>et</strong>ers β and of the residues ε entering in the <strong>de</strong>finition of the factor<br />
mo<strong>de</strong>l (6) is performed for each ass<strong>et</strong> by regressing the stocks r<strong>et</strong>urns on the mark<strong>et</strong> r<strong>et</strong>urn. The<br />
coefficient β is thus given by the ordinary least square estim<strong>at</strong>or, which is consistent as long as<br />
the residues are weak white noise and with zero mean and finite variance. The idiosyncr<strong>at</strong>ic noise<br />
ε is obtained by substracting β times the mark<strong>et</strong> r<strong>et</strong>urn to the stock r<strong>et</strong>urn. Table 2 presents<br />
the results for the three periods we consi<strong>de</strong>r. For each period, we give the value of the estim<strong>at</strong>ed<br />
coefficient β (first columns of table 2 for each time interval). We then calcul<strong>at</strong>e the correl<strong>at</strong>ion<br />
coefficient b<strong>et</strong>ween the mark<strong>et</strong> r<strong>et</strong>urns and the estim<strong>at</strong>ed idiosyncr<strong>at</strong>ic noise. All of them are less<br />
than 10 −8 , so th<strong>at</strong> none of them is significantly different from zero, which allows us to conclu<strong>de</strong><br />
th<strong>at</strong> there is no linear correl<strong>at</strong>ion b<strong>et</strong>ween the factor and the residues. To check one step further<br />
the in<strong>de</strong>pen<strong>de</strong>nce hypothesis, we have estim<strong>at</strong>ed the correl<strong>at</strong>ion coefficient b<strong>et</strong>ween the square of<br />
the factor and the square of the error-terms. In table 2, their values are given in the second of the<br />
pair of columns presented for each period. A Fisher’s test shows th<strong>at</strong>, <strong>at</strong> the 95% confi<strong>de</strong>nce level,<br />
all these correl<strong>at</strong>ion coefficients are significantly different from zero. This result is not surprising<br />
and shows the existence of small but significant correl<strong>at</strong>ions b<strong>et</strong>ween the mark<strong>et</strong> vol<strong>at</strong>ility and<br />
the idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility. However, this will not invalid<strong>at</strong>e the empirical tests of our theor<strong>et</strong>ical<br />
results, since they hold even in presence of weakly <strong>de</strong>pen<strong>de</strong>nt factor and noise.<br />
The coefficient β’s we obtain by regressing each ass<strong>et</strong> r<strong>et</strong>urns on the Standard & Poor’s 500 r<strong>et</strong>urns<br />
are very close to within their uncertainties to the β’s given by the CRSP d<strong>at</strong>abase, which are<br />
estim<strong>at</strong>ed by regressing the ass<strong>et</strong>s r<strong>et</strong>urns on the value-weighted mark<strong>et</strong> portfolio. Thus, the choice<br />
of the Standard and Poor’s 500 in<strong>de</strong>x to represent the whole mark<strong>et</strong> portfolio is reasonable.<br />
3.3 Estim<strong>at</strong>ion of the tail in<strong>de</strong>xes<br />
Assuming th<strong>at</strong> the distributions of stocks and mark<strong>et</strong> r<strong>et</strong>urns are asymptotically power laws (Longin<br />
(1996), Lux (1996), Pagan (1996) or Gopikrishnan, Meyer, Amaral, and Stanley (1998)), we now<br />
estim<strong>at</strong>e the tail in<strong>de</strong>x of the distribution of each stock and their corresponding residue by the factor<br />
mo<strong>de</strong>l, both for the positive and neg<strong>at</strong>ive tails. Each tail in<strong>de</strong>x α is given by Hill’s estim<strong>at</strong>or:<br />
ˆα =<br />
⎡<br />
⎣ 1<br />
k<br />
⎤<br />
k<br />
log xj,N − log xk,N⎦<br />
j=1<br />
−1<br />
, (21)<br />
where x1,N ≥ x2,N ≥ · · · ≥ xN,N <strong>de</strong>notes the or<strong>de</strong>red st<strong>at</strong>istics of the sample containing N in<strong>de</strong>pen<strong>de</strong>nt<br />
and i<strong>de</strong>ntically distributed realiz<strong>at</strong>ions of the variable X.<br />
Hill’s estim<strong>at</strong>or is asymptotically normally distributed with mean α and variance α 2 /k. But,<br />
for finite k, it is known th<strong>at</strong> the estim<strong>at</strong>or is biased. As the range k increases, the variance of the<br />
estim<strong>at</strong>or <strong>de</strong>creases while its bias increases. The comp<strong>et</strong>ition b<strong>et</strong>ween these two effects implies th<strong>at</strong><br />
there is an optimal choice for k = k ∗ which minimizes the mean squared error of the estim<strong>at</strong>or.<br />
To select this value k ∗ , one can apply the Danielsson and <strong>de</strong> Vries (1997)’s algorithm which is an<br />
improvement over the Hall (1990)’s subsample bootstrap procedure. One can also prefer the more<br />
recent Danielsson, <strong>de</strong> Haan, Peng, and <strong>de</strong> Vries (2001)’s algorithm for the sake of parsimony. We<br />
have tested all three algorithms to d<strong>et</strong>ermine the optimal k ∗ . It turns out th<strong>at</strong> the Danielsson,<br />
<strong>de</strong> Haan, Peng, and <strong>de</strong> Vries (2001)’s algorithm <strong>de</strong>veloped for high frequency d<strong>at</strong>a is not well<br />
adapted to samples containing less than 100,000 d<strong>at</strong>a points, as is the case here. Thus, we have<br />
focused on the two other algorithms. An accur<strong>at</strong>e d<strong>et</strong>ermin<strong>at</strong>ion of k ∗ is r<strong>at</strong>her difficult with any of<br />
12
312 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
them, but in every case, we found th<strong>at</strong> the relevant range for the tail in<strong>de</strong>x estim<strong>at</strong>ion was b<strong>et</strong>ween<br />
the 1% and 5% quantiles. Tables 3 and 4 give the estim<strong>at</strong>ed tail in<strong>de</strong>x for each ass<strong>et</strong> and residues<br />
<strong>at</strong> the 1%, 2.5% and 5% quantile, for both the positive and the neg<strong>at</strong>ive tails for the two time<br />
sub-intervals. The second time interval from January 1980 to December 2000 is characterized by<br />
values of the tail in<strong>de</strong>xes th<strong>at</strong> are homogeneous over the various quantiles and range b<strong>et</strong>ween 3 and<br />
4 for the neg<strong>at</strong>ive tails and b<strong>et</strong>ween 3 and 5 for the positive tails. There is slightly more dispersions<br />
in the first time interval from July 1962 to December 1979.<br />
For each ass<strong>et</strong> and their residue of the regression on the mark<strong>et</strong> factor, we tested wh<strong>et</strong>her the<br />
hypothesis, according to which the tail in<strong>de</strong>x measured for each ass<strong>et</strong> and each residue is the same<br />
as the tail in<strong>de</strong>x of the Standard & Poor’s 500 in<strong>de</strong>x, can be rejected <strong>at</strong> the 95% confi<strong>de</strong>nce level,<br />
for a given quantile. Before proceding with the present<strong>at</strong>ion of our tests, two cave<strong>at</strong>s have to<br />
be accounted for. First, due to the phenomenon of vol<strong>at</strong>ility clustering in financial time series,<br />
extremes are more likely to occur tog<strong>et</strong>her. In this situ<strong>at</strong>ion, Hill’s estim<strong>at</strong>or is no more normally<br />
distributed with variance α 2 /k. In fact, for weakly <strong>de</strong>pen<strong>de</strong>nt time series, it can only be asserted<br />
th<strong>at</strong> the estim<strong>at</strong>or remains consistent (see Rootzén and <strong>de</strong> Haan (1998)). Moreover, as shown by<br />
Kearns and Pagan (1997) for h<strong>et</strong>eroskedastic time series, the variance of the estim<strong>at</strong>ed tail in<strong>de</strong>x<br />
can be seven times larger than the variance given by the asymptic normality assumption. Second,<br />
the idiosyncr<strong>at</strong>ic noise is estim<strong>at</strong>ed by substracting β times the factor from the ass<strong>et</strong> r<strong>et</strong>urn. Thus,<br />
even when the factor and the error-term are in<strong>de</strong>pen<strong>de</strong>nt, the empirically estim<strong>at</strong>ed residues <strong>de</strong>pend<br />
on the realiz<strong>at</strong>ions of the factor. As a consequence, the tail in<strong>de</strong>x estim<strong>at</strong>ors for the factor and for<br />
the idiosyncr<strong>at</strong>ic noise are correl<strong>at</strong>ed. This correl<strong>at</strong>ion obviously <strong>de</strong>pends on the exact form of the<br />
distributions of the factor and the indiosyncr<strong>at</strong>ic noise. Even without the knowledge of the true<br />
test st<strong>at</strong>istics, for both problem<strong>at</strong>ic points, we can assert th<strong>at</strong> the fluctu<strong>at</strong>ions of the estim<strong>at</strong>ors<br />
are larger than those given by the asymptotically normal st<strong>at</strong>isics for i.i.d realiz<strong>at</strong>ions. Thus,<br />
performing the test un<strong>de</strong>r the asymptotic normality assumption is more constraining than un<strong>de</strong>r<br />
the true (but unknown) test st<strong>at</strong>istics, so th<strong>at</strong> the non-rejection of the equality hypothesis un<strong>de</strong>r<br />
the assumption of a normally distributed estim<strong>at</strong>or ensures th<strong>at</strong> we would not be able to reject this<br />
hypothesis un<strong>de</strong>r the real st<strong>at</strong>istics of the estim<strong>at</strong>or.<br />
The values which reject the equality hypothesis are indic<strong>at</strong>ed by a star in the tables 3 and 4. During<br />
the second time interval from January 1980 to December 2000, only four residues have a tail in<strong>de</strong>x<br />
significantly different from th<strong>at</strong> of th<strong>at</strong> Standard & Poor’s 500, and only in the neg<strong>at</strong>ive tail. The<br />
situ<strong>at</strong>ion is not as good during the first time interval, especially for the neg<strong>at</strong>ive tail, for which<br />
no less than 13 ass<strong>et</strong>s and 10 residues out of 20 have a tail in<strong>de</strong>x significantly different from the<br />
Standard & Poor’s 500 ones, for the 5% quantile. Recall th<strong>at</strong> the the equality tests have been<br />
performed un<strong>de</strong>r the assumption of a normally distributed estim<strong>at</strong>or with variance α 2 /k which, as<br />
explained above, is too strong an hypothesis. As a consequence, a rejection un<strong>de</strong>r the normality<br />
hypothesis does not imply necessarily th<strong>at</strong> the equality hypothesis would have been rejected un<strong>de</strong>r<br />
the true st<strong>at</strong>istics. While providing a note of caution, this st<strong>at</strong>ement is nevertheless not very useful<br />
from a practical point of view. More importantly, we stress th<strong>at</strong> the equality of the tail indices of<br />
the distribution of the factor and of the idiosyncr<strong>at</strong>ic noise is not crucial. In<strong>de</strong>ed, we shall propose<br />
below two different estim<strong>at</strong>ors for the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce. One of them does not rely on<br />
the equality of these two tail indices and thus remains oper<strong>at</strong>ional even when they are different<br />
and in particular when the tail in<strong>de</strong>x of the idiosyncr<strong>at</strong>ic noise appears larger than the tail of the<br />
factor.<br />
To summarize, our tests confirm th<strong>at</strong> the tail in<strong>de</strong>xes of most stock r<strong>et</strong>urn distributions range<br />
b<strong>et</strong>ween three and four, even though no b<strong>et</strong>ter precision can be given with good significance.<br />
Moreover, in most cases, we can assume th<strong>at</strong> both the ass<strong>et</strong>, the factor and the residue have the<br />
13
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 313<br />
same tail in<strong>de</strong>x. We can also add th<strong>at</strong>, as asserted by Lor<strong>et</strong>an and Phillips (1994) or Longin (1996),<br />
we cannot reject the hypothesis th<strong>at</strong> the tail in<strong>de</strong>x remains the same over time. Nevertheless, it<br />
seems th<strong>at</strong> during the first period from July 1962 to December 1979, the tail in<strong>de</strong>xes were sightly<br />
larger than during the second period from January 1980 to <strong>de</strong>cember 2000.<br />
3.4 D<strong>et</strong>ermin<strong>at</strong>ion of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
Using the just established empirical fact th<strong>at</strong> we cannot reject the hypothesis th<strong>at</strong> the ass<strong>et</strong>s, the<br />
mark<strong>et</strong> and the residues have the same tail in<strong>de</strong>x, we can use the theorem of Appendix A and<br />
its second corollary st<strong>at</strong>ed in section 2. This allows us to conclu<strong>de</strong> th<strong>at</strong> one cannot reject the<br />
hypothesis of a non-vanishing tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s and the mark<strong>et</strong>.<br />
In addition, the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is given by equ<strong>at</strong>ions (12) and (13). These equ<strong>at</strong>ions<br />
provi<strong>de</strong> two ways for estim<strong>at</strong>ing the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce: non-param<strong>et</strong>ric with (12) and<br />
param<strong>et</strong>ric with (13). The first one is more general since it only requires the hypothesis of a regular<br />
vari<strong>at</strong>ion, while the second one explicitly assumes th<strong>at</strong> the factor and the residues have distributions<br />
with power law tails.<br />
To estim<strong>at</strong>e the tail <strong>de</strong>pen<strong>de</strong>nce according equ<strong>at</strong>ion (12), we need only to d<strong>et</strong>ermine the constant l<br />
<strong>de</strong>fined (8). Consi<strong>de</strong>r N sorted realiz<strong>at</strong>ions of X and Y <strong>de</strong>noted by x1,N ≥ x2,N ≥ · · · ≥ xN,N and<br />
y1,N ≥ y2,N ≥ · · · ≥ yN,N, the quantile of F −1<br />
−1<br />
X (u) and of FY (u) are estim<strong>at</strong>ed by<br />
−1 −1<br />
FX<br />
ˆ (u) = x[(1−u)·N],N and FY<br />
ˆ (u) = y[(1−u)·N],N, (22)<br />
where [·] <strong>de</strong>notes the integer part. Thus, the constant l is non-param<strong>et</strong>rically estim<strong>at</strong>ed by<br />
ˆ lk = xk,N<br />
yk,N<br />
as k → 0 or N. (23)<br />
As u goes to zero or one (or k goes to zero or N), the number of observ<strong>at</strong>ions <strong>de</strong>creases dram<strong>at</strong>ically.<br />
However, we observe a large interval of small or large k’s such th<strong>at</strong> the r<strong>at</strong>io of the empirical<br />
quantiles remains remarkably stable and thus allows for an accur<strong>at</strong>e estim<strong>at</strong>ion of l. A more<br />
precise estim<strong>at</strong>ion could be performed with a kernel-based quantile estim<strong>at</strong>or (see Shealter and<br />
Marron (1990) or Pagan and Ullah (1999) for instance). A non-param<strong>et</strong>ric estim<strong>at</strong>or for λ is then<br />
obtained by replacing l by its estim<strong>at</strong>ed value in equ<strong>at</strong>ion (12)<br />
ˆλNP =<br />
1<br />
<br />
max 1, ˆ 1<br />
α = <br />
l<br />
ˆβ max 1, xk,N<br />
α . (24)<br />
ˆβ·yk,N<br />
It can also be advantageous to follow a param<strong>et</strong>ric approach, which generally allows for a more<br />
accur<strong>at</strong>e estim<strong>at</strong>ion of (the r<strong>at</strong>io of) the quantiles, provi<strong>de</strong>d th<strong>at</strong> the assumed param<strong>et</strong>ric form of<br />
the distributions is not too far from the true one. For this purpose, we will use formula (13) which<br />
requires the estim<strong>at</strong>ion of the scale factors for the different ass<strong>et</strong>s. To g<strong>et</strong> the scale factors, we<br />
proceed as follows. Consi<strong>de</strong>r a variable X which asymptotically follows a power law distribution<br />
Pr{X > x} ∼ C · x −α . Given a rank or<strong>de</strong>red sample x1,N ≥ x2,N ≥ · · · ≥ xN,N, the scale factor C<br />
can be consistently estim<strong>at</strong>ed from the k largest realiz<strong>at</strong>ions by<br />
Ĉ = k<br />
N · (xk,N) α . (25)<br />
The estim<strong>at</strong>ed value of the scale factor must not <strong>de</strong>pends on the rank k for k large enough, in or<strong>de</strong>r<br />
for the param<strong>et</strong>eriz<strong>at</strong>ion of the distribution in terms of a power law to hold true. Thus, <strong>de</strong>noting<br />
14
314 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
by ĈY and Ĉε the estim<strong>at</strong>ed scale factors of the factor Y and of the noise ε <strong>de</strong>fined in equ<strong>at</strong>ion (6),<br />
the estim<strong>at</strong>or of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is<br />
ˆλ =<br />
1<br />
1 + ˆ β −α · ĈY<br />
Ĉε<br />
=<br />
1 +<br />
1<br />
εk,N<br />
ˆβ·yk,N<br />
α , (26)<br />
where ˆ β <strong>de</strong>notes the estim<strong>at</strong>ed coefficient β. Since the estim<strong>at</strong>ors ĈY , Ĉε and ˆ β are consistent and<br />
using the continuous mapping theorem, we can assert th<strong>at</strong> the estim<strong>at</strong>or ˆ λ is also consistent.<br />
Since the tail indices α are impossible to d<strong>et</strong>ermine with sufficient accuracy other than saying th<strong>at</strong><br />
the α probably fall in the interval 3 − 4 as we have seen above, our str<strong>at</strong>egy is to d<strong>et</strong>ermine the<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce using (24) and (26) for three different common values α = 3, 3.5 and<br />
4. This procedure allows us to test for the sensitivity of the scale factor and therefore of the tail<br />
coefficient with respect to the uncertain value of the tail in<strong>de</strong>x.<br />
Tables 5 and 6 give the values of the coefficients of lower tail <strong>de</strong>pen<strong>de</strong>nce over the whole time<br />
interval from July 1962 to December 2000, un<strong>de</strong>r the assumption th<strong>at</strong> the tail in<strong>de</strong>x α equals 3,<br />
for the non-param<strong>et</strong>ric estim<strong>at</strong>or (table 5) and the param<strong>et</strong>ric one (table 6). For each table, the<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is estim<strong>at</strong>ed over the first centile, the first quintile and the first <strong>de</strong>cile<br />
to also test for any possible sensitivity on the tail asymptotics. For each of these quantiles, the<br />
mean values, their standard <strong>de</strong>vi<strong>at</strong>ions and their minimum and maximum values are given. We<br />
first remark th<strong>at</strong> the standard <strong>de</strong>vi<strong>at</strong>ion of the tail <strong>de</strong>pen<strong>de</strong>nce coefficient remains small compared<br />
with its average value and th<strong>at</strong> the minimum and maximum values cluster closely around its mean<br />
value. This shows th<strong>at</strong> the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is well-estim<strong>at</strong>ed by its mean over a given<br />
quantile. Secondly, we find th<strong>at</strong> these estim<strong>at</strong>ed coefficients of tail <strong>de</strong>pen<strong>de</strong>nce exhibit a good<br />
stability over the various quantiles. These two observ<strong>at</strong>ions enable us to conclu<strong>de</strong> th<strong>at</strong> the average<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce over the first centile is sufficient to provi<strong>de</strong> a good estim<strong>at</strong>e of the<br />
true coefficient of tail <strong>de</strong>pen<strong>de</strong>nce.<br />
Note th<strong>at</strong> the two estim<strong>at</strong>ors yield essentially equivalent results, even if the coefficients of tail<br />
<strong>de</strong>pen<strong>de</strong>nce given by the non-param<strong>et</strong>ric estim<strong>at</strong>or exhibit a system<strong>at</strong>ic ten<strong>de</strong>ncy to be slightly<br />
smaller than the estim<strong>at</strong>es provi<strong>de</strong>d by the param<strong>et</strong>ric estim<strong>at</strong>or. Since the results given by these<br />
two estim<strong>at</strong>ors are very close to each other, we choose to present below only those given by the<br />
param<strong>et</strong>ric one. This choice has also been gui<strong>de</strong>d by the lower sensibility of this last estim<strong>at</strong>or<br />
to small changes of the tail exponent α. In<strong>de</strong>ed, since the evalu<strong>at</strong>ion of the scale factors ĈY and<br />
Ĉε by the formula (25) involves the tail exponent α, the small <strong>de</strong>vi<strong>at</strong>ions from its true value are<br />
compens<strong>at</strong>ed by the estim<strong>at</strong>ed scale factors. This explains the observ<strong>at</strong>ion th<strong>at</strong> the param<strong>et</strong>ric<br />
estim<strong>at</strong>or appears more robust than the non-param<strong>et</strong>ric one with respect to small changes in α.<br />
Tables 7, 8 and 9 summarize the different values of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce for both the<br />
positive and the neg<strong>at</strong>ive tails, un<strong>de</strong>r the assumptions th<strong>at</strong> the tail in<strong>de</strong>x α equals 3, 3.5 and 4<br />
respectively, over the three consi<strong>de</strong>red time intervals. Overall, we find th<strong>at</strong> the coefficients of tail<br />
<strong>de</strong>pen<strong>de</strong>nce are almost equal for both the neg<strong>at</strong>ive and the positive tail and th<strong>at</strong> they are not very<br />
sensitive to the value of the tail in<strong>de</strong>x in the interval consi<strong>de</strong>red. More precisely, during the first<br />
time interval from July 1962 to December 1979 (table 7), the tail <strong>de</strong>pen<strong>de</strong>nce is symm<strong>et</strong>ric in both<br />
the upper and the lower tail. During the second time interval from January 1980 to December 2000<br />
and over the whole time interval (tables 8 and 9), the coefficient of lower tail <strong>de</strong>pen<strong>de</strong>nce is slightly<br />
but system<strong>at</strong>ically larger than the upper one. Moreover, since these coefficients of tail <strong>de</strong>pen<strong>de</strong>nce<br />
are all less than 1/2, they <strong>de</strong>crease when the tail in<strong>de</strong>x α increases and the smaller the coefficient<br />
of tail <strong>de</strong>pen<strong>de</strong>nce, the larger the <strong>de</strong>cay.<br />
During the first time interval, most of the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce range b<strong>et</strong>ween 0.15 and<br />
15
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 315<br />
0.35 in both tails, while during the second time interval, almost all range b<strong>et</strong>ween 0.10 and 0.25 in<br />
the lower tail and b<strong>et</strong>ween 0.10 and 0.20 in the upper one. Thus, the tail <strong>de</strong>pen<strong>de</strong>nce is smaller<br />
during the last period compared to the first one. This result is interesting because it is in agreement<br />
and confirms the recent studies by Campbell, L<strong>et</strong>tau, Malkiel, and Xu (2001) and Xu and Malkiel<br />
(2002), showing th<strong>at</strong> the idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility of each stocks have increased rel<strong>at</strong>ive to the mark<strong>et</strong><br />
vol<strong>at</strong>ility. And, as already discussed, the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce given by equ<strong>at</strong>ion (10) must<br />
<strong>de</strong>crease when the idiosynchr<strong>at</strong>ic vol<strong>at</strong>ility of the stocks increases rel<strong>at</strong>ive to the mark<strong>et</strong> vol<strong>at</strong>ility.<br />
The strong similarity of the tail <strong>de</strong>pen<strong>de</strong>ncies in the upper and lower tails is an interesting empirical<br />
finding which suggests th<strong>at</strong> extreme co-movements reflect behaviors of agents which are more<br />
sensitive to large amplitu<strong>de</strong>s r<strong>at</strong>her than to a specific direction (loss or gain). Pictorially, the<br />
specific mechanism triggering co-movements of extreme amplitu<strong>de</strong>s may well be different for losses<br />
compared to gains, such as fear for the former and greed for the l<strong>at</strong>ter, but the resulting large<br />
co-movements have similar frequencies of occurrence.<br />
The observed lack of st<strong>at</strong>ionarity of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce in the two time sub-intervals<br />
suggests th<strong>at</strong> it could be necessary to have a mo<strong>de</strong>l where the tail <strong>de</strong>pen<strong>de</strong>nce in<strong>de</strong>x is not constant<br />
but varies as a function of past shocks (just as the vol<strong>at</strong>ility varies with time in a GARCH mo<strong>de</strong>l)<br />
in or<strong>de</strong>r to investig<strong>at</strong>e wh<strong>et</strong>her large recent common shocks lead to higher future tail <strong>de</strong>pen<strong>de</strong>nce.<br />
This point is beyond the scope of the present study, but we will provi<strong>de</strong> in our concluding remarks<br />
some ways for explicitely accounting for this lack of st<strong>at</strong>ionarity.<br />
3.5 Comparison with the historical extremes<br />
Our d<strong>et</strong>ermin<strong>at</strong>ion of the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce provi<strong>de</strong>s predictions on the probability<br />
th<strong>at</strong> future large moves of stocks may be simultaneous to large moves of the mark<strong>et</strong>. This begs for<br />
a check over the available historical period to d<strong>et</strong>ermine wh<strong>et</strong>her our estim<strong>at</strong>ed coefficients of tail<br />
<strong>de</strong>pen<strong>de</strong>nce are comp<strong>at</strong>ible with the realized historical extremes.<br />
For this, we consi<strong>de</strong>r the ten largest losses of the Standard & Poor’s 500 in<strong>de</strong>x during the two time<br />
sub-intervals5 . Since λ− is by <strong>de</strong>finition equal to the probability th<strong>at</strong> a given ass<strong>et</strong> incurs a large<br />
loss (say, one of its ten largest losses) conditional on the occurrence of one of the ten largest losses<br />
of the Standard & Poor’s 500 in<strong>de</strong>x, the probability, for this ass<strong>et</strong>, to un<strong>de</strong>rgo n of its ten largest<br />
losses simultaneously with any of the ten largest losses of the Standard & Poor’s 500 in<strong>de</strong>x is given<br />
by the binomial law with param<strong>et</strong>er λ−:<br />
Pλ− (n) =<br />
<br />
10<br />
λ−<br />
n<br />
n (1 − λ−) (10−n) . (27)<br />
We stress th<strong>at</strong> our consi<strong>de</strong>r<strong>at</strong>ion of only the ten largest drops ensures th<strong>at</strong> the present test is not<br />
embodied in the d<strong>et</strong>ermin<strong>at</strong>ion of the tail <strong>de</strong>pen<strong>de</strong>nce coefficient, which has been d<strong>et</strong>ermined on<br />
a robust procedure over the 1%, 5% and 10% quantiles. We checked th<strong>at</strong> removing these then<br />
largest drops does not modify the d<strong>et</strong>ermin<strong>at</strong>ion of λ−. Our present test can thus be consi<strong>de</strong>red as<br />
“out-of-sample,” in this sense.<br />
Table 10 presents, for the two time sub-intervals, the number of extreme losses among the ten<br />
largest losses incured by a given ass<strong>et</strong> which occured simultaneously with one of the ten largest<br />
losses of the standard & Poor’s 500 in<strong>de</strong>x. For each ass<strong>et</strong>, we give the probability of occurence of<br />
such a realis<strong>at</strong>ion, according to (27). We notice th<strong>at</strong> during the first time interval, only two ass<strong>et</strong>s<br />
5 We do not consi<strong>de</strong>r the whole time interval since the ten largest losses over the whole period coinci<strong>de</strong> with the<br />
ten largest ones over the second time subinterval, which would bias the st<strong>at</strong>istics towards the second time interval.<br />
16
316 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
are incomp<strong>at</strong>ible, <strong>at</strong> the 95% confi<strong>de</strong>nce level, with the value of λ− previously d<strong>et</strong>ermined: Du Pont<br />
(E.I.) <strong>de</strong> Nemours & Co. and Texas Instruments Inc. In contrast, during the second time interval,<br />
four ass<strong>et</strong>s reject the value of λ−: Coca Cola Corp., Pepsico Inc., Pharmicia Corp. and Texaco Inc.<br />
These results are very encouraging. However, there is a noticeable system<strong>at</strong>ic bias. In<strong>de</strong>ed, during<br />
the first time interval, 17 out of the 20 ass<strong>et</strong>s have a realized number of large losses lower than<br />
their expected number (according to the estim<strong>at</strong>ed λ−), while during the second time interval, 19<br />
out of the 20 ass<strong>et</strong>s have a realized number of large losses larger than their expected one. Thus, it<br />
seems th<strong>at</strong> during the first time interval the number of large losses is overestim<strong>at</strong>ed by λ− while it<br />
is un<strong>de</strong>restim<strong>at</strong>ed during the second time interval.<br />
We propose to explain the un<strong>de</strong>restim<strong>at</strong>ion of the number of large losses b<strong>et</strong>ween January 1980<br />
and December 2000 by a possible comonitonicity th<strong>at</strong> occurred during the October 1987 crash.<br />
In<strong>de</strong>ed, on October 19, 1987, 12 out of the 20 consi<strong>de</strong>red ass<strong>et</strong>s incurred their most severe loss,<br />
which strongly suggests a comonotonic effect. Table 11 shows the same results as in table 10 but<br />
corrected by substracting this comonotonic effect to the number of large losses. The comp<strong>at</strong>ibility<br />
b<strong>et</strong>ween the number of large losses and the estim<strong>at</strong>ed λ− becomes significantly b<strong>et</strong>ter since only<br />
Pepsico Inc. and Pharmicia Corp. are still rejected, and only 16 ass<strong>et</strong>s out of 20 are un<strong>de</strong>restim<strong>at</strong>ed,<br />
representing a slight <strong>de</strong>crease of the bias.<br />
Previous works have shown th<strong>at</strong>, in period of crashes, the mark<strong>et</strong> conditions change, herding effects<br />
may become more important and almost dominant, so th<strong>at</strong> the mark<strong>et</strong> enters an unusual regime,<br />
which can be characterized by outliers present in the distribution of drawdowns Johansen and<br />
Sorn<strong>et</strong>te (2002). Our d<strong>et</strong>ection of an anomalous comonotonicity can thus be consi<strong>de</strong>red as an<br />
in<strong>de</strong>pen<strong>de</strong>nt confirm<strong>at</strong>ion of the existence of this abnormal regime.<br />
Another explain<strong>at</strong>ion for this slight discrepancy may be ascribed to a limit<strong>at</strong>ion of the CAPM.<br />
In<strong>de</strong>ed, the CAPM is known to explain the rel<strong>at</strong>ion b<strong>et</strong>ween the expected r<strong>et</strong>urn on an ass<strong>et</strong> and<br />
its amount of system<strong>at</strong>ic risk. But, it is questionable wh<strong>et</strong>her extreme system<strong>at</strong>ic risks as those<br />
measured by the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce are really accounted for by the economic agents and<br />
then effectively priced.<br />
Concerning the overestim<strong>at</strong>ion of the number of large losses during the first time interval, it can<br />
obviously not be ascribed to the comonotonicity of very large events, which in fact only occurred<br />
once for the Coca-Cola Corp. This overestim<strong>at</strong>ion is probably linked with the low “vol<strong>at</strong>ility” of<br />
the mark<strong>et</strong> during this period, which can have two effects. The first one is to lead to a less accur<strong>at</strong>e<br />
estim<strong>at</strong>ion of the scale factor of the power-law distribution of the ass<strong>et</strong>s. The second one is th<strong>at</strong><br />
a mark<strong>et</strong> with smaller vol<strong>at</strong>ility produces fewer large losses. As a consequence, the asymptotic<br />
regime for which the rel<strong>at</strong>ion Pr{X < FX −1 (u)|Y < FY −1 (u)} λ− holds may not be reached<br />
in the sample, and the number of recor<strong>de</strong>d large losses remain lower than th<strong>at</strong> asymptotically<br />
expected.<br />
4 Concluding remarks<br />
We have used the framework offered by factor mo<strong>de</strong>ls in or<strong>de</strong>r to <strong>de</strong>rive a general theor<strong>et</strong>ical<br />
expression for the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween an ass<strong>et</strong> and any of its explan<strong>at</strong>ory factor<br />
or b<strong>et</strong>ween any two ass<strong>et</strong>s. The coefficient of tail <strong>de</strong>pen<strong>de</strong>nce represents the probability th<strong>at</strong> a given<br />
ass<strong>et</strong> incurs a large loss (say), assuming th<strong>at</strong> the mark<strong>et</strong> (or another ass<strong>et</strong>) has also un<strong>de</strong>rgone a<br />
large loss. We find th<strong>at</strong> factors characterized by rapidly varying distributions, such as Normal<br />
or exponential distributions, always lead to a vanishing coefficient of tail <strong>de</strong>pen<strong>de</strong>nce with other<br />
17
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 317<br />
stocks. In contrast, factors with regularly varying distributions, such as power-law distributions,<br />
can exhibit tail <strong>de</strong>pen<strong>de</strong>nce with other stocks, provi<strong>de</strong>d th<strong>at</strong> the idiosyncr<strong>at</strong>ic noise distributions<br />
of the corresponding stocks are not f<strong>at</strong>ter-tailed than the factor.<br />
Applying this general result to individual daily stock r<strong>et</strong>urns, we have been able to estim<strong>at</strong>e the<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the r<strong>et</strong>urns of each stock and those of the mark<strong>et</strong>. This<br />
d<strong>et</strong>ermin<strong>at</strong>ion of the tail <strong>de</strong>pen<strong>de</strong>nce relies only on the simple estim<strong>at</strong>ion of the param<strong>et</strong>ers of the<br />
un<strong>de</strong>rlying factor mo<strong>de</strong>l and on the tail param<strong>et</strong>ers of the distribution of the factor and of the<br />
idiosyncr<strong>at</strong>ic noise of each stock. As a consequence, the two strong advantages of our approach are<br />
the following.<br />
- The coefficients of tail <strong>de</strong>pen<strong>de</strong>nce are estim<strong>at</strong>ed non-param<strong>et</strong>rically. In<strong>de</strong>ed, we never specify<br />
any explicit expression of the <strong>de</strong>pen<strong>de</strong>nce structure, contrary to most previous works (see<br />
Longin and Solnik (2001), Malevergne and Sorn<strong>et</strong>te (2001) or P<strong>at</strong>ton (2001) for instance);<br />
- Our theor<strong>et</strong>ical result enables us to estim<strong>at</strong>e an extreme param<strong>et</strong>er, not accessible by a direct<br />
st<strong>at</strong>istical inference. This is achieved by the measurement of param<strong>et</strong>ers whose estim<strong>at</strong>ion<br />
involves a significant part of the d<strong>at</strong>a with sufficient st<strong>at</strong>istics.<br />
Having performed this estim<strong>at</strong>ion, we have checked the compt<strong>at</strong>ibility of these estim<strong>at</strong>ed coefficients<br />
of tail <strong>de</strong>pen<strong>de</strong>nce with the historically realized extreme losses observed in the empirical time<br />
series. A good agreement is found, notwithstanding a slight bias which leads to an overestim<strong>at</strong>e<br />
of the occurence of large events during the period from July 1962 to December 1979 and to an<br />
un<strong>de</strong>restim<strong>at</strong>e during the time interval from January 1980 to December 2000.<br />
This bias can be explained by the low vol<strong>at</strong>ility of the mark<strong>et</strong> during the first period and by a<br />
comonotonicity effect, due to the October 1987 crach, during the second period. In<strong>de</strong>ed, from july<br />
1962 to December 1979, the vol<strong>at</strong>ility was so low th<strong>at</strong> the distributions of r<strong>et</strong>urns have probably not<br />
sampled their tails sufficiently for the probability of large conditional losses to be represented by<br />
its asymptotic expression given by the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce. The situ<strong>at</strong>ion is very different<br />
for the period from january 1980 to December 2000. On October 19, 1987, many ass<strong>et</strong>s incurred<br />
their largest loss ever. This is presumably the manifest<strong>at</strong>ion of an ‘abnormal’ regime probably<br />
due to herding effects and irr<strong>at</strong>ional behaviors and has been previously characterized as yielding<br />
sign<strong>at</strong>ures in the form of outliers in the distribution of drawdowns.<br />
Finally, the observed lack of st<strong>at</strong>ionarity exhibited by the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce across the<br />
two time sub-intervals suggests the importance of going beyond a st<strong>at</strong>ionary view of tail <strong>de</strong>pen<strong>de</strong>nce<br />
and of studying its dynamics. This question, which could be of gre<strong>at</strong> interest in the context of the<br />
contagion problem, could be easily tre<strong>at</strong>ed with the new conditional quantile dynamics proposed<br />
by Engle and Manganelli (1999). Moreover, it should be interesting to account for the change of<br />
the β’s with incoming bad or good news, as shown by Cho and Engle (2000), for instance. These<br />
points are left for a future work.<br />
From a practical point of view, we stress th<strong>at</strong> the coefficient λ studied here can be seen as a<br />
generaliz<strong>at</strong>ion or a tool complementary to the CAPM’s β. These two coefficients have in common<br />
th<strong>at</strong> they probe the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween a given stock and the mark<strong>et</strong>. However, the coefficient β<br />
quantifies only the correl<strong>at</strong>ions b<strong>et</strong>ween mo<strong>de</strong>r<strong>at</strong>e movements of both an ass<strong>et</strong> and the mark<strong>et</strong>. In<br />
contrast, the coefficient λ offers a measure of extreme co-movements, which is particularly useful<br />
in period of high mark<strong>et</strong> vol<strong>at</strong>ility. In such periods, a pru<strong>de</strong>nt fund manager should overweight its<br />
portofolio with ass<strong>et</strong>s whose λ is very small such as Texaco or Walgreen, for instance.<br />
Moreover, the observed <strong>de</strong>crease of the tail <strong>de</strong>pen<strong>de</strong>nce during the last year concomitant with the<br />
18
318 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
increase of the idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility suggests th<strong>at</strong> the main source of risk in such a period does<br />
not consist in the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s but r<strong>at</strong>her in their instrinsic fluctu<strong>at</strong>ions measured<br />
by the idiosyncr<strong>at</strong>ic volality.<br />
Our study has focused on the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween different risks. In fact, our theorem can obviously<br />
be applied to extreme temporal <strong>de</strong>pen<strong>de</strong>nces, when the variable follows an autoregressive process.<br />
This should provi<strong>de</strong> an estim<strong>at</strong>e of the probability th<strong>at</strong> a large loss (respectively gain) is followed<br />
by another large loss (resp. gain) in the following period. Such inform<strong>at</strong>ion is very interesting in<br />
investment and hedging str<strong>at</strong>egies.<br />
19
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 319<br />
A Proof of the theorem<br />
A.1 Tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween an ass<strong>et</strong> and the factor<br />
A.1.1 St<strong>at</strong>ement<br />
We consi<strong>de</strong>r two random variables X and Y , rel<strong>at</strong>ed by the rel<strong>at</strong>ion<br />
X = β · Y + ε, (28)<br />
where ε is a random variable in<strong>de</strong>pen<strong>de</strong>nt of Y and β a non-random positive coefficient.<br />
L<strong>et</strong> PY and FY <strong>de</strong>note respectively the <strong>de</strong>nsity with respect to the Lebesgue measure and the<br />
distribution function of the variable Y . L<strong>et</strong> FX <strong>de</strong>notes the distribution function of X and Fε the<br />
distribution function of ε. We st<strong>at</strong>e the following theorem:<br />
Theorem 1<br />
Assuming th<strong>at</strong><br />
H0: The variables Y and ε have distribution functions with infinite support,<br />
H1: For all x ∈ [1, ∞),<br />
lim<br />
t→∞<br />
t PY (tx)<br />
¯<br />
FY (t)<br />
= f(x), (29)<br />
H2: There are real numbers t0 > 0, δ > 0 and A > 0, such th<strong>at</strong> for all t ≥ t0 and all x ≥ 1<br />
H3: There is a constant l ∈ R+, such th<strong>at</strong><br />
¯FY (tx)<br />
¯FY (t)<br />
then, the coefficient of (upper) tail <strong>de</strong>pen<strong>de</strong>nce of (X, Y ) is given by<br />
A.1.2 Proof<br />
λ =<br />
A<br />
≤ , (30)<br />
xδ FX<br />
lim<br />
u→1<br />
−1 (u)<br />
FY −1 = l, (31)<br />
(u)<br />
∞<br />
max{1, l<br />
β }<br />
dx f(x). (32)<br />
We first give a general expression for the probability for X to be larger than F −1<br />
X (u) knowing th<strong>at</strong><br />
(u) :<br />
Y is larger than F −1<br />
Y<br />
Lemma 1<br />
The probability th<strong>at</strong> X is larger than F −1<br />
−1<br />
X (u) knowing th<strong>at</strong> Y is larger than FY (u) is given by :<br />
Pr X > F −1<br />
−1<br />
X (u)|Y > FY (u) =<br />
F −1<br />
Y (u)<br />
1 − u<br />
∞<br />
dx PY<br />
1<br />
20<br />
−1<br />
FY (u) x · ¯ <br />
Fε FX −1 (u) − βF −1<br />
Y (u) x .<br />
(33)
320 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Proof :<br />
Pr{X > FX −1 (u), Y > FY −1 (u)} =<br />
<br />
E 1 {X>FX −1 (u)} · 1 {Y >FY −1 <br />
(u)}<br />
<br />
= E E 1 {X>FX −1 (u)} · 1 {Y >FY −1 (u)} |Y<br />
<br />
<br />
= E 1 {Y >FY −1 <br />
(u)} · E 1 {X>FX −1 (u)} |Y<br />
<br />
<br />
= E 1 {Y >FY −1 <br />
(u)} · E 1 {ε>FX −1 <br />
=<br />
(u)−βY }<br />
<br />
E 1 {Y >FY −1 (u)} · ¯ Fε(FX −1 <br />
(u) − βY )<br />
Assuming th<strong>at</strong> the variable Y admits a <strong>de</strong>nsity PY with respect to the Lebesgue measure, this<br />
yields<br />
Pr{X > FX −1 (u), Y > FY −1 ∞<br />
(u)} =<br />
F −1<br />
Y (u)<br />
(34)<br />
(35)<br />
(36)<br />
(37)<br />
(38)<br />
dy PY (y) · ¯ Fε[FX −1 (u) − βy] . (39)<br />
Performing the change of variable y = FY −1 (u) · x, in the equ<strong>at</strong>ion above, we obtain<br />
Pr{X > FX −1 (u), Y > FY −1 (u)} = F −1<br />
Y (u)<br />
∞<br />
1<br />
dx PY (F −1<br />
and, dividing by ¯<br />
<br />
FY FY −1 (u) = 1 − u, this conclu<strong>de</strong>s the proof. <br />
L<strong>et</strong> us now <strong>de</strong>fine the function<br />
fu(x) =<br />
We can st<strong>at</strong>e the following result<br />
−1<br />
FY (u)<br />
1 − u PY (F −1<br />
Y (u) x) · ¯ Fε[FX −1 (u) − βF −1<br />
Y<br />
Lemma 2<br />
Un<strong>de</strong>r assumption H1 and H3, for all x ∈ [1, ∞),<br />
almost everywhere, as u goes to 1.<br />
fu(x) −→ 1x> l<br />
Proof: L<strong>et</strong> us apply the assumption H1. We have<br />
lim<br />
u→1<br />
F −1<br />
Y (u)<br />
Applying now the assumption H3, we have<br />
lim<br />
u→1 FX −1 (u) − βF −1<br />
Y<br />
Y (u) x) · ¯ Fε[FX −1 (u) − βF −1<br />
Y<br />
(u) x] ,<br />
(40)<br />
(u) x] . (41)<br />
β · f(x), (42)<br />
1 − u PY (F −1<br />
Y (u) x) =<br />
t PY (t x)<br />
lim<br />
t→∞ FY<br />
¯<br />
,<br />
(t)<br />
(43)<br />
= f(x). (44)<br />
(u) x = lim<br />
=<br />
21<br />
u→1<br />
βF −1<br />
Y (u)<br />
<br />
−∞ if x > l<br />
β ,<br />
∞ if x < l<br />
β ,<br />
FX −1 (u)<br />
βF −1 − x<br />
Y (u)<br />
<br />
(45)<br />
(46)<br />
(47)
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 321<br />
which gives<br />
and finally<br />
lim<br />
u→1<br />
lim<br />
u→1 fu(x) = lim<br />
u→1<br />
= 1x> l<br />
which conclu<strong>de</strong>s the proof. <br />
¯Fε[FX −1 (u) − βF −1<br />
β Y (u) x] = 1x><br />
l , (48)<br />
F −1<br />
Y (u)<br />
1 − u PY (F −1<br />
Y (u) x) · lim ¯Fε[FX<br />
u→1<br />
−1 (u) − βF −1<br />
Y (u) x], (49)<br />
β · f(x), (50)<br />
L<strong>et</strong> us now prove th<strong>at</strong> there exists an integrable function g(x) such th<strong>at</strong>, for all t ≥ t0 and all x ≥ 1,<br />
we have ft(x) ≤ g(x). In<strong>de</strong>ed, l<strong>et</strong> us write<br />
t PY (tx)<br />
¯FY (t) = t PY (tx)<br />
¯FY (tx) · ¯ FY (tx)<br />
¯FY (t)<br />
For the leftmost factor in the right-hand-si<strong>de</strong> of equ<strong>at</strong>ion (51), we easily obtain<br />
∀t, ∀x ≥ 1,<br />
t PY (tx)<br />
¯FY (tx) ≤ x∗ PY (x ∗ )<br />
¯FY (x ∗ )<br />
. (51)<br />
· 1<br />
, (52)<br />
x<br />
where x∗ <strong>de</strong>notes the point where the function x PY (x)<br />
¯FY<br />
reaches its maximum. The rightmost factor<br />
(x)<br />
in the right-hand-si<strong>de</strong> of (51) is smaller than A/xδ by assumption H2, so th<strong>at</strong><br />
Posing<br />
∀t ≥ t0, ∀x ≥ 1,<br />
t PY (tx)<br />
¯FY (t) ≤ x∗ PY (x ∗ )<br />
¯FY (x ∗ )<br />
g(x) = x∗ PY (x ∗ )<br />
¯FY (x ∗ )<br />
· A<br />
. (53)<br />
x1+δ · A<br />
, (54)<br />
x1+δ and recalling th<strong>at</strong>, for all ε ∈ R, ¯ Fε(ε) ≤ 1, we have found an integrable function such th<strong>at</strong> for<br />
some u0 ≥ 0, we have<br />
∀u ∈ [u0, 1), ∀x ≥ 1, fu(x) ≤ g(x) . (55)<br />
Thus, applying Lebesgue’s theorem of domin<strong>at</strong>ed convergence, we can assert th<strong>at</strong><br />
∞<br />
∞<br />
lim dx fu(x) = dx 1x> u→1 1<br />
1<br />
l β · f(x). (56)<br />
Since<br />
lim<br />
u→1<br />
∞<br />
the proof of theorem 1 is conclu<strong>de</strong>d. <br />
1<br />
dx fu(x) = lim Pr<br />
u→1 X > F −1<br />
−1<br />
X (u)|Y > FY (u) , (57)<br />
= λ, (58)<br />
Remark: This result still holds in the presence of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the factor and the idiosyncr<strong>at</strong>ic<br />
noise. In<strong>de</strong>ed, <strong>de</strong>noting by ¯ Fε|Y the survival distribution of ε conditional on Y , lemma 1 can<br />
easily be generalized:<br />
Pr X > F −1<br />
−1<br />
X (u)|Y > FY (u) =<br />
F −1<br />
Y (u)<br />
1 − u<br />
∞<br />
dx PY<br />
1<br />
22<br />
F −1<br />
Y (u) x · ¯ F −1<br />
ε|Y =FY (u) x<br />
FX −1 (u) − βF −1<br />
Y (u) x ,<br />
(59)
322 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
where the only change in (59) compared to (33) is to replace ¯ Fε(·) by ¯ F −1<br />
ε|Y =FY (u) x(·). L<strong>et</strong> us<br />
now assume th<strong>at</strong> the function ¯ Fε|Y =y(x) admits a uniform limit when x and y tend to ±∞. Then,<br />
equ<strong>at</strong>ion (48) still holds and lemma 2 remains true.<br />
As an example, l<strong>et</strong> F <strong>de</strong>note any one-dimensional distribution fonction. Then, one can easily check<br />
th<strong>at</strong>, for any conditional distribution whose form is<br />
¯F ε|Y =y(x) = ¯ <br />
y2 F x , (60)<br />
+ y2<br />
the uniform limit condition is s<strong>at</strong>isfied and theorem 1 and lemma 2 still hold. In contrast, conditional<br />
distributions of the form<br />
¯F ε|Y =y(x) = ¯ F (x − ρy) (61)<br />
do not fulfill the uniform limit condition, so th<strong>at</strong> the result given by theorem 1 does not hold.<br />
The full un<strong>de</strong>rstanding of the impact of more general <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween the factor and the<br />
idiosyncr<strong>at</strong>ic noise on the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce requires a full-fledge investig<strong>at</strong>ion th<strong>at</strong> we<br />
<strong>de</strong>fer to a future work. Our goal here has been to show th<strong>at</strong> one can reasonably expect our results<br />
to survice in the presence of weak <strong>de</strong>pen<strong>de</strong>nce.<br />
A.2 Tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween two ass<strong>et</strong>s<br />
A.2.1 St<strong>at</strong>ement<br />
We consi<strong>de</strong>r three random variables X1, X2 and Y , rel<strong>at</strong>ed by the rel<strong>at</strong>ions<br />
y 2 0<br />
X1 = β1 · Y + ε1 (62)<br />
X2 = β2 · Y + ε2, (63)<br />
where ε1 and ε2 are two random variables in<strong>de</strong>pen<strong>de</strong>nt of Y and β1, β2 two non-random positive<br />
coefficients.<br />
L<strong>et</strong> PY and FY <strong>de</strong>note respectively the <strong>de</strong>nsity with respect to the Lebesgue measure and the<br />
distribution function of the variable Y . L<strong>et</strong> F1, (resp. F2) <strong>de</strong>notes the distribution function of<br />
X1 (resp. X2) and Fε1 (resp. Fε2 ) the marginal distribution function of ε1 (resp. ε2). L<strong>et</strong> Fε1,ε2<br />
<strong>de</strong>notes the joined distribution of (ε1, ε2). We st<strong>at</strong>e the following theorem:<br />
Theorem 2<br />
Assuming th<strong>at</strong><br />
H0: The variables Y , ε1 and ε2 have distribution functions with infinite support,<br />
H1: For all x ∈ [1, ∞),<br />
lim<br />
t→∞<br />
t PY (tx)<br />
¯<br />
FY (t)<br />
= f(x), (64)<br />
H2: There are real numbers t0 > 0, δ > 0 and A > 0, such th<strong>at</strong> for all t ≥ t0 and all x ≥ 1<br />
¯FY (tx)<br />
¯FY (t)<br />
23<br />
A<br />
≤ , (65)<br />
xδ
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 323<br />
H3: There is two constant (l1, l2) ∈ R+ × R+, such th<strong>at</strong><br />
F1<br />
lim<br />
u→1<br />
−1 (u)<br />
FY −1 (u) = l1,<br />
F2<br />
and lim<br />
u→1<br />
−1 (u)<br />
FY −1 (u) = l2, (66)<br />
then, the coefficient of (upper) tail <strong>de</strong>pen<strong>de</strong>nce of (X, Y ) is given by<br />
A.2.2 Proof<br />
λ =<br />
∞<br />
maxl 1<br />
β1 , l 2<br />
β2 dx f(x). (67)<br />
We first give a general expression for the probability for X to be larger than F −1<br />
X (u) knowing th<strong>at</strong><br />
(u) :<br />
Y is larger than F −1<br />
Y<br />
Lemma 3<br />
The probability th<strong>at</strong> X is larger than F −1<br />
X<br />
F −1<br />
Y (u)<br />
<br />
1 − u<br />
dx PY<br />
−1<br />
(u) knowing th<strong>at</strong> Y is larger than F (u) is given by :<br />
Pr X > F −1<br />
−1<br />
X (u)|Y > FY (u) =<br />
−1<br />
FY (u) x · ¯ <br />
Fε1,ε2 F1 −1 (u) − β1F −1<br />
Y (u) x, F2 −1 (u) − β1F −1<br />
Y (u) x . (68)<br />
Proof : The proof is the same for lemma 1. <br />
L<strong>et</strong> us now <strong>de</strong>fine the function<br />
F −1<br />
Y (u)<br />
fu(x) =<br />
1 − u PY (F −1<br />
We can st<strong>at</strong>e the following result<br />
Lemma 4<br />
Un<strong>de</strong>r assumption H1 and H3, for all x ∈ [1, ∞),<br />
almost everywhere, as u goes to 1.<br />
Y (u) x) · ¯ Fε1,ε2 [F1 −1 (u) − β1F −1<br />
Y (u) x, F2 −1 (u) − β2F −1<br />
Y<br />
fu(x) −→ 1x>maxl 1<br />
β1 , l 2<br />
Proof: Applying the assumption H3, we have<br />
and<br />
lim<br />
u→1 F1 −1 (u) − β1F −1<br />
Y<br />
lim<br />
u→1 F2 −1 (u) − β2F −1<br />
Y<br />
(u) x = lim<br />
=<br />
u→1<br />
(u) x = lim<br />
=<br />
Y<br />
(u) x] . (69)<br />
β2 · f(x), (70)<br />
β1F −1<br />
Y (u)<br />
<br />
−∞ if x > l1<br />
β1 ,<br />
∞ if x < l1<br />
β1 ,<br />
u→1<br />
β2F −1<br />
Y (u)<br />
<br />
−∞ if x > l2<br />
β2 ,<br />
∞ if x < l2<br />
β2 ,<br />
24<br />
F1 −1 (u)<br />
β1F −1 − x<br />
Y (u)<br />
F2 −1 (u)<br />
β2F −1 − x<br />
Y (u)<br />
<br />
<br />
(71)<br />
(72)<br />
(73)<br />
(74)<br />
(75)<br />
(76)
324 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
which give<br />
lim ¯Fε1,ε2<br />
u→1<br />
[F1 −1 (u) − β1F −1<br />
Y (u) x, F2 −1 (u) − β2F −1<br />
Y (u) x] = 1x>maxl 1 , β1 l2 and following the same calcul<strong>at</strong>ions as in part A.1, it conclu<strong>de</strong>s the proof. <br />
β2 , (77)<br />
We can now apply Lebesgue’s theorem of domin<strong>at</strong>ed convergence (see part A.1 for the justific<strong>at</strong>ion),<br />
which allows us to assert th<strong>at</strong><br />
<br />
<br />
lim dx fu(x) =<br />
u→1<br />
Since<br />
<br />
lim<br />
u→1<br />
the proof of theorem 2 is conclu<strong>de</strong>d. <br />
dx 1x>maxl 1 , β1 l β2 2 · f(x). (78)<br />
dx fu(x) = lim<br />
u→1 Pr X1 > F −1<br />
1 (u)|X2 > F −1<br />
2 (u) , (79)<br />
= λ, (80)<br />
25
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 325<br />
B Proofs of the corollaries<br />
B.1 First corollary<br />
Corollary 1<br />
If the random variable Y has a rapidly varying distribution function, then λ = 0.<br />
Proof : L<strong>et</strong> us write<br />
For a rapidly varying function ¯ FY , we have<br />
t PY (tx)<br />
FY<br />
¯ (t) = t PY (tx)<br />
¯<br />
FY (tx)<br />
∀x > 1, lim<br />
t→∞<br />
FY (tx)<br />
FY (t)<br />
¯<br />
·<br />
¯<br />
¯FY (tx)<br />
¯FY (t)<br />
. (81)<br />
= 0, (82)<br />
while the leftmost factor of the right-hand-si<strong>de</strong> of equ<strong>at</strong>ion (81) remains boun<strong>de</strong>d as t goes to<br />
infinity, so th<strong>at</strong><br />
t PY (tx) FY<br />
¯ (tx)<br />
lim<br />
t→∞ FY<br />
¯<br />
·<br />
(tx) FY<br />
¯<br />
= f(x) = 0 . (83)<br />
(t)<br />
Since f(x) = 0, we can apply lemma 2 without the hypothesis H3, which conclu<strong>de</strong>s the proof. <br />
B.2 Second corollary<br />
Corollary 2<br />
L<strong>et</strong> Y be regularly varying with in<strong>de</strong>x (−α), and assume th<strong>at</strong> hypothesis H3 is s<strong>at</strong>isfied. Then, the<br />
coefficient of (upper) tail <strong>de</strong>pen<strong>de</strong>nce is<br />
λ =<br />
1<br />
<br />
max<br />
1, l<br />
β<br />
where l <strong>de</strong>notes the limit, when u → 1, of the r<strong>at</strong>io FX −1 (u)/FY −1 (u).<br />
α , (84)<br />
Proof : Karam<strong>at</strong>a’s theorem (see Embrechts, Kluppelberg, and Mikosh (1997, p 567)) ensures th<strong>at</strong><br />
H1 is s<strong>at</strong>isfied with f(x) = α<br />
xα+1 , which is sufficient to prove the corollary. To go one step further,<br />
l<strong>et</strong> us <strong>de</strong>fine<br />
where L1(·) and L2(·) are slowly varying functions.<br />
¯Fy(y) = y −α · L1(y), (85)<br />
¯Fε(ε) = ε −α · L2(ε), (86)<br />
Using the proposition st<strong>at</strong>ed in Feller (1971, p 278), we obtain, for the distribution of the variable<br />
X<br />
¯FX(x) ∼ x −α<br />
<br />
β α for large x.<br />
<br />
x<br />
· L1 + L2(x) ,<br />
β<br />
(87)<br />
Assuming now, for simplicity, th<strong>at</strong> L1 (resp. L2) goes to a constant C1 (resp. C2), this implies th<strong>at</strong><br />
H3 is s<strong>at</strong>istified, since<br />
FX<br />
l = lim<br />
u→1<br />
−1 (u)<br />
FY −1 <br />
= β 1 +<br />
(u) C2<br />
βα 1<br />
α<br />
C1<br />
This allows us to obtain the equ<strong>at</strong>ions (10) and (13). <br />
26<br />
. (88)
326 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
References<br />
Bigham, N.H., C.M. Goldie, and J.L. Teugel, 1987, Regular Vari<strong>at</strong>ion. (Cambridge university press<br />
Cambridge).<br />
Bouchaus, J.P., D. Sorn<strong>et</strong>te, C. Walter and J.P. Aguilar, 1998, Taming large events: Optimal<br />
portfolio theory for strongly fluctu<strong>at</strong>ing ass<strong>et</strong>s, Intern<strong>at</strong>ional Journal of Theor<strong>et</strong>ical and Applied<br />
Finance, 1, 25-41.<br />
Boyer, B.H., M.S. Gibson, and M. Laur<strong>et</strong>an, 1997, Pitfalls in tests for changes in correl<strong>at</strong>ions,<br />
Working paper, Intern<strong>at</strong>ional Finance Discussion Paper 597, Board of the Governors of the<br />
Fe<strong>de</strong>ral Reserve System.<br />
Brennan, M.J., and E.J. Schwarz, 1978, A continuous time approach to the pricing of bonds,<br />
Journal of Banking and Finance 3, 133–155.<br />
Campbell, J.Y., M. L<strong>et</strong>tau, B. G. Malkiel, and Y. Xu, 2001, Have individual stocks become more<br />
vol<strong>at</strong>ile? An empirical explor<strong>at</strong>ion of idiosyncr<strong>at</strong>ic risk, Journal of Finance 56, 1–43.<br />
Carey, M., 1998, Credit risk in priv<strong>at</strong>e <strong>de</strong>bt portfolio, Journal of Finance 53, 56–61.<br />
Cho, Y.-H., and R.F. Engle, 2000, Time-varying b<strong>et</strong>as and asymm<strong>et</strong>ric effects of news : empirical<br />
analysis of blue chip stocks, Working paper, University of California, San Diego.<br />
Coles, S., J. Heffernan, and J. Tawn, 1999, Depen<strong>de</strong>nce measures for extreme value analyses,<br />
Extremes 2, 339–365.<br />
Cox, J.C, J.E. Ingersoll, and S.A. Ross, 1985, A theory of the term structure of interest r<strong>at</strong>es,<br />
Econom<strong>et</strong>rica 51, 385–408.<br />
Danielsson, J., L. <strong>de</strong> Haan, L. Peng, and C.G. <strong>de</strong> Vries, 2001, Using a bootstrap m<strong>et</strong>hod to choose<br />
the optimal sample fraction in tail in<strong>de</strong>x estim<strong>at</strong>ion, Journal of Multivari<strong>at</strong>e analysis 76, 226–248.<br />
Danielsson, J., and C.G. <strong>de</strong> Vries, 1997, Tail in<strong>de</strong>x and quantile estim<strong>at</strong>ion with very high frequency<br />
d<strong>at</strong>a, Journal of Empirical Finance 4, 241–257.<br />
<strong>de</strong> Haan, L., S.I. Resnick, H. Rootzen, and C.G. <strong>de</strong> Vries, 1989, Extremal behaviour of solutions to<br />
a stochastic difference equ<strong>at</strong>ion with applic<strong>at</strong>ion to ARCH processes, Stochastic Processes and<br />
their Applic<strong>at</strong>ions 32, 213–224.<br />
Embrechts, P., C.P. Kluppelberg, and T. Mikosh, 1997, Mo<strong>de</strong>lling Extremal Events. (Springer-<br />
Verlag Berlin).<br />
Embrechts, P., A.J. McNeil, and D. Straumann, 1999, Correl<strong>at</strong>ion: Pitfalls and Altern<strong>at</strong>ives, Risk<br />
pp. 69–71.<br />
Embrechts, P., A.J. McNeil, and D. Straumann, 2001, Correl<strong>at</strong>ion and Depen<strong>de</strong>ncy in Risk Management:<br />
Properties and Pitfalls, in M. Dempster, eds.: Value <strong>at</strong> Risk and Beyond (Cambridge<br />
University Press, ).<br />
Engle, R.F., and S. Manganelli, 1999, CAViaR: Conditional autoregressive Value-<strong>at</strong>-Risk by regression<br />
quantiles, Working paper, University of California, San Diego.<br />
Fama, E., and J. Mc B<strong>et</strong>h, 1973, Risk, r<strong>et</strong>urn and equilibrium: empirical tests, Journal of Political<br />
Economy 81, 607–636.<br />
27
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 327<br />
Feller, W., 1971, An introduction to probability theory and its applic<strong>at</strong>ions II. (Wiley New-York).<br />
Frees, E., and E. Val<strong>de</strong>z, 1998, Un<strong>de</strong>rstanding rel<strong>at</strong>ionships using copulas, North Americam Actuarial<br />
Journal 2, 1–25.<br />
Gopikrishnan, P., M. Meyer, L.A.N. Amaral, and H.E. Stanley, 1998, Inverse Cubic Law for the<br />
Distribution of Stock Price Vari<strong>at</strong>ions, European Physical Journal B 3, 139 –140.<br />
Gorby, M.B., 2000, A compar<strong>at</strong>ive an<strong>at</strong>omy of credit risk mo<strong>de</strong>ls, Journal of Banking and Finance<br />
24, 119–149.<br />
Hall, P., 1990, Using the bootstrap m<strong>et</strong>hod to estim<strong>at</strong>e mean squared error and select smoothing<br />
param<strong>et</strong>er in non param<strong>et</strong>ric problems, Journal of Multivari<strong>at</strong>e Analysis 32, 177–203.<br />
Joe, H., 1997, Multivari<strong>at</strong>e mo<strong>de</strong>ls and <strong>de</strong>pen<strong>de</strong>nce concepts. (Chapman & Hall London).<br />
Johansen, A., and D. Sorn<strong>et</strong>te, 2002, Large stock mark<strong>et</strong> price drawdowns are outliers, Journal of<br />
Risk 4, 69–110.<br />
Jones, C.P., and J.W.Wilson, 1989, Is stock price vol<strong>at</strong>ility increasing?, Financial Analysts Journal<br />
45, 20–26.<br />
Kan<strong>de</strong>l, S., and R. Staumbaugh, 1987, On correl<strong>at</strong>ions and the sensitivity of inference about meanvariance<br />
efficiency, Journal of Financial Economics 18, 61–90.<br />
Kearns, P., and A.R. Pagan, 1997, Estim<strong>at</strong>ing the <strong>de</strong>nsity tail in<strong>de</strong>x for financial time series, Review<br />
of Economics and St<strong>at</strong>istics 79, 171–175.<br />
Lindskog, F., 2000, Mo<strong>de</strong>lling Depen<strong>de</strong>nce with Copulas, Working paper, RiskLab, http :<br />
//www.risklab.ch/P apers.html#MT Lindskog.<br />
Lintner, J., 1965, The valu<strong>at</strong>ion of risk ass<strong>et</strong>s and the selection of risky investments in stock<br />
portfolios and the capital bug<strong>et</strong>s, Review of Economics and St<strong>at</strong>itics 41, 13–37.<br />
Longin, F.M., 1996, The asymptotic distribution of extrem stock mark<strong>et</strong> r<strong>et</strong>urns, Journal of Business<br />
96, 383–408.<br />
Longin, F.M., and B. Solnik, 2001, Extreme Correl<strong>at</strong>ion of Intern<strong>at</strong>ional Equity Mark<strong>et</strong>s, Journal<br />
of Finance 56, 649–676.<br />
Lor<strong>et</strong>an, M., and P.C.B. Phillips, 1994, Testing the covariance st<strong>at</strong>ionarity of heavy-tailed times<br />
series, Journal of Empirical Finance 1, 211–248.<br />
Lucas, A., P. Klaassen, P. Spreij, and S. Stra<strong>et</strong>mans, 2001, An analytic approach to credit risk of<br />
large corpor<strong>at</strong>e bond and loan portfolios, Journal of Banking and Finance 25, 1635–1664.<br />
Lux, T., 1996, The stable Par<strong>et</strong>ian hypothesis and the frequency of large r<strong>et</strong>urns: an examin<strong>at</strong>ion<br />
of major German stocks, Applied Financial Economics 6, 463–475.<br />
Malevergne, Y., and D. Sorn<strong>et</strong>te, 2001, Testing the Gaussian copula hypothesis for financial ass<strong>et</strong>s<br />
<strong>de</strong>pen<strong>de</strong>nces, Working paper, (e-print available <strong>at</strong> http://papers.ssrn.com/abstract=291140).<br />
Malevergne, Y., and D. Sorn<strong>et</strong>te, 2002, Hedging Extreme Co-Movements, submitted to RISK,<br />
preprint <strong>at</strong> http://arXiv.org/abs/cond-m<strong>at</strong>/0205636.<br />
Mossin, J., 1966, Equilibrium in capital ass<strong>et</strong> mark<strong>et</strong>, Econom<strong>et</strong>rica 35, 768–783.<br />
28
328 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Nelsen, R.B., 1998, An introduction to copulas. (Springer Verlag New York).<br />
Pagan, A., 1996, The econom<strong>et</strong>rics of financial mark<strong>et</strong>s, Journal of Empirical Finance 3, 15 – 102.<br />
Pagan, A., and A. Ullah, 1999, Non param<strong>et</strong>rics econom<strong>et</strong>rics. (Cambridge University Press Cambridge).<br />
P<strong>at</strong>ton, J.A., 2001, Estim<strong>at</strong>ion of copula mo<strong>de</strong>ls for time series of possibly different lengths, Working<br />
paper, University of California, Econ. Disc. Paper No. 2001-17.<br />
Roll, R., 1988, The Intern<strong>at</strong>ional Crash of October 1987, Financial Analysts Journal pp. 19–35.<br />
Rootzén, H., M.R. Leadb<strong>et</strong>ter, and L. <strong>de</strong> Haan, 1998, On the distribution tail array sums for<br />
strongly mixing st<strong>at</strong>ionary sequences, Annals of Applied Probability 8, 868–885.<br />
Ross, S.A., 1976, The arbitrage theory of capital ass<strong>et</strong> pricing, Journal of Economic Theory 17,<br />
254–286.<br />
Sharpe, W., 1964, Capital ass<strong>et</strong>s prices: a theory of mark<strong>et</strong> equilibrium un<strong>de</strong>r conditions of risk,<br />
Journal of Finance 19, 425–442.<br />
Shealter, S.J., and J.S. Marron, 1990, Kernel quantile estim<strong>at</strong>ors, Journal of the Americam St<strong>at</strong>istical<br />
Associ<strong>at</strong>ion 85, 410–415.<br />
Sklar, A., 1959, Fonction <strong>de</strong> répartition à n dimensions <strong>et</strong> leurs marges, Publ. Inst. St<strong>at</strong>ist. Univ.<br />
Paris 8, 229–231.<br />
Starica, C., 1999, Multivari<strong>at</strong>e extremes for mo<strong>de</strong>ls with constant conditional correl<strong>at</strong>ions, Journal<br />
of Empirical Finance 6, 515–553.<br />
Vasicek, O., 1977, An equilibrium characteris<strong>at</strong>ion of the term structure of interest r<strong>at</strong>es, Journal<br />
of Financial Economics 5, 177–188.<br />
Xu, Y., and B.G. Malkiel, 2002, Investig<strong>at</strong>ing the behavior of idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility, forthcoming<br />
Journal of Business.<br />
29
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 329<br />
July 1962 - December 1979 January 1980 - December 2000 July 1962 - December 2000<br />
Mean Std. Skew. Kurt. Mean Std. Skew. Kurt. Mean Std. Skew. Kurt.<br />
Abbott Labs 0.6677 0.0154 0.2235 2.192 0.9217 0.0174 -0.0434 2.248 0.8066 0.0165 0.0570 2.300<br />
American Home Products Corp. 0.4755 0.0136 0.2985 3.632 0.8486 0.0166 0.1007 8.519 0.6803 0.0154 0.1717 7.557<br />
Boeing Co. 0.8460 0.0228 0.6753 4.629 0.7752 0.0193 0.1311 4.785 0.8068 0.0209 0.4495 4.901<br />
Bristol-Myers Squibb Co. 0.5342 0.0152 -0.0811 2.808 0.9353 0.0175 -0.3437 16.733 0.7546 0.0165 -0.2485 12.573<br />
Chevron Corp. 0.4916 0.0134 0.2144 2.442 0.6693 0.0169 0.0491 4.355 0.5885 0.0154 0.1033 4.209<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0.2193 0.0126 0.3493 2.754 0.6792 0.0172 -0.1021 4.731 0.4715 0.0153 0.0231 4.937<br />
Disney (Walt) Co. 0.9272 0.0215 0.2420 2.762 0.8759 0.0195 -0.6661 17.655 0.8997 0.0204 -0.1881 9.568<br />
General Motors Corp. 0.3547 0.0126 0.4138 4.302 0.5338 0.0183 -0.0128 5.373 0.4538 0.0160 0.0872 6.164<br />
Hewl<strong>et</strong>t-Packard Co. 0.7823 0.0199 0.0212 3.063 0.8913 0.0238 0.0254 4.921 0.8420 0.0221 0.0256 4.624<br />
Coca-Cola Co. 0.4829 0.0138 0.0342 5.436 0.9674 0.0170 -0.1012 14.377 0.7483 0.0157 -0.0513 12.611<br />
Minnesota Mining & MFG Co. 0.3459 0.0139 0.3016 2.997 0.6885 0.0150 -0.7861 20.609 0.5333 0.0145 -0.3550 14.066<br />
Philip Morris Cos Inc. 0.7930 0.0153 0.2751 2.799 0.9664 0.0180 -0.2602 10.954 0.8863 0.0169 -0.0784 8.790<br />
Pepsico Inc. 0.4982 0.0147 0.2380 2.867 0.9443 0.0180 0.1372 4.594 0.7431 0.0166 0.1786 4.413<br />
Procter & Gamble Co. 0.3569 0.0115 0.3911 4.343 0.7916 0.0164 -1.6610 46.916 0.5947 0.0144 -1.2408 44.363<br />
Pharmacia Corp. 0.3801 0.0145 0.2699 3.508 0.9027 0.0191 -0.6133 13.587 0.6666 0.0172 -0.3773 12.378<br />
Schering-Plough Corp. 0.6328 0.0163 0.2619 3.112 1.0663 0.0192 0.1781 7.9979 0.8703 0.0179 0.2139 6.757<br />
Texaco Inc. 0.3416 0.0134 0.2656 2.596 0.6644 0.0166 0.1192 6.477 0.5197 0.0152 0.1725 5.829<br />
Texas Instruments Inc. 0.6839 0.0198 0.2076 3.174 1.0299 0.0268 0.1595 7.848 0.8726 0.0239 0.1831 7.737<br />
United Technologies Corp 0.5801 0.0185 0.3397 2.826 0.7752 0.0170 0.0396 3.190 0.6876 0.0177 0.1933 3.034<br />
Walgreen Co. 0.5851 0.0165 0.3530 3.030 1.1996 0.0185 0.1412 3.316 0.9217 0.0176 0.2260 3.295<br />
30<br />
Standart & Poor’s 500 0.1783 0.0075 0.2554 3.131 0.5237 0.0101 -1.6974 36.657 0.3674 0.0090 -1.2236 32.406<br />
Table 1: This table gives the main st<strong>at</strong>istical fe<strong>at</strong>ures of the three samples we have consi<strong>de</strong>red. The columns Mean, Std., Skew. and Kurt.<br />
respectively give the average r<strong>et</strong>urn multiplied by one thousand, the standard <strong>de</strong>vi<strong>at</strong>ion, the skewness and the excess kurtosis of each ass<strong>et</strong><br />
over the time intervals form July 1962 to December 1979, January 1980 to Decemeber 2000 and July 1962 to December 2000. The excess<br />
kurtosis is given as indic<strong>at</strong>ive of the rel<strong>at</strong>ive weight of large r<strong>et</strong>urn amplitu<strong>de</strong>s, and can always be calcul<strong>at</strong>ed over a finite time series even if<br />
it may not be asymptotically <strong>de</strong>fined for power tails with exponents less than 4.
330 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
July 1962 - December 1979 January 1979 - December 2000 July 1962 - December 2000<br />
β ρY 2 ,ε2 β ρY 2 ,ε2 β ρY 2 ,ε2 Abbott Labs 0.8994 0.0879 0.9122 0.1879 0.9081 0.1597<br />
American Home Products Corp. 0.9855 0.1253 0.8102 0.0587 0.8652 0.0736<br />
Boeing Co. 1.4416 0.1196 0.9036 0.0928 1.0715 0.1279<br />
Bristol-Myers Squibb Co. 1.0832 0.1056 1.0435 0.0457 1.0559 0.0481<br />
Chevron Corp. 1.0062 0.1191 0.8333 0.0776 0.8873 0.0906<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 1.0818 0.0960 0.9451 0.0433 0.9880 0.0595<br />
Disney (Walt) Co. 1.5530 0.0960 1.0016 0.1304 1.1736 0.0641<br />
General Motors Corp. 1.0945 0.1531 1.0112 0.0400 1.0371 0.0563<br />
Hewl<strong>et</strong>t-Packard Co. 1.3910 0.1023 1.3074 0.0739 1.3332 0.0832<br />
Coca-Cola Co. 1.0347 0.2146 0.9833 0.1254 0.9995 0.1238<br />
Minnesota Mining & MFG Co. 1.1339 0.1203 0.8756 0.2605 0.9564 0.1706<br />
Philip Morris Cos Inc. 1.0894 0.0723 0.8598 0.0340 0.9314 0.0545<br />
Pepsico Inc. 0.9587 0.1233 0.9004 0.3294 0.9187 0.3169<br />
Procter & Gamble Co. 0.8293 0.1873 0.8938 0.1188 0.8738 0.1287<br />
Pharmacia Corp. 1.0750 0.0783 0.8824 0.0373 0.9429 0.0357<br />
Schering-Plough Corp. 1.1244 0.1284 1.0480 0.0494 1.0720 0.0540<br />
Texaco Inc. 1.4578 0.1410 1.3811 0.0674 1.4049 0.0766<br />
Texas Instruments Inc. 0.9414 0.1354 0.6600 0.0823 0.7481 0.1053<br />
United Technologies Corp 1.1336 0.1243 0.9049 0.1175 0.9763 0.1098<br />
Walgreen Co. 0.6354 0.1052 0.8554 0.1087 0.7869 0.0798<br />
31<br />
Table 2: This table presents the estim<strong>at</strong>ed coefficient β for the factor mo<strong>de</strong>l (6) and the correl<strong>at</strong>ion coefficient ρY 2 ,ε2 b<strong>et</strong>ween the square of<br />
factor and the square of the estim<strong>at</strong>ed idiosyncr<strong>at</strong>ic noise, for the different time intervals we have consi<strong>de</strong>red. A Fisher’s test shows th<strong>at</strong><br />
these correl<strong>at</strong>ion coefficients are all significantly different from zero.
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 331<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
q = 1% q = 2.5% q = 5% q = 1% q = 2.5% q = 5%<br />
Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε<br />
Abbott Labs 5.54 5.31 3.94 4.02 3.27 3.31 5.10 4.50 4.09 3.71 3.53∗ 3.14<br />
American Home Products Corp. 4.58 5.11 3.89 3.81 3.02∗ 3.21∗ 3.64 4.66 3.60 3.81 3.11 3.15<br />
Boeing Co. 6.07 4.90 4.57 3.74 3.32 3.49 4.04 4.27 3.95 4.19 3.35∗ 2.93<br />
Bristol-Myers Squibb Co. 4.32 4.27 3.31 3.95 2.99∗ 3.16∗ 5.96∗ 5.19 3.94 4.82∗ 3.62∗ 4.03∗ Chevron Corp. 5.24 4.78 3.75 3.29 2.91 3.12∗ 5.21 5.15 3.90 4.26 3.25∗ 3.07<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 5.26 4.36 3.69 3.76 3.17∗ 3.23∗ 5.35 5.15 4.00 3.37 3.13 3.04<br />
Disney (Walt) Co. 3.59 4.23 3.59 3.84 3.08∗ 3.22∗ 4.90 4.34 4.26 3.73 3.33∗ 3.29∗ General Motors Corp. 4.82 3.50 3.72 3.66 2.94∗ 3.36 3.91 4.78 3.64 3.86 2.94 3.07<br />
Hewl<strong>et</strong>t-Packard Co. 3.76 3.89 3.12∗ 3.05∗ 2.81∗ 3.00∗ 4.64 5.08 4.08 4.20 3.41∗ 3.42∗ Coca-Cola Co. 3.45 3.45 3.05∗ 3.71 2.75∗ 3.17∗ 3.91 4.26 3.16 3.61 2.81 3.16<br />
Minnesota Mining & MFG Co. 5.16 4.86 4.06 4.35 3.43 3.71 4.35 4.47 3.96 3.31 3.14 3.06<br />
Philip Morris Cos Inc. 4.63 3.79 3.82 3.90 3.38 3.48 4.10 4.64 3.59 3.85 3.03 3.06<br />
Pepsico Inc. 4.89 5.35 3.93 4.49 3.02∗ 3.27 4.07 4.67 3.49 3.86 3.15 3.21∗ Procter & Gamble Co. 4.42 3.77 3.77 3.74 3.13∗ 3.42 4.14 5.39 3.59 3.73 2.97 3.40∗ Pharmacia Corp. 4.73 4.24 4.05 3.45 2.88∗ 3.34 4.46 3.72 3.95 3.90 3.14 2.99<br />
Schering-Plough Corp. 4.59 4.70 4.20 3.87 3.37 3.33 4.60 5.88∗ 3.50 3.91 3.07 3.22∗ Texaco Inc. 5.34 4.59 3.99 3.84 3.07∗ 3.19∗ 3.83 4.10 3.94 3.67 3.14 2.98<br />
Texas Instruments Inc. 4.08 4.54 3.36 3.13∗ 3.22∗ 2.87∗ 4.52 4.20 3.67 3.79 3.16 3.07<br />
United Technologies Corp 4.00 4.49 3.52 3.92 3.27 3.46 4.78 4.97 3.73 3.98 3.26∗ 3.49∗ Walgreen Co. 4.63 6.50 3.85 4.26 2.94∗ 3.18∗ 5.16 4.56 3.47 3.30 3.15 2.82<br />
32<br />
Standart & Poor’s 500 5.17 - 4.16 - 3.91 - 3.74 - 3.34 - 2.64 -<br />
Table 3: This table gives the estim<strong>at</strong>ed value of the tail in<strong>de</strong>x for the twenty consi<strong>de</strong>red ass<strong>et</strong>s, the Standard & Poor’s 500 in<strong>de</strong>x and the<br />
residues ε obtained by regressing each ass<strong>et</strong> on the Standard & Poor’s 500 in<strong>de</strong>x, for both the neg<strong>at</strong>ive and the positive tails, during the<br />
time interval from July 1962 to December 1979. The tail in<strong>de</strong>xes are estimed by Hill’s estim<strong>at</strong>or <strong>at</strong> the quantile 1%, 2.5% and 5% which are<br />
the optimal quantiles given by the Hall (1990) and Danielsson and <strong>de</strong> Vries (1997)’s algorithms. The values <strong>de</strong>corr<strong>at</strong>ed with stars represent<br />
the tail in<strong>de</strong>xes which cannot be consi<strong>de</strong>red equal to the Standard & Poor’s 500 in<strong>de</strong>x’s tail in<strong>de</strong>x <strong>at</strong> the 95% confi<strong>de</strong>nce level.
332 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
q = 1% q = 2.5% q = 5% q = 1% q = 2.5% q = 5%<br />
Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε Ass<strong>et</strong> ε<br />
Abbott Labs 3.59 3.60 3.35 3.62 3.22 3.39 5.14 4.60 4.16 3.76 3.77 3.07<br />
American Home Products Corp. 3.03 3.07 3.11 2.78 2.73 2.49∗ 4.01 3.47 3.28 3.02 2.87 2.79<br />
Boeing Co. 3.39 3.97 3.23 3.53 3.02 3.21 4.86 3.65 3.45 3.16 3.13 3.23<br />
Bristol-Myers Squibb Co. 3.21 3.15 2.90 3.41 2.80 3.16 2.98 3.74 3.35 3.12 3.20 2.75<br />
Chevron Corp. 4.13 4.48 3.99 3.91 3.30 3.45 5.16 4.53 3.88 3.81 3.01 3.06<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 3.99 3.49 3.76 3.23 3.02 3.04 5.36 4.33 4.31 3.35 3.44 2.76<br />
Disney (Walt) Co. 2.83 3.24 2.76 2.97 2.85 2.83 3.97 3.70 3.68 3.33 3.15 2.87<br />
General Motors Corp. 4.44 4.79 3.88 4.27∗ 3.44 3.56 5.76 5.32 4.45 3.86 3.43 3.22<br />
Hewl<strong>et</strong>t-Packard Co. 3.73 3.45 3.52 3.12 3.00 2.73 4.31 3.40 3.47 3.29 3.24 2.99<br />
Coca-Cola Co. 3.01 3.76 3.14 3.48 2.99 2.86 4.06 3.47 3.45 3.16 3.37 2.87<br />
Minnesota Mining & MFG Co. 3.52 3.38 3.21 3.39 2.88 3.04 3.76 3.46 3.95 3.22 3.10 2.76<br />
Philip Morris Cos Inc. 3.58 3.34 3.33 3.12 2.68 2.53∗ 3.42 3.16 3.70 3.07 2.85 2.81<br />
Pepsico Inc. 4.14 4.46 3.39 3.60 2.99 3.27 4.00 3.87 3.61 3.34 3.44 3.31<br />
Procter & Gamble Co. 2.65 2.46 3.29 3.19 3.19 2.87 4.35 3.90 3.48 3.20 3.14 2.91<br />
Pharmacia Corp. 2.96 3.20 3.09 2.79 2.80 2.70 4.12 4.70 3.44 3.50 3.31 2.89<br />
Schering-Plough Corp. 4.22 5.20∗ 3.29 3.68 3.11 3.05 3.23 3.51 3.45 3.08 3.06 2.87<br />
Texaco Inc. 3.09 3.20 3.10 3.15 2.88 2.84 3.65 3.36 3.20 3.04 2.86 2.70<br />
Texas Instruments Inc. 3.49 3.53 3.35 3.31 2.89 2.99 4.00 3.42 3.36 3.30 2.97 3.06<br />
United Technologies Corp 4.21 3.98 3.82 3.46 3.34 3.18 5.39 4.50 4.00 3.80 3.51 3.26<br />
Walgreen Co. 4.06 4.35 3.81 4.04 3.20 3.40 4.60 5.12 3.79 3.54 3.20 3.07<br />
33<br />
Standart & Poor’s 500 3.16 - 3.17 - 3.16 - 4.00 - 3.65 - 3.19 -<br />
Table 4: This table gives the estim<strong>at</strong>ed value of the tail in<strong>de</strong>x for the twenty consi<strong>de</strong>red ass<strong>et</strong>s, the Standard & Poor’s 500 in<strong>de</strong>x and the<br />
residues ε obtained by regressing each ass<strong>et</strong> on the Standard & Poor’s 500 in<strong>de</strong>x, for both the neg<strong>at</strong>ive and the positive tails, during the time<br />
interval from January 1980 to December 2000. The tail in<strong>de</strong>xes are estimed by Hill’s estim<strong>at</strong>or <strong>at</strong> the quantile 1%, 2.5% and 5% which are<br />
the optimal quantiles given by the Hall (1990) and Danielsson and <strong>de</strong> Vries (1997)’s algorithms. The values <strong>de</strong>corr<strong>at</strong>ed with stars represent<br />
the tail in<strong>de</strong>xes whose value cannot be consi<strong>de</strong>red equal to the Standard & Poor’s 500 in<strong>de</strong>x’s tail in<strong>de</strong>x <strong>at</strong> the 95% confi<strong>de</strong>nce level.
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 333<br />
First Centile First Quintile First Decile<br />
mean std. min. max. mean std. min. max. mean std. min. max.<br />
Abbott Labs 0.1264 0.0106 0.1039 0.1745 0.1232 0.0058 0.1039 0.1745 0.1185 0.0069 0.1039 0.1745<br />
American Home Products Corp. 0.1181 0.0091 0.0796 0.1288 0.1349 0.0111 0.0796 0.1494 0.1401 0.0096 0.0796 0.1505<br />
Boeing Co. 0.1116 0.0128 0.0954 0.1653 0.1090 0.0074 0.0954 0.1653 0.1066 0.0066 0.0954 0.1653<br />
Bristol-Myers Squibb Co. 0.1927 0.0171 0.1407 0.2175 0.2220 0.0220 0.1407 0.2474 0.2180 0.0180 0.1407 0.2474<br />
Chevron Corp. 0.1566 0.0194 0.1334 0.2368 0.1407 0.0121 0.1265 0.2368 0.1365 0.0101 0.1257 0.2368<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0.2089 0.0142 0.1785 0.2424 0.2067 0.0105 0.1785 0.2424 0.2021 0.0106 0.1785 0.2424<br />
Disney (Walt) Co. 0.1317 0.0160 0.0930 0.1613 0.1587 0.0161 0.0930 0.1754 0.1566 0.0123 0.0930 0.1754<br />
General Motors Corp. 0.2210 0.0149 0.2045 0.2996 0.2109 0.0096 0.1947 0.2996 0.2020 0.0120 0.1808 0.2996<br />
Hewl<strong>et</strong>t-Packard Co. 0.1455 0.0103 0.1188 0.1769 0.1615 0.0116 0.1188 0.1776 0.1603 0.0095 0.1188 0.1776<br />
Coca-Cola Co. 0.1870 0.0259 0.1199 0.2522 0.2159 0.0204 0.1199 0.2522 0.2160 0.0164 0.1199 0.2522<br />
Minnesota Mining & MGF Co. 0.2311 0.0254 0.1851 0.3268 0.2262 0.0126 0.1851 0.3268 0.2218 0.0115 0.1851 0.3268<br />
Philip Morris Cos Inc. 0.1251 0.0078 0.1050 0.1526 0.1334 0.0089 0.1050 0.1526 0.1340 0.0070 0.1050 0.1526<br />
Pepsico Inc. 0.1263 0.0111 0.1014 0.1706 0.1242 0.0071 0.1014 0.1706 0.1239 0.0057 0.1014 0.1706<br />
Procter & Gamble Co. 0.1980 0.0325 0.1131 0.2351 0.2017 0.0153 0.1131 0.2351 0.1972 0.0131 0.1131 0.2351<br />
Pharmacia Corp. 0.1180 0.0172 0.0596 0.1378 0.1291 0.0109 0.0596 0.1432 0.1337 0.0091 0.0596 0.1434<br />
Schering-Plough Corp. 0.1759 0.0143 0.1320 0.2403 0.1733 0.0078 0.1320 0.2403 0.1663 0.0102 0.1320 0.2403<br />
Texaco Inc. 0.0214 0.0017 0.0170 0.0269 0.0246 0.0023 0.0170 0.0278 0.0248 0.0018 0.0170 0.0278<br />
Texas Instruments Inc. 0.5657 0.0687 0.3897 0.6572 0.6162 0.0438 0.3897 0.6623 0.6168 0.0368 0.3897 0.6761<br />
United Technologies Corp 0.1300 0.0123 0.0990 0.1872 0.1254 0.0073 0.0990 0.1872 0.1192 0.0086 0.0990 0.1872<br />
Walgreen Co. 0.0739 0.0109 0.0664 0.1185 0.0682 0.0058 0.0629 0.1185 0.0674 0.0044 0.0629 0.1185<br />
34<br />
Table 5: This table gives the average (mean), the standard <strong>de</strong>vi<strong>at</strong>ion (std.), the minimum (min.) and the maximum (max.) values of the<br />
coefficient of lower tail <strong>de</strong>pen<strong>de</strong>nce estim<strong>at</strong>ed over the first centile, quintile and <strong>de</strong>cile during the entire time interval from July 1962 to<br />
December 2000, un<strong>de</strong>r the assumption th<strong>at</strong> the tail of the distributions of the ass<strong>et</strong>s and the mark<strong>et</strong> are regularly varying with a in<strong>de</strong>x equal<br />
to three.
334 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
First Centile First Quintile First Decile<br />
mean std. min. max. mean std. min. max. mean std. min. max.<br />
Abbott Labs 0.1670 0.0127 0.1442 0.2137 0.1633 0.0071 0.1442 0.2137 0.1540 0.0120 0.1331 0.2137<br />
American Home Products Corp. 0.1423 0.0207 0.0910 0.1720 0.1728 0.0205 0.091 0.1963 0.1823 0.0175 0.0910 0.2020<br />
Boeing Co. 0.1372 0.0127 0.1101 0.1804 0.1349 0.0064 0.1101 0.1804 0.1289 0.0078 0.1101 0.1804<br />
Bristol-Myers Squibb Co. 0.2720 0.0231 0.1878 0.3052 0.2751 0.0115 0.1878 0.3052 0.2696 0.0110 0.1878 0.3052<br />
Chevron Corp. 0.1853 0.0188 0.1656 0.2564 0.1790 0.0105 0.1634 0.2564 0.1748 0.0096 0.1606 0.2564<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0.2547 0.0148 0.2127 0.2871 0.2695 0.0117 0.2127 0.2876 0.2685 0.0103 0.2127 0.2876<br />
Disney (Walt) Co. 0.1772 0.0149 0.1368 0.1957 0.1938 0.0123 0.1368 0.2094 0.1900 0.0109 0.1368 0.2094<br />
General Motors Corp. 0.2641 0.0259 0.2393 0.3652 0.2565 0.0138 0.2349 0.3652 0.2545 0.0108 0.2349 0.3652<br />
Hewl<strong>et</strong>t-Packard Co. 0.1701 0.0096 0.1389 0.1914 0.2018 0.0230 0.1389 0.2303 0.2039 0.0176 0.1389 0.2303<br />
Coca-Cola Co. 0.2343 0.0223 0.1686 0.2719 0.2576 0.0163 0.1686 0.2731 0.2579 0.0123 0.1686 0.2731<br />
Minnesota Mining & MFG Co. 0.2844 0.0196 0.2399 0.3407 0.2873 0.0099 0.2399 0.3407 0.2802 0.0117 0.2399 0.3407<br />
Philip Morris Cos Inc. 0.1369 0.0168 0.0983 0.1673 0.1700 0.0206 0.0983 0.1919 0.1729 0.0155 0.0983 0.1919<br />
Pepsico Inc. 0.1634 0.0132 0.1483 0.2106 0.1535 0.0083 0.1448 0.2106 0.1512 0.0067 0.1434 0.2106<br />
Procter & Gamble Co. 0.2284 0.0292 0.1434 0.2673 0.2461 0.0169 0.1434 0.2673 0.2413 0.0141 0.1434 0.2673<br />
Pharmacia Corp. 0.1279 0.0104 0.0863 0.1432 0.1588 0.0192 0.0863 0.1822 0.1643 0.0149 0.0863 0.1822<br />
Schering-Plough Corp. 0.2195 0.0190 0.1920 0.2863 0.2179 0.0103 0.1920 0.2863 0.2107 0.0123 0.1877 0.2863<br />
Texaco Inc. 0.0327 0.0027 0.0243 0.0369 0.0369 0.0033 0.0243 0.0414 0.0371 0.0027 0.0243 0.0414<br />
Texas Instruments Inc. 0.4355 0.0195 0.3389 0.4906 0.4500 0.0142 0.3389 0.4906 0.4515 0.011 0.3389 0.4906<br />
United Technologies Corp 0.1570 0.0153 0.1298 0.2182 0.1562 0.0075 0.1298 0.2182 0.1511 0.0084 0.1298 0.2182<br />
Walgreen Co. 0.0937 0.0112 0.0808 0.1384 0.0837 0.0071 0.0776 0.1384 0.0786 0.0078 0.0669 0.1384<br />
35<br />
Table 6: This table gives the average (mean), the standard <strong>de</strong>vi<strong>at</strong>ion (std.), the minimum (min.) and the maximum (max.) values of the<br />
coefficient of lower tail <strong>de</strong>pen<strong>de</strong>nce estim<strong>at</strong>ed over the first centile, quintile and <strong>de</strong>cile during the entire time interval from July 1962 to<br />
December 2000, un<strong>de</strong>r the assumption th<strong>at</strong> the tail of the distributions of the ass<strong>et</strong>s and the mark<strong>et</strong> are power laws with an exponent equal<br />
to three.
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 335<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
α = 3 α = 3.5 α = 4 α = 3 α = 3.5 α = 4<br />
Abbott Labs 0.12 0.09 0.06 0.11 0.08 0.06<br />
American Home Products Corp. 0.22 0.18 0.15 0.25 0.22 0.19<br />
Boeing Co. 0.16 0.13 0.10 0.13 0.10 0.07<br />
Bristol-Myers Squibb Co. 0.22 0.19 0.16 0.28 0.25 0.23<br />
Chevron Corp. 0.21 0.17 0.14 0.26 0.23 0.20<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0.38 0.37 0.35 0.37 0.35 0.33<br />
Disney (Walt) Co. 0.24 0.20 0.17 0.23 0.19 0.16<br />
General Motors Corp. 0.39 0.37 0.35 0.48 0.47 0.47<br />
Hewl<strong>et</strong>t-Packard Co. 0.15 0.12 0.09 0.23 0.20 0.17<br />
Coca-Cola Co. 0.26 0.22 0.19 0.26 0.23 0.20<br />
Minnesota Mining & MFG Co. 0.35 0.32 0.30 0.35 0.33 0.31<br />
Philip Morris Cos Inc. 0.25 0.22 0.19 0.20 0.17 0.14<br />
Pepsico Inc. 0.15 0.12 0.09 0.17 0.14 0.11<br />
Procter & Gamble Co. 0.23 0.19 0.16 0.24 0.21 0.18<br />
Pharmacia Corp. 0.23 0.19 0.16 0.26 0.23 0.20<br />
Schering-Plough Corp. 0.21 0.18 0.15 0.20 0.17 0.14<br />
Texaco Inc. 0.06 0.04 0.03 0.07 0.05 0.03<br />
Texas Instruments Inc. 0.47 0.46 0.46 0.49 0.49 0.49<br />
United Technologies Corp 0.13 0.10 0.07 0.13 0.10 0.07<br />
Walgreen Co. 0.03 0.02 0.01 0.02 0.01 0.01<br />
Table 7: This table summarizes the mean values over the first centile of the distribution of the<br />
coefficients of (upper or lower) tail <strong>de</strong>pen<strong>de</strong>nce for the positive and neg<strong>at</strong>ive tails during the time<br />
interval from July 1962 to December 1979, for three values of the tail in<strong>de</strong>x α = 3, 3.5, 4.<br />
36
336 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
α = 3 α = 3.5 α = 4 α = 3 α = 3.5 α = 4<br />
Abbott Labs 0.20 0.17 0.14 0.16 0.13 0.10<br />
American Home Products Corp. 0.12 0.09 0.06 0.10 0.08 0.05<br />
Boeing Co. 0.14 0.11 0.08 0.10 0.07 0.05<br />
Bristol-Myers Squibb Co. 0.32 0.29 0.26 0.25 0.21 0.19<br />
Chevron Corp. 0.18 0.14 0.11 0.13 0.09 0.07<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0.23 0.20 0.17 0.16 0.13 0.10<br />
Disney (Walt) Co. 0.16 0.13 0.10 0.15 0.12 0.09<br />
General Motors Corp. 0.26 0.22 0.19 0.20 0.16 0.13<br />
Hewl<strong>et</strong>t-Packard Co. 0.19 0.15 0.13 0.21 0.18 0.15<br />
Coca-Cola Co. 0.24 0.20 0.18 0.20 0.17 0.14<br />
Minnesota Mining & MFG Co. 0.26 0.23 0.20 0.20 0.17 0.14<br />
Philip Morris Cos Inc. 0.11 0.08 0.06 0.11 0.08 0.06<br />
Pepsico Inc. 0.17 0.14 0.11 0.14 0.11 0.09<br />
Procter & Gamble Co. 0.24 0.21 0.18 0.20 0.16 0.13<br />
Pharmacia Corp. 0.10 0.08 0.05 0.10 0.07 0.05<br />
Schering-Plough Corp. 0.23 0.20 0.17 0.16 0.13 0.10<br />
Texaco Inc. 0.02 0.01 0.01 0.02 0.01 0.01<br />
Texas Instruments Inc. 0.43 0.42 0.41 0.31 0.28 0.26<br />
United Technologies Corp 0.20 0.16 0.14 0.18 0.14 0.11<br />
Walgreen Co. 0.15 0.12 0.09 0.09 0.07 0.05<br />
Table 8: This table summarizes the mean values over the first centile of the distribution of the<br />
coefficients of (upper or lower) tail <strong>de</strong>pen<strong>de</strong>nce for the positive and neg<strong>at</strong>ive tails during the time<br />
interval from January 1980 to December 2000, for three values of the tail in<strong>de</strong>x α = 3, 3.5, 4.<br />
37
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 337<br />
Neg<strong>at</strong>ive Tail Positive Tail<br />
α = 3 α = 3.5 α = 4 α = 3 α = 3.5 α = 4<br />
Abbott Labs 0.17 0.13 0.11 0.15 0.12 0.09<br />
American Home Products Corp. 0.14 0.11 0.08 0.15 0.11 0.09<br />
Boeing Co. 0.14 0.10 0.08 0.10 0.07 0.05<br />
Bristol-Myers Squibb Co. 0.27 0.24 0.21 0.27 0.24 0.21<br />
Chevron Corp. 0.19 0.15 0.12 0.17 0.13 0.10<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0.25 0.22 0.19 0.23 0.19 0.16<br />
Disney (Walt) Co. 0.18 0.14 0.11 0.17 0.13 0.11<br />
General Motors Corp. 0.26 0.23 0.20 0.24 0.21 0.18<br />
Hewl<strong>et</strong>t-Packard Co. 0.17 0.14 0.11 0.23 0.19 0.16<br />
Coca-Cola Co. 0.23 0.20 0.17 0.23 0.20 0.17<br />
Minnesota Mining & MFG Co. 0.28 0.25 0.23 0.25 0.22 0.19<br />
Philip Morris Cos Inc. 0.14 0.10 0.08 0.14 0.11 0.08<br />
Pepsico Inc. 0.16 0.13 0.10 0.16 0.12 0.10<br />
Procter & Gamble Co. 0.23 0.20 0.17 0.22 0.18 0.15<br />
Pharmacia Corp. 0.13 0.10 0.07 0.14 0.10 0.08<br />
Schering-Plough Corp. 0.22 0.19 0.16 0.19 0.15 0.12<br />
Texaco Inc. 0.03 0.02 0.01 0.03 0.02 0.01<br />
Texas Instruments Inc. 0.44 0.42 0.41 0.37 0.35 0.33<br />
United Technologies Corp 0.16 0.12 0.10 0.15 0.12 0.09<br />
Walgreen Co. 0.09 0.07 0.05 0.06 0.04 0.03<br />
Table 9: This table summarizes the mean values over the first centile of the distribution of the<br />
coefficients of (upper or lower) tail <strong>de</strong>pen<strong>de</strong>nce for the positive and neg<strong>at</strong>ive tails during the time<br />
interval from July 1962 to December 2000, for three values of the tail in<strong>de</strong>x α = 3, 3.5, 4.<br />
38
338 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
July 1962 - Dec. 1979 Jan.1980 - Dec. 2000<br />
Extremes λ− p-value Extremes λ− p-value<br />
Abbott Labs 0 0.12 0.2937 4 0.20 0.0904<br />
American Home Products Corp. 1 0.22 0.2432 2 0.12 0.2247<br />
Boeing Co. 0 0.16 0.1667 3 0.14 0.1176<br />
Bristol-Myers Squibb Co. 2 0.22 0.2987 4 0.32 0.2144<br />
Chevron Corp. 3 0.21 0.2112 4 0.18 0.0644<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0 0.38 0.0078 4 0.23 0.1224<br />
Disney (Walt) Co. 2 0.24 0.2901 2 0.16 0.2873<br />
General Motors Corp. 2 0.39 0.1345 4 0.26 0.1522<br />
Hewl<strong>et</strong>t-Packard Co. 0 0.15 0.1909 2 0.19 0.3007<br />
Coca-Cola Co. 2 0.26 0.2765 5 0.24 0.0494<br />
Minnesota Mining & MFG Co. 2 0.35 0.1784 4 0.26 0.1571<br />
Philip Morris Cos Inc. 1 0.25 0.1841 2 0.11 0.2142<br />
Pepsico Inc. 2 0.15 0.2795 5 0.17 0.0141<br />
Procter & Gamble Co. 1 0.23 0.2245 3 0.24 0.2447<br />
Pharmacia Corp. 2 0.23 0.2956 4 0.10 0.0128<br />
Schering-Plough Corp. 0 0.21 0.0946 4 0.23 0.1224<br />
Texaco Inc. 0 0.06 0.5222 2 0.02 0.0212<br />
Texas Instruments Inc. 1 0.47 0.0161 3 0.43 0.1862<br />
United Technologies Corp 1 0.13 0.3728 4 0.20 0.0870<br />
Walgreen Co. 1 0.03 0.2303 3 0.15 0.1373<br />
Table 10: This table gives, for the time intervals from July 1962 to December 1979 and from<br />
January 1980 to December 2000, the number of losses within the ten largest losses incured by an<br />
ass<strong>et</strong> which have occured tog<strong>et</strong>her with one of the ten largest losses of the Standard & Poor’s 500<br />
in<strong>de</strong>x during the same time interval. The probabilty of occurence of such a realis<strong>at</strong>ion is given by<br />
the p-value <strong>de</strong>rived from the binomial law (27) with param<strong>et</strong>er λ−.<br />
39
9.2. Estim<strong>at</strong>ion du coefficient <strong>de</strong> dépendance <strong>de</strong> queue 339<br />
July 1962 - Dec. 1979 Jan.1980 - Dec. 2000<br />
Extremes λ− p-value Extremes λ− p-value<br />
Abbott Labs 0 0.12 0.2937 4 0.20 0.0904<br />
American Home Products Corp. 1 0.22 0.2432 1 0.12 0.3828<br />
Boeing Co. 0 0.16 0.1667 3 0.14 0.1176<br />
Bristol-Myers Squibb Co. 2 0.22 0.2987 3 0.32 0.2653<br />
Chevron Corp. 3 0.21 0.2112 3 0.18 0.1708<br />
Du Pont (E.I.) <strong>de</strong> Nemours & Co. 0 0.38 0.0078 3 0.23 0.2342<br />
Disney (Walt) Co. 2 0.24 0.2901 1 0.16 0.3300<br />
General Motors Corp. 2 0.39 0.1345 3 0.26 0.2536<br />
Hewl<strong>et</strong>t-Packard Co. 0 0.15 0.1909 1 0.19 0.2880<br />
Coca-Cola Co. 1 0.26 0.1782 4 0.24 0.1318<br />
Minnesota Mining & MFG Co. 2 0.35 0.1784 3 0.26 0.2561<br />
Philip Morris Cos Inc. 1 0.25 0.1841 2 0.11 0.2142<br />
Pepsico Inc. 2 0.15 0.2795 5 0.17 0.0141<br />
Procter & Gamble Co. 1 0.23 0.2245 3 0.24 0.2447<br />
Pharmacia Corp. 2 0.23 0.2956 4 0.10 0.0128<br />
Schering-Plough Corp. 0 0.21 0.0946 3 0.23 0.2342<br />
Texaco Inc. 0 0.06 0.5222 1 0.02 0.1922<br />
Texas Instruments Inc. 1 0.47 0.0161 3 0.43 0.1862<br />
United Technologies Corp 1 0.13 0.3728 3 0.20 0.2001<br />
Walgreen Co. 1 0.03 0.2303 3 0.15 0.1373<br />
Table 11: This table gives, for the time intervals from July 1962 to December 1979 and from<br />
January 1980 to December 2000, the number of losses within the ten largest losses incured by an<br />
ass<strong>et</strong> which have occured tog<strong>et</strong>her with one of the ten largest losses of the Standard & Poor’s 500<br />
in<strong>de</strong>x during the same time interval, provi<strong>de</strong>d th<strong>at</strong> the losses are not both the largest of each series.<br />
The probabilty of occurence of such a realis<strong>at</strong>ion is given by the p-value <strong>de</strong>rived from the binomial<br />
law (27) with param<strong>et</strong>er λ−.<br />
40
340 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
9.3 Synthèse <strong>de</strong> la <strong>de</strong>scription <strong>de</strong> la dépendance entre actifs financiers<br />
Les valeurs <strong>de</strong> dépendance <strong>de</strong> queue présentées dans la partie précé<strong>de</strong>nte entre un actif <strong>et</strong> le facteur <strong>de</strong><br />
marché (l’indice) perm<strong>et</strong>tent aisément d’en déduire la dépendance <strong>de</strong> queue entre <strong>de</strong>ux actifs, comme<br />
étant le minimum <strong>de</strong> la dépendance <strong>de</strong> queue entre chacun <strong>de</strong>s actifs <strong>et</strong> l’indice. On en déduit que si l’indice<br />
à une distribution régulièrement variable, les actifs présentent une dépendance <strong>de</strong> queue non nulle.<br />
En conséquence, l’hypothèse <strong>de</strong> copule gaussienne faite au chapitre 7 ne peut être considérée comme<br />
une approxim<strong>at</strong>ion valable, puisque nous rappelons qu’elle n’adm<strong>et</strong> pas <strong>de</strong> dépendance <strong>de</strong> queue. Cependant,<br />
comme nous l’avons montré au chapitre 3, la distribution <strong>de</strong> ren<strong>de</strong>ments <strong>de</strong>s actifs n’est peutêtre<br />
pas régulièrement variable, mais rapi<strong>de</strong>ment variable, si l’on considère les distributions exponentielles<br />
étirées. Dans ce cas, <strong>et</strong> pour autant que l’indice <strong>de</strong> marché reste le facteur principal, les actifs ne<br />
présentent pas <strong>de</strong> dépendance <strong>de</strong> queue, <strong>et</strong> la <strong>de</strong>scription <strong>de</strong> la dépendance en terme <strong>de</strong> copule gaussienne<br />
re<strong>de</strong>vient acceptable.<br />
Ceci étant, l’examen direct <strong>de</strong> la répartition <strong>de</strong>s ren<strong>de</strong>ments <strong>de</strong>s actifs en fonction <strong>de</strong>s ren<strong>de</strong>ments <strong>de</strong><br />
l’indice, au cours <strong>de</strong> la pério<strong>de</strong> 1980-2000, suggère clairement l’existence d’une dépendance <strong>de</strong> queue<br />
comme le montre la figure 9.1. De plus, c<strong>et</strong>te représent<strong>at</strong>ion confirme que la dépendance <strong>de</strong> queue pour<br />
Texaco (panneau <strong>de</strong> gauche) est beaucoup plus faible que pour United Technologies (panneau <strong>de</strong> droite)<br />
par exemple. En eff<strong>et</strong>, on observe que pour les ren<strong>de</strong>ments nég<strong>at</strong>ifs, les extrêmes sont regroupés en fuseau<br />
pour United Technologies alors qu’ils sont beaucoup plus dispersés pour Texaco. Ceci est parfaitement<br />
conforme aux résult<strong>at</strong>s énoncés au paragraphe précé<strong>de</strong>nt où les valeurs <strong>de</strong> dépendance <strong>de</strong> queue mesurées<br />
étaient respectivement <strong>de</strong> 2% <strong>et</strong> 20%.<br />
Texaco<br />
4<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
−4 −3 −2 −1 0<br />
S&P 500<br />
1 2 3 4<br />
United Technologies<br />
4<br />
3<br />
2<br />
1<br />
0<br />
−1<br />
−2<br />
−3<br />
−4<br />
−4 −3 −2 −1 0<br />
S&P 500<br />
1 2 3 4<br />
FIG. 9.1 – Ren<strong>de</strong>ments <strong>de</strong> Texaco (panneau <strong>de</strong> gauche) <strong>et</strong> <strong>de</strong> United Technologies (panneau <strong>de</strong> droite)<br />
en fonction <strong>de</strong>s ren<strong>de</strong>ments du Standard & Poor’s 500 durant la pério<strong>de</strong> 1980-2000. Les distributions<br />
marginales ont été proj<strong>et</strong>ées sur <strong>de</strong>s distributions gaussiennes pour perm<strong>et</strong>tre une meilleure comparaison<br />
<strong>et</strong> m<strong>et</strong>tre en lumière l’eff<strong>et</strong> <strong>de</strong> la copule.<br />
Pour compléter c<strong>et</strong>te étu<strong>de</strong>, nous avons utilisé la métho<strong>de</strong> d’estim<strong>at</strong>ion non paramétrique <strong>de</strong> Coles, Heffernan<br />
<strong>et</strong> Tawn (1999) mise en œuvre par Poon, Rockinger <strong>et</strong> Tawn (2001) que nous avons présentée au<br />
paragraphe 9.1. C<strong>et</strong>te métho<strong>de</strong> est très délic<strong>at</strong>e <strong>et</strong> nous semble assez peu précise puisqu’elle repose sur<br />
l’estim<strong>at</strong>ion d’un exposant <strong>de</strong> queue <strong>et</strong> d’un facteur d’échelle. Nous avons déjà évoqué ces problèmes<br />
au chapitre 1, nous n’y revenons donc pas. Malgré (ou plutôt à cause <strong>de</strong>) ces imprécisions, nous n’avons<br />
jamais pu rej<strong>et</strong>er, à 95% <strong>de</strong> confiance, l’hypothèse <strong>de</strong> l’existence <strong>de</strong> dépendance <strong>de</strong> queue entre les actifs.<br />
En outre, même si les valeurs numériques obtenues sont entachées d’une gran<strong>de</strong> incertitu<strong>de</strong>, elles
9.3. Synthèse <strong>de</strong> la <strong>de</strong>scription <strong>de</strong> la dépendance entre actifs financiers 341<br />
Modèle à facteur<br />
0.35<br />
0.3<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
0<br />
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35<br />
Modèle elliptique<br />
FIG. 9.2 – Coefficient <strong>de</strong> dépendance <strong>de</strong> queue estimé à l’ai<strong>de</strong> du modèle à facteur en fonction du<br />
coefficient <strong>de</strong> dépendance <strong>de</strong> queue estimé sous hypothèse <strong>de</strong> copule elliptique. L’indice <strong>de</strong> vari<strong>at</strong>ion<br />
régulière du facteur est <strong>de</strong> 3 <strong>et</strong> celui <strong>de</strong> la copule elliptique <strong>de</strong> 4. Ce couple <strong>de</strong> valeurs est celui qui donne<br />
le meilleur accord entre les résult<strong>at</strong>s <strong>de</strong>s <strong>de</strong>ux modèles.<br />
<strong>de</strong>meurent du même ordre <strong>de</strong> gran<strong>de</strong>ur que celles trouvées précé<strong>de</strong>mment à l’ai<strong>de</strong> du modèle à facteur.<br />
Nous avons également estimé la dépendance <strong>de</strong> queue sous l’hypothèse <strong>de</strong> copule elliptique régulièrement<br />
variable, dont l’expression a été dérivée par Hult <strong>et</strong> Lindskog (2001) <strong>et</strong> qui n’est fonction que <strong>de</strong><br />
l’indice <strong>de</strong> vari<strong>at</strong>ion régulière <strong>et</strong> du coefficient <strong>de</strong> corrél<strong>at</strong>ion. Ce <strong>de</strong>rnier a été estimé <strong>de</strong> manière non<br />
paramétrique par l’intermédiaire du τ du Kendall, puis obtenu grâce à la rel<strong>at</strong>ion<br />
ρ = sin<br />
π · τ<br />
2<br />
<br />
, (9.1)<br />
bien connue pour la distribution gaussienne mais aussi valable pour la classe <strong>de</strong>s distributions elliptiques<br />
comme l’ont récemment montré Lindskog, McNeil <strong>et</strong> Schmock (2001). Là encore, les valeurs<br />
numériques sont qualit<strong>at</strong>ivement en accord avec les résult<strong>at</strong>s précé<strong>de</strong>nts. Plus précisément, on peut toujours<br />
trouver un indice <strong>de</strong> vari<strong>at</strong>ion régulière ν <strong>de</strong> la copule elliptique tel que les <strong>de</strong>ux modèles donnent<br />
exactement la même valeur <strong>de</strong> dépendance <strong>de</strong> queue, ce qui revient à accepter que la copule elliptique<br />
soit différente d’un couple d’actifs à l’autre, mais ceci nous semble guère s<strong>at</strong>isfaisant. Si par contre l’on<br />
considère <strong>de</strong>s copules elliptiques <strong>de</strong> même indice <strong>de</strong> queue pour toutes les paires, <strong>de</strong>s différences importantes<br />
sont observées par rapport aux estim<strong>at</strong>ions données par le modèle à facteur, comme le montre<br />
la figure 9.2, où sont représentés les coefficients <strong>de</strong> dépendance <strong>de</strong> queue estimés à l’ai<strong>de</strong> du modèle à<br />
facteur <strong>et</strong> sous hypothèse <strong>de</strong> copule elliptique.<br />
En conclusion, nous pensons pouvoir affirmer que les actifs auxquels nous nous sommes intéressés<br />
présentent une dépendance <strong>de</strong> queue, <strong>et</strong> que celle-ci est <strong>de</strong> l’ordre <strong>de</strong> 5% à 30% selon les cas. Ceci<br />
à d’importantes conséquences pr<strong>at</strong>iques du point <strong>de</strong> vue <strong>de</strong> la <strong>gestion</strong> <strong>de</strong>s risques. En eff<strong>et</strong>, reprenant<br />
le p<strong>et</strong>it calcul présenté au point 1.2 du paragraphe 9.2, nous considérons la probabilité que <strong>de</strong>ux actifs<br />
d’un <strong>portefeuille</strong> chutent ensemble au-<strong>de</strong>là <strong>de</strong> leur VaR journalière à 99% par exemple. Négligeant<br />
la dépendance <strong>de</strong> queue, un tel événement à une probabilité extrêmement faible <strong>de</strong> se produire (temps<br />
<strong>de</strong> récurrence typique <strong>de</strong> 40 ans), alors que si l’on considère une dépendance <strong>de</strong> queue <strong>de</strong> 30% un tel<br />
événement se produit typiquement tous les 16 mois (contre huit ans pour λ = 5%).
342 9. Mesure <strong>de</strong> la dépendance extrême entre <strong>de</strong>ux actifs financiers<br />
Pour ce qui est <strong>de</strong>s implic<strong>at</strong>ions théoriques, les résult<strong>at</strong>s que nous venons d’exposer ren<strong>de</strong>nt l’approxim<strong>at</strong>ion<br />
<strong>de</strong> copule gaussienne non valable, à strictement parler. Toutefois, pour <strong>de</strong>s actifs dont la dépendance<br />
<strong>de</strong> queue n’est que <strong>de</strong> l’ordre <strong>de</strong> 5%, ce type d’approxim<strong>at</strong>ion peut être considéré comme non déraisonnable,<br />
comme vient <strong>de</strong> le montrer le p<strong>et</strong>it calcul ci-<strong>de</strong>ssus <strong>et</strong> dans la mesure où bien d’autres sources<br />
d’erreurs - notamment sur la <strong>de</strong>scription <strong>de</strong>s marginales - sont à prendre en compte dans l’estim<strong>at</strong>ion<br />
globale <strong>de</strong> l’incertitu<strong>de</strong> associée à l’estim<strong>at</strong>ion <strong>de</strong> la distribution jointe dans son ensemble.<br />
Par ailleurs, dans les cas où la dépendance <strong>de</strong> queue est trop importante pour que la copule gaussienne<br />
puisse être r<strong>et</strong>enue, il convient aussi <strong>de</strong> s’interroger sur la pertinence <strong>de</strong> la modélis<strong>at</strong>ion en terme <strong>de</strong><br />
copule elliptique que nous avions initialement choisie. En eff<strong>et</strong>, nous venons <strong>de</strong> montrer que l’accord<br />
entre ce type <strong>de</strong> représent<strong>at</strong>ion <strong>et</strong> le modèle à facteur, qui a <strong>de</strong> soli<strong>de</strong>s points d’ancrage dans la <strong>théorie</strong><br />
financière, n’est pas très s<strong>at</strong>isfaisant. De plus, les résult<strong>at</strong>s du paragraphe 9.2 montrent que durant la<br />
pério<strong>de</strong> 1980-2000, la dépendance <strong>de</strong> queue est légèrement plus importante dans la queue inférieure que<br />
dans la queue supérieure - ce qui a été confirmé par l’approche non paramétrique - même si dans la<br />
majeure partie <strong>de</strong>s cas, c<strong>et</strong>te différence n’est pas <strong>st<strong>at</strong>istique</strong>ment signific<strong>at</strong>ive. Cela dit, c<strong>et</strong>te différence<br />
systém<strong>at</strong>iquement en faveur <strong>de</strong> la queue nég<strong>at</strong>ive ne peut être négligée <strong>et</strong> est dès lors en contradiction<br />
avec l’hypothèse <strong>de</strong> copules elliptiques qui sont <strong>de</strong>s copules symétriques dans leurs parties positive <strong>et</strong><br />
nég<strong>at</strong>ive <strong>et</strong> présentent donc une même dépendance <strong>de</strong> queue supérieure <strong>et</strong> inférieure.
Troisième partie<br />
Mesures <strong>de</strong>s risques extrêmes <strong>et</strong><br />
applic<strong>at</strong>ion à la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong><br />
343
Chapitre 10<br />
La mesure du risque<br />
Dans ce chapitre, nous souhaitons présenter quelques <strong>théorie</strong>s ayant permis la quantific<strong>at</strong>ion du risque<br />
associé à un actif financier ou a un <strong>portefeuille</strong>. Les sources <strong>de</strong> risque sont en fait aussi diverses que<br />
variées. Nous pouvons citer par exemple le risque <strong>de</strong> défaut lié au fait qu’une contrepartie ne puisse<br />
faire face à ses oblig<strong>at</strong>ions ou bien le risque <strong>de</strong> liquidité lié à la capacité limitée <strong>de</strong>s marchés à pouvoir<br />
absorber un afflux massif <strong>de</strong> titres à la vente ou à l’ach<strong>at</strong>, mais il nous semble que la source principale du<br />
risque est le risque <strong>de</strong> marché, c’est-à-dire le risque associé aux fluctu<strong>at</strong>ions <strong>de</strong>s actifs financiers. C’est<br />
pourquoi nous nous bornerons à décrire comment quantifier ce type <strong>de</strong> risque.<br />
Depuis le milieu du vingtième siècle, plusieurs <strong>théorie</strong>s ont vu le jour pour tenter <strong>de</strong> cerner le comportement<br />
<strong>de</strong>s individus face à l’incertain en vue d’en déduire <strong>de</strong>s règles <strong>de</strong> décisions <strong>et</strong> <strong>de</strong>s mesures <strong>de</strong> risque.<br />
On peut citer tout d’abord les travaux <strong>de</strong> von Neumann <strong>et</strong> Morgenstern (1947) <strong>et</strong> Savage (1954) sur la<br />
formalis<strong>at</strong>ion <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> l’utilité espérée, visant à décrire les préférences <strong>de</strong>s agents économiques,<br />
<strong>et</strong> l’introduction <strong>de</strong> la notion d’aversion au risque. Puis, au vu <strong>de</strong>s incomp<strong>at</strong>ibilités <strong>de</strong> c<strong>et</strong>te <strong>théorie</strong> avec<br />
certains comportements observés par Allais (1953) ou Ellsberg (1961), <strong>de</strong> nouvelles approches ont vu le<br />
jour, telle la “<strong>théorie</strong> <strong>de</strong> la perspective” <strong>de</strong> Kahneman <strong>et</strong> Tversky (1979), qui perm<strong>et</strong> <strong>de</strong> rendre compte<br />
d’une part du fait que les agents s’<strong>at</strong>tachent plus aux perspectives d’évolution <strong>de</strong> leur richesse qu’à leur<br />
richesse elle-même, <strong>et</strong> d’autre part qu’ils traitent leurs gains <strong>et</strong> leurs pertes <strong>de</strong> manière disymétriques : ils<br />
sont risquophobes face à <strong>de</strong>s gains potentiels mais risquophiles vis-à-vis <strong>de</strong> pertes à venir. Une autre altern<strong>at</strong>ive<br />
plus générale est ensuite apparue (Quiggin 1982, Gilboa <strong>et</strong> Schmeidler 1989) <strong>et</strong> s’est développée<br />
autour <strong>de</strong>s modèles dits non-additifs, c’est-à-dire <strong>de</strong>s modèles où les probabilités ne jouissent plus <strong>de</strong><br />
la propriété d’additivité, <strong>et</strong> sont donc remplacées par <strong>de</strong>s capacités, ce qui perm<strong>et</strong> notamment <strong>de</strong> rendre<br />
compte du phénomène <strong>de</strong> distorsion <strong>de</strong>s probabilités souvent observé chez la plupart <strong>de</strong>s agents.<br />
Récemment, dans un cadre plus restreint que celui <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> la décision, Artzner, Delbaen, Eber<br />
<strong>et</strong> He<strong>at</strong>h (1999) ont proposé une nouvelle approche <strong>de</strong> la notion <strong>de</strong> risque, reposant sur les propriétés<br />
minimales que l’on est en droit d’<strong>at</strong>tendre d’une mesure <strong>de</strong> risque. Ceci a donné naissance à la notion <strong>de</strong><br />
mesures <strong>de</strong> risque cohérentes, qui a ensuite été étendue par He<strong>at</strong>h (2000) puis Föllmer <strong>et</strong> Schied (2002a)<br />
à la notion <strong>de</strong> mesures <strong>de</strong> risque convexes. Cependant, comme nous le verrons, ce type <strong>de</strong> mesure <strong>de</strong><br />
risque ne semble pas toujours être le mieux adapté à la quantific<strong>at</strong>ion <strong>de</strong>s risques d’un <strong>portefeuille</strong>. En<br />
eff<strong>et</strong>, il nous semble qu’il convient <strong>de</strong> bien différencier <strong>de</strong>ux types <strong>de</strong> risques :<br />
– premièrement, ce que l’on peut appeler la mesure du risque en terme <strong>de</strong> capital économique, c’est-àdire<br />
la somme d’argent dont doit disposer un <strong>gestion</strong>naire <strong>de</strong> <strong>portefeuille</strong> / une institution pour pouvoir<br />
faire face à ses oblig<strong>at</strong>ions <strong>et</strong> ainsi éviter la ruine,<br />
– <strong>et</strong> <strong>de</strong>uxièmement le risque lié aux fluctu<strong>at</strong>ions <strong>st<strong>at</strong>istique</strong>s <strong>de</strong> la richesse ou du ren<strong>de</strong>ment d’un <strong>portefeuille</strong><br />
autour <strong>de</strong> l’objectif <strong>de</strong> rentabilité préalablement fixé.<br />
345
346 10. La mesure du risque<br />
C’est pour rendre compte <strong>de</strong> c<strong>et</strong>te <strong>de</strong>uxième c<strong>at</strong>égorie <strong>de</strong> risques que nous proposons une autre approche<br />
dans laquelle il est fait usage <strong>de</strong>s moments <strong>de</strong> la distribution <strong>de</strong>s ren<strong>de</strong>ments d’un actif pour quantifier le<br />
risque associé aux fluctu<strong>at</strong>ions <strong>de</strong> c<strong>et</strong> actif.<br />
10.1 La <strong>théorie</strong> <strong>de</strong> l’utilité<br />
Notre objectif dans ce paragraphe est <strong>de</strong> passer en revue les avancées récentes <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> la décision,<br />
dont on sait qu’elle joue un rôle très important en économie <strong>et</strong> en finance au travers <strong>de</strong> la notion d’utilité.<br />
Notre présent<strong>at</strong>ion ne fait qu’effleurer la surface <strong>de</strong> ce vaste problème <strong>et</strong> nous renvoyons le lecteur à<br />
l’article <strong>de</strong> Cohen <strong>et</strong> Tallon (2000) notamment pour <strong>de</strong> plus amples détails.<br />
10.1.1 Théorie <strong>de</strong> l’utilité en environnement certain<br />
La <strong>théorie</strong> <strong>de</strong> l’utilité trouve ses fon<strong>de</strong>ments dans le courant <strong>de</strong> pensée sociale développé <strong>de</strong> la fin du<br />
18ème siècle au milieu du 19ème par J. Bentham <strong>et</strong> J.S. Mill, fond<strong>at</strong>eurs <strong>de</strong> la philosophie <strong>de</strong> l’utilitarisme<br />
1 <strong>et</strong> selon laquelle on doit juger une action à ses résult<strong>at</strong>s <strong>et</strong> conséquences sur le bien-être <strong>de</strong>s<br />
individus. En cela, l’utilitarisme s’oppose farouchement au rigorisme moral prôné, à la même époque,<br />
par Kant, pour qui la valeur d’une action ne peut se juger qu’à l’aune <strong>de</strong>s principes <strong>et</strong> intentions qui en<br />
sont à l’origine.<br />
Loin <strong>de</strong> ces querelles philosophiques, les économistes se sont rapi<strong>de</strong>ment emparés du principe selon lequel<br />
les individus agissent en vue <strong>de</strong> maximiser leur bien-être <strong>et</strong> en ont fait le moteur <strong>de</strong>s comportements<br />
individuels <strong>de</strong>s agents économique. On peut d’ailleurs faire remonter à Adam Smith (1776) l’introduction<br />
<strong>de</strong> la notion d’utilité en économie du fait <strong>de</strong> la distinction qu’il introduit entre les concepts <strong>de</strong> “valeur<br />
à l’échange” (le prix) <strong>et</strong> <strong>de</strong> “valeur à l’usage” (l’utilité) d’un bien <strong>et</strong> qui sont à la base <strong>de</strong> la loi <strong>de</strong> l’offre<br />
<strong>et</strong> <strong>de</strong> la <strong>de</strong>man<strong>de</strong> <strong>et</strong> <strong>de</strong> la notion d’équilibre <strong>de</strong> marchés. La formalis<strong>at</strong>ion m<strong>at</strong>hém<strong>at</strong>ique <strong>de</strong> la <strong>théorie</strong> <strong>de</strong><br />
l’utilité n’interviendra que près <strong>de</strong> <strong>de</strong>ux siècles plus tard <strong>et</strong> repose (en univers certain) sur <strong>de</strong>ux axiomes<br />
simples décrivant les capacités <strong>de</strong>s agents à déterminer leurs préférences.<br />
Considérons à partir <strong>de</strong> maintenant, l’ensemble B <strong>de</strong>s actifs financiers (ou plus généralement <strong>de</strong>s biens)<br />
accessibles aux agents économiques, <strong>et</strong> postulons <strong>de</strong>ux axiomes simples, un axiome <strong>de</strong> comparaison <strong>et</strong><br />
un axiome <strong>de</strong> continuité :<br />
AXIOME 1 (COMPARABILITÉ)<br />
Un agent économique est capable d’établir une préférence entre tous les actifs <strong>de</strong> B. Cela revient à dire<br />
qu’il existe un préordre compl<strong>et</strong> “l’actif X est préféré à l’actif Y ”, notée X Y , entre tous les actifs<br />
X, Y ∈ B.<br />
AXIOME 2 (CONTINUITÉ)<br />
La rel<strong>at</strong>ion d’ordre est continue. Cela signifie qu’étant donné trois actifs X, Y, Z ∈ B il existe toujours<br />
<strong>de</strong>ux reéls α, β ∈]0, 1[ tels que d’une part, le <strong>portefeuille</strong> composé d’une proportion α <strong>de</strong> l’actif X <strong>et</strong><br />
(1 − α) <strong>de</strong> l’actif Y est strictement préféré à l’actif Z <strong>et</strong> d’autre part, l’actif Z est strictement préféré au<br />
<strong>portefeuille</strong> composé d’une proportion β <strong>de</strong> l’actif X <strong>et</strong> (1 − β) <strong>de</strong> l’actif Y :<br />
α X + (1 − α) Y ≻ Z <strong>et</strong> Z ≻ β X + (1 − β) Y. (10.1)<br />
1 En toute rigueur les prémices <strong>de</strong> la philosophie <strong>de</strong> l’utilité remontent plutôt à la fin du 17ème siècle avec Hobbes puis avec<br />
Helvétius au début 18ème (voire, si l’on veut aller aussi loin dans le temps à Epicure).
10.1. La <strong>théorie</strong> <strong>de</strong> l’utilité 347<br />
Moyennant ces <strong>de</strong>ux axiomes <strong>et</strong> quelques hypothèses à caractère purement technique que nous om<strong>et</strong>tons,<br />
il est possible <strong>de</strong> montrer qu’il existe une fonction dite fonction d’utilité telle que :<br />
DÉFINITION 5 (FONCTION D’UTILITÉ)<br />
La rel<strong>at</strong>ion <strong>de</strong> préférence sur l’ensemble <strong>de</strong>s actifs B peut être représentée par une fonction d’utilité<br />
U : B → R, telle que :<br />
∀(X, Y ) ∈ B 2 , X Y ⇐⇒ U(X) ≥ U(Y ). (10.2)<br />
C<strong>et</strong>te définition signifie simplement que l’actif X est préféré à l’actif Y si <strong>et</strong> seulement si l’utilité (ou<br />
valeur à l’usage) <strong>de</strong> l’actif X est supérieure à l’utilité <strong>de</strong> l’actif Y . Bien évi<strong>de</strong>mment, la fonction d’utilité<br />
n’est pas unique, puisque pour toute fonction g : R → R strictement croissante, la fonction V = g ◦ U<br />
est aussi une fonction d’utilité. La fonction U ainsi définie n’a donc qu’une valeur ordinale. Si l’on<br />
souhaite pouvoir considérer qu’un accroissement <strong>de</strong> la fonction d’utilité mesure une augment<strong>at</strong>ion <strong>de</strong><br />
la s<strong>at</strong>isfaction <strong>de</strong> l’agent économique, il faut adm<strong>et</strong>tre que la fonction d’utilité a une valeur cardinale,<br />
<strong>et</strong> elle n’est alors plus définie qu’à une tranform<strong>at</strong>ion affine croissante près. Dans toute la suite nous<br />
considérerons uniquement <strong>de</strong>s fonctions d’utilité cardinales.<br />
Si l’on s’intéresse, comme le plus fréquemment, à l’utilité <strong>de</strong> la richesse W d’un individu, on déduit<br />
aisément du comportement <strong>de</strong>s agents économiques que la fonction U(W ) est croissante, ce qui traduit<br />
l’appât du gain ou ins<strong>at</strong>iabilité, <strong>et</strong> généralement concave ce qui exprime la décroissance marginale <strong>de</strong><br />
l’utilité <strong>de</strong> la richesse : cent euros ne représentent pas la même utilité pour un agent possédant en tout <strong>et</strong><br />
pour tout mille euros ou un million d’euros.<br />
10.1.2 Théorie <strong>de</strong> la décision face au risque<br />
Considérons maintenant que l’on s’intéresse au comportement <strong>de</strong>s agents économiques à l’égard d’actifs<br />
dont la valeur future n’est pas parfaitement connue <strong>et</strong> dépend, à l’instant futur T , <strong>de</strong> l’ét<strong>at</strong> <strong>de</strong> la n<strong>at</strong>ure<br />
dans lequel se trouvera l’univers en T . Nous sommes alors confrontés à un problème <strong>de</strong> décision en<br />
univers risqué. Le premier exemple <strong>de</strong> résolution d’un tel problème remonte à Daniel Bernoulli (1738)<br />
qui apparaît comme le précurseur <strong>de</strong> l’introduction <strong>de</strong> l’utilité espérée, dont il fit usage pour résoudre le<br />
célèbre paradoxe <strong>de</strong> Saint-Pétersbourg.<br />
Dans ce paradoxe, un individu se voit offrir la possibilité <strong>de</strong> jouer au jeu suivant : on lance une pièce<br />
parfaitement équilibrée autant <strong>de</strong> fois que nécessaire pour voir le coté pile apparaître. A ce momentlà,<br />
le jeu s’arrête <strong>et</strong> le joueur reçoit 2 n euros, n étant le nombre <strong>de</strong> fois que la pièce a été lancée. La<br />
question est alors <strong>de</strong> savoir combien est prêt à payer l’individu pour pouvoir participer à ce jeu. Un<br />
simple calcul d’espérance m<strong>at</strong>hém<strong>at</strong>ique montre qu’en moyenne ce jeu offre un gain infini 2 . Donc, pour<br />
être équitable, le joueur <strong>de</strong>vrait accepter <strong>de</strong> payer une mise sinon infinie, du moins colossale. Or en<br />
réalité, on observe que les joueurs n’acceptent guère <strong>de</strong> payer plus <strong>de</strong> quelques euros pour participer au<br />
jeu, d’où le paradoxe.<br />
La solution proposée par Bernoulli (1738) consiste à supposer que les joueurs ne s’intéressent pas à la<br />
valeur moyenne <strong>de</strong>s gains espérés mais plutôt à l’espérance du logarithme <strong>de</strong>s gains, <strong>et</strong> l’on obtient alors :<br />
∞<br />
2 −n ln (2 n ) = 2 ln 2. (10.3)<br />
n=1<br />
Donc, pour reprendre la terminologie d’Adam Smith, les joueurs ne s’intéresse pas à la “valeur à l’échange”<br />
- ici les gains espérés - du jeu, mais plutôt à sa “valeur à l’usage” <strong>et</strong> donc à l’espérance du logarithme<br />
2 n 1 n.<br />
La probabilité <strong>de</strong> gagner 2 euros est égale à la probabilité d’obtenir n fois <strong>de</strong> suite le coté face <strong>de</strong> la pièce, soit 2
348 10. La mesure du risque<br />
<strong>de</strong>s gains. C<strong>et</strong>te approche coïnci<strong>de</strong> exactement avec la <strong>théorie</strong> <strong>de</strong> l’utilité espérée dont von Neumann<br />
<strong>et</strong> Morgenstern (1947) poseront les bases plus <strong>de</strong> <strong>de</strong>ux siècles plus tard <strong>et</strong> que nous allons maintenant<br />
exposer.<br />
Pour cela considérons l’ensemble Ω <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure <strong>et</strong> F une tribu sur Ω <strong>de</strong> sorte que l’espace<br />
(Ω, F) soit mesurable. A chaque actif X ∈ B est associé une loi <strong>de</strong> probabilité PX sur (Ω, F)<br />
représentant la distribution <strong>de</strong> la valeur future <strong>de</strong> l’actif X. Par abus <strong>de</strong> langage <strong>et</strong> pour alléger les not<strong>at</strong>ions,<br />
la variable alé<strong>at</strong>oire donnant la valeur future <strong>de</strong> l’actif X ∈ B sera elle même notée X (mais c<strong>et</strong>te<br />
fois, X ∈ (Ω, F, PX)). Comme précé<strong>de</strong>mment en environnement certain, nous supposons que les agents<br />
économiques sont capables d’établir un préordre total sur l’ensemble <strong>de</strong>s actifs ou <strong>de</strong>s valeurs futures<br />
<strong>de</strong>s actifs <strong>et</strong> que ces préférences sont continues. Nous adm<strong>et</strong>tons <strong>de</strong> plus que<br />
AXIOME 3 (INDÉPENDENCE)<br />
Pour tout actif X, Y, Z ∈ B <strong>et</strong> tout reél α ∈]0, 1],<br />
X Y ⇐⇒ α X + (1 − α) Z α Y + (1 − α) Z, (10.4)<br />
ce qui suppose que l’adjonction d’un même actif ne modifie pas l’ordre <strong>de</strong>s préférences. C<strong>et</strong> axiome est<br />
central dans la <strong>théorie</strong> <strong>de</strong> l’utilité espérée <strong>de</strong> von Neumann <strong>et</strong> Morgenstern (1947). C’est en eff<strong>et</strong> grâce à<br />
lui que l’on peut montrer que l’utilité U d’un actif X dont la valeur future est risquée s’exprime comme<br />
U(X) = E[u(X)], (10.5)<br />
où u(·) est une fonction continue, croissante <strong>et</strong> définie à une transform<strong>at</strong>ion affine croissante près. La<br />
fonction u(·) est donc elle même une fonction d’utilité, <strong>de</strong> sorte que l’utilité U d’un bien risqué apparait<br />
comme la moyenne <strong>de</strong> l’utilité u <strong>de</strong> ce même bien dans chacun <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure.<br />
La fonction u perm<strong>et</strong> <strong>de</strong> définir la notion d’aversion pour le risque. Selon Rotschild <strong>et</strong> Stiglitz (1970),<br />
un actif Y est plus risqué qu’un actif X si pour toute fonction croissante <strong>et</strong> concave u(·), E[u(X)] ≥<br />
E[u(Y )]. Ceci est en fait équivalent (Levy 1998) au fait que X domine Y au sens <strong>de</strong> la dominance<br />
stochastique d’ordre <strong>de</strong>ux :<br />
∀t ∈ R,<br />
t<br />
FY (y) dy ≥<br />
−∞<br />
t<br />
−∞<br />
FX(x) dx. (10.6)<br />
Ainsi donc, un individu qui préfère l’actif X à l’actif Y présente <strong>de</strong> l’aversion pour le risque, <strong>et</strong> une<br />
fonction d’utilité u(·) concave caractérise un individu risquophobe. Le coefficient absolu d’aversion<br />
pour le risque <strong>de</strong> c<strong>et</strong> individu est alors défini par<br />
a = − u′′<br />
. (10.7)<br />
u ′<br />
La notion <strong>de</strong> dominance stochastique d’ordre <strong>de</strong>ux peut être étendue à un ordre n quelconque, <strong>et</strong> un<br />
individu présentant une aversion pour le risque au sens <strong>de</strong> la dominance stochastique d’ordre n possè<strong>de</strong><br />
une fonction d’utilité u(·), telle que pour tout x, <strong>et</strong> tout k ≤ n :<br />
(−1) k u (k) (x) ≤ 0, (10.8)<br />
(Levy 1998). De telles fonctions d’utilité caractèrisent une aversion pour le risque dite standard selon<br />
la terminologie <strong>de</strong> Kimball (1993) <strong>et</strong> pour n = 4 par exemple, un individu ayant une fonction d’utilité<br />
vérifiant (10.8) est dit ins<strong>at</strong>iable (u ′ > 0), risquophobe (u ′′ < 0), pru<strong>de</strong>nt (u (3) > 0) <strong>et</strong> tempéré (u (4) <<br />
0).
10.1. La <strong>théorie</strong> <strong>de</strong> l’utilité 349<br />
C<strong>et</strong>te formul<strong>at</strong>ion simple <strong>et</strong> parcimonieuse <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> la décision face au rique a contribué à rendre<br />
l’approche <strong>de</strong> von Neumann <strong>et</strong> Morgenstern (1947) très populaire. Cependant, c<strong>et</strong>te simplicité ne va<br />
pas sans soulever quelques difficultés <strong>et</strong> incohérences, au premier rang <strong>de</strong>squelles se trouve la viol<strong>at</strong>ion<br />
du postul<strong>at</strong> d’indépendance, clé <strong>de</strong> voute <strong>de</strong> c<strong>et</strong>te <strong>théorie</strong>. En eff<strong>et</strong>, Allais (1953) a montré par <strong>de</strong>s tests<br />
simples, consistant à proposer une série d’altern<strong>at</strong>ives à <strong>de</strong>s suj<strong>et</strong>s, que la majorité <strong>de</strong> ceux-ci ém<strong>et</strong>tent <strong>de</strong>s<br />
choix en contradiction avec c<strong>et</strong> axiome d’in<strong>de</strong>pendance. En particulier, les agents semblent très sensibles<br />
aux p<strong>et</strong>ites vari<strong>at</strong>ions <strong>de</strong> probabilités au voisinage du certain : ils accor<strong>de</strong>nt beaucoup d’importance au<br />
passage d’une probabilité <strong>de</strong> 0 à 0.01 ou <strong>de</strong> 0.99 à 1, alors qu’un changement <strong>de</strong> 5 à 10 pour-cent au<br />
voisinage d’un niveau <strong>de</strong> probabilité <strong>de</strong> 0.50 les laissent bien souvent indifférents. Ainsi, les agents sont<br />
suj<strong>et</strong>s à une distorsion <strong>de</strong> leur perception <strong>de</strong>s probabilités. D’autre part, en plus <strong>de</strong> c<strong>et</strong>te contradiction<br />
empirique vient s’ajouter une limit<strong>at</strong>ion théorique : la fonction u(·) joue un double rôle. Elle quantifie en<br />
même temps l’aversion pour le risque du déci<strong>de</strong>ur <strong>et</strong> la décroissance marginale <strong>de</strong> l’utilité <strong>de</strong> la richesse,<br />
<strong>de</strong> sorte qu’il est impossible <strong>de</strong> modéliser un agent qui est à la fois risquophile <strong>et</strong> dont l’utilité marginale<br />
décroit.<br />
En fait, ces <strong>de</strong>ux contradictions peuvent être levées en affaiblissant le postul<strong>at</strong> d’indépendance, qui sera<br />
remplacé par :<br />
AXIOME 4 (CHOSE SÛRE COMONOTONE DANS LE RISQUE)<br />
Soit <strong>de</strong>ux actifs X, Y ∈ B dont les valeurs futures sont données par les variables alé<strong>at</strong>oires (supposées<br />
discrètes pour simplifier) : X = (x1, p1; · · · ; xk, pk; · · · ; xn, pn) <strong>et</strong> Y = (y1, p1; · · · ; yk, pk; · · · ; yn, pn),<br />
telles que x1 ≤ · · · ≤ xk ≤ · · · ≤ xn <strong>et</strong> y1 ≤ · · · ≤ yk ≤ · · · ≤ yn avec xk = yk. Soient alors les actifs<br />
X ′ , Y ′ ∈ B obtenus en remplaçant xk par x ′ k dans les actifs X <strong>et</strong> Y <strong>de</strong> sorte que xk−1 ≤ x ′ k ≤ xk+1 <strong>et</strong><br />
yk−1 ≤ x ′ k = y′ k ≤ yk+1. Alors<br />
X Y ⇐⇒ X ′ Y ′ . (10.9)<br />
Cela veut simplement dire que l’on ne modifie pas l’ordre <strong>de</strong> préférence <strong>de</strong> <strong>de</strong>ux actifs lorsque l’on<br />
modifie leur commune valeur future sans changer le rang <strong>de</strong> celle-ci, ce qui fait toute la différence avec<br />
l’axiome d’indépendance. Moyennant cela, les paradoxes d’Allais (1953) sont levés <strong>et</strong> l’on est amené<br />
à généraliser la <strong>théorie</strong> <strong>de</strong> l’utilité espérée par la <strong>théorie</strong> <strong>de</strong> l’utilité dépendante du rang originellement<br />
développée par Quiggin (1982). En eff<strong>et</strong>, il est alors possible <strong>de</strong> montrer que l’on peut caractériser le<br />
comportement <strong>de</strong> tout agent économique par <strong>de</strong>ux fonctions croissantes. La première, u : B → R,<br />
définie à une transform<strong>at</strong>ion affine croissante près, joue le rôle <strong>de</strong> fonction d’utilité dans le certain. La<br />
secon<strong>de</strong>, ϕ : [0, 1] → [0, 1] est unique <strong>et</strong> représente la fonction <strong>de</strong> transform<strong>at</strong>ion (ou <strong>de</strong> distorsion) <strong>de</strong>s<br />
probabilités. Ainsi, l’utilité <strong>de</strong> l’actif X est :<br />
<br />
U(X) = − u(x) dϕ(Pr{X > x}) = Eϕ◦PX [u(x)]. (10.10)<br />
C<strong>et</strong>te intégrale est en fait une intégrale <strong>de</strong> Choqu<strong>et</strong>, c’est-à-dire une intégrale par rapport à une mesure<br />
généralement non-additive. Dans le cas particulier où ϕ(x) = x, on r<strong>et</strong>rouve bien évi<strong>de</strong>mment l’expression<br />
<strong>de</strong> l’utilité espérée <strong>de</strong> von Neumann <strong>et</strong> Morgenstern (1947), qui est donc englobée par la <strong>théorie</strong> <strong>de</strong><br />
l’utilité espérée dépendante du rang.<br />
L’intérêt <strong>de</strong> c<strong>et</strong>te nouvelle formul<strong>at</strong>ion est <strong>de</strong> complètement découpler les notions <strong>de</strong> décroissance marginale<br />
<strong>de</strong> l’utilité, mesurée par la concavité <strong>de</strong> la fonction u(·), <strong>et</strong> d’aversion pour le risque, entièrement<br />
caractérisée par la fonction <strong>de</strong> transform<strong>at</strong>ion <strong>de</strong>s probabilités ϕ(x) : un agent dont la fonction <strong>de</strong> transform<strong>at</strong>ion<br />
<strong>de</strong>s probabilités est telle que ϕ(x) ≤ x sera dit pessimiste dans le risque. En eff<strong>et</strong>, reprenant<br />
l’exemple discr<strong>et</strong> <strong>de</strong> l’axiome 4, l’équ<strong>at</strong>ion (10.10) <strong>de</strong>vient :<br />
U(X) = u(x1) + ϕ(p2 + · · · + pn) · [u(x2) − u(x1)] + · · · + ϕ(pn) · [u(xn) − u(xn−1)], (10.11)
350 10. La mesure du risque<br />
ce qui montre qu’un tel agent commence par calculer l’utilité minimale que peut lui procurer l’actif X,<br />
soit u(x1), puis il ajoute les accroissements possibles <strong>de</strong> l’utilité u(xk) − u(xk−1) qu’il peut recevoir en<br />
les pondérant non pas par leur probabilités d’occurence mais par sa fonction <strong>de</strong> distorsion <strong>de</strong>s probabiltés.<br />
Ainsi, lorsque ϕ(x) ≤ x, il sous-estime la probabilité d’événements favorables <strong>et</strong> sous-pondère les<br />
accroissements d’utilité qu’il peut en r<strong>et</strong>irer.<br />
10.1.3 Théorie <strong>de</strong> la décision face à l’incertain<br />
Nous venons d’exposer les bases <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> la décision face au risque, c’est-à-dire lorsque le<br />
déci<strong>de</strong>ur connait <strong>de</strong> façon objective les probabilités associées aux différents ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure Ω. Cependant,<br />
dans la plupart <strong>de</strong>s situ<strong>at</strong>ions économiques <strong>et</strong> financières, celles-ci ne sont que partiellement<br />
révélées voire totalement inconnues. Il convient donc <strong>de</strong> s’intéresser à c<strong>et</strong>te situ<strong>at</strong>ion, qualifiée <strong>de</strong> <strong>théorie</strong><br />
<strong>de</strong> la décision dans l’incertain, par opposition à la <strong>théorie</strong> <strong>de</strong> la décision face au risque, où l’on suppose<br />
données les probabilités sur les ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure.<br />
L’approche classique (ou bayésienne) <strong>de</strong> ce problème est celle <strong>de</strong> Savage (1954) qui consiste à réduire<br />
le problème <strong>de</strong> décision dans l’incertain à un problème <strong>de</strong> décision face au risque, à l’ai<strong>de</strong> <strong>de</strong> la notion<br />
<strong>de</strong> probabilités subjectives. Ces probabilités, dites subjectives, diffèrent <strong>de</strong>s probabilités objectives <strong>de</strong><br />
la même manière que les courses <strong>de</strong> chevaux diffèrent du jeu <strong>de</strong> roul<strong>et</strong>te au casino : à la roul<strong>et</strong>te, la<br />
table étant parfaitement équilibrée, tous les joueurs connaissent la probabilité que sorte le trois, rouge,<br />
impair <strong>et</strong> passe, alors qu’au tiercé, nul ne connaît avec exactitu<strong>de</strong> la probabilité que tel ou tel cheval a <strong>de</strong><br />
l’emporter.<br />
De plus, les probabilités objectives ont une interprét<strong>at</strong>ion très simple dans la mesure où elles sont reliées<br />
à la fréquence typique d’occurrence d’un événement. En eff<strong>et</strong>, la probabilité d’obtenir face en j<strong>et</strong>ant une<br />
pièce parfaitement équilibrée est <strong>de</strong> un <strong>de</strong>mi, tout simplement parce ce qu’en répétant un grand nombre<br />
<strong>de</strong> fois ce lancer <strong>de</strong> pièce, on observe que celle-ci tombe sur face la moitié du temps, <strong>et</strong> ce quelle que<br />
soit la personne effectuant les lancers. Donc, la probabilité objective est une propriété intrinsèque <strong>de</strong><br />
l’obj<strong>et</strong> (ici, la pièce) ou <strong>de</strong> l’événement (ici, tomber sur face) considéré. Au contraire, les probabilités<br />
subjectives mesurent un <strong>de</strong>gré <strong>de</strong> croyance en la vraisemblance d’un événement. Quelle est la probabilité<br />
qu’existe une vie extra-terrestre ? Nous ne pouvons pas faire d’expériences à ce suj<strong>et</strong>. Donc, la probabilité<br />
accordée à ce type d’événement ne peut être que fonction <strong>de</strong> l’opinion <strong>de</strong> chacun sur la question.<br />
Ainsi une probabilité subjective n’a pas une valeur unique <strong>et</strong> dépend <strong>de</strong> chaque individu. Pour autant, ces<br />
probabilités subjectives obéissent aux mêmes règles que les probabilités objectives en vertu du théorème<br />
du “ dutch book” (<strong>de</strong> Fin<strong>et</strong>ti 1937). Selon ce théorème, tout pari basé sur un ensemble <strong>de</strong> probabilités<br />
subjectives est équitable (<strong>et</strong> ne peut donc conduire à un gain certain) si <strong>et</strong> seulement si la probabilité subjective<br />
<strong>at</strong>tribuée à un événement certain vaut un, ainsi que la somme <strong>de</strong>s probabilités <strong>de</strong> <strong>de</strong>ux événements<br />
complémentaires.<br />
Moyennant c<strong>et</strong>te nouvelle interprét<strong>at</strong>ion <strong>de</strong>s probabilités, les problèmes <strong>de</strong> décision dans l’incertain<br />
peuvent être ramenés à <strong>de</strong> “simples” problèmes <strong>de</strong> décision face au risque, ce qui, outre l’axiome <strong>de</strong><br />
préordre total, repose selon Savage (1954), sur l’axiome suivant :<br />
AXIOME 5 (PRINCIPE DE LA CHOSE SÛRE)<br />
Etant donné un sous-ensemble ˜ Ω ∈ Ω <strong>de</strong> l’ensemble <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure <strong>et</strong> <strong>de</strong>s actifs X, X ′ , Y, Y ′ ∈ B<br />
tels que ∀ω ∈ ˜ Ω, X(ω) = X ′ (ω), Y (ω) = Y ′ (ω) <strong>et</strong> ∀ω ∈ ˜ Ω, X(ω) = Y (ω), X ′ (ω) = Y ′ (ω), alors<br />
X Y ⇐⇒ X ′ Y ′ . (10.12)<br />
Ceci signifie qu’une modific<strong>at</strong>ion commune <strong>de</strong> la partie commune <strong>de</strong> <strong>de</strong>ux actifs ne modifie pas l’ordre
10.1. La <strong>théorie</strong> <strong>de</strong> l’utilité 351<br />
<strong>de</strong>s préférences. Ajouté aux axiomes <strong>de</strong> comparabilité <strong>et</strong> <strong>de</strong> continuité, c<strong>et</strong> axiome perm<strong>et</strong> d’affirmer<br />
qu’il existe une unique probabilité P sur (Ω, F) <strong>et</strong> une fonction u(·) continue <strong>et</strong> croissante (définie à une<br />
fonction affine croissante près) telle que l’utilité U <strong>de</strong> l’actif X est donnée par<br />
U(X) = EP[u(X)], (10.13)<br />
où u(·) joue comme d’habitu<strong>de</strong> le rôle <strong>de</strong> fonction d’utilité dans le certain. C<strong>et</strong>te expression est analogue<br />
à celle obtenue par la <strong>théorie</strong> <strong>de</strong> l’utilité espérée <strong>de</strong> von Neumann <strong>et</strong> Morgenstern (1947), à la différence<br />
notoire qu’ici, la mesure <strong>de</strong> probabilité P est subjective <strong>et</strong> non pas objective.<br />
L’axiome <strong>de</strong> la chose sûre est extrêmement fort car il perm<strong>et</strong> <strong>de</strong> traiter tout problème <strong>de</strong> décision dans<br />
l’incertain comme un problème <strong>de</strong> décision face au risque. Ceci n’est en fait pas très réaliste, si bien<br />
que c<strong>et</strong>te approche est rej<strong>et</strong>ée sur le plan théorique comme sur le plan pr<strong>at</strong>ique. En eff<strong>et</strong>, <strong>de</strong> même<br />
qu’Allais (1953) avait prouvé que l’axiome d’indépendance était contredit empiriquement par la majorité<br />
<strong>de</strong>s agents, Ellsberg (1961) a pu montrer que dans <strong>de</strong>s situ<strong>at</strong>ions très simple <strong>de</strong> choix dans l’incertain<br />
l’axiome <strong>de</strong> la chose sûre ne résistait pas non plus à l’expérience. En fait, la plupart <strong>de</strong>s agents présentent<br />
une aversion pour l’ambiguïté, dans le sens où, pour une même mise, ils préférent parier pour ou contre<br />
un événement <strong>de</strong> probabilité P connue plutôt que pour ou contre un événement dont ils savent seulement<br />
que sa probabilité est comprise P − ε <strong>et</strong> P + ε. On peut alors montrer que les agents s<strong>at</strong>isfaisant au<br />
modèle <strong>de</strong> Savage (1954) sont indifférents à l’ambiguité, puisqu’ils ne peuvent pas faire la différence<br />
entre ces <strong>de</strong>ux paris.<br />
Pour dépasser c<strong>et</strong>te contradiction empirique il convient d’affaiblir l’axiome <strong>de</strong> la chose sûre. Pour cela,<br />
Schmeidler (1989) a proposé l’altern<strong>at</strong>ive suivante :<br />
AXIOME 6 (CHOSE SÛRE COMONOTONE)<br />
Soit une partition {Ωk} n k=1<br />
<strong>de</strong> l’ensemble <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure Ω <strong>et</strong> <strong>de</strong>ux actifs X, Y ∈ B dont les<br />
valeurs futures sont données par les variables alé<strong>at</strong>oires : X = (x1, Ω1; · · · ; xk, Ωk; · · · ; xn, Ωn) <strong>et</strong><br />
Y = (y1, Ω1; · · · ; yk, Ωk; · · · ; yn, Ωn), telles que x1 ≤ · · · ≤ xk ≤ · · · ≤ xn <strong>et</strong> y1 ≤ · · · ≤ yk ≤ · · · ≤<br />
dans les actifs X<br />
yn avec xk = yk. Soient alors les actifs X ′ , Y ′ ∈ B obtenus en remplaçant xk par x ′ k<br />
<strong>et</strong> Y <strong>de</strong> sorte que xk−1 ≤ x ′ k ≤ xk+1 <strong>et</strong> yk−1 ≤ x ′ k = y′ k ≤ yk+1. Alors<br />
X Y ⇐⇒ X ′ Y ′ . (10.14)<br />
C<strong>et</strong> axiome est très proche <strong>de</strong> l’axiome <strong>de</strong> la chose sûre comonotone dans le risque, si ce n’est que c<strong>et</strong>te<br />
fois, les probabilités pi <strong>de</strong>s ét<strong>at</strong>s Ωi ne sont pas connues.<br />
A l’ai<strong>de</strong> <strong>de</strong> cela, on montre qu’il existe non plus une unique probabilité P sur {Ω, F} mais une unique<br />
capacité 3 v sur {Ω, F} <strong>et</strong> une fonction u(·) croissante <strong>et</strong> continue (définie à une transform<strong>at</strong>ion affine<br />
3<br />
Une capacité v est une fonction d’ensemble <strong>de</strong> {Ω, F} dans [0, 1] telle que :<br />
– v(∅) = 0,<br />
– v(Ω) = 1,<br />
– ∀A, B ∈ F, A ⊂ B =⇒ v(A) ≤ v(B).<br />
Rappelons qu’une mesure <strong>de</strong> probabilité P (addititive) vérifierait en plus<br />
Une capacité est dite convexe si<br />
∀A, B ∈ F, P (A ∪ B) = P (A) + P (B) − P (A ∩ B).<br />
∀A, B ∈ F, v(A) + v(B) ≤ v(A ∪ B) + v(A ∩ B).<br />
Toute capacité convexe a un noyau non vi<strong>de</strong>, où le noyau <strong>de</strong> v est<br />
core(v) = {P ∈ P | ∀A ∈ F, P (A) ≥ v(A)} ,<br />
<strong>et</strong> P est l’ensemble <strong>de</strong>s mesures <strong>de</strong> probabilités additives sur {Ω, F}.
352 10. La mesure du risque<br />
croissante près) telles que l’utilité <strong>de</strong> l’actif X est<br />
<br />
U(X) =<br />
u(X) dv, (10.15)<br />
qui est une intégrale <strong>de</strong> Choqu<strong>et</strong> par rapport à la mesure non-additive (capacité) v. On peut noter la<br />
très forte ressemblance <strong>de</strong> c<strong>et</strong>te expression avec celle obtenue pour l’utilité dépendante du rang (cf<br />
équ<strong>at</strong>ion (10.10)). En fait, dans (10.10), l’expression ϕ◦P est une capacité. De plus, si dans le modèle <strong>de</strong><br />
Schmeidler (1989), il existe une probabilité objective P sur {Ω, F}, la capacité v peut s’exprimer comme<br />
v = ϕ ◦ P, où ϕ est unique <strong>et</strong> v est convexe si <strong>et</strong> seulement si ϕ l’est aussi.<br />
Dans le cadre <strong>de</strong> ce modèle, Montessano <strong>et</strong> Giovannoni (1996) définissent la notion d’aversion pour l’incertitu<strong>de</strong>,<br />
à savoir qu’un agent présente <strong>de</strong> l’aversion pour l’incertitu<strong>de</strong>, s’il existe une loi <strong>de</strong> probabilité<br />
P telle que quelque soit X ∈ B, u(X) dv ≤ EP[u(X)], ce qui implique que le noyau <strong>de</strong> la capacité<br />
v contient P <strong>et</strong> est donc non vi<strong>de</strong>. Réciproqement, on peut donc affirmer que tout agent caracterisé par<br />
une capacité convexe (donc à noyau non vi<strong>de</strong>) a <strong>de</strong> l’aversion pour l’incertitu<strong>de</strong>. Intuitivement, un agent<br />
averse à l’incertitu<strong>de</strong> affectera toujours un événement <strong>de</strong> la “probabilité” la moins favorable parmi toutes<br />
les probabilités <strong>at</strong>tribuées à c<strong>et</strong> événement par l’ensemble <strong>de</strong>s lois présentes dans le noyau <strong>de</strong> la capacité.<br />
Schmeidler (1986) fournit une interpr<strong>et</strong><strong>at</strong>ion <strong>de</strong> ce modèle en terme <strong>de</strong> croyances. En eff<strong>et</strong>, sous l’hypothèse<br />
que la capacité v est convexe, son noyau est non vi<strong>de</strong> <strong>et</strong><br />
∀X ∈ B,<br />
<br />
u(X) dv = min<br />
P∈core(v) EP[u(X)], (10.16)<br />
donc l’utilité U(X) est donnée par l’espérance minimale <strong>de</strong> u(X) calculée sur un ensemble <strong>de</strong> scénarii.<br />
Ceci a conduit Gilboa <strong>et</strong> Schmeidler (1989), Nakamura (1990), Ch<strong>at</strong>eauneuf (1991) ou encore Casa<strong>de</strong>sus-<br />
Masanell, Klibanoff <strong>et</strong> Oz<strong>de</strong>noren (2000) à développer <strong>de</strong>s modèles dits multi-prior, où les agents se<br />
donnent, a priori, un ensemble <strong>de</strong> distributions <strong>de</strong> probabilités P (ou scénarii) <strong>et</strong> définissent l’utilité<br />
comme<br />
∀X ∈ B, U(X) = min<br />
P∈P EP[u(X)]. (10.17)<br />
Il faut bien remarquer que les <strong>de</strong>ux approches ne sont pas équivalentes, car tout ensemble (fermé <strong>et</strong><br />
convexe) P n’est pas nécessairement le noyau d’une capacité v. De plus, c<strong>et</strong>te <strong>de</strong>rnière approche peut<br />
paraitre excessivement pessimiste puisqu’elle ne r<strong>et</strong>ient que la plus p<strong>et</strong>ite utilité possible, vu l’ensemble<br />
<strong>de</strong> scénarii considérés. En tout cas, elle est beaucoup plus pessimiste que l’utilité dérivée du modèle <strong>de</strong><br />
Schmeidler (1986) puisque Jaffray <strong>et</strong> Philippe (1997) ont montré que l’intégrale <strong>de</strong> Choqu<strong>et</strong> pouvait toujours<br />
s’exprimer comme la somme pondérée <strong>de</strong> <strong>de</strong>ux termes : le minimum <strong>et</strong> le maximum <strong>de</strong> l’espérance<br />
d’utilité par rapport à un ensemble <strong>de</strong> distributions <strong>de</strong> probabilité, le poids rel<strong>at</strong>if <strong>de</strong> ces <strong>de</strong>ux termes<br />
perm<strong>et</strong>tant <strong>de</strong> définir un indice <strong>de</strong> pessimisme <strong>de</strong> l’agent.<br />
Cependant, malgré les récentes avancées <strong>de</strong> la <strong>théorie</strong> <strong>de</strong> la décision que nous venons <strong>de</strong> présenter, il<br />
faut reconnaitre qu’une <strong>de</strong> ses limit<strong>at</strong>ions fondamentales <strong>de</strong>meure, à savoir comment m<strong>et</strong>tre en œuvre<br />
concrètement <strong>et</strong> pr<strong>at</strong>iquement c<strong>et</strong>te <strong>théorie</strong>. En eff<strong>et</strong>, dans la mesure où chaque agent possè<strong>de</strong> une fonction<br />
d’utilité différente, il est très difficile <strong>de</strong> déci<strong>de</strong>r <strong>de</strong> manière objective laquelle employer. Donc, il va<br />
s’avérer utile <strong>de</strong> considérer d’autres outils <strong>de</strong> décision <strong>et</strong> d’autres moyens <strong>de</strong> mesurer les risques.
10.2. Les mesures <strong>de</strong> risque cohérentes 353<br />
10.2 Les mesures <strong>de</strong> risque cohérentes<br />
10.2.1 Définition<br />
Selon Artzner <strong>et</strong> al. (1999), le risque associé aux vari<strong>at</strong>ions <strong>de</strong> la valeur d’une position est mesuré par la<br />
somme d’argent qui doit être investie dans un actif sans risque pour que dans le futur, c<strong>et</strong>te position reste<br />
acceptable, c’est-à-dire pour que les pertes éventuelles liées à la valeur future <strong>de</strong> la position ne m<strong>et</strong>tent pas<br />
en péril les proj<strong>et</strong>s du <strong>gestion</strong>naire <strong>de</strong> fond, <strong>de</strong> l’entreprise ou plus généralement la personne / l’organisme<br />
qui garantit la position. En ce sens, une mesure <strong>de</strong> risque constitue pour Artzner <strong>et</strong> al. (1999) une mesure<br />
<strong>de</strong> capital économique. Donc, la mesure <strong>de</strong> risque, notée ρ dans toute la suite, pourra être positive ou<br />
nég<strong>at</strong>ive selon que, respectivement, il faille augmenter la somme investie dans l’actif sans risque pour<br />
garantir la position risquée ou bien que l’on puisse réduire c<strong>et</strong>te somme tout en maintenant c<strong>et</strong>te garantie.<br />
Une mesure <strong>de</strong> risque sera dite cohérente au sens <strong>de</strong> Artzner <strong>et</strong> al. (1999) si elle vérifie les qu<strong>at</strong>res<br />
propriétés ou axiomes que nous allons exposer ci-<strong>de</strong>ssous. Décidons tout d’abord d’appeler G l’espace<br />
<strong>de</strong>s risques. Si l’espace <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure Ω est supposé fini (hypothèse faite par Artzner <strong>et</strong> al. (1999)),<br />
G est isomorphe à R N <strong>et</strong> une position risquée X n’est alors rien d’autre qu’un vecteur <strong>de</strong> R N . Une mesure<br />
<strong>de</strong> risque ρ est alors une applic<strong>at</strong>ion <strong>de</strong> R N dans R. Une généralis<strong>at</strong>ion à d’autres espaces <strong>de</strong> risque G a<br />
été proposée par Delbaen (2000).<br />
Soit donc une position risquée X <strong>et</strong> une somme d’argent α investie dans l’actif sans risque en début <strong>de</strong><br />
pério<strong>de</strong>, <strong>de</strong> sorte qu’en fin <strong>de</strong> pério<strong>de</strong>, le montant investi dans l’actif sans risque est α · (1 + r), où r<br />
représente le taux d’intérêt sans risque, alors :<br />
AXIOME 7 (INVARIANCE PAR TRANSLATION)<br />
∀X ∈ G <strong>et</strong> ∀α ∈ R, ρ(X + α · (1 + r)) = ρ(X) − α. (10.18)<br />
Ceci signifie simplement qu’investir une somme α dans l’actif sans risque, diminue le risque <strong>de</strong> la même<br />
quantité α. En particulier, pour toute position risquée X, ρ(X + ρ(X) · (1 + r)) = 0.<br />
Considérons maintenant <strong>de</strong>ux positions risquées X1 <strong>et</strong> X2, représentant par exemple les positions <strong>de</strong><br />
<strong>de</strong>ux tra<strong>de</strong>rs dans une salle <strong>de</strong> marché. Il est commo<strong>de</strong> pour le superviseur <strong>de</strong> c<strong>et</strong>te salle <strong>de</strong> marché que<br />
le risque agrégé <strong>de</strong> tous les tra<strong>de</strong>rs soit inférieur ou égal à la somme <strong>de</strong>s risques <strong>de</strong> chaque tra<strong>de</strong>r, donc<br />
en particulier il est souhaitable que le risque associé à la position (X1 + X2) soit inférieur ou égal à la<br />
somme <strong>de</strong>s risques associés aux positions X1 <strong>et</strong> X2 séparément :<br />
AXIOME 8 (SOUS-ADDITIVITÉ)<br />
∀(X1, X2) ∈ G × G, ρ(X1 + X2) ≤ ρ(X1) + ρ(X2). (10.19)<br />
De plus, la sous-additivité garantit qu’un <strong>gestion</strong>naire <strong>de</strong> <strong>portefeuille</strong> a interêt à agréger ses diverses<br />
positions afin d’en diminuer le risque par diversific<strong>at</strong>ion.<br />
Le troisième axiome est un axiome d’homogénéité (ou d’extensivité pour reprendre le language <strong>de</strong>s<br />
physiciens) :<br />
AXIOME 9 (POSITIVE HOMOGÉNÉITÉ)<br />
∀X ∈ G <strong>et</strong> ∀λ ≥ 0, ρ(λ · X) = λ · ρ(X), (10.20)
354 10. La mesure du risque<br />
ce qui signifie simplement que le risque d’une position croît avec la taille <strong>de</strong> c<strong>et</strong>te position, <strong>et</strong> plus<br />
précisément, le risque est ici supposé proportionnel à la taille <strong>de</strong> la position risquée. Nous reviendrons<br />
un peu plus loin sur l’hypothèse sous-jacente que suggère un tel axiome.<br />
Enfin, sachant que dans tous les ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure, le risque X conduit à une perte supérieure à l’actif Y<br />
(c’est-à-dire que toutes les composantes du vecteur X <strong>de</strong> R N sont toujours inférieures ou égales à celles<br />
du vecteur Y ), la mesure <strong>de</strong> risque ρ(X) doit être supérieure ou égale à ρ(Y ) :<br />
AXIOME 10 (MONOTONIE)<br />
∀X, Y ∈ G tel que X ≤ Y, ρ(X) ≥ ρ(Y ). (10.21)<br />
Ainsi posés, ces qu<strong>at</strong>re axiomes définissent ce que l’on appelle les mesures <strong>de</strong> risques cohérentes.<br />
10.2.2 Quelques exemples <strong>de</strong> mesures <strong>de</strong> risque cohérentes<br />
Beaucoup <strong>de</strong> mesures <strong>de</strong> risque communément utilisées aussi bien dans la recherche académique que par<br />
les professionels s’avèrent être non cohérentes au sens <strong>de</strong> Artzner <strong>et</strong> al. (1999). En eff<strong>et</strong>, il est évi<strong>de</strong>nt<br />
que la variance - dont l’utilis<strong>at</strong>ion comme mesure <strong>de</strong> risque remonte à Markovitz (1959) - ne s<strong>at</strong>isfait pas<br />
à l’axiome <strong>de</strong> monotonie. De même, il est aisé <strong>de</strong> montrer que la Value-<strong>at</strong>-Risk n’est généralement pas<br />
sous-additive. Nous rappelons que la Value-<strong>at</strong>-Risk, calculée au seuil <strong>de</strong> confiance α est définie par<br />
DÉFINITION 6 (VALUE-AT-RISK)<br />
Soit X ∈ G supposé à distribution continue. La Value-<strong>at</strong>-Risk, calculée au seuil <strong>de</strong> confiance α ∈ [0, 1]<br />
<strong>et</strong> notée VaRα est<br />
Pr[X + (1 + r) · VaRα ≥ 0] = α. (10.22)<br />
A ce suj<strong>et</strong>, la classe <strong>de</strong>s actifs dont la distribution jointe est elliptique constitue une exception notable<br />
pour laquelle Embrechts <strong>et</strong> al. (2002) ont montré que la VaR est sous-additive, <strong>et</strong> <strong>de</strong>meure donc une<br />
mesure cohérente <strong>de</strong> risque.<br />
Vue l’utilis<strong>at</strong>ion très répandue <strong>de</strong> la VaR dans le milieu professionel, il était souhaitable d’essayer <strong>de</strong><br />
construire une mesure <strong>de</strong> risque cohérente se rapprochant le plus possible <strong>de</strong> la VaR <strong>et</strong> qui <strong>de</strong> plus<br />
complète l’inform<strong>at</strong>ion fournie par celle-ci, à savoir : conditionnée au fait <strong>de</strong> subir une perte dépassant<br />
la VaR, quelle est, en moyenne, la perte observée. Ceci a conduit à la définition <strong>de</strong> l’Expected Shortfall :<br />
DÉFINITION 7 (EXPECTED SHORTFALL)<br />
Soit X ∈ G supposé à distribution continue. L’Expected Shortfall, calculée au niveau <strong>de</strong> confiance α est<br />
<br />
<br />
X <br />
ESα = −E <br />
X<br />
1 + r ≤ −V aRα . (10.23)<br />
1 + r<br />
Dans le cas où la distribution <strong>de</strong> X n’est pas supposée continue, l’expression <strong>de</strong> l’Expected Shortfall est<br />
un peu plus compliquée (voir Acerbi <strong>et</strong> Tasche (2002) ou Tasche (2002) par exemple).<br />
Lorsque l’on souhaite tenir compte <strong>de</strong>s grands risques, une autre approche, sur laquelle nous reviendrons<br />
en détail dans la section 10.3, consiste à prendre en compte l’eff<strong>et</strong> <strong>de</strong>s moments d’ordres supérieurs<br />
(à <strong>de</strong>ux). La question concernant la cohérence <strong>de</strong> ce type <strong>de</strong> mesures <strong>de</strong> risques contruites à partir <strong>de</strong>s<br />
moments est abordée par Delbaen (2000) <strong>et</strong> plus particulièrement par Fisher (2001). Ce <strong>de</strong>rnier montre<br />
que toute mesure <strong>de</strong> risque<br />
ρ(X) = −E[X] + a · σp, (10.24)
10.2. Les mesures <strong>de</strong> risque cohérentes 355<br />
où 0 ≤ a ≤ 1 <strong>et</strong><br />
σp = (E [max{(E[X] − X) p , 0}]) 1/p<br />
(10.25)<br />
est le semi-moment centré (inférieur) d’ordre p, est une mesure cohérente <strong>de</strong> risque. Plus généralement,<br />
du fait que toute somme convexe <strong>de</strong> mesures cohérentes <strong>de</strong> risque est une mesure cohérente <strong>de</strong> risque, la<br />
mesure<br />
∞<br />
ρ(X) = −E[X] + ap · σp, (10.26)<br />
avec<br />
est cohérente.<br />
∞<br />
p=1<br />
p=1<br />
ap ≤ 1 <strong>et</strong> ap ≥ 0, (10.27)<br />
Ceci perm<strong>et</strong> d’obtenir <strong>de</strong> façon simple les équivalents cohérents <strong>de</strong> certaines mesures <strong>de</strong> risque ou fonctions<br />
d’utilité. Par exemple, la mesure <strong>de</strong> risque ρ(X) = −E[X] + a · σ2 peut être considérée comme la<br />
généralis<strong>at</strong>ion cohérente <strong>de</strong> la fonction d’utilité moyenne-variance.<br />
10.2.3 Représent<strong>at</strong>ion <strong>de</strong>s mesures <strong>de</strong> risque cohérentes<br />
Les axiomes présentés au paragraphe 10.2.1 ainsi que les quelques exemples que nous venons <strong>de</strong> donner<br />
montrent qu’il existe une gran<strong>de</strong> variété <strong>de</strong> mesures cohérentes : les axiomes ne sont pas assez contraignants<br />
pour spécifier complétement une (unique) mesure <strong>de</strong> risque cohérente. Aussi est-il intéressant <strong>de</strong><br />
rechercher une représent<strong>at</strong>ion <strong>de</strong> ce type <strong>de</strong> mesures. Artzner <strong>et</strong> al. (1999) ont montré que<br />
THÉORÈME 2 (REPRÉSENTATION DES MESURES COHÉRENTES DE RISQUE)<br />
Une mesure <strong>de</strong> risque ρ est cohérente si <strong>et</strong> seulement si il existe une famille P <strong>de</strong> mesures <strong>de</strong> probabilité<br />
sur l’espace <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure telle que :<br />
ρ(X) = sup<br />
P∈P<br />
EP<br />
<br />
− X<br />
<br />
. (10.28)<br />
1 + r<br />
Ainsi, une mesure cohérente <strong>de</strong> risque apparait comme l’espérance <strong>de</strong> perte maximale sur un ensemble<br />
<strong>de</strong> scénarii réalisables. Il est alors évi<strong>de</strong>nt que plus l’ensemble <strong>de</strong> scénarii considérés sera grand, plus<br />
ρ(X) sera grand lui aussi, toutes choses égales par ailleurs. Ainsi, plus l’ensemble <strong>de</strong> scénarii considérés<br />
est grand, plus la mesure <strong>de</strong> risque est conserv<strong>at</strong>rice.<br />
C<strong>et</strong>te expression m<strong>at</strong>hém<strong>at</strong>ique n’est pas sans rappeler certaines formules que nous avons rencontrées<br />
concernant la <strong>théorie</strong> <strong>de</strong> la décision dans l’incertain (voir les équ<strong>at</strong>ions 10.16 <strong>et</strong> 10.17). Ceci est en<br />
fait très n<strong>at</strong>urel car les axiomes choisis par Artzner <strong>et</strong> al. (1999) pour définir les mesures <strong>de</strong> risques<br />
cohérentes sont tout-à-fait similaires à ceux dont découlent le modèle d’utilité <strong>de</strong> Schmeidler (1986). La<br />
légère différence entre les expressions (10.16) <strong>et</strong> (10.28), c’est-à-dire le changement du min en sup <strong>et</strong> le<br />
passage d’une fonction d’utilité u(·) croissante à la fonction décroissante − ·<br />
1+r<br />
vient simplement du fait<br />
que le déci<strong>de</strong>ur ou le <strong>gestion</strong>naire <strong>de</strong> risques tend à maximiser son utilité alors qu’il cherche à minimiser<br />
sa prise <strong>de</strong> risque. Ceci a comme conséquence que dans le modèle multi-prior, l’utilité est une quantité<br />
super-additive <strong>et</strong> non pas sous-additive comme les mesures <strong>de</strong> risque cohérentes. De plus, la spécific<strong>at</strong>ion<br />
<strong>de</strong> la fonction − ·<br />
1+r apparaissant dans (10.28) vient <strong>de</strong> l’axiome d’invariance par transl<strong>at</strong>ion qui impose<br />
que pour un investissement α dans l’actif sans risque ρ(α (1 + r)) = −α.<br />
Aussi générale soit-elle, la représent<strong>at</strong>ion <strong>de</strong>s mesures <strong>de</strong> risque cohérentes fournie par le théorème 2<br />
n’est pas d’une utilis<strong>at</strong>ion ou d’une mise en oeuvre très simple. En eff<strong>et</strong>, s’il est facile, dans la pr<strong>at</strong>ique,
356 10. La mesure du risque<br />
d’utiliser l’Expected Shortfall comme mesure cohérente <strong>de</strong> risque ou toute autre mesure cohérente ayant<br />
une expression analytique simple, comment doit-on choisir l’ensemble <strong>de</strong>s scénarii réalisables si l’on<br />
veut en rester au <strong>de</strong>gré <strong>de</strong> généralité proposé par le théorème <strong>de</strong> représent<strong>at</strong>ion ? C<strong>et</strong>te question n’adm<strong>et</strong><br />
guère <strong>de</strong> réponse s<strong>at</strong>isfaisante, <strong>et</strong> dans l’optique d’une mise en œuvre pr<strong>at</strong>ique, le point fondamental est<br />
plutôt <strong>de</strong> savoir si la mesure <strong>de</strong> risque que l’on souhaite utiliser peut être estimée à partir <strong>de</strong>s données<br />
empiriques. De telles mesure <strong>de</strong> risques sont dites law-invariant, <strong>et</strong> nous nous restreindrons désormais à<br />
la seule étu<strong>de</strong> <strong>de</strong> ce type <strong>de</strong> mesures cohérentes. Si <strong>de</strong> plus, on ne s’intéresse qu’aux mesures comonotoniquement<br />
additives 4 , on définit alors ce qu’Acerbi (2002) qualifie <strong>de</strong> mesures spectrales, pour lesquelles<br />
Kusuoka (2001) puis Tasche (2002) ont démontré le théorème <strong>de</strong> représent<strong>at</strong>ion suivant :<br />
THÉORÈME 3 (REPRÉSENTATION DES MESURES SPECTRALES)<br />
Soit F une fonction <strong>de</strong> distribution continue <strong>et</strong> convexe <strong>et</strong> un réel p ∈ [0, 1]. La mesure <strong>de</strong> risque ρ est une<br />
mesure spectrale, c’est-à-dire cohérente, law-invariant <strong>et</strong> comonotoniquement additive, si <strong>et</strong> seulement<br />
si elle adm<strong>et</strong> la représent<strong>at</strong>ion<br />
ρ(X) = p<br />
1<br />
0<br />
VaRu(X) F (du) + (1 − p)VaR1(X). (10.29)<br />
Si <strong>de</strong> plus F adm<strong>et</strong> une <strong>de</strong>nsité φ par rapport à la mesure <strong>de</strong> Lebesgue, (<strong>et</strong> en supposant p = 1 pour<br />
simplifier) alors<br />
ρ(X) =<br />
1<br />
0<br />
VaRu(X) φ(u) du, (10.30)<br />
<strong>et</strong> ρ(X) apparait comme la somme pondérée par φ(u) <strong>de</strong>s VaRu, ce qui justifie, suivant Acerbi (2002),<br />
que l’on puisse qualifier φ <strong>de</strong> “fonction d’aversion pour le risque”, puisque φ quantifie l’importance<br />
accordée aux différents niveaux <strong>de</strong> risque quantifiés par le seuil <strong>de</strong> confiance u. Dans le cas où φ(u) =<br />
α −1 · 1 (u
10.2. Les mesures <strong>de</strong> risque cohérentes 357<br />
Le premier point assure l’absence d’encaisse oisive. Le second est un principe <strong>de</strong> diversific<strong>at</strong>ion : mieux<br />
vaut agréger les risques. Le troisième point stipule que seul le risque est un paramètre pertinent pour<br />
l’alloc<strong>at</strong>ion <strong>de</strong> capital. Enfin, le qu<strong>at</strong>rième point exprime simplement que le capital investi dans l’actif<br />
sans risque ne peut l’être ailleurs.<br />
Considérant maintenant que la mesure <strong>de</strong> risque ρ est cohérente, Denault (2001) a prouvé, par analogie<br />
avec la <strong>théorie</strong> <strong>de</strong>s jeux, l’existence, <strong>et</strong> parfois l’unicité, d’alloc<strong>at</strong>ions cohérentes. Ces résult<strong>at</strong>s étant<br />
avant tout théorique <strong>et</strong> pour l’heure sans applic<strong>at</strong>ion pr<strong>at</strong>ique directe, nous n’irons pas plus avant sur<br />
c<strong>et</strong>te voie.<br />
En vue d’une mise en oeuvre pr<strong>at</strong>ique, mais aussi sur le plan théorique, il est intéressant <strong>de</strong> noter que <strong>de</strong><br />
par les axiomes <strong>de</strong> sous-additivité <strong>et</strong> <strong>de</strong> positive homogénéité, il est évi<strong>de</strong>nt que les mesures <strong>de</strong> risque<br />
cohérentes sont convexes, ce qui assure aux problèmes d’optimis<strong>at</strong>ion <strong>de</strong> <strong>portefeuille</strong> l’existence d’une<br />
unique solution optimale. Cependant, malgré le bon comportement m<strong>at</strong>hém<strong>at</strong>ique <strong>de</strong> la fonction à minimiser,<br />
la mise en oeuvre pr<strong>at</strong>ique se révèle parfois délic<strong>at</strong>e. Dans le cas particulier <strong>de</strong> l’Expected<br />
Shortfall, Pflug (2000) <strong>et</strong> Rockafellar <strong>et</strong> Uryasev (2000) ont établi un algorithme d’optimis<strong>at</strong>ion efficace<br />
par l’introdution <strong>de</strong> variables auxiliaires qui perm<strong>et</strong>tent <strong>de</strong> rendre le problème linéaire par morceau. C<strong>et</strong><br />
algorithme a ensuite été généralisé au cas <strong>de</strong>s mesures spectrales par Acerbi <strong>et</strong> Simon<strong>et</strong>ti (2002).<br />
Enfin, il peut être utile <strong>de</strong> calculer le risque marginal associé à chaque actif au sein d’un <strong>portefeuille</strong>. Etant<br />
donné N actifs X1, · · · , XN <strong>et</strong> w1, · · · , wN le nombre (ou le poids) <strong>de</strong> chaque actif dans le <strong>portefeuille</strong>,<br />
le risque marginal <strong>de</strong> l’actif i est la contribution d’une unité <strong>de</strong> c<strong>et</strong> actif au risque du <strong>portefeuille</strong>. D’après<br />
l’axiome d’homogénéité <strong>et</strong> en conséquence du théorème d’Euler sur les fonctions homogènes, le risque<br />
ρw = ρ(w1 · X1 + · · · + wN · XN) du <strong>portefeuille</strong> peut s’écrire<br />
ρw =<br />
N<br />
i=1<br />
wi · ∂ρw<br />
, (10.31)<br />
∂wi<br />
sous l’hypothèse que la mesure <strong>de</strong> risque est différentiable par rapport aux wi. En conséquence, il apparait<br />
clairement que la contribution marginale <strong>de</strong> l’actif i au risque du <strong>portefeuille</strong> est<br />
ρi = ∂ρw<br />
. (10.32)<br />
∂wi<br />
Ceci perm<strong>et</strong> en outre, généralisant l’approche <strong>de</strong> Gouriéroux, Laurent <strong>et</strong> Scaill<strong>et</strong> (2000) concernant la<br />
Value-<strong>at</strong>-Risk <strong>et</strong> <strong>de</strong> Scaill<strong>et</strong> (2000a) pour ce qui est <strong>de</strong> l’Expected-Shortfall, d’étudier la sensibilité du<br />
risque du <strong>portefeuille</strong> par rapport à l’alloc<strong>at</strong>ion <strong>de</strong>s différents actifs.<br />
10.2.5 Critique <strong>de</strong>s mesures cohérentes <strong>de</strong> risque<br />
Le mérite <strong>de</strong> l’axiom<strong>at</strong>ique <strong>de</strong>s mesures <strong>de</strong> risque cohérentes est <strong>de</strong> proposer une approche m<strong>at</strong>hém<strong>at</strong>ique<br />
rigoureuse <strong>de</strong> la problèm<strong>at</strong>ique associée à la notion <strong>de</strong> risque. Cependant, comme toute approche axiom<strong>at</strong>ique,<br />
certaines <strong>de</strong>s hypothèses <strong>de</strong> base peuvent être soumises à discussion. Si les axiomes d’invariance<br />
par transl<strong>at</strong>ion <strong>et</strong> <strong>de</strong> monotonie ne semblent pas souffrir la critique, on peut cependant légitimement<br />
s’interroger sur la validité <strong>et</strong> les limites <strong>de</strong>s <strong>de</strong>ux autres axiomes.<br />
Commencons par discuter l’axiome d’homogénéité. Comme nous l’avons précé<strong>de</strong>mment écrit, c<strong>et</strong> axiome<br />
n’est rien d’autre qu’une propriété d’extensivité <strong>de</strong> la mesure du risque : le risque associé à une position<br />
est proportionel à la taille <strong>de</strong> c<strong>et</strong>te position. Cependant, cela suppose qu’à tout moment l’on soit capable<br />
<strong>de</strong> liqui<strong>de</strong>r c<strong>et</strong>te position, quelle qu’en soit sa taille. Sous l’hypothèse d’un marché parfaitement liqui<strong>de</strong>,<br />
c<strong>et</strong>te hypothèse est tout-à-fait raisonnable. Mais, dès que l’on souhaite considérer un marché réel, <strong>et</strong> donc
358 10. La mesure du risque<br />
à liquidité limitée, il est alors bien évi<strong>de</strong>nt que la taille <strong>de</strong> la position <strong>de</strong>vient un facteur <strong>de</strong> risque, aucun<br />
marché n’étant en mesure d’absorber, sans fluctu<strong>at</strong>ion <strong>de</strong> cours, n’importe quelle taille d’ordre. Il peut<br />
donc sembler a priori raisonnable <strong>de</strong> reformuler l’axiome d’homogénéité <strong>de</strong> la façon suivante :<br />
AXIOME 11 (POSITIVE HOMOGÉNÉITÉ EN MARCHÉ ILLIQUIDE)<br />
Il existe une constante β > 1, telle que<br />
∀X ∈ G <strong>et</strong> ∀λ ≥ 0, ρ(λ · X) = λ β · ρ(X) , (10.33)<br />
<strong>et</strong> plus la constante β est gran<strong>de</strong>, plus les positions <strong>de</strong> gran<strong>de</strong>s tailles sont pénalisées.<br />
C<strong>et</strong> axiome suppose en fait seulement que l’impact <strong>de</strong> la liquidité limitée du marché est la même pour<br />
tous les actifs. Ceci n’est peut-être pas rigoureusement vrai, mais <strong>de</strong>meure une très bonne approxim<strong>at</strong>ion<br />
pour <strong>de</strong>s compagnies <strong>de</strong> taille comparable (Lillo <strong>et</strong> al. 2002).<br />
Il faut cependant noter qu’en tant que tel, c<strong>et</strong> axiome n’est pas comp<strong>at</strong>ible avec l’axiome d’invariance<br />
par transl<strong>at</strong>ion. En eff<strong>et</strong>, considérons le risque ρ(λ(X +α·(1+r))), avec X ∈ G <strong>et</strong> α <strong>et</strong> λ <strong>de</strong>ux réels. En<br />
appliquant tout d’abord l’axiome d’invariance par transl<strong>at</strong>ion, puis l’axiome d’homogénéité en marché<br />
illiqui<strong>de</strong>, on obtient :<br />
ρ(λ(X + α · (1 + r))) = ρ(λX + λα · (1 + r)), (10.34)<br />
= ρ(λX) − λα, (10.35)<br />
= λ β · ρ(X) − λα. (10.36)<br />
Si maintenant on fait usage <strong>de</strong> ces <strong>de</strong>ux axiomes dans l’ordre inverse, il vient :<br />
ρ(λ(X + α · (1 + r))) = λ β · ρ(X + α · (1 + r)), (10.37)<br />
= λ β · ρ(X) − λ β α, (10.38)<br />
ce qui est en contradiction avec le résult<strong>at</strong> précé<strong>de</strong>nt donné par l’équ<strong>at</strong>ion (10.36)<br />
Donc, si l’on veut restaurer la comp<strong>at</strong>iblité entre l’axiome d’invariance par transl<strong>at</strong>ion <strong>et</strong> l’axiome d’homogénéité<br />
en marche illiqui<strong>de</strong>, il faut restreindre ce <strong>de</strong>rnier aux positions purement risquées, ce qui<br />
revient à adm<strong>et</strong>tre la parfaite liquidité <strong>de</strong> l’actif sans risque. Une autre altern<strong>at</strong>ive consiste à modifier<br />
l’axiome d’invariance par transl<strong>at</strong>ion <strong>de</strong> sorte que pour un investissement <strong>de</strong> taille α dans l’actif sans<br />
risque ρ(α · (1 + r)) = −α β , auquel cas, on tient aussi compte du risque d’illiquidité pour l’actif sans<br />
risque.<br />
C<strong>et</strong>te approche n’est donc pas très s<strong>at</strong>isfaisante. Une meilleure solution a tout d’abord été proposée par<br />
He<strong>at</strong>h (2000) puis par Föllmer <strong>et</strong> Schied (2002a). Leur idée consiste pour commencer à remarquer que<br />
l’axiome <strong>de</strong> sous-additivité peut être remplacé par un axiome <strong>de</strong> convexité :<br />
AXIOME 12 (CONVEXITÉ)<br />
∀(X1, X2) ∈ G × G <strong>et</strong> ∀λ ∈ [0, 1], ρ(λ X1 + (1 − λ) X2) ≤ λ ρ(X1) + (1 − λ) ρ(X2). (10.39)<br />
C<strong>et</strong>te substitution est parfaitement légitime puisque pour les fonctions homogènes, convexité <strong>et</strong> sousadditivité<br />
sont équivalentes. Notons que comme l’axiome <strong>de</strong> sous-additivité, l’axiome <strong>de</strong> convexité garantit<br />
que l’agrég<strong>at</strong>ion <strong>de</strong> positions risquées assure leur diversific<strong>at</strong>ion.<br />
Ainsi, les mesures <strong>de</strong> risque cohérentes peuvent être définies par un ensemble d’axiomes équivalents à<br />
ceux énoncés en section 10.2.1 <strong>et</strong> qui sont les axiomes d’invariance par transl<strong>at</strong>ion, d’homogénéité, <strong>de</strong><br />
convexité <strong>et</strong> <strong>de</strong> monotonie.
10.3. Les mesures <strong>de</strong> fluctu<strong>at</strong>ions 359<br />
Pour prendre en compte le risque <strong>de</strong> liquidité, He<strong>at</strong>h (2000) <strong>et</strong> Föllmer <strong>et</strong> Schied (2002a) proposent<br />
<strong>de</strong> rej<strong>et</strong>er l’axiome d’homogénéité. Moyennant cela, ils définissent un nouvel ensemble <strong>de</strong> mesures <strong>de</strong><br />
risque dites convexes, qui englobent les mesures <strong>de</strong> risque cohérentes, <strong>et</strong> pour lesquelles ils donnent un<br />
théorème <strong>de</strong> représent<strong>at</strong>ion :<br />
THÉORÈME 4 (REPRÉSENTATION DES MESURES DE RISQUE CONVEXES)<br />
Une mesure <strong>de</strong> risque ρ est convexe si <strong>et</strong> seulement si il existe une famille Q <strong>de</strong> mesures <strong>de</strong> probabilité<br />
sur l’espace <strong>de</strong>s ét<strong>at</strong>s <strong>de</strong> la n<strong>at</strong>ure <strong>et</strong> une fonctionnelle α sur Q telle que :<br />
où la fonctionnelle α est donnée par<br />
<strong>et</strong><br />
ρ(X) = sup<br />
Q∈Q<br />
α(Q) = sup<br />
X∈Aρ<br />
(EQ [−X] − α(Q)) , (10.40)<br />
EQ[−X], (10.41)<br />
Aρ = {X ∈ G | ρ(X) ≤ 0} . (10.42)<br />
Dans l’énoncé du théorème, nous avons omis le facteur d’actualis<strong>at</strong>ion, afin d’alléger l’écriture, mais il<br />
peut être réintroduit <strong>de</strong> manière évi<strong>de</strong>nte.<br />
Ici encore, comme souligné par Föllmer <strong>et</strong> Schied (2002b), le lien avec la <strong>théorie</strong> <strong>de</strong> la décision dans<br />
l’incertain est immédi<strong>at</strong>, <strong>et</strong> perm<strong>et</strong> d’ancrer ces mesures <strong>de</strong> risque convexes dans la <strong>théorie</strong> <strong>de</strong> l’utilité <strong>et</strong><br />
ainsi <strong>de</strong> leur donner un sens économique très n<strong>et</strong>. Cependant, nous avons vu que c<strong>et</strong>te <strong>théorie</strong> conduisait<br />
à prendre <strong>de</strong>s décisions extrêmement pessismistes puisqu’elle ne r<strong>et</strong>ient que l’utilité minimale que peut<br />
r<strong>et</strong>irer un agent étant donné l’ensemble <strong>de</strong>s situ<strong>at</strong>ions qu’il considère. Ce même pessimisme excessif<br />
affecte les mesures <strong>de</strong> risque convexes (<strong>et</strong> donc cohérentes) puisque là aussi, le <strong>gestion</strong>naire n’est sensible<br />
qu’à la plus gran<strong>de</strong> perte qu’il peut subir.<br />
Enfin, il convient <strong>de</strong> signaler que la mesure du risque en terme <strong>de</strong> capital économique <strong>de</strong>meure insuffisante.<br />
Certes elle garantit, avec un certain niveau <strong>de</strong> confiance déterminé, que le <strong>portefeuille</strong> ou l’entreprise<br />
évitera la ruine, ce qui est fondamental du point <strong>de</strong> vue du régul<strong>at</strong>eur, mais si l’on se place du<br />
point <strong>de</strong> vue du <strong>gestion</strong>naire <strong>de</strong> fond ou d’un éventuel investisseur, cela ne suffit pas. Il faut aussi être<br />
capable <strong>de</strong> mesurer les fluctu<strong>at</strong>ions autour <strong>de</strong> l’objectif <strong>de</strong> rentabilité fixé, c’est-à-dire <strong>de</strong> la richesse du<br />
<strong>portefeuille</strong> autour <strong>de</strong> la richesse espérée (ou richesse moyenne). En eff<strong>et</strong>, la qualité d’un <strong>portefeuille</strong> se<br />
juge aussi à la régularité <strong>de</strong> ses performances.<br />
10.3 Les mesures <strong>de</strong> fluctu<strong>at</strong>ions<br />
Comme nous venons <strong>de</strong> l’exposer, la mesure du risque en terme <strong>de</strong> capital économique, pour nécessaire<br />
qu’elle soit - elle constitue en fait la première <strong>de</strong>s exigences - ne suffit cependant pas. Il semble en<br />
eff<strong>et</strong> tout-à-fait souhaitable <strong>de</strong> pouvoir mesurer les fluctu<strong>at</strong>ions d’un actif ou d’un portfeuille autour<br />
<strong>de</strong> sa valeur moyenne ou plus généralement autour d’un objectif <strong>de</strong> rentabilité préalablement établi.<br />
Les qualités du <strong>portefeuille</strong> seront alors d’autant meilleures que les fluctu<strong>at</strong>ions seront plus p<strong>et</strong>ites. Il<br />
convient donc <strong>de</strong> rechercher les propriétés minimales que doit s<strong>at</strong>isfaire une mesure <strong>de</strong> fluctu<strong>at</strong>ion ρ.<br />
Ceci est exposé dans Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2002c) que nous présenterons au chapitre 14 section 1.2,<br />
<strong>et</strong> dont nous donnons ici un résumé.<br />
En premier lieu, nous requèrons qu’une mesure <strong>de</strong> fluctu<strong>at</strong>ion soit positive :
360 10. La mesure du risque<br />
AXIOME 13 (POSITIVITÉ)<br />
Soit X ∈ G une gran<strong>de</strong>ur risquée, alors ρ(X) ≥ 0. De plus, ρ(X) = 0 si <strong>et</strong> seulement si X est non<br />
risquée (ou certain).<br />
En particulier, tout actif sans risque a une mesure <strong>de</strong> fluctu<strong>at</strong>ion égale à zéro, ce qui est bien n<strong>at</strong>urel. De<br />
plus, l’ajout d’une quantité certaine à une valeur risquée ne modifiant en rien les fluctu<strong>at</strong>ions <strong>de</strong> c<strong>et</strong>te<br />
<strong>de</strong>rnière, nous <strong>de</strong>vons avoir :<br />
AXIOME 14 (INVARIANCE PAR TRANSLATION)<br />
∀X ∈ G <strong>et</strong> ∀µ ∈ R, ρ(X + α) = ρ(X). (10.43)<br />
Enfin, nous <strong>de</strong>mandons que la mesure <strong>de</strong> fluctu<strong>at</strong>ion soit une fonction croissante, <strong>et</strong> plus spécifiquement<br />
homogéne, <strong>de</strong> la taille <strong>de</strong> la position :<br />
AXIOME 15 (POSITIVE HOMOGÉNÉITÉ)<br />
Il existe une constante β ≥ 1, telle que<br />
∀X ∈ G <strong>et</strong> ∀λ ≥ 0, ρ(λ · X) = λ β · ρ(X). (10.44)<br />
Dans le cas où β égale un, la mesure <strong>de</strong> fluctu<strong>at</strong>ions est extensive par rapport à la taille <strong>de</strong> la position<br />
mais ne prend alors pas en compte le risque <strong>de</strong> liquidité.<br />
Les mesures <strong>de</strong> fluctu<strong>at</strong>ions s<strong>at</strong>isfaisant les axiomes 14 <strong>et</strong> 15 sont connues sous le nom <strong>de</strong> semi-invariants.<br />
Ils en existent <strong>de</strong> très nombreux, parmi lesquels on peut citer par exemples les moments centrés<br />
ou les cumulants<br />
µn(X) = E [(X − E[X]) n ] , (10.45)<br />
Cn(X) = 1<br />
i n · n!<br />
d E eikX <br />
<br />
<br />
<br />
dk<br />
k=0<br />
. (10.46)<br />
L’utilis<strong>at</strong>ion <strong>de</strong>s moments centrés comme mesure du risque associé aux fluctu<strong>at</strong>ions d’un actif n’est pas<br />
nouvelle. Elle remonte au moins à Markovitz (1959) qui choisit d’utiliser la variance (moment centré<br />
d’ordre <strong>de</strong>ux) comme mesure du risque <strong>de</strong>s actifs financiers. Plus tard, Rubinstein (1973) montrera que<br />
les moments centrés d’ordre supérieur à <strong>de</strong>ux - pour autant qu’ils existent (cf. chapitres 1 <strong>et</strong> 3) - interviennent<br />
<strong>de</strong> façon n<strong>at</strong>urelle pour quantifier <strong>de</strong>s risques plus grands que ceux pris en compte par la<br />
variance, en les m<strong>et</strong>tant en rel<strong>at</strong>ion avec la <strong>théorie</strong> <strong>de</strong> l’utilité espérée <strong>de</strong> von Neumann <strong>et</strong> Morgenstern<br />
(1947). Dans le cas général, c’est-à-dire pour les cumulants ou pour toute autre mesure <strong>de</strong> fluctu<strong>at</strong>ions,<br />
il ne semble cependant pas que l’on puisse trouver <strong>de</strong> lien avec la <strong>théorie</strong> <strong>de</strong> l’utilité.<br />
L’axiome <strong>de</strong> positivité perm<strong>et</strong> <strong>de</strong> restreindre les semi-invariants acceptables. En eff<strong>et</strong>, par définition, les<br />
moments centrés d’ordre pairssont positifs, mais ce n’est pas nécessairement le cas pour ceux d’ordre<br />
impair. La situ<strong>at</strong>ion est beaucoup plus floue pour ce qui est <strong>de</strong>s cumulants, puisqu’aucun résult<strong>at</strong> général<br />
ne peut être donné concernant leur positivité. En fait tout dépend <strong>de</strong> la distribution <strong>de</strong> la variable alé<strong>at</strong>oire<br />
X.<br />
Partant <strong>de</strong>s moments centrés, il est en fait facile <strong>de</strong> construire une mesure <strong>de</strong> fluctu<strong>at</strong>ion qui s<strong>at</strong>isfait les<br />
trois axiomes, quelque soit la valeur <strong>de</strong> β. En eff<strong>et</strong>, il suffit <strong>de</strong> considérer les moments absolus centrés :<br />
<br />
¯µβ(X) = E |X − E[X]| β<br />
, (10.47)
10.4. Conclusion 361<br />
<strong>et</strong> <strong>de</strong> manière plus générale : ¯µp β/p . Ceci perm<strong>et</strong> d’ailleurs <strong>de</strong> construire <strong>de</strong> façon très simple d’autres<br />
mesures <strong>de</strong> fluctu<strong>at</strong>ions. En eff<strong>et</strong>, il est aisé <strong>de</strong> montrer que toute somme (positive mais non nécessairement<br />
convexe) <strong>de</strong> mesures <strong>de</strong> fluctu<strong>at</strong>ions <strong>de</strong> même <strong>de</strong>gré d’homogénéité β est une mesure <strong>de</strong> fluctu<strong>at</strong>ion <strong>de</strong><br />
<strong>de</strong>gré d’homogénéité β. Donc, dans l’esprit <strong>de</strong>s mesures spectrales d’Acerbi (2002), nous pouvons définir<br />
<br />
ρ(X) = dα φ(α) E [|X − E[X]| α ] β/α , (10.48)<br />
pourvu que l’intégrale existe. Là encore, la fonction φ perm<strong>et</strong> <strong>de</strong> quantifier l’aversion du <strong>gestion</strong>naire <strong>de</strong><br />
risque vis-à-vis <strong>de</strong>s gran<strong>de</strong>s fluctu<strong>at</strong>ions.<br />
10.4 Conclusion<br />
Les avancées récentes en m<strong>at</strong>ière <strong>de</strong> <strong>théorie</strong> <strong>de</strong> la décision <strong>et</strong> <strong>de</strong> mesure <strong>de</strong> risque, que nous venons<br />
d’exposer, nous perm<strong>et</strong>trons au chapitre suivant <strong>de</strong> montrer comment m<strong>et</strong>tre en œuvre une <strong>gestion</strong> <strong>de</strong><br />
<strong>portefeuille</strong> efficace vis-à-vis <strong>de</strong>s grands risques. Toutefois, il convient <strong>de</strong> gar<strong>de</strong>r à l’esprit qu’une <strong>de</strong>s<br />
limites les plus importantes <strong>de</strong>s mesures <strong>de</strong> risque que nous venons <strong>de</strong> présenter est <strong>de</strong> se restreindre à une<br />
prise en compte mono-périodique <strong>de</strong>s risques <strong>et</strong> <strong>de</strong> négliger l’approche inter-temporelle, qui est pourtant<br />
fondamentale dans la mesure où les contraintes <strong>de</strong> risque ont souvent pour objectif d’être intégrées dans<br />
un problème d’optimis<strong>at</strong>ion <strong>de</strong> <strong>portefeuille</strong> qui généralement est un problème dynamique.<br />
Malheureusement, l’étu<strong>de</strong> dynamique <strong>de</strong>s risques est beaucoup moins avancée car beaucoup plus délic<strong>at</strong>e<br />
à formaliser que l’étu<strong>de</strong> st<strong>at</strong>ique. On peut cependant présenter quelques tent<strong>at</strong>ives ayant permis d’appréhen<strong>de</strong>r<br />
ce problème. Tout d’abord citons les approches <strong>de</strong> Dacorogna, Gençay, Müller <strong>et</strong> Pict<strong>et</strong><br />
(2001) <strong>et</strong> Muzy, Sorn<strong>et</strong>te, Delour <strong>et</strong> Arnéodo (2001) notamment qui, partant <strong>de</strong> mesures <strong>de</strong> risques<br />
mono-périodiques, montrent qu’il est possible <strong>de</strong> construire <strong>de</strong>s mesures “multi-échelles” en moyennant<br />
les mesures <strong>de</strong> risque mono-périodiques calculées à différentes échelles <strong>de</strong> temps. Ceci peut paraître<br />
une métho<strong>de</strong> purement ad hoc, mais a d’une part révélé d’intéressants résult<strong>at</strong>s <strong>et</strong> d’autre part repose<br />
tout <strong>de</strong> même sur l’existence d’une casca<strong>de</strong> causale entre les différentes échelles temporelles (Arnéodo<br />
<strong>et</strong> al. 1998, Muzy <strong>et</strong> al. 2001). Plus généralement, Wang (1999) a proposé un ensemble d’axiomes que<br />
les mesures <strong>de</strong> risques dynamiques doivent s<strong>at</strong>isfaire <strong>et</strong> a complété ainsi les travaux précé<strong>de</strong>nts <strong>de</strong> Shapiro<br />
<strong>et</strong> Basak (2000) concernant la maximis<strong>at</strong>ion <strong>de</strong> l’utilité en temps continu ou <strong>de</strong> Ahn, Boudoukh,<br />
Richardson <strong>et</strong> Whitelaw (1999) sur l’optimis<strong>at</strong>ion <strong>de</strong> la VaR en temps continu. Enfin, signalons certaines<br />
mesures <strong>de</strong> risques telles les drawdowns, qui mesurent les pertes cumulées indépendamment <strong>de</strong><br />
leur durée, <strong>et</strong> qui ont suscité récemment un certain regain d’intérêt (Grossman <strong>et</strong> Zhou 1993, Cvitanic <strong>et</strong><br />
Kar<strong>at</strong>zas 1995, Chekhlov, Uryasev <strong>et</strong> Zabarankin 2000, Johansen <strong>et</strong> Sorn<strong>et</strong>te 2002).
362 10. La mesure du risque
Chapitre 11<br />
Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong><br />
marché<br />
On doit à Markovitz (1959) la première formalis<strong>at</strong>ion m<strong>at</strong>hém<strong>at</strong>ique <strong>de</strong> la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>. Celle-ci<br />
est basée sur la nécessité d’accepter un compromis entre l’obtention d’un ren<strong>de</strong>ment espéré le plus élevé<br />
possible <strong>et</strong> d’une quantité <strong>de</strong> risques encourus la plus faible possible, ce qui conduit tout n<strong>at</strong>urellement<br />
à s’intéresser aux <strong>portefeuille</strong>s dits optimaux, c’est-à-dire les <strong>portefeuille</strong>s tels que pour un niveau <strong>de</strong><br />
risque donné <strong>et</strong> un jeu <strong>de</strong> contraintes à s<strong>at</strong>isfaire - telle que l’absence <strong>de</strong> vente à découvert, par exemple<br />
- il n’existe pas <strong>de</strong> <strong>portefeuille</strong>s <strong>de</strong> ren<strong>de</strong>ment supérieur au ren<strong>de</strong>ment du <strong>portefeuille</strong> optimal. La courbe<br />
représentant l’ensemble <strong>de</strong>s <strong>portefeuille</strong>s optimaux dans le plan risque / ren<strong>de</strong>ment définit la frontière<br />
efficiente au-<strong>de</strong>ssus <strong>de</strong> laquelle aucun couple risque / ren<strong>de</strong>ment ne peut être <strong>at</strong>teint.<br />
Dans l’approche initiale <strong>de</strong> Markovitz (1959), le risque est mesuré par la variance (ou dévi<strong>at</strong>ion standard)<br />
du ren<strong>de</strong>ment <strong>de</strong>s actifs. Cela revient à adm<strong>et</strong>tre soit que leur distribution (multivariée) est gaussienne<br />
- puisque c’est le seul cas où la variance caractérise complètement les fluctu<strong>at</strong>ions <strong>de</strong>s actifs - soit que<br />
la fonction d’utilité <strong>de</strong>s agents est quadr<strong>at</strong>ique, auquel cas les décisions <strong>de</strong>s agents ne sont régies que<br />
par les <strong>de</strong>ux premiers moments <strong>de</strong> la distribution <strong>de</strong>s actifs. Moyennant ces hypothèses, il est possible<br />
<strong>de</strong> dériver analytiquement la composition <strong>de</strong>s <strong>portefeuille</strong>s efficients en fonction du seul vecteur <strong>de</strong>s<br />
ren<strong>de</strong>ments espérés <strong>de</strong>s actifs <strong>et</strong> <strong>de</strong> leur m<strong>at</strong>rice <strong>de</strong> covariance (Elton <strong>et</strong> Gruber 1995, par exemple).<br />
Cependant, il est désormais bien admis que la variance ne saurait suffire à quantifier convenablement les<br />
risques puisqu’elle ne rend compte que <strong>de</strong>s p<strong>et</strong>ites fluctu<strong>at</strong>ions du ren<strong>de</strong>ment <strong>de</strong>s actifs autour <strong>de</strong> leur<br />
valeur moyenne, négligeant ainsi totalement les grands risques dont l’impact est généralement le plus<br />
conséquent. Aussi, est-il important <strong>de</strong> se tourner vers d’autres mesures <strong>de</strong> risques ou d’autres critères<br />
d’optimis<strong>at</strong>ion.<br />
L’une <strong>de</strong>s premières altern<strong>at</strong>ives proposées à l’analyse moyenne-variance a été <strong>de</strong> considérer les <strong>portefeuille</strong>s<br />
dont la moyenne non pas arithmétique mais géométrique <strong>de</strong>s ren<strong>de</strong>ments est la plus gran<strong>de</strong>,<br />
car ce sont eux qui ont la probabilité la plus élevée <strong>de</strong> dépasser un niveau donné <strong>de</strong> rentabilité, quelque<br />
soit l’intervalle <strong>de</strong> temps considéré (Brieman 1960, Hakansson 1971, Roll 1973). C<strong>et</strong>te approche n’est<br />
cependant comp<strong>at</strong>ible avec la <strong>théorie</strong> <strong>de</strong> la décision que pour <strong>de</strong>s fonctions d’utilité logarithmiques. Une<br />
secon<strong>de</strong> altern<strong>at</strong>ive, dite “Saf<strong>et</strong>y First”, consiste à m<strong>et</strong>tre l’accent sur les pertes qu’encourt le <strong>portefeuille</strong>.<br />
De nombreux critères ont vu le jour, tel le critère <strong>de</strong> Roy (1952) qui consiste à minimiser la probabilité<br />
<strong>de</strong> subir une perte supérieure à une valeur prédéterminée, ou encore les critères <strong>de</strong> K<strong>at</strong>aoka ou <strong>de</strong> Telser<br />
(voir Elton <strong>et</strong> Gruber (1995)) qui sont en fait très proches <strong>de</strong> critères d’optimis<strong>at</strong>ion sous contrainte <strong>de</strong><br />
VaR. De manière générale, toute optimis<strong>at</strong>ion sous contrainte <strong>de</strong> capital économique <strong>et</strong> donc utilisant notamment<br />
les mesures <strong>de</strong> risques cohérentes s’apparente à c<strong>et</strong>te approche. Enfin, une troisième altern<strong>at</strong>ive<br />
363
364 11. Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong> marché<br />
consiste, pour palier directement aux limit<strong>at</strong>ions <strong>de</strong> l’approche moyenne-variance (Samuelson 1958), à<br />
prendre en compte l’eff<strong>et</strong> <strong>de</strong>s moments d’ordre supérieur, tels la skewness (Arditti 1967, Krauss <strong>et</strong> Lintzenburger<br />
1976) ou plus généralement les mesures <strong>de</strong> fluctu<strong>at</strong>ions que nous avons présentées au chapitre<br />
précé<strong>de</strong>nt.<br />
Outre l’intérêt pr<strong>at</strong>ique <strong>de</strong>s métho<strong>de</strong>s d’optimis<strong>at</strong>ion <strong>de</strong> <strong>portefeuille</strong>s, celles-ci ont aussi un intérêt théorique,<br />
dans la mesure où la composition <strong>de</strong>s <strong>portefeuille</strong>s optimaux perm<strong>et</strong> <strong>de</strong> déduire <strong>de</strong>s rel<strong>at</strong>ions entre les<br />
prix <strong>de</strong>s actifs <strong>et</strong> le prix du <strong>portefeuille</strong> <strong>de</strong> marché à l’équilibre. Ceci perm<strong>et</strong> alors <strong>de</strong> généraliser le CAPM<br />
<strong>de</strong> Sharpe (1964), Lintner (1965) <strong>et</strong> Mossin (1966) dérivé dans un univers <strong>de</strong> Markovitz <strong>et</strong> donc soumis<br />
aux mêmes limites quant à sa validité.<br />
11.1 Les limites <strong>de</strong> l’approche moyenne - variance<br />
Dans l’approche <strong>de</strong> Markovitz (1959), le vecteur <strong>de</strong>s ren<strong>de</strong>ments espérés <strong>et</strong> la m<strong>at</strong>rice <strong>de</strong> corrél<strong>at</strong>ion<br />
jouent un rôle crucial. Or, si l’estim<strong>at</strong>ion <strong>de</strong>s ren<strong>de</strong>ments moyens peut être réalisée avec une assez bonne<br />
précision dans la mesure où les distributions <strong>de</strong>s actifs décroissent plus vite qu’une loi <strong>de</strong> puissance d’exposant<br />
<strong>de</strong>ux (cf chapitre 1), l’estim<strong>at</strong>ion <strong>de</strong> la m<strong>at</strong>rice <strong>de</strong> covariance pose beaucoup plus <strong>de</strong> problèmes,<br />
car son estim<strong>at</strong>ion correcte nécessite que la distribution <strong>de</strong> ren<strong>de</strong>ment décroisse plus rapi<strong>de</strong>ment qu’une<br />
loi <strong>de</strong> puissance d’exposant qu<strong>at</strong>re dans la région intermédiaire, ce qui n’est pas le cas.<br />
En eff<strong>et</strong>, lorsque l’on s’intéresse à <strong>de</strong>s <strong>portefeuille</strong>s <strong>de</strong> gran<strong>de</strong> taille, la m<strong>at</strong>rice <strong>de</strong> covariance empirique<br />
est à telle point bruitée, que ses propriétés sont très proches <strong>de</strong>s propriétés universelles <strong>de</strong> certains<br />
ensembles <strong>de</strong> m<strong>at</strong>rices alé<strong>at</strong>oires. Plus précisément, Laloux, Cizeau, Bouchaud <strong>et</strong> Potters (1999), Laloux,<br />
Cizeau, Bouchaud <strong>et</strong> Potters (2000) ou encore Plerou, Gopikrishnan, Rosenow, Amaral <strong>et</strong> Stanley<br />
(1999) ont montré que les distributions <strong>de</strong> valeurs propres <strong>et</strong> vecteurs propres <strong>de</strong> ces m<strong>at</strong>rices, ainsi<br />
que la distribution <strong>de</strong>s écarts entre valeurs propres, étaient très proches <strong>de</strong> celles <strong>de</strong>s m<strong>at</strong>rices <strong>de</strong> l’ensemble<br />
<strong>de</strong> Wishart (1928), à l’exception <strong>de</strong>s quelques plus gran<strong>de</strong>s valeurs propres qui semblent pouvoir<br />
être associées à <strong>de</strong>s facteurs comme le marché ou certains secteurs d’activité. Cependant, comme<br />
suggéré par les résult<strong>at</strong>s <strong>de</strong> Meerschaert <strong>et</strong> Scheffler (2001), l’ensemble <strong>de</strong> Whishart n’est peut-être pas<br />
le plus adapté. En eff<strong>et</strong>, l’ensemble <strong>de</strong> Whishart est l’ensemble <strong>de</strong>s m<strong>at</strong>rices <strong>de</strong> corrél<strong>at</strong>ions empiriques<br />
dérivées d’échantillons gaussiens. Or, si l’on adm<strong>et</strong> que les queues <strong>de</strong> distributions <strong>de</strong>s ren<strong>de</strong>ments sont<br />
en lois <strong>de</strong> puissance (ou régulièrement variables) d’exposant <strong>de</strong> queue inférieur à qu<strong>at</strong>re, les ensembles<br />
<strong>de</strong> m<strong>at</strong>rice <strong>de</strong> Lévy (Burda, Janik, Jurkiewicz, Nowak, Papp <strong>et</strong> Zahed 2002) semblent plus indiqués. Les<br />
résult<strong>at</strong>s <strong>de</strong> Burda, Jurkiewicz, Nowak, Papp <strong>et</strong> Zahed (2001a) <strong>et</strong> Burda, Jurkiewicz, Nowak, Papp <strong>et</strong> Zahed<br />
(2001b) montrent alors que les m<strong>at</strong>rices <strong>de</strong> covariances estimées sont encore plus bruitées que prévu<br />
par comparaison à l’ensemble <strong>de</strong> Whishart, puisque même les plus gran<strong>de</strong>s valeurs ne paraissent pas signific<strong>at</strong>ives.<br />
En conséquence, le contenu <strong>de</strong>s m<strong>at</strong>rices <strong>de</strong> covariances estimées <strong>de</strong> gran<strong>de</strong>s tailles semble<br />
peu inform<strong>at</strong>if. En fait, nous montrons en annexe <strong>de</strong> ce chapitre que les plus gran<strong>de</strong>s valeurs propres<br />
peuvent être estimées avec une bonne précision tandis que le cœur <strong>de</strong> la distribution <strong>de</strong> valeurs propres<br />
s’écarte sensiblement <strong>de</strong> la distribution <strong>de</strong> Wishart. En outre, nous justifions, à l’ai<strong>de</strong> <strong>de</strong> la <strong>théorie</strong> <strong>de</strong>s<br />
m<strong>at</strong>rices alé<strong>at</strong>oires, comment émerge <strong>de</strong> tout système <strong>de</strong> gran<strong>de</strong> taille, dont la corrél<strong>at</strong>ion moyenne entre<br />
éléments est non nulle, un facteur (ou valeur propre) dominant qui perm<strong>et</strong> à lui seul <strong>de</strong> rendre compte<br />
<strong>de</strong> la plus gran<strong>de</strong> partie <strong>de</strong>s interactions (ou corrél<strong>at</strong>ions) entre les éléments du système (Malevergne <strong>et</strong><br />
Sorn<strong>et</strong>te 2002a).<br />
Ceci a d’importantes conséquences quant à la composition <strong>et</strong> la stabilité dans le temps <strong>de</strong>s <strong>portefeuille</strong>s<br />
optimaux obtenus à l’ai<strong>de</strong> <strong>de</strong> ces m<strong>at</strong>rices <strong>de</strong> corrél<strong>at</strong>ions empiriques (Rosenow, Plerou, Gopikrishnan<br />
<strong>et</strong> Stanley 2001). En fait, l’importance du bruit dépend du contexte. D’une part, elle semble assez faible<br />
pour les <strong>portefeuille</strong>s optimisés sous contraintes linéaires plutôt que sous contraintes non linéaires (Pafka
11.2. Prise en compte <strong>de</strong>s grands risques 365<br />
<strong>et</strong> Kondor 2001). D’autre part, c<strong>et</strong>te influence dépend du rapport r = N/T , où N est le nombre d’actifs<br />
dans le <strong>portefeuille</strong> <strong>et</strong> T la taille <strong>de</strong>s séries temporelles ayant servi à estimer la m<strong>at</strong>rice <strong>de</strong> covariance.<br />
Pafka <strong>et</strong> Kondor (2002) ont montré que pour un rapport supérieur ou <strong>de</strong> l’ordre <strong>de</strong> 0.6, le bruit à une<br />
influence primordiale, alors que pour r inférieur à 0.2, son impact <strong>de</strong>vient négligeable. Ceci implique<br />
qu’il est nécessaire <strong>de</strong> disposer <strong>de</strong> séries temporelles dont la longueur est au moins cinq fois supérieure à<br />
la taille du <strong>portefeuille</strong> considéré. Ainsi, pour un <strong>portefeuille</strong> d’une centaine d’actifs gérés sur la base <strong>de</strong><br />
données journalières, cela requièrt <strong>de</strong>s échantillons <strong>de</strong> cinq cents points, soit <strong>de</strong>ux années <strong>de</strong> cot<strong>at</strong>ions, ce<br />
qui reste très raisonnable. En revanche, pour un <strong>portefeuille</strong> d’un millier d’actifs, il faut alors <strong>de</strong>s séries<br />
temporelles <strong>de</strong> cinq mille points, ce qui représente une vingtaine d’années <strong>de</strong> cot<strong>at</strong>ions, <strong>et</strong> se posent alors<br />
d’autres problèmes tels que celui <strong>de</strong> la st<strong>at</strong>ionnarité <strong>de</strong> ces données.<br />
Quand bien même ces problèmes pr<strong>at</strong>iques ne se poseraient pas, il faut gar<strong>de</strong>r à l’esprit que l’approche<br />
moyenne - variance <strong>de</strong> Markovitz (1959) ne se révèle adaptée que dans l’hypothèse où les actifs sont<br />
conjointement gaussiens ou dans la mesure où les agents forment <strong>de</strong>s préférences quadr<strong>at</strong>iques dans leur<br />
richesse, ce qui dans un cas comme dans l’autre traduit que l’on ne s’intéresse qu’à <strong>de</strong>s risques <strong>de</strong> faibles<br />
amplitu<strong>de</strong>s. Dès lors que l’on souhaite intégrer d’autre dimension du risque, c’est-à-dire <strong>de</strong>s risques<br />
associés à <strong>de</strong> gran<strong>de</strong>s pertes ou <strong>de</strong> gran<strong>de</strong>s fluctu<strong>at</strong>ions, il convient <strong>de</strong> se tourner vers d’autres mesures<br />
<strong>de</strong> risques que la variance.<br />
11.2 Prise en compte <strong>de</strong>s grands risques<br />
L’utilis<strong>at</strong>ion <strong>de</strong> nouvelles mesures <strong>de</strong> risques perm<strong>et</strong>tant <strong>de</strong> prendre en compte les grands risques est<br />
nécessaire à l’obtention <strong>de</strong> <strong>portefeuille</strong>s moins sensibles que les <strong>portefeuille</strong>s moyenne - variance aux<br />
gran<strong>de</strong>s vari<strong>at</strong>ions <strong>de</strong> cours. Pour cela, nous <strong>de</strong>vons nous intéresser à <strong>de</strong>s mesures <strong>de</strong> risques accordant<br />
plus d’importance aux événements rares <strong>et</strong> <strong>de</strong> gran<strong>de</strong>s amplitu<strong>de</strong>s. De telles mesures ont été présentées<br />
au chapitre précé<strong>de</strong>nt, <strong>et</strong> selon Tasche (2000) <strong>et</strong> Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2002c) peuvent être divisées en<br />
<strong>de</strong>ux classes :<br />
– premièrement les mesures <strong>de</strong> risques associées au capital économique pour lesquelles peuvent être<br />
requises <strong>de</strong>s conditions <strong>de</strong> cohérence (Artzner <strong>et</strong> al. 1999) ou <strong>de</strong> convexité (He<strong>at</strong>h 2000, Föllmer <strong>et</strong><br />
Schied 2002a),<br />
– <strong>et</strong> <strong>de</strong>uxièmement les mesures <strong>de</strong>s fluctu<strong>at</strong>ions du ren<strong>de</strong>ment autour <strong>de</strong> sa valeur espérée (Malevergne <strong>et</strong><br />
Sorn<strong>et</strong>te 2002c), ce qui historiquement est la première approche à avoir vu le jour puisque la variance<br />
est une mesure <strong>de</strong> fluctu<strong>at</strong>ion <strong>et</strong> pas une mesure <strong>de</strong> capital économique.<br />
En fait, ces <strong>de</strong>ux classes ne suffisent pas à englober toutes les mesures <strong>de</strong> grands risques : en restant dans<br />
un cadre strictement mono - périodique, on peut au moins citer le coefficient <strong>de</strong> dépendance <strong>de</strong> queue qui<br />
perm<strong>et</strong> <strong>de</strong> quantifier les co-mouvements extrêmes entre actifs (Malevergne <strong>et</strong> Sorn<strong>et</strong>te 2002b), ou encore<br />
la “covariance <strong>de</strong> queue” utilisée par Bouchaud, Sorn<strong>et</strong>te, Walter <strong>et</strong> Aguilar (1998) pour quantifier le<br />
risque d’un <strong>portefeuille</strong> d’actifs dont les ren<strong>de</strong>ments suivent <strong>de</strong>s lois <strong>de</strong> puissances, généralisant ainsi<br />
l’approche <strong>de</strong> Fama (1965b) valable uniquement pour <strong>de</strong>s actifs distribués selon <strong>de</strong>s lois stables. Si l’on<br />
s’autorise à considérer <strong>de</strong>s mesures inter - temporelles, les “drawdowns” (Chekhlov <strong>et</strong> al. 2000), par<br />
exemple peuvent être pris en compte.<br />
11.2.1 Optimis<strong>at</strong>ion sous contrainte <strong>de</strong> capital économique<br />
L’optimis<strong>at</strong>ion d’un <strong>portefeuille</strong> par rapport à <strong>de</strong>s contraintes portant sur le capital économique conduit<br />
n<strong>at</strong>urellement à s’intéresser aux <strong>portefeuille</strong>s VaR-efficients (Consigli 2002, Huisman, Koedijk <strong>et</strong> Pownall<br />
2001, Kaplanski <strong>et</strong> Kroll 2001a) ou Expected Shortfall-efficients (Frey <strong>et</strong> McNeil 2002). En eff<strong>et</strong>,
366 11. Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong> marché<br />
ces <strong>de</strong>ux mesures <strong>de</strong> risques sont, à l’heure actuelle, les plus employées. Comme il est légitime <strong>de</strong> le<br />
penser, l’alloc<strong>at</strong>ion optimale selon <strong>de</strong>s critères moyenne - VaR ou moyenne - ES est très différente <strong>de</strong><br />
celle obtenue selon le critère moyenne variance (Alexan<strong>de</strong>r <strong>et</strong> Baptista 2002), ce <strong>de</strong>rnier n’étant pas à<br />
même <strong>de</strong> tenir compte <strong>de</strong>s grands risques.<br />
D’un point <strong>de</strong> vue pr<strong>at</strong>ique, l’optimis<strong>at</strong>ion selon <strong>de</strong>s critères <strong>de</strong> VaR est délic<strong>at</strong>e pour <strong>de</strong>ux raisons.<br />
D’une part du fait <strong>de</strong> sa non convexité, qui impose d’avoir recours à <strong>de</strong>s algorithmes <strong>de</strong> minimis<strong>at</strong>ion<br />
non standards (algorithmes génétiques, par exemple) <strong>et</strong> d’autre part, son estim<strong>at</strong>ion dans un cadre non<br />
- paramétrique est généralement difficile car sensible à la métho<strong>de</strong> utilisée <strong>et</strong> grosse consomm<strong>at</strong>rice<br />
<strong>de</strong> temps <strong>de</strong> calcul 1 . Aussi <strong>de</strong>s métho<strong>de</strong>s <strong>de</strong> calculs approxim<strong>at</strong>ifs (Tasche <strong>et</strong> Tibil<strong>et</strong>ti 2001) basées<br />
notamment sur l’applic<strong>at</strong>ion <strong>de</strong> la <strong>théorie</strong> <strong>de</strong>s valeurs extrêmes (Longin 2000, Danielson <strong>et</strong> <strong>de</strong> Vries<br />
2000, Consigli, Frascella <strong>et</strong> Sartorelli 2001) ou <strong>de</strong>s approches paramétriques (Malevergne <strong>et</strong> Sorn<strong>et</strong>te<br />
2002d), prennent tout leur sens. L’Expected-Shortfall présente quant à elle l’avantage <strong>de</strong> s<strong>at</strong>isfaire aux<br />
contraintes <strong>de</strong> cohérences 2 <strong>de</strong> Artzner <strong>et</strong> al. (1999). Ainsi, elle conduit à <strong>de</strong>s problèmes d’optimis<strong>at</strong>ion<br />
bien conditionnés pour lesquels <strong>de</strong>s algorithmes <strong>de</strong> minimis<strong>at</strong>ion particulièrement simples <strong>et</strong> efficaces<br />
existent (Rockafellar <strong>et</strong> Uryasev 2002).<br />
11.2.2 Optimis<strong>at</strong>ion sous contrainte <strong>de</strong> fluctu<strong>at</strong>ions autour du ren<strong>de</strong>ment espéré<br />
Le capital économique n’est pas la seule quantité à minimiser <strong>et</strong> les fluctu<strong>at</strong>ions du <strong>portefeuille</strong> autour <strong>de</strong><br />
son ren<strong>de</strong>ment moyen ou <strong>de</strong> tout autre objectif <strong>de</strong> rentabilité sont aussi à prendre en compte. La variance<br />
réalise cela, mais elle se focalise uniquement sur les écarts à la moyenne <strong>de</strong> faibles amplitu<strong>de</strong>s <strong>et</strong> néglige<br />
donc complètement les grands risques. C’est pourquoi il convient d’utiliser d’autres quantités partageant<br />
certaines <strong>de</strong>s propriétés <strong>de</strong> la variance mais m<strong>et</strong>tant l’emphase sur les gran<strong>de</strong>s fluctu<strong>at</strong>ions. Rubinstein<br />
(1973) fut l’un <strong>de</strong>s premiers à s’intéresser à c<strong>et</strong>te approche, suggérant que les moments centrés d’ordre<br />
supérieur à <strong>de</strong>ux ne <strong>de</strong>vaient être négligés puisqu’ils apparaissent n<strong>at</strong>urellement dans le développement<br />
en série <strong>de</strong> la fonction d’utilité. Plus récemment, Sorn<strong>et</strong>te, An<strong>de</strong>rsen <strong>et</strong> Simon<strong>et</strong>ti (2000) <strong>et</strong> Sorn<strong>et</strong>te,<br />
Simon<strong>et</strong>ti <strong>et</strong> An<strong>de</strong>rsen (2000) ont émis l’idée que les cumulants pouvaient aussi fournir d’utiles mesures<br />
<strong>de</strong> fluctu<strong>at</strong>ions, perm<strong>et</strong>tant notamment <strong>de</strong> rendre compte du comportement <strong>de</strong> certains agents globalement<br />
risquophobes dans le sens où ils cherchent à éviter les grands risques mais sont près à accepter un<br />
certain niveau <strong>de</strong> p<strong>et</strong>its risques (An<strong>de</strong>rsen <strong>et</strong> Sorn<strong>et</strong>te 2002a, Malevergne <strong>et</strong> Sorn<strong>et</strong>te 2002c). Dans le cas<br />
où les distributions marginales <strong>de</strong>s actifs suivent <strong>de</strong>s lois exponentielles étirées (cf chapitre 3) <strong>et</strong> où la<br />
copule décrivant leur dépendance est gaussienne (cf chapitres 7 <strong>et</strong> 8), <strong>de</strong>s expressions analytiques ont pu<br />
être dérivées pour l’expression <strong>de</strong>s moments <strong>et</strong> cumulants <strong>de</strong>s <strong>portefeuille</strong>s constitués <strong>de</strong> tels actifs (voir<br />
chapitre 14).<br />
Ces quelques exemples sont <strong>de</strong>s cas particuliers <strong>de</strong>s mesures <strong>de</strong> fluctu<strong>at</strong>ions définies au chapitre précé<strong>de</strong>nt<br />
<strong>et</strong> perm<strong>et</strong>tent <strong>de</strong> dériver simplement la plupart <strong>de</strong>s propriétés générales <strong>de</strong>s <strong>portefeuille</strong>s optimaux. Ces<br />
propriétés sont en fait <strong>de</strong>s généralis<strong>at</strong>ions immédi<strong>at</strong>es, aux cas <strong>de</strong>s grands risques, <strong>de</strong>s propriétés dont<br />
jouissent les <strong>portefeuille</strong>s moyenne - variance efficients.<br />
11.2.3 Optimis<strong>at</strong>ion sous d’autres contraintes<br />
Lorsque l’on souhaite considérer les risques extrêmes systém<strong>at</strong>iques, c’est-à-dire les mouvements extrêmes<br />
que subissent les actifs conjointement avec le marché, le coefficient <strong>de</strong> dépendance <strong>de</strong> queue λ peut<br />
1 Ce point est discuté en détail dans Chabaane, Duclos, Laurent, Malevergne <strong>et</strong> Turpin (2002)<br />
2 Voir Acerbi <strong>et</strong> Tasche (2002) pour une discussion <strong>de</strong>s propriétés <strong>de</strong> cohérence <strong>de</strong> l’Expected-Shortfall selon la définition<br />
adoptée.
11.3. Equilibre <strong>de</strong> marché 367<br />
s’avérer utile. En eff<strong>et</strong>, comme nous le montrons dans Malevergne <strong>et</strong> Sorn<strong>et</strong>te (2002b), les <strong>portefeuille</strong>s<br />
constitués d’actifs <strong>de</strong> faibles λ présentent globalement beaucoup moins <strong>de</strong> dépendance dans les extrêmes<br />
que les <strong>portefeuille</strong>s d’actifs ayant individuellement <strong>de</strong> grands λ.<br />
Lorsque l’on abandonne le strict cadre mono - périodique, les choses se compliquent. L’approche st<strong>at</strong>ique<br />
<strong>de</strong> Markovitz (1959) peut certes être généralisée dans un cadre dynamique (Merton 1992), mais cela nous<br />
ramène à la seule prise en compte <strong>de</strong>s p<strong>et</strong>its risques. Quelques tent<strong>at</strong>ives ont été menées pour tenter <strong>de</strong><br />
concilier grand risques <strong>et</strong> approche inter - temporelle telle par exemple la minimis<strong>at</strong>ion <strong>de</strong>s “drawdowns”<br />
(Grossman <strong>et</strong> Zhou 1993, Cvitanic <strong>et</strong> Kar<strong>at</strong>zas 1995, Chekhlov <strong>et</strong> al. 2000), ou encore l’utilis<strong>at</strong>ion <strong>de</strong>s<br />
cumulants à différentes échelles temporelles (Muzy <strong>et</strong> al. 2001).<br />
11.3 Equilibre <strong>de</strong> marché<br />
L’alloc<strong>at</strong>ion <strong>de</strong> capital <strong>et</strong> le choix <strong>de</strong>s actifs effectués par les agents a bien évi<strong>de</strong>mment une influence sur<br />
le prix <strong>de</strong> marché <strong>de</strong> ces actifs. Sous les hypothèses que le marché est parfaitement liqui<strong>de</strong> <strong>et</strong> en l’absence<br />
<strong>de</strong> taxe <strong>de</strong> quelque sorte que ce soit, il est possible <strong>de</strong> dériver <strong>de</strong>s équ<strong>at</strong>ions reliant les ren<strong>de</strong>ments<br />
espérés <strong>de</strong> chaque actif au ren<strong>de</strong>ment du marché (à l’équilibre). Le premier modèle d’équilibre, ancré<br />
dans l’univers <strong>de</strong> Markovitz, est le CAPM dérivé par Sharpe (1964), Lintner (1965) <strong>et</strong> Mossin (1966).<br />
Très tôt, <strong>de</strong> nombreuses généralis<strong>at</strong>ions ont été obtenues pour tenir compte notamment <strong>de</strong> l’eff<strong>et</strong> <strong>de</strong>s<br />
moments d’ordre supérieur à <strong>de</strong>ux (Jurczenko <strong>et</strong> Maill<strong>et</strong> 2002, <strong>et</strong> les références citées dans c<strong>et</strong> article),<br />
<strong>et</strong> ainsi tenter <strong>de</strong> résoudre l’ “equity premium puzzle”. Les résult<strong>at</strong>s n’étant pas très concluants, d’autres<br />
voies ont été explorées pour rendre compte notamment <strong>de</strong>s implic<strong>at</strong>ions économiques <strong>de</strong> l’optimis<strong>at</strong>ion<br />
<strong>de</strong> <strong>portefeuille</strong> sous d’autres contraintes que la variance, telle la VaR par exemple (Alexan<strong>de</strong>r <strong>et</strong> Baptista<br />
2002, Kaplanski <strong>et</strong> Kroll 2001b), sans malheureusement beaucoup plus <strong>de</strong> succès.<br />
Pour notre part, nous avons montré que la rel<strong>at</strong>ion standard du CAPM reste valable pour les mesures <strong>de</strong><br />
fluctu<strong>at</strong>ions que nous avons considérées ainsi que dans le cas où les agents n’utilisent pas tous la même<br />
mesure <strong>de</strong> risque, <strong>et</strong> donc lorsque le marché est hétérogène (Malevergne <strong>et</strong> Sorn<strong>et</strong>te 2002c), reprenant <strong>et</strong><br />
généralisant certains travaux antérieurs <strong>de</strong> Lintner (1969) ou Gone<strong>de</strong>s (1976) notamment.<br />
Toutes ces approches <strong>de</strong>meurent dans le cadre <strong>de</strong> l’étu<strong>de</strong> mono-périodique <strong>de</strong> l’équilibre <strong>de</strong>s marchés.<br />
Mais, <strong>de</strong> même que le problème <strong>de</strong> choix <strong>de</strong> <strong>portefeuille</strong> a reçu certaines extensions multi-périodiques,<br />
<strong>de</strong>s généralis<strong>at</strong>ions inter-temporelles du CAPM ont vu le jour dont celles proposées par Fama (1970) ou<br />
Merton (1973) pour ne citer que les plus célèbres, que nous ne faisons que mentionner puisque nous ne<br />
nous y sommes pas du tout intéressés. Comme pour le problème <strong>de</strong> sélection <strong>de</strong> <strong>portefeuille</strong>, nous nous<br />
sommes restreints au cas mono-périodiques.<br />
11.4 Conclusion<br />
Nous avons synthétisé dans ce chapitre les résult<strong>at</strong>s antérieurs obtenus dans le domaine <strong>de</strong> la <strong>gestion</strong> <strong>de</strong><br />
<strong>portefeuille</strong> quantit<strong>at</strong>ive <strong>et</strong> les conséquences théoriques qui en découlaient concernant les équilibres <strong>de</strong><br />
marchés. Ceci nous a permis <strong>de</strong> situer les résult<strong>at</strong>s que nous avons obtenus sur ce suj<strong>et</strong> par rapport à ceux<br />
déjà établis.<br />
Ces résult<strong>at</strong> seront présentés en détail dans les chapitres suivants, en commençant par les conséquences <strong>de</strong><br />
la prise en compte <strong>de</strong>s risques grands <strong>et</strong> extrêmes, notamment au travers du coefficient <strong>de</strong> dépendance <strong>de</strong><br />
queue (chapitre 12), puis les implic<strong>at</strong>ions <strong>et</strong> difficultés <strong>de</strong> l’optimis<strong>at</strong>ion <strong>de</strong> <strong>portefeuille</strong> sous contraintes<br />
<strong>de</strong> mesures <strong>de</strong> risques (cohérentes ou non) associées au capital économique (chapitre 13) <strong>et</strong> enfin nous
368 11. Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong> marché<br />
étudierons les <strong>portefeuille</strong>s efficients <strong>et</strong> équilibres <strong>de</strong> marchés qui en découlent sous contraintes <strong>de</strong> fluctu<strong>at</strong>ions<br />
autour d’un objectif <strong>de</strong> rentabilité sur le ren<strong>de</strong>ment espéré (chapitre 14).<br />
11.5 Annexe<br />
A l’ai<strong>de</strong> <strong>de</strong> calculs simples <strong>et</strong> <strong>de</strong> simul<strong>at</strong>ions numériques, nous démontrons l’existence générique d’un<br />
ét<strong>at</strong> macroscopique auto-organisé dans tout système <strong>de</strong> gran<strong>de</strong> taille présentant une corrél<strong>at</strong>ion moyenne<br />
non-nulle entre une fraction finie <strong>de</strong> toutes ces paires d’éléments. Nous montrons que la coexistence<br />
d’un spectre <strong>de</strong> valeurs propres, prédit par la <strong>théorie</strong> <strong>de</strong>s m<strong>at</strong>rices alé<strong>at</strong>oires, <strong>et</strong> <strong>de</strong> quelques très gran<strong>de</strong>s<br />
valeurs propres dans les m<strong>at</strong>rices <strong>de</strong> corrél<strong>at</strong>ion empiriques <strong>de</strong> gran<strong>de</strong> taille résulte d’un eff<strong>et</strong> collectif<br />
<strong>de</strong>s séries temporelles sous-jacentes plutôt que <strong>de</strong> l’impact <strong>de</strong> facteurs. Nos résult<strong>at</strong>s, en excellent accord<br />
avec <strong>de</strong> précé<strong>de</strong>ntes étu<strong>de</strong>s menées sur les m<strong>at</strong>rices <strong>de</strong> corrél<strong>at</strong>ion financières, montrent également que le<br />
cœur du spectre <strong>de</strong> valeurs propres contient une part signific<strong>at</strong>ive d’inform<strong>at</strong>ion <strong>et</strong> r<strong>at</strong>ionalise la présence<br />
<strong>de</strong> facteurs <strong>de</strong> marchés jusqu’ici introduits <strong>de</strong> manière ad hoc.
11.5. Annexe 369<br />
Collective Origin of the Coexistence of Apparent RMT Noise<br />
and Factors in Large Sample Correl<strong>at</strong>ion M<strong>at</strong>rices<br />
Y. Malevergne1, 2 1, 3<br />
and D. Sorn<strong>et</strong>te<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x, France<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095, USA<br />
Through simple analytical calcul<strong>at</strong>ions and numerical simul<strong>at</strong>ions, we <strong>de</strong>monstr<strong>at</strong>e the generic<br />
existence of a self-organized macroscopic st<strong>at</strong>e in any large multivari<strong>at</strong>e system possessing nonvanishing<br />
average correl<strong>at</strong>ions b<strong>et</strong>ween a finite fraction of all pairs of elements. The coexistence of<br />
an eigenvalue spectrum predicted by random m<strong>at</strong>rix theory (RMT) and a few very large eigenvalues<br />
in large empirical correl<strong>at</strong>ion m<strong>at</strong>rices is shown to result from a bottom-up collective effect of the<br />
un<strong>de</strong>rlying time series r<strong>at</strong>her than a top-down impact of factors. Our results, in excellent agreement<br />
with previous results obtained on large financial correl<strong>at</strong>ion m<strong>at</strong>rices, show th<strong>at</strong> there is relevant<br />
inform<strong>at</strong>ion also in the bulk of the eigenvalue spectrum and r<strong>at</strong>ionalize the presence of mark<strong>et</strong> factors<br />
previously introduced in an ad hoc manner.<br />
Since Wigner’s seminal i<strong>de</strong>a to apply random m<strong>at</strong>rix<br />
theory (RMT) to interpr<strong>et</strong> the complex spectrum of energy<br />
levels in nuclear physics [1], RMT has ma<strong>de</strong> enormous<br />
progress [2] with many applic<strong>at</strong>ions in physical sciences<br />
and elsewhere such as in m<strong>et</strong>eorology [3] and image<br />
processing [4]. A new applic<strong>at</strong>ion was proposed a<br />
few years ago to the problem of correl<strong>at</strong>ions b<strong>et</strong>ween financial<br />
ass<strong>et</strong>s and to the portfolio optimiz<strong>at</strong>ion problem.<br />
It was shown th<strong>at</strong>, among the eigenvalues and principal<br />
components of the empirical correl<strong>at</strong>ion m<strong>at</strong>rix of the<br />
r<strong>et</strong>urns of hundreds of ass<strong>et</strong> on the New York Stock Exchange<br />
(NYSE), apart from the few highest eigenvalues,<br />
the marginal distribution of the other eigenvalues and<br />
eigenvectors closely resembles the spectral distribution<br />
of a positive symm<strong>et</strong>ric random m<strong>at</strong>rix with maximum<br />
entropy, suggesting th<strong>at</strong> the correl<strong>at</strong>ion m<strong>at</strong>rix does not<br />
contain any specific inform<strong>at</strong>ion beyond these few largest<br />
eigenvalues and eigenvectors [5]. These results apparently<br />
invalid<strong>at</strong>e the standard mean-variance portfolio optimiz<strong>at</strong>ion<br />
theory [6] consecr<strong>at</strong>ed by the financial industry<br />
[7] and seemingly support the r<strong>at</strong>ionale behind factor<br />
mo<strong>de</strong>ls such as the capital ass<strong>et</strong> pricing mo<strong>de</strong>l (CAPM)<br />
[8] and the arbitrage pricing theory (APT) [9], where the<br />
correl<strong>at</strong>ions b<strong>et</strong>ween a large number of ass<strong>et</strong>s are represented<br />
through a small number of so-called mark<strong>et</strong> factors.<br />
In<strong>de</strong>ed, if the spectrum of eigenvalues of the empirical<br />
covariance or correl<strong>at</strong>ion m<strong>at</strong>rices are predicted<br />
by RMT, it seems n<strong>at</strong>ural to conclu<strong>de</strong> th<strong>at</strong> there is no<br />
usable inform<strong>at</strong>ion in these m<strong>at</strong>rices and th<strong>at</strong> empirical<br />
covariance m<strong>at</strong>rices should not be used for portfolio optimiz<strong>at</strong>ion.<br />
In contrast, if one d<strong>et</strong>ects <strong>de</strong>vi<strong>at</strong>ions b<strong>et</strong>ween<br />
the universal – and therefore non-inform<strong>at</strong>ive – part of<br />
the spectral properties of empirically estim<strong>at</strong>ed covariance<br />
and correl<strong>at</strong>ion m<strong>at</strong>rices and those of the relevant<br />
ensemble of random m<strong>at</strong>rices [10], this may quantify the<br />
amount of real inform<strong>at</strong>ion th<strong>at</strong> can be used in portfolio<br />
optimiz<strong>at</strong>ion from the “noise” th<strong>at</strong> should be discar<strong>de</strong>d.<br />
More generally, in many different scientific fields, one<br />
needs to d<strong>et</strong>ermine the n<strong>at</strong>ure and amount of inform<strong>at</strong>ion<br />
contained in large covariance and correl<strong>at</strong>ion m<strong>at</strong>rices.<br />
This occurs as soon as one <strong>at</strong>tempts to estim<strong>at</strong>e<br />
very large covariance and correl<strong>at</strong>ion m<strong>at</strong>rices in multivari<strong>at</strong>e<br />
dynamics of systems exhibiting non-Gaussian<br />
fluctu<strong>at</strong>ions with f<strong>at</strong> tails and/or long-range time correl<strong>at</strong>ions<br />
with intermittency. In such cases, the convergence<br />
of the estim<strong>at</strong>ors of the large covariance and correl<strong>at</strong>ion<br />
m<strong>at</strong>rices is often too slow for all practical purposes. The<br />
problem becomes even more complex with time-varying<br />
variances and covariances as occurs in systems with h<strong>et</strong>eroskedasticity<br />
[11] or with regime-switching [12]. A<br />
prominent example where such difficulties arise is the<br />
d<strong>at</strong>a-assimil<strong>at</strong>ion problem in engineering and in m<strong>et</strong>eorology<br />
where forecasting is combined with observ<strong>at</strong>ions<br />
iter<strong>at</strong>ively through the Kalman filter, based on the estim<strong>at</strong>ion<br />
and forward prediction of large covariance m<strong>at</strong>rices<br />
[13].<br />
As we said in the context of financial time series, the<br />
rescuing str<strong>at</strong>egy is to invoke the existence of a few dominant<br />
factors, such as an overall mark<strong>et</strong> factor and the<br />
factors rel<strong>at</strong>ed to firm size, firm industry and book-tomark<strong>et</strong><br />
equity, thought to embody most of the relevant<br />
<strong>de</strong>pen<strong>de</strong>nce structure b<strong>et</strong>ween the studied time series<br />
[14]. In<strong>de</strong>ed, there is no doubt th<strong>at</strong> observed equity prices<br />
respond to a wi<strong>de</strong> vari<strong>et</strong>y of unanticip<strong>at</strong>ed factors, but<br />
there is much weaker evi<strong>de</strong>nce th<strong>at</strong> expected r<strong>et</strong>urns are<br />
higher for equities th<strong>at</strong> are more sensitive to these factors,<br />
as required by Markowitz’s mean-variance theory,<br />
by the CAPM and the APT [15]. This severe failure of<br />
the most fundamental finance theories could conceivably<br />
be <strong>at</strong>tributable to an inappropri<strong>at</strong>e proxy for the mark<strong>et</strong><br />
portfolio, but nobody has been able to show th<strong>at</strong> this is<br />
really the correct explan<strong>at</strong>ion. This remark constitutes
370 11. Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong> marché<br />
the crux of the problem: the factors invoked to mo<strong>de</strong>l the<br />
cross-sectional <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s are not known<br />
in general and are either postul<strong>at</strong>ed based on economic<br />
intuition in financial studies or obtained as black box results<br />
in the recent analyses using RMT [5].<br />
Here, we show th<strong>at</strong> the existence of factors results from<br />
a collective effect of the ass<strong>et</strong>s, similar to the emergence<br />
of a macroscopic self-organiz<strong>at</strong>ion of interacting microscopic<br />
constituents. For this, we unravel the general<br />
physical origin of the large eigenvalues of large covariance<br />
and correl<strong>at</strong>ion m<strong>at</strong>rices and provi<strong>de</strong> a compl<strong>et</strong>e<br />
un<strong>de</strong>rstanding of the coexistence of fe<strong>at</strong>ures resembling<br />
properties of random m<strong>at</strong>rices and of large “anomalous”<br />
eigenvalues. Through simple analytical calcul<strong>at</strong>ions and<br />
numerical simul<strong>at</strong>ions, we <strong>de</strong>monstr<strong>at</strong>e the generic existence<br />
of a self-organized macroscopic st<strong>at</strong>e in any large<br />
system possessing non-vanishing average correl<strong>at</strong>ions b<strong>et</strong>ween<br />
a finite fraction of all pairs of elements.<br />
L<strong>et</strong> us first consi<strong>de</strong>r a large system of size N with correl<strong>at</strong>ion<br />
m<strong>at</strong>rix C in which every non-diagonal pairs of<br />
elements exhibits the same correl<strong>at</strong>ion coefficient Cij = ρ<br />
for i = j and Cii = 1. Its eigenvalues are<br />
λ1 = 1 + (N − 1)ρ and λi≥2 = 1 − ρ (1)<br />
with multiplicity N − 1 and with ρ ∈ (0, 1) in or<strong>de</strong>r for<br />
the correl<strong>at</strong>ion m<strong>at</strong>rix to remain positive <strong>de</strong>finite. Thus,<br />
in the thermodynamics limit N → ∞, even for a weak<br />
positive correl<strong>at</strong>ion ρ → 0 (with ρN ≫ 1), a very large<br />
eigenvalue appears, associ<strong>at</strong>ed with the <strong>de</strong>localized eigenvector<br />
v1 = (1/ √ N)(1, 1, · · · , 1), which domin<strong>at</strong>es compl<strong>et</strong>ely<br />
the correl<strong>at</strong>ion structure of the system. This trivial<br />
example stresses th<strong>at</strong> the key point for the emergence<br />
of a large eigenvalue is not the strength of the correl<strong>at</strong>ions,<br />
provi<strong>de</strong>d th<strong>at</strong> they do not vanish, but the large size<br />
N of the system.<br />
This result (1) still holds qualit<strong>at</strong>ively when the correl<strong>at</strong>ion<br />
coefficients are all distinct. To see this, it is convenient<br />
to use a perturb<strong>at</strong>ion approach. We thus add a<br />
small random component to each correl<strong>at</strong>ion coefficient:<br />
Cij = ρ + ɛ · aij for i = j , (2)<br />
where the coefficients aij = aji have zero mean, variance<br />
σ 2 and are in<strong>de</strong>pen<strong>de</strong>ntly distributed (There are additional<br />
constraints on the support of the distribution of<br />
the aij’s in or<strong>de</strong>r for the m<strong>at</strong>rix Cij to remain positive<br />
<strong>de</strong>finite with probability one). The d<strong>et</strong>ermin<strong>at</strong>ion of the<br />
eigenvalues and eigenfunctions of Cij is performed using<br />
the perturb<strong>at</strong>ion theory <strong>de</strong>veloped in quantum mechanics<br />
[16] up to the second or<strong>de</strong>r in ɛ. We find th<strong>at</strong> the<br />
largest eigenvalue becomes<br />
E[λ1] = (N −1)ρ+1+<br />
(N − 1)(N − 2)<br />
N 2 · ɛ2σ2 ρ +O(ɛ3 ) (3)<br />
while, <strong>at</strong> the same or<strong>de</strong>r, the corresponding eigenvector<br />
v1 remains unchanged. The <strong>de</strong>generacy of the eigenvalue<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
0<br />
0 0.5 1 1.5 2 2.5 3<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
largest<br />
eigenvalue<br />
0<br />
0 20 40 60<br />
FIG. 1: Spectrum of eigenvalues of a random correl<strong>at</strong>ion m<strong>at</strong>rix<br />
with average correl<strong>at</strong>ion coefficient ρ = 0.14 and standard<br />
<strong>de</strong>vi<strong>at</strong>ion of the correl<strong>at</strong>ion coefficients σ = 0.345/ √ N. The<br />
size N = 406 of the m<strong>at</strong>rix is the same as in previous studies<br />
[5] for the sake of comparison. The continuous curve is<br />
the theor<strong>et</strong>ical transl<strong>at</strong>ed semi-circle distribution of eigenvalues<br />
<strong>de</strong>scribing the bulk of the distribution which passes the<br />
Kolmogorov test. The center value λ = 1 − ρ ensures the<br />
conserv<strong>at</strong>ion of the trace equal to N. There is no adjustable<br />
param<strong>et</strong>er. The ins<strong>et</strong> represents the whole spectrum with the<br />
largest eigenvalue whose size is in agreement with the prediction<br />
ρN = 56.8.<br />
λ = 1 − ρ is broken and leads to a complex s<strong>et</strong> of smaller<br />
eigenvalues <strong>de</strong>scribed below.<br />
In fact, this result (3) can be generalized to the nonperturb<strong>at</strong>ive<br />
domain of any correl<strong>at</strong>ion m<strong>at</strong>rix with in<strong>de</strong>pen<strong>de</strong>nt<br />
random coefficients Cij, provi<strong>de</strong>d th<strong>at</strong> they have<br />
the same mean value ρ and variance σ 2 . In<strong>de</strong>ed, it has<br />
been shown [17] th<strong>at</strong>, in such a case, the expect<strong>at</strong>ions of<br />
the largest and second largest eigenvalues are<br />
E[λ1] = (N − 1) · ρ + 1 + σ 2 /ρ + o(1) , (4)<br />
E[λ2] ≤ 2σ √ N + O(N 1/3 log N) . (5)<br />
Moreover, the st<strong>at</strong>istical fluctu<strong>at</strong>ions of these two largest<br />
eigenvalues are asymptotically (for large fluctu<strong>at</strong>ions t ><br />
O( √ N)) boun<strong>de</strong>d by a Gaussian distribution according<br />
to the following large <strong>de</strong>vi<strong>at</strong>ion theorem<br />
Pr{|λ1,2 − E[λ1,2]| ≥ t} ≤ e −c1,2t2<br />
2<br />
, (6)<br />
for some positive constant c1,2 [18].<br />
This result is very different from th<strong>at</strong> obtained when<br />
the mean value ρ vanishes. In such a case, the distribution<br />
of eigenvalues of the random m<strong>at</strong>rix C is given by<br />
the semi-circle law [2]. However, due to the presence of<br />
the ones on the main diagonal of the correl<strong>at</strong>ion m<strong>at</strong>rix<br />
C, the center of the circle is not <strong>at</strong> the origin but <strong>at</strong> the
11.5. Annexe 371<br />
point λ = 1. Thus, the distribution of the eigenvalues of<br />
random correl<strong>at</strong>ion m<strong>at</strong>rices with zero mean correl<strong>at</strong>ion<br />
coefficients is a semi-circle of radius 2σ √ N centered <strong>at</strong><br />
λ = 1.<br />
The result (4) is <strong>de</strong>eply rel<strong>at</strong>ed to the so-called “friendship<br />
theorem” in m<strong>at</strong>hem<strong>at</strong>ical graph theory, which<br />
st<strong>at</strong>es th<strong>at</strong>, in any finite graph such th<strong>at</strong> any two vertices<br />
have exactly one common neighbor, there is one<br />
and only one vertex adjacent to all other vertices [19].<br />
A more heuristic but equivalent st<strong>at</strong>ement is th<strong>at</strong>, in a<br />
group of people such th<strong>at</strong> any pair of persons have exactly<br />
one common friend, there is always one person (the<br />
“politician”) who is the friend of everybody. The connection<br />
is established by taking the non-diagonal entries Cij<br />
(i = j) equal to Bernouilli random variable with param<strong>et</strong>er<br />
ρ, th<strong>at</strong> is, P r[Cij = 1] = ρ and P r[Cij = 0] = 1 − ρ.<br />
Then, the m<strong>at</strong>rix Cij − I, where I is the unit m<strong>at</strong>rix,<br />
becomes nothing but the adjacency m<strong>at</strong>rix of the random<br />
graph G(N, ρ) [18]. The proof of [19] of the “friendship<br />
theorem” in<strong>de</strong>ed relies on the N-<strong>de</strong>pen<strong>de</strong>nce of the<br />
largest eigenvalue and on the √ N-<strong>de</strong>pen<strong>de</strong>nce of the second<br />
largest eigenvalue of Cij as given by (4) and (5).<br />
Figure 1 shows the distribution of eigenvalues of a<br />
random correl<strong>at</strong>ion m<strong>at</strong>rix. The ins<strong>et</strong> shows the largest<br />
eigenvalue lying <strong>at</strong> the predicting size ρN = 56.8, while<br />
the bulk of the eigenvalues are much smaller and are <strong>de</strong>scribed<br />
by a modified semi-circle law centered on λ =<br />
1 − ρ, in the limit of large N. The result on the largest<br />
eigenvalue emerging from the collective effect of the crosscorrel<strong>at</strong>ion<br />
b<strong>et</strong>ween all N(N −1)/2 pairs provi<strong>de</strong>s a novel<br />
perspective to the observ<strong>at</strong>ion [20] th<strong>at</strong> the only reasonable<br />
explan<strong>at</strong>ion for the simultaneous crash of 23 stock<br />
mark<strong>et</strong>s worldwi<strong>de</strong> in October 1987 is the impact of a<br />
world mark<strong>et</strong> factor: according to our <strong>de</strong>monstr<strong>at</strong>ion,<br />
the simultaneous occurrence of significant correl<strong>at</strong>ions<br />
b<strong>et</strong>ween the mark<strong>et</strong>s worldwi<strong>de</strong> is bound to lead to the<br />
existence of an extremely large eigenvalue, the world mark<strong>et</strong><br />
factor constructed by ... a linear combin<strong>at</strong>ion of the<br />
23 stock mark<strong>et</strong>s! Wh<strong>at</strong> our result shows is th<strong>at</strong> invoking<br />
factors to explain the cross-sectional structure of stock r<strong>et</strong>urns<br />
is cursed by the chicken-and-egg problem: factors<br />
exist because stocks are correl<strong>at</strong>ed; stocks are correl<strong>at</strong>ed<br />
because of common factors impacting them [24].<br />
Figure 2 shows the eigenvalues distribution of the sample<br />
correl<strong>at</strong>ion m<strong>at</strong>rix reconstructed by sampling N =<br />
406 time series of length T = 1309 gener<strong>at</strong>ed with a given<br />
correl<strong>at</strong>ion m<strong>at</strong>rix C with theor<strong>et</strong>ical spectrum shown in<br />
figure 1. The largest eigenvalue is again very close to<br />
the prediction ρN = 56.8 while the bulk of the distribution<br />
<strong>de</strong>parts very strongly from the semi-circle law and<br />
is not far from the Wishart prediction, as expected from<br />
the <strong>de</strong>finition of the Wishart ensemble as the ensemble of<br />
sample covariance m<strong>at</strong>rices of Gaussian distributed time<br />
series with unit variance and zero mean. A Kolmogorov<br />
test shows however th<strong>at</strong> the bulk of the spectrum (renormalized<br />
so as to take into account the presence of the<br />
frequency<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
0<br />
0 20 40 60<br />
λ<br />
0<br />
0 0.5 1 1.5 2 2.5<br />
λ<br />
3 3.5 4 4.5 5<br />
frequency<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
largest<br />
eigenvalue<br />
FIG. 2: Spectrum estim<strong>at</strong>ed from the sample correl<strong>at</strong>ion m<strong>at</strong>rix<br />
obtained from N = 406 time series of length T = 1309<br />
(the same length as in [5]) with the same theor<strong>et</strong>ical correl<strong>at</strong>ion<br />
m<strong>at</strong>rix as th<strong>at</strong> presented in figure 1.<br />
outlier eigenvalue) is not in the Wishart class, in contradiction<br />
with previous claims lacking formal st<strong>at</strong>istical<br />
tests [5]. This result holds for different simul<strong>at</strong>ions of the<br />
sample correl<strong>at</strong>ion m<strong>at</strong>rix and different realiz<strong>at</strong>ions of the<br />
theor<strong>et</strong>ical correl<strong>at</strong>ion m<strong>at</strong>rix with the same param<strong>et</strong>ers<br />
(ρ, σ). The st<strong>at</strong>istically significant <strong>de</strong>parture from the<br />
Wishart prediction implies th<strong>at</strong> there is actually some<br />
inform<strong>at</strong>ion in the bulk of the spectrum of eigenvalues,<br />
which can be r<strong>et</strong>rieved using Marsili’s procedure [10]. We<br />
have also checked th<strong>at</strong> these results remain robust for<br />
non-Gaussian distribution of r<strong>et</strong>urns as long as the second<br />
moments exist. In<strong>de</strong>ed, correl<strong>at</strong>ed time series with<br />
multivari<strong>at</strong>e Gaussian or Stu<strong>de</strong>nt distributions with three<br />
<strong>de</strong>grees of freedom (which provi<strong>de</strong> more acceptable proxies<br />
for financial time series [21]) give no discernible differences<br />
in the spectrum of eigenvalues. This is surprising<br />
as the estim<strong>at</strong>or of a correl<strong>at</strong>ion coefficient is asymptotically<br />
Gaussian for time series with finite fourth moment<br />
and Lévy stable otherwise [22].<br />
Empirically [5], a few other eigenvalues below the<br />
largest one have an amplitu<strong>de</strong> of the or<strong>de</strong>r of 5 − 10<br />
th<strong>at</strong> <strong>de</strong>vi<strong>at</strong>e significantly from the bulk of the distribution.<br />
Our analysis provi<strong>de</strong>s a very simple constructive<br />
mechanism for them, justifying the postul<strong>at</strong>ed mo<strong>de</strong>l of<br />
Ref.[23]. The solution consists in consi<strong>de</strong>ring, as a first<br />
approxim<strong>at</strong>ion, the block diagonal m<strong>at</strong>rix C ′ with diagonal<br />
elements ma<strong>de</strong> of the m<strong>at</strong>rices A1, · · · , Ap of sizes<br />
N1, · · · , Np with Ni = N, constructed according to<br />
(2) such th<strong>at</strong> each m<strong>at</strong>rix Ai has the average correl<strong>at</strong>ion<br />
coefficient ρi. When the coefficients of the m<strong>at</strong>rix C ′<br />
outsi<strong>de</strong> the m<strong>at</strong>rices Ai are zero, the spectrum of C ′ is<br />
given by the union of all the spectra of the Ai’s, which<br />
3
372 11. Portefeuilles optimaux <strong>et</strong> équilibre <strong>de</strong> marché<br />
frequency<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
0<br />
0 20 40 60<br />
λ<br />
0<br />
0 0.5 1 1.5<br />
λ<br />
2 2.5 3<br />
frequency<br />
15<br />
10<br />
5<br />
largest<br />
eigenvalue<br />
FIG. 3: Spectrum of eigenvalues estim<strong>at</strong>ed from the sample<br />
correl<strong>at</strong>ion m<strong>at</strong>rix of N = 406 time series of length T = 1309.<br />
The times series have been constructed from a multivari<strong>at</strong>e<br />
Gaussian distribution with a correl<strong>at</strong>ion m<strong>at</strong>rix ma<strong>de</strong> of three<br />
block-diagonal m<strong>at</strong>rices of sizes respectively equal to 130, 140<br />
and 136 and mean correl<strong>at</strong>ion coefficients equal to 0.18 for all<br />
of them. The off-diagonal elements are all equal to 0.1. The<br />
same results hold if the off-diagonal elements are random.<br />
are each domin<strong>at</strong>ed by a large eigenvalue λ1,i ρi · Ni.<br />
The spectrum of C ′ then exhibits p large eigenvalues.<br />
Each block Ai can be interpr<strong>et</strong>ed as a sector of the<br />
economy, including all the companies belonging to a same<br />
industrial branch and the eigenvector associ<strong>at</strong>ed with<br />
each largest eigenvalue represents the main factor driving<br />
this sector of activity [25]. For similar sector sizes Ni and<br />
average correl<strong>at</strong>ion coefficients ρi, the largest eigenvalues<br />
are of the same or<strong>de</strong>r of magnitu<strong>de</strong>. In or<strong>de</strong>r to recover<br />
a very large unique eigenvalue, we reintroduce some coupling<br />
constants outsi<strong>de</strong> the block diagonal m<strong>at</strong>rices. A<br />
well-known result of perturb<strong>at</strong>ion theory in quantum mechanics<br />
st<strong>at</strong>es th<strong>at</strong> such coupling leads to a repulsion b<strong>et</strong>ween<br />
the eigenst<strong>at</strong>es, which can be observed in figure 3<br />
where C ′ has been constructed with three block m<strong>at</strong>rices<br />
A1, A2 and A3 and non-zero off-diagonal coupling <strong>de</strong>scribed<br />
in the figure caption. These values allow us to<br />
quantit<strong>at</strong>ively replic<strong>at</strong>e the empirical finding of Laloux<br />
<strong>et</strong> al. in [5], where the three first eigenvalues are approxim<strong>at</strong>ely<br />
λ1 57, λ2 10 and λ3 8. The bulk of the<br />
spectrum (which exclu<strong>de</strong>s the three largest eigenvalues)<br />
is similar to the Wishart distribution but again st<strong>at</strong>istically<br />
different from it as tested with a Kolmogorov test.<br />
As a final remark, expressions (3,4) and our numerical<br />
tests for a large vari<strong>et</strong>y of correl<strong>at</strong>ion m<strong>at</strong>rices show th<strong>at</strong><br />
the <strong>de</strong>localized eigenvector v1 = (1/ √ N)(1, 1, · · · , 1), associ<strong>at</strong>ed<br />
with the largest eigenvalue is extremely robust<br />
and remains (on average) the same for any large system.<br />
Thus, even for time-varying correl<strong>at</strong>ion m<strong>at</strong>rices<br />
(see Drozdz <strong>et</strong> al. in [5]) – as in finance with important<br />
h<strong>et</strong>eroskedastic effects – the composition of the main<br />
factor remains almost the same. This can be seen as a<br />
generalized limit theorem reflecting the bottom-up organiz<strong>at</strong>ion<br />
of broadly correl<strong>at</strong>ed time series.<br />
[1] Wigner, E.P., Ann. M<strong>at</strong>h. 53, 36 (1951).<br />
[2] Mehta, M.L., Random m<strong>at</strong>rices, 2nd ed. (Boston: Aca<strong>de</strong>mic<br />
Press, 1991).<br />
[3] Santhanam, M.S. and P. K. P<strong>at</strong>ra, Phys. Rev. E 64,<br />
016102 (2001).<br />
[4] S<strong>et</strong>punga, A.M. and P.P. Mitra, Phys. Rev E 60, 3389<br />
(1999).<br />
[5] Laloux, L. <strong>et</strong> al., Phys. Rev. L<strong>et</strong>t. 83, 1467 (1999);<br />
Plerou, V., <strong>et</strong> al., Phys. Rev. L<strong>et</strong>t. 83, 1471 (1999);<br />
Maslov, S., Physica A 301, 397 (2001); Drozdz, S. <strong>et</strong> al.,<br />
Physica A 287, 440 (2000); Physica A 294, 226 (2001);<br />
Plerou, V. <strong>et</strong> al., Phys. Rev E 6506 066126 (2002).<br />
[6] Markowitz, H., Portfolio selection (John Wiley and Sons,<br />
New York, 1959).<br />
[7] RiskM<strong>et</strong>rics Group, RiskM<strong>et</strong>rics (Technical Document,<br />
NewYork: J.P. Morgan/Reuters, 1996).<br />
[8] Sharpe, W.F., J. Finance (September), 425 (1964); Lintner,<br />
J., Rev. Econ. St<strong>at</strong>. (February), 13 (1965); Mossin,<br />
J., Econom<strong>et</strong>rica (October), 768 (1966); Black, F., J.<br />
Business (July), 444 (1972).<br />
[9] Ross, S.A., J. Economic Theory (December), 341 (1976).<br />
[10] Marsili, M., cond-m<strong>at</strong>/0003241; Giada, L. and Marsili,<br />
M., Phys. Rev. E art. no. 061101, 6306 N6 PT1:1101,U17-<br />
U23 (2001); T. Guhr and B. Kalber, cond-m<strong>at</strong>/0206577.<br />
[11] Engle, R.F. and K. Sheppard, NBER Working Paper No.<br />
W8554 (2001).<br />
[12] Schaller, H. and van Nor<strong>de</strong>n, S., Appl. Financial Econ.<br />
7, 177 (1997).<br />
[13] Brammer, K., Kalman-Bucy filters (Gerhard Siffling,<br />
Norwood, MA: Artech House, 1989).<br />
[14] Fama, E.F. and Kenn<strong>et</strong>h R., J. Finance 51, 55 (1996); J.<br />
Financial Econ. 33, 3 (1993); Fama, E.F. <strong>et</strong> al., Financial<br />
Analysts J. 49, 37 (1993).<br />
[15] Roll, R., Financial Management 23, 69 (1994).<br />
[16] Cohen-Tannoudji, C., B. Diu and F. Laloe, Quantum<br />
mechanics (New York: Wiley, 1977).<br />
[17] F¨redi Z. and J. Komlós, Combin<strong>at</strong>orica 1, 233-241 (1981).<br />
[18] Krivelevich, M. and V. H. Vu, m<strong>at</strong>h-ph/0009032 (2000).<br />
[19] Erdos, P. <strong>et</strong> al., Studia Sci. M<strong>at</strong>h. 1, 215 (1966).<br />
[20] R. Roll, Financial Analysts J. 44, 19 (1988).<br />
[21] Gopikrishnan, P. <strong>et</strong> al., Eur. Phys. Journal B 3 ¯ , 139<br />
(1998); Guillaume, D.M., <strong>et</strong> al., Finance and Stochastics<br />
1, 95 (1997); Lux, L., Appl. Financial Economics 6,<br />
463 (1996); Pagan, A., J. Emp. Fin. 3, 15 (1996).<br />
[22] Davis, R.A. and J.E. Marengo, Commun. St<strong>at</strong>ist.-<br />
Stochastic Mo<strong>de</strong>ls 6, 483 (1990); Meerschaert, M.M. and<br />
H.P. Scheffler, J. Time Series Anal. 22, 481 (2001).<br />
[23] Noh, J.D., Phys.Rev.E 61, 5981 (2000).<br />
[24] see D. Sorn<strong>et</strong>te <strong>et</strong> al., in press in Risk, condm<strong>at</strong>/0204626,<br />
for a mechanism and empirical results contrasting<br />
the endogenous character of the October 1987<br />
crash and other large endogenous mark<strong>et</strong> moves.<br />
[25] Mantegna, R.N., Eur. Phys. J. 11, 193 (1999); Marsili,<br />
M., Quant. Fin. 2, 297 (2002).<br />
4
Chapitre 12<br />
Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
Dans la première partie <strong>de</strong> ce chapitre, nous exposons nos idées concernant la <strong>gestion</strong> <strong>de</strong>s risques<br />
extrêmes mais aussi <strong>de</strong>s risques “intermédiaires”, si l’on peut dire, que nous qualifions <strong>de</strong> grands risques<br />
dans le sens où leur impact est bien plus important que celui associé aux risques quantifiés par la variance<br />
mais reste très inférieur <strong>de</strong> part leur conséquences aux risques extrêmes. Cela nous perm<strong>et</strong> donc<br />
<strong>de</strong> discuter <strong>de</strong>s moyens d’appréhen<strong>de</strong>r toute la gamme <strong>de</strong>s risques : <strong>de</strong>s plus p<strong>et</strong>its aux plus extrêmes.<br />
La <strong>de</strong>uxième partie du chapitre est exclusivement consacrée aux risques extrêmes <strong>et</strong> nous nous interrogeons<br />
sur la façon <strong>de</strong> s’en prémunir. En fait, à cause <strong>de</strong> l’existence d’une dépendance <strong>de</strong> queue<br />
entre les actifs, il n’est pas possible <strong>de</strong> diversifier parfaitement les risques extrêmes par agrég<strong>at</strong>ion. On<br />
peut néanmoins construire <strong>de</strong>s <strong>portefeuille</strong>s dont les actifs ont chacun <strong>de</strong> très faibles coefficients <strong>de</strong><br />
dépendance <strong>de</strong> queue, ce qui assure au <strong>portefeuille</strong> une assez faible sensibilité aux grands mouvements<br />
<strong>de</strong> ses constituants. Pour autant, l’absence <strong>de</strong>s risques extrêmes est un objectif qui semble hors d’<strong>at</strong>teinte.<br />
Nous nous sommes focalisés sur <strong>de</strong>s <strong>portefeuille</strong>s “traditionnels”, c’est-à-dire où les ventes à découvert<br />
ne sont pas autorisées. Il est cependant intéressant <strong>de</strong> se <strong>de</strong>man<strong>de</strong>r si <strong>de</strong>s str<strong>at</strong>égies mixtes consistant à<br />
détenir <strong>de</strong>s positions longues sur certains actifs <strong>et</strong> courtes sur d’autres ne perm<strong>et</strong>traient pas <strong>de</strong> se prémunir<br />
contre les risques extrêmes dans la mesure où les très gran<strong>de</strong>s baisses <strong>de</strong>s uns seraient compensées par<br />
les très gran<strong>de</strong>s baisses (concomitantes) <strong>de</strong>s autres. C<strong>et</strong>te str<strong>at</strong>égie fonctionne très bien dans c<strong>et</strong>te configur<strong>at</strong>ion,<br />
mais fait apparaître une autre situ<strong>at</strong>ion critique : celle où les actifs en position longue baissent<br />
<strong>et</strong> où ceux en position courte montent. Or, <strong>de</strong>s pertes extrêmes concomitantes <strong>de</strong> hausses extrêmes sont<br />
tout à fait envisageables : que l’on se remémore la figure 7.1 page 187 <strong>et</strong> la figure 1 page 214 concernant<br />
la copule <strong>de</strong> Stu<strong>de</strong>nt <strong>et</strong> son coefficient <strong>de</strong> dépendance <strong>de</strong> queue. On y observe que les gran<strong>de</strong>s<br />
hausses concomitantes <strong>de</strong> gran<strong>de</strong>s pertes ont une probabilité d’occurrence non nulle, même si elle est<br />
systém<strong>at</strong>iquement plus faible que la probabilité d’occurrence simultanée <strong>de</strong> <strong>de</strong>ux pertes (ou hausses)<br />
extrêmes.<br />
Ainsi, contrairement à l’approche moyenne-variance où les str<strong>at</strong>égies mixtes perm<strong>et</strong>tent <strong>de</strong> complètement<br />
décorréler le <strong>portefeuille</strong> du marché, par exemple, elles ne peuvent apporter <strong>de</strong> réelle solution face au<br />
problème <strong>de</strong>s risques extrêmes, même s’il semble qu’elles présentent une amélior<strong>at</strong>ion par rapport aux<br />
str<strong>at</strong>égies exclusivement en positions longues.<br />
373
374 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
12.1 Comprendre <strong>et</strong> gérer les risques grands <strong>et</strong> extrêmes<br />
L’impact <strong>de</strong>s risques grands <strong>et</strong> extrêmes sur l’activité financière <strong>et</strong> le secteur <strong>de</strong> l’assurance est <strong>de</strong>venu<br />
si important qu’il ne peut plus être ignoré <strong>de</strong>s <strong>gestion</strong>naires <strong>de</strong> <strong>portefeuille</strong>. C’est pourquoi nous nous<br />
proposons <strong>de</strong> synthétiser ici les étapes successives conduisant à une <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> rigoureuse<br />
visant à prendre en compte (1) le comportement sous-exponentiel <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ment, (2)<br />
les dépendances non-gaussiennes entre actifs <strong>et</strong> (3) les dépendances temporelles intermittentes amenant<br />
les gran<strong>de</strong>s pertes tout en déployant (4) le concept <strong>de</strong> risque sur ses différentes dimensions allant <strong>de</strong>s<br />
“p<strong>et</strong>its” risques jusqu’aux risques “extrêmes” afin <strong>de</strong> définir <strong>de</strong>s fonctions <strong>de</strong> décision cohérentes <strong>et</strong><br />
pr<strong>at</strong>iques perm<strong>et</strong>tant d’établir <strong>de</strong>s <strong>portefeuille</strong>s optimaux.<br />
Reprint from : J.V An<strong>de</strong>rsen, Y. Malevergne <strong>et</strong> D.Sorn<strong>et</strong>te, 2002, Comprendre <strong>et</strong> gérer les risques grands<br />
<strong>et</strong> extêmes, Risque 49, 105-110.
12.1. Comprendre <strong>et</strong> gérer les risques grands <strong>et</strong> extrêmes 375<br />
COMPRENDRE ET GÉRER<br />
LES RISQUES GRANDS ET EXTRÊMES<br />
Jorgen V. An<strong>de</strong>rsen<br />
Chargé <strong>de</strong> recherche à l’université Paris X-Nanterre (Thema)<br />
<strong>et</strong> à l’université <strong>de</strong> Nice Sophia-Antipolis<br />
Yannick Malevergne<br />
Doctorant à l’université <strong>de</strong> Nice Sophia-Antipolis<br />
<strong>et</strong> à l’université Lyon I (<strong>ISFA</strong>)<br />
Didier Sorn<strong>et</strong>te<br />
Directeur <strong>de</strong> recherche à l’université <strong>de</strong> Nice Sophia-Antipolis<br />
<strong>et</strong> professeur à l’université <strong>de</strong> Californie, Los Angeles (Ucla)<br />
L’impact <strong>de</strong>s risques grands <strong>et</strong> extrêmes sur l’activité financière <strong>et</strong> le secteur<br />
<strong>de</strong> l’assurance est <strong>de</strong>venu si important qu’il ne peut plus être ignoré <strong>de</strong>s <strong>gestion</strong>naires<br />
<strong>de</strong> <strong>portefeuille</strong>. C’est pourquoi nous nous proposons <strong>de</strong> synthétiser<br />
ici les étapes successives conduisant à une <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong> rigoureuse<br />
visant à prendre en compte (1) le comportement sous-exponentiel <strong>de</strong>s distributions<br />
<strong>de</strong> ren<strong>de</strong>ment, (2) les dépendances non-gaussiennes entre actifs <strong>et</strong> (3)<br />
les dépendances temporelles intermittentes amenant les gran<strong>de</strong>s pertes tout en<br />
déployant (4) le concept <strong>de</strong> risque sur ses différentes dimensions allant <strong>de</strong>s<br />
“p<strong>et</strong>its” risques jusqu’aux risques “extrêmes” afin <strong>de</strong> définir <strong>de</strong>s fonctions <strong>de</strong><br />
décision cohérentes <strong>et</strong> pr<strong>at</strong>iques perm<strong>et</strong>tant d’établir <strong>de</strong>s <strong>portefeuille</strong>s optimaux.<br />
La capitalis<strong>at</strong>ion totale <strong>de</strong>s marchés financiers à<br />
travers le mon<strong>de</strong> a considérablement augmenté<br />
<strong>de</strong>puis le début <strong>de</strong>s années 1980. En eff<strong>et</strong>, alors<br />
qu’elle ne représente que 3380 milliards <strong>de</strong> dollars<br />
en 1983, soit 4 fois le budg<strong>et</strong> annuel <strong>de</strong>s<br />
Ét<strong>at</strong>s-Unis d’Amérique, elle <strong>at</strong>teint, en 1999, le<br />
chiffre <strong>de</strong> 38700 milliards <strong>de</strong> dollars, soit 22 fois<br />
budg<strong>et</strong> annuel <strong>de</strong>s Ét<strong>at</strong>s-Unis, pour c<strong>et</strong>te annéelà.<br />
Ainsi, en moins <strong>de</strong> vingt ans, la capitalis<strong>at</strong>ion<br />
boursière mondiale est passée <strong>de</strong> 4 fois à 22<br />
fois le budg<strong>et</strong> <strong>de</strong>s Et<strong>at</strong>s-Unis ! Rien qu’en ce qui<br />
concerne la <strong>de</strong>rnière décennie, la capitalis<strong>at</strong>ion<br />
boursière <strong>et</strong> les volumes échangés ont triplé, alors<br />
que le volume d’actions émises a été multiplié par<br />
six. De plus, la vol<strong>at</strong>ilité a connu une croissance<br />
signific<strong>at</strong>ive <strong>de</strong>puis le début <strong>de</strong>s années 1990,<br />
surtout pour les marchés intégrant <strong>de</strong>s sociétés<br />
fortement centrées sur le secteur <strong>de</strong>s technologies<br />
<strong>de</strong> l’inform<strong>at</strong>ion (indice américain Nasdaq ou<br />
finlandais Helsinki General (Hex) par exemple).<br />
La même tendance, certes moins prononcée, est<br />
également visible sur <strong>de</strong>s marchés plus traditionnels<br />
comme par exemple le Cac40 (bourse <strong>de</strong> Paris),<br />
le Dow Jones (Bourse <strong>de</strong> New York) ou le<br />
Ftse100 (bourse <strong>de</strong> Londres).<br />
C<strong>et</strong>te intense <strong>et</strong> lucr<strong>at</strong>ive activité financière est<br />
cependant tempérée par quelques rares mais très<br />
violentes secousses. En eff<strong>et</strong>, l’écl<strong>at</strong>ement <strong>de</strong>s<br />
bulles spécul<strong>at</strong>ives <strong>de</strong> la fin <strong>de</strong>s années 1990,
376 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
ainsi que <strong>de</strong>ux années <strong>de</strong> tourmentes sur les<br />
marchés financiers, ont fait fondre la capitalis<strong>at</strong>ion<br />
boursière mondiale <strong>de</strong> plus <strong>de</strong> 30% par rapport<br />
à son niveau <strong>de</strong> 1999, pour la ramener à<br />
un montant <strong>de</strong> 25100 milliards <strong>de</strong> dollars. Un<br />
autre crash d’une telle ampleur, se déclenchant<br />
simultanément (comme en octobre 1987) dans la<br />
plupart <strong>de</strong>s bourses mondiales, amènerait encore<br />
une perte quasi-instantanée <strong>de</strong> près <strong>de</strong> 7500 milliards<br />
<strong>de</strong> dollars. Ainsi, <strong>de</strong> par les sommes astronomiques<br />
qu’ils engloutissent, les crashs financiers<br />
peuvent anéantir en quelques instants les<br />
plus gros fonds d’investissement, ruinant, par la<br />
même, <strong>de</strong>s années d’épargne <strong>et</strong> <strong>de</strong> financement<br />
<strong>de</strong> r<strong>et</strong>raite. Ce pourrait-il même qu’ils soient,<br />
comme en 1929-33 après le grand crash d’octobre<br />
1929, les précurseurs ou les déclencheurs <strong>de</strong><br />
récessions majeures ? Voire, qu’ils puissent mener<br />
à un écroulement général <strong>de</strong>s systèmes financiers<br />
<strong>et</strong> bancaires qui semblent y avoir échappé <strong>de</strong> justesse<br />
déjà quelques fois dans le passé ?<br />
Les gran<strong>de</strong>s crises <strong>et</strong> les crashs financiers sont<br />
également fascinants parce-qu’ils personnifient<br />
une classe <strong>de</strong> phénomènes appelés “phénomènes<br />
extrêmes”. Des recherches récentes en physique,<br />
psychologie, en <strong>théorie</strong> <strong>de</strong> jeux ou<br />
encore en sciences cognitives au sens large<br />
suggèrent qu’ils sont les caractéristiques incontournables<br />
<strong>de</strong> systèmes complexes autoorganisés.<br />
Marchés turbulents, crises, crashs exposent<br />
donc un investisseur à <strong>de</strong> grands risques<br />
dont la compréhension précise <strong>de</strong>vient essentielle.<br />
Compte-tenu <strong>de</strong> l’évolution <strong>de</strong>s marchés <strong>et</strong> <strong>de</strong><br />
leurs caractéristiques citées plus haut, il est plus<br />
que jamais dans l’intérêt <strong>de</strong>s <strong>gestion</strong>naires <strong>de</strong> <strong>portefeuille</strong>s<br />
<strong>et</strong> <strong>de</strong>s investisseurs en général <strong>de</strong> comprendre<br />
<strong>et</strong> <strong>de</strong> gérer les risques extrêmes.<br />
Distributions<br />
<strong>de</strong>s ren<strong>de</strong>ments<br />
à queues épaisses<br />
Le premier pas vers une quantific<strong>at</strong>ion <strong>de</strong>s grands<br />
risques est d’adm<strong>et</strong>tre que les <strong>st<strong>at</strong>istique</strong>s <strong>de</strong>s<br />
risques – que ce soient les risques <strong>de</strong> marchés associés<br />
aux fluctu<strong>at</strong>ions <strong>de</strong>s actions ou la distri-<br />
bution <strong>de</strong>s remboursements survenant à la suite<br />
<strong>de</strong> sinistres ou <strong>de</strong>s sinistres eux-mêmes – suivent<br />
<strong>de</strong>s distributions à queues épaisses. Ainsi, le paradigme<br />
gaussien, en vogue en finance jusqu’à une<br />
pério<strong>de</strong> rel<strong>at</strong>ivement récente, n’est plus <strong>de</strong> mise<br />
aujourd’hui. Sa disparition a laissé le champ libre<br />
à diverses modélis<strong>at</strong>ions possibles <strong>de</strong>s risques,<br />
par exemple la modélis<strong>at</strong>ion Parétienne, très appliquée<br />
en finance, <strong>et</strong> la modélis<strong>at</strong>ion à l’ai<strong>de</strong> <strong>de</strong>s<br />
distributions dites “exponentielles étirées” que<br />
nous avons développée ces <strong>de</strong>rnières années.<br />
Ces <strong>de</strong>ux classes <strong>de</strong> distributions sont qualifiées<br />
<strong>de</strong> sous-exponentielles, c’est-à-dire que la probabilité<br />
d’occurrence d’événements extrêmes est<br />
plus probable qu’avec une distribution exponentielle.<br />
Cela a pour conséquence immédi<strong>at</strong>e que<br />
<strong>de</strong> telles distributions n’adm<strong>et</strong>tent pas <strong>de</strong> moment<br />
exponentiel, ou pour adopter le langage <strong>de</strong> la<br />
<strong>théorie</strong> <strong>de</strong> la ruine, ces distributions ne s<strong>at</strong>isfont<br />
pas à la condition <strong>de</strong> Cramér-Lundberg.<br />
Sous l’hypothèse que les ren<strong>de</strong>ments sont distribués<br />
<strong>de</strong> manière i<strong>de</strong>ntique <strong>et</strong> indépendante ou<br />
ne possè<strong>de</strong>nt qu’une faible dépendance, ces distributions<br />
caractérisent complètement les risques.<br />
Cela a l’énorme avantage <strong>de</strong> perm<strong>et</strong>tre d’établir<br />
<strong>de</strong>s lois <strong>de</strong> comportement universelles, liées à<br />
certains théorèmes <strong>de</strong> convergence m<strong>at</strong>hém<strong>at</strong>ique<br />
tels que la loi <strong>de</strong>s grands nombres, le théorème <strong>de</strong><br />
la limite centrale, la <strong>théorie</strong> <strong>de</strong>s valeurs extrêmes<br />
<strong>et</strong> <strong>de</strong>s gran<strong>de</strong>s dévi<strong>at</strong>ions. L’immense majorité <strong>de</strong>s<br />
<strong>théorie</strong>s sur la <strong>gestion</strong> <strong>de</strong>s risques, établies aussi<br />
bien en finance qu’en assurance, est fondée sur<br />
c<strong>et</strong>te hypothèse d’indépendance.<br />
Dépendance temporelle<br />
intermittente à l’origine<br />
<strong>de</strong>s gran<strong>de</strong>s pertes<br />
Ce premier pas s’avère en fait très insuffisant<br />
pour apprécier toute la dimension <strong>de</strong>s risques<br />
réels encourus. En eff<strong>et</strong>, <strong>de</strong>s étu<strong>de</strong>s récentes<br />
indiquent que l’hypothèse d’indépendance <strong>de</strong>s<br />
ren<strong>de</strong>ments sur <strong>de</strong>s pério<strong>de</strong>s successives (par<br />
exemple journalières) tombe en défaut lors <strong>de</strong><br />
grands mouvements qui s’avèrent persistants :
12.1. Comprendre <strong>et</strong> gérer les risques grands <strong>et</strong> extrêmes 377<br />
la distribution <strong>de</strong>s drawdowns (ou somme <strong>de</strong><br />
pertes quotidiennes successives) d’un actif, que<br />
ce soit d’un indice financier, du taux <strong>de</strong> change<br />
entre <strong>de</strong>ux monnaies ou <strong>de</strong> la côte d’une action,<br />
présente un comportement anormal pour les très<br />
grands drawdowns. Autrement dit, les très grands<br />
drawdowns n’appartiennent pas à la même popul<strong>at</strong>ion<br />
que le reste <strong>de</strong> la <strong>st<strong>at</strong>istique</strong> observée <strong>et</strong> font<br />
apparaître d’importantes corrél<strong>at</strong>ions sérielles qui<br />
les rend beaucoup plus probables.<br />
Les distributions Parétiennes ou en exponentielles<br />
étirées sont insuffisantes pour quantifier<br />
ces grands risques intermittents, que nous appelons<br />
“outliers”, pour faire référence au vocable<br />
<strong>st<strong>at</strong>istique</strong> désignant <strong>de</strong>s occurrences anormales,<br />
distinctes du reste <strong>de</strong> la popul<strong>at</strong>ion. La<br />
n<strong>at</strong>ure outlier <strong>de</strong>s évènements extrêmes semble<br />
ne pas se confiner aux systèmes financiers<br />
mais a été proposée également pour la rupture<br />
c<strong>at</strong>astrophique <strong>de</strong> m<strong>at</strong>ériaux <strong>et</strong> <strong>de</strong> structures<br />
industrielles, les tremblements <strong>de</strong> terre,<br />
les c<strong>at</strong>astrophes météorologiques <strong>et</strong> enfin divers<br />
phénomènes biologiques <strong>et</strong> sociaux. Les crises<br />
extrêmes semblent donc résulter <strong>de</strong> mécanismes<br />
amplific<strong>at</strong>eurs spécifiques signalant probablement<br />
<strong>de</strong>s phénomènes coopér<strong>at</strong>ifs. Ainsi, <strong>de</strong>s étu<strong>de</strong>s<br />
comportementales, dans lesquelles l’économie<br />
dite “cognitive” tient un grand rôle, perm<strong>et</strong>tent<br />
d’associer ces corrél<strong>at</strong>ions sérielles intermittentes<br />
concomitantes <strong>de</strong>s gran<strong>de</strong>s pertes à certains comportements<br />
<strong>de</strong>s acteurs économiques, tels que <strong>de</strong>s<br />
eff<strong>et</strong>s <strong>de</strong> paniques <strong>et</strong>/ou d’imit<strong>at</strong>ions.<br />
Dans ce contexte, les mesures <strong>de</strong> risques réalisées<br />
à partir d’outils standards comme la VaR<br />
(Value-<strong>at</strong>-Risk) peuvent se révéler totalement<br />
inadéqu<strong>at</strong>es. En eff<strong>et</strong>, une perte journalière <strong>de</strong> 2 %<br />
ou 3 % sur les marchés financiers n’est pas rare,<br />
<strong>et</strong> ne constitue pas un événement extrême. Mais si<br />
une perte d’une telle ampleur vient à se reproduire<br />
plusieurs jours <strong>de</strong> suite, qui plus est en s’amplifiant,<br />
la perte cumulée peut alors <strong>at</strong>teindre 10%,<br />
20% ou même beaucoup plus, ce qui entraîne <strong>de</strong>s<br />
conséquences bien plus dram<strong>at</strong>iques que ne l’indique<br />
la VaR à l’échelle quotidienne. La prise<br />
en compte <strong>de</strong> ce type <strong>de</strong> dépendances sérielles<br />
intermittentes nécessite impér<strong>at</strong>ivement le calcul<br />
d’indic<strong>at</strong>eurs <strong>de</strong> risques à plusieurs échelles temporelles<br />
perm<strong>et</strong>tant <strong>de</strong> couvrir la distribution <strong>de</strong>s<br />
durées <strong>de</strong> drawdowns, comme le suggère le Comité<br />
<strong>de</strong> Basle quand il recomman<strong>de</strong> <strong>de</strong> calculer la<br />
VaR sur un intervalle <strong>de</strong> dix jours. La distribution<br />
<strong>de</strong>s drawdowns fournit un tel indic<strong>at</strong>eur parmi<br />
d’autres. Notons aussi que <strong>de</strong> nombreux investisseurs<br />
professionnels <strong>at</strong>tachent une gran<strong>de</strong> importance<br />
aux drawdowns pour caractériser leurs<br />
risques <strong>et</strong> la qualité d’une str<strong>at</strong>égie ou d’un <strong>portefeuille</strong>.<br />
Dépendance <strong>de</strong> queue<br />
<strong>et</strong> contagion<br />
Cependant, une <strong>gestion</strong> <strong>de</strong>s risques digne <strong>de</strong> ce<br />
nom ne peut se réduire à une étu<strong>de</strong> individuelle <strong>de</strong><br />
chaque actif. En eff<strong>et</strong>, l’épine dorsale <strong>de</strong> la <strong>gestion</strong><br />
<strong>de</strong>s risques est la diversific<strong>at</strong>ion par la constitution<br />
<strong>de</strong> <strong>portefeuille</strong>s <strong>de</strong> risques, aussi bien en assurance<br />
qu’en finance. De la même manière que<br />
le paradigme gaussien est inadéqu<strong>at</strong> pour quantifier<br />
les grands risques <strong>de</strong>s distributions marginales<br />
<strong>de</strong>s ren<strong>de</strong>ments, la covariance intervenant dans la<br />
<strong>théorie</strong> standard du <strong>portefeuille</strong> ne donne qu’une<br />
idée très limitée <strong>de</strong>s grands risques collectifs. Ces<br />
grands risques <strong>de</strong> <strong>portefeuille</strong> résultent en eff<strong>et</strong> <strong>de</strong><br />
la conjonction d’eff<strong>et</strong>s non-gaussien dans les distributions<br />
marginales à queue épaisse <strong>et</strong> dans les<br />
dépendances entre actifs.<br />
On peut donner une idée <strong>de</strong> l’importance <strong>de</strong>s eff<strong>et</strong>s<br />
<strong>de</strong> dépendance non-gaussienne en étudiant la<br />
“dépendance <strong>de</strong> queue” λ c’est-à-dire la probabilité<br />
pour que l’actif X subisse une perte plus<br />
gran<strong>de</strong> que Xq, associée au quantile q tendant vers<br />
zéro, conditionnée à la réalis<strong>at</strong>ion d’une perte <strong>de</strong><br />
l’actif Y plus gran<strong>de</strong> que Yq associée au même<br />
quantile q. Il se trouve que λ est nul pour les actifs<br />
à dépendance gaussienne ! Par contre, dans le<br />
cadre <strong>de</strong>s modèles à facteurs, nous avons montré<br />
que seules les distributions sous-exponentielles<br />
(<strong>et</strong> plus particulièrement Parétiennes) présentent<br />
une dépendance <strong>de</strong> queue asymptotique (pour q<br />
tendant vers 0). Il est possible d’accé<strong>de</strong>r au paramètre<br />
<strong>de</strong> dépendance <strong>de</strong> queue entre <strong>de</strong>ux actifs<br />
par <strong>de</strong>s métho<strong>de</strong>s qui ne font pas appel à une<br />
détermin<strong>at</strong>ion <strong>st<strong>at</strong>istique</strong> directe. Nos tests empiriques<br />
trouvent alors un bon accord entre le ca-
378 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
librage du coefficient <strong>de</strong> dépendance <strong>de</strong> queue <strong>et</strong><br />
les gran<strong>de</strong>s pertes réalisées entre les années 1962<br />
<strong>et</strong> 2000 pour <strong>de</strong>s actions principales <strong>et</strong> <strong>de</strong> grands<br />
indices <strong>de</strong> marché. Conditionné à un grand mouvement<br />
du marché, on peut ainsi déduire la probabilité<br />
que tel ou tel actif subisse une perte du<br />
même ordre.<br />
N<strong>at</strong>ure multidimensionnelle<br />
<strong>de</strong>s risques<br />
Le Graal est <strong>de</strong> conjuguer la <strong>de</strong>scription <strong>de</strong>s<br />
distributions marginales sous-exponentielles avec<br />
les dépendances inter-actifs non-gaussiennes <strong>et</strong><br />
idéalement les dépendances temporelles intermittentes<br />
amenant les gran<strong>de</strong>s pertes pour établir un<br />
<strong>portefeuille</strong> optimal. Le problème est alors que la<br />
notion d’“optimalité” n’est pas évi<strong>de</strong>nte à définir<br />
en pr<strong>at</strong>ique : si la <strong>théorie</strong> économique nous dit<br />
<strong>de</strong> maximiser la fonction d’utilité <strong>de</strong> l’investisseur,<br />
en réalité nous ne la connaissons pas avec<br />
précision. Le problème se complique par les multiples<br />
dimensions du risque introduites par la n<strong>at</strong>ure<br />
non-gaussienne <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ment<br />
<strong>et</strong> <strong>de</strong>s dépendances .<br />
Dans une série d’articles, nous avons développé<br />
une <strong>théorie</strong> du <strong>portefeuille</strong> reposant sur la caractéris<strong>at</strong>ion<br />
<strong>de</strong>s risques par les cumulants <strong>de</strong><br />
la distribution <strong>de</strong>s ren<strong>de</strong>ments du <strong>portefeuille</strong>.<br />
Les cumulants, notés cn, s’expriment comme <strong>de</strong>s<br />
combinaisons <strong>de</strong> moments, <strong>et</strong> quantifient notamment<br />
l’écart à la gaussienne. De même que les<br />
moments, les cumulants n’existent pas tous pour<br />
les distributions Parétiennes mais sont définis à<br />
tout ordre pour les distributions exponentielles<br />
étirées. En particulier, les cumulants d’ordres un<br />
<strong>et</strong> <strong>de</strong>ux sont respectivement la moyenne <strong>et</strong> la variance<br />
<strong>de</strong>s ren<strong>de</strong>ments, tandis que les cumulants<br />
d’ordre trois <strong>et</strong> qu<strong>at</strong>re (après normalis<strong>at</strong>ion par<br />
l’écart type) perm<strong>et</strong>tent <strong>de</strong> définir la skewness<br />
<strong>et</strong> la kurtosis. De façon générale, les cumulants<br />
d’ordres pairs quantifient <strong>de</strong>s risques d’autant plus<br />
grands que l’ordre du cumulant considéré est<br />
élevé, tandis que les cumulants d’ordres impairs<br />
caractérisent la dissymétrie entre les queues positives<br />
<strong>et</strong> nég<strong>at</strong>ives <strong>de</strong> la distribution <strong>de</strong>s ren<strong>de</strong>-<br />
ments. Plus l’ordre n du cumulant cn considéré<br />
est grand, plus celui-ci accor<strong>de</strong> d’importance aux<br />
événements extrêmes. L’ordre n <strong>de</strong>s cumulants allant<br />
<strong>de</strong> 1 à l’infini, varier n revient à étaler ou<br />
développer toutes les dimensions du risque : les<br />
“p<strong>et</strong>its” risques quantifiés par c2 <strong>et</strong> les “grands”<br />
risques quantifiés par c4 <strong>et</strong> les cumulants d’ordres<br />
plus élevés.<br />
Notre <strong>théorie</strong> du <strong>portefeuille</strong> utilise les distributions<br />
marginales <strong>de</strong> la famille <strong>de</strong>s exponentielles<br />
étirées <strong>et</strong> la dépendance entre actif est décrite par<br />
la copule gaussienne. Si on souhaite créer un <strong>portefeuille</strong><br />
qui évite les grands risques, on choisit le<br />
poids <strong>de</strong>s actifs <strong>de</strong> telle manière que les cumulants<br />
c4, c6, c8, <strong>et</strong>c. soient tous proches <strong>de</strong> leur<br />
minimum, tout en laissant libre la variance c2 (p<strong>et</strong>its<br />
risques). C<strong>et</strong>te approche est très différente <strong>de</strong><br />
l’approche standard <strong>de</strong> Markowitz qui se focalise<br />
sur c2 <strong>et</strong> <strong>de</strong> plus construit une frontière efficiente<br />
dans l’espace (ren<strong>de</strong>ment-variance).<br />
P<strong>et</strong>its risques, grands<br />
risques <strong>et</strong> ren<strong>de</strong>ment<br />
Pour illustrer l’impact <strong>de</strong> la décomposition<br />
<strong>de</strong>s risques en “p<strong>et</strong>its” <strong>et</strong> “grands” risques,<br />
considérons le cas simple d’un <strong>portefeuille</strong> avec<br />
seulement <strong>de</strong>ux actifs : l’action Chevron <strong>et</strong> la <strong>de</strong>vise<br />
malaise : le Ringgit. Ces <strong>de</strong>ux actifs ont <strong>de</strong>s<br />
caractéristiques très différentes <strong>et</strong> illustrent admirablement<br />
un eff<strong>et</strong> surprenant a priori. La figure 1a<br />
montre le ren<strong>de</strong>ment quotidien d’un <strong>portefeuille</strong><br />
dont la proportion w1, investie dans l’action Chevron,<br />
a été obtenue en minimisant la variance. La<br />
figure 1b donne la solution <strong>de</strong> la minimis<strong>at</strong>ion <strong>de</strong><br />
cn vis-à-vis du poids w1. Les lignes <strong>de</strong>s points<br />
horizontaux sont les valeurs maximales du ren<strong>de</strong>ment<br />
quotidien dans le cas où l’on optimise c2n,<br />
pour n>1. Les ren<strong>de</strong>ments quotidiens pour le <strong>portefeuille</strong><br />
<strong>de</strong> la figure 1b surpassent ces limites,<br />
i.e. le <strong>portefeuille</strong> <strong>de</strong> la figure 1a subit plus <strong>de</strong><br />
fluctu<strong>at</strong>ions <strong>de</strong> gran<strong>de</strong>s amplitu<strong>de</strong>s.Ces <strong>de</strong>ux figures<br />
illustrent clairement le fait que minimiser<br />
<strong>de</strong>s p<strong>et</strong>its risques peut faire augmenter les grands<br />
risques ! De plus, le gain cumul<strong>at</strong>if <strong>de</strong> la figure 1c<br />
montre que le <strong>portefeuille</strong> <strong>de</strong> la figure 1b voit son
12.1. Comprendre <strong>et</strong> gérer les risques grands <strong>et</strong> extrêmes 379<br />
gain s’accroître considérablement par rapport au<br />
<strong>portefeuille</strong> standard à la Markovitz. Autrement<br />
dit “on peut avoir le beurre <strong>et</strong> l’argent du beur-<br />
re” : diminuer les grands risques <strong>et</strong> augmenter le<br />
profit !<br />
Ren<strong>de</strong>ments quotidiens annualisés (en pourcentage) <strong>et</strong> gain cumulé pour les <strong>de</strong>ux <strong>portefeuille</strong>s<br />
correspondants au minimum <strong>de</strong> la variance (poids Chevron w1 = 0,095) <strong>et</strong> au minimum <strong>de</strong>s cumulants<br />
c2n, d’ordre 2n > 2 (poids Chevron w1 = 0,38).<br />
Ren<strong>de</strong>ment quotiedien<br />
Ren<strong>de</strong>ment quotiedien<br />
Richesse cummulée<br />
10<br />
¤<br />
20<br />
20<br />
10<br />
0<br />
¤<br />
20<br />
10<br />
¤<br />
20<br />
10<br />
0<br />
£<br />
6<br />
¦<br />
4<br />
¤<br />
2<br />
0<br />
0<br />
0<br />
0<br />
¡<br />
1000 2000 3000<br />
¤<br />
1000 2000<br />
¡<br />
3000<br />
¡<br />
1000 2000 3000<br />
Le mécanisme <strong>de</strong> c<strong>et</strong> eff<strong>et</strong> remarquable est<br />
simple : le Ringgit malais contribue le plus au<br />
cumulant d’ordres élevés (grands risques) <strong>et</strong> a <strong>de</strong><br />
plus un ren<strong>de</strong>ment très faible par rapport à celui <strong>de</strong><br />
Chevron. Par contre, sa distribution étroite dans la<br />
zone <strong>de</strong>s p<strong>et</strong>its ren<strong>de</strong>ments lui donne une faible<br />
variance. L’optimis<strong>at</strong>ion à la Markowitz m<strong>et</strong>tra<br />
donc plus <strong>de</strong> poids sur le Ringgit qui semble apporter<br />
une diversific<strong>at</strong>ion intéressante du point <strong>de</strong><br />
vue <strong>de</strong> la variance. Mais cela est une illusion dangereuse,<br />
car le risque réel du Ringgit est beaucoup<br />
plus grand que ne le fait croire sa variance. Le cumulant<br />
c4, par exemple, le quantifie clairement <strong>et</strong><br />
sa minimis<strong>at</strong>ion conduit en conséquence à réduire<br />
le poids <strong>de</strong> la <strong>de</strong>vise malaisienne dans le porte-<br />
¥<br />
w1=0.095<br />
w1=0.38<br />
¢<br />
4000 5000<br />
¦<br />
4000<br />
¢<br />
5000<br />
¢<br />
4000 5000<br />
£<br />
6000<br />
£<br />
6000<br />
w1=0.38<br />
w1=0.095<br />
£<br />
6000<br />
7000<br />
§<br />
7000<br />
7000<br />
feuille. Du coup, les grands risques sont réduits.<br />
Comme le Ringgit n’a que peu <strong>de</strong> ren<strong>de</strong>ment, on<br />
gagne alors sur les <strong>de</strong>ux tableaux car alors le ren<strong>de</strong>ment<br />
augmente n<strong>et</strong>tement.<br />
Bibliographie<br />
An<strong>de</strong>rsen, J. V. and D. Sorn<strong>et</strong>te (2001) Have your<br />
cake and e<strong>at</strong> it too : increasing r<strong>et</strong>urns while lowering<br />
large risks ! Journal of Risk Finance 2 (3),<br />
70-82.<br />
Embrechts P., C. Klüppelberg <strong>et</strong> T. Mikosch<br />
(1997), Mo<strong>de</strong>lling Extremal Events for Insurance<br />
(a)<br />
(b)<br />
(c)
380 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
and Finance (Springer, New York).<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te (2001) General<br />
framework for a portfolio theory with non-<br />
Gaussian risks and non-linear correl<strong>at</strong>ions, paper<br />
presented <strong>at</strong> the 18th Intern<strong>at</strong>ional conference in<br />
Finance, June 2001, Namur, Belgium (e-print <strong>at</strong><br />
http : //arXiv.org/abs/cond−m<strong>at</strong>/0103020).<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te (2002) Tail Depen<strong>de</strong>nce<br />
of Factor Mo<strong>de</strong>ls, working paper (http :<br />
//arXiv.org/abs/cond − m<strong>at</strong>/0202356).<br />
Johansen, A. <strong>et</strong> D. Sorn<strong>et</strong>te (2002), “Large<br />
Price Drawdowns Are Outliers”, Journal of<br />
Risk 4 (2), (http : //arXiv.org/abs/cond −<br />
m<strong>at</strong>/0010050).<br />
Sorn<strong>et</strong>te, D. (1999) Complexity, c<strong>at</strong>astrophe and<br />
physics, Physics World 12 (N12), 57-57.<br />
Sorn<strong>et</strong>te, D. (2002) Predictability of c<strong>at</strong>astrophic<br />
events : m<strong>at</strong>erial rupture, earthquakes, turbulence,<br />
financial crashes and human birth, Proceedings<br />
of the N<strong>at</strong>ional Aca<strong>de</strong>my of Sciences USA, vol.<br />
4, (e-print <strong>at</strong> http : //arXiv.org/abs/cond −<br />
m<strong>at</strong>/0107173).<br />
Sorn<strong>et</strong>te, D., P. Simon<strong>et</strong>ti and J. V. An<strong>de</strong>rsen<br />
(2000) φ q -field theory for Portfolio optimiz<strong>at</strong>ion :<br />
``f<strong>at</strong> tails” and non-linear correl<strong>at</strong>ions, Physics<br />
Report 335 (2), 19-92.<br />
Zaj<strong>de</strong>nweber, D., Economie <strong>de</strong>s extrêmes (Flammarion,<br />
2000).
12.2. Minimiser l’impact <strong>de</strong>s grands co-mouvements 381<br />
12.2 Minimiser l’impact <strong>de</strong>s grands co-mouvements<br />
A l’ai<strong>de</strong> <strong>de</strong>s résult<strong>at</strong>s exposés au chapitre 9 section 9.2, nous montrons comment le coefficient <strong>de</strong><br />
dépendance <strong>de</strong> queue entre un actif <strong>et</strong> un <strong>de</strong> ces facteurs explic<strong>at</strong>ifs (le marché par exemple) ou entre<br />
<strong>de</strong>ux actifs, peut facilement être calibré. Nous construisons ensuite <strong>de</strong>s <strong>portefeuille</strong>s composés d’actifs<br />
ayant <strong>de</strong> très faibles coefficients <strong>de</strong> dépendance <strong>de</strong> queue avec le marché <strong>et</strong> montrons qu’ils présentent<br />
une corrél<strong>at</strong>ion remarquablement plus faible que <strong>de</strong>s <strong>portefeuille</strong>s composés d’actifs ayant une plus forte<br />
dépendance <strong>de</strong> queue avec le marché, <strong>et</strong> ce sans dégrad<strong>at</strong>ion <strong>de</strong> la performance mesurée par le r<strong>at</strong>io <strong>de</strong><br />
Sharpe.<br />
Reprint from : Y. Malevergne <strong>et</strong> D.Sorn<strong>et</strong>te (2002), Minimizing extremes, RISK 15(11), 129-134.
382 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
M<br />
ore than 100 years ago, Vilfred Par<strong>et</strong>o discovered a st<strong>at</strong>istical rel<strong>at</strong>ionship,<br />
now known as the 80-20 rule, th<strong>at</strong> manifests itself over<br />
and over in large systems: “In any series of elements to be controlled,<br />
a selected small fraction, in terms of numbers of elements, always<br />
accounts for a large fraction in terms of effect.” The stock mark<strong>et</strong> is no exception:<br />
events occurring over a very small fraction of the total invested<br />
time may account for most of the gains and/or losses. Diversifying away<br />
such large risks requires novel approaches to portfolio management, which<br />
must take into account the non-Gaussian f<strong>at</strong> tail structure of distributions<br />
of r<strong>et</strong>urns and their <strong>de</strong>pen<strong>de</strong>nce. Recent economic shocks and crashes<br />
have shown th<strong>at</strong> standard portfolio diversific<strong>at</strong>ion works well in normal<br />
times but may break down in stressful times, precisely when diversific<strong>at</strong>ion<br />
is most important. One could say th<strong>at</strong> diversific<strong>at</strong>ion works when one<br />
does not really need it and may fail severely when it is most nee<strong>de</strong>d.<br />
Technically, the question boils down to wh<strong>et</strong>her large price movements<br />
occur mainly in an isol<strong>at</strong>ed manner or in a co-ordin<strong>at</strong>ed way. This question<br />
is vital for fund managers who take advantage of the diversific<strong>at</strong>ion<br />
to minimise their risks. Here, we introduce a new technique to quantify<br />
and empirically estim<strong>at</strong>e the propensity for ass<strong>et</strong>s to exhibit extreme comovements,<br />
through the use of the so-called coefficient of tail <strong>de</strong>pen<strong>de</strong>nce.<br />
Using a factor mo<strong>de</strong>l framework and tools from extreme value<br />
theory, we provi<strong>de</strong> novel analytical formulas for the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween arbitrary ass<strong>et</strong>s, which yields an efficient non-param<strong>et</strong>ric<br />
estim<strong>at</strong>or. We then construct portfolios of stocks with minimal tail<br />
<strong>de</strong>pen<strong>de</strong>nce with the mark<strong>et</strong> represented by the S&P 500, and show th<strong>at</strong><br />
their superior behaviour in stressed times comes tog<strong>et</strong>her with qualities<br />
in terms of Sharpe r<strong>at</strong>io and standard quality measures th<strong>at</strong> are <strong>at</strong> least as<br />
good as standard portfolios.<br />
Assessing large co-movements<br />
Standard estim<strong>at</strong>ors of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s inclu<strong>de</strong> the correl<strong>at</strong>ion<br />
coefficient and the Spearman’s rank correl<strong>at</strong>ion. However, as stressed<br />
by Embrechts, McNeil & Straumann (1999), these kind of <strong>de</strong>pen<strong>de</strong>nce measures<br />
suffer from many <strong>de</strong>ficiencies. Moreover, their values are mostly controlled<br />
by rel<strong>at</strong>ively small moves of the ass<strong>et</strong> prices around their mean. To<br />
solve this problem, it has been proposed to use the correl<strong>at</strong>ion coefficients<br />
conditioned on large movements of the ass<strong>et</strong>s. But Boyer, Gibson & Laur<strong>et</strong>an<br />
(1997) have emphasised th<strong>at</strong> this approach suffers also from a severe<br />
system<strong>at</strong>ic bias leading to spurious str<strong>at</strong>egies: the conditional<br />
correl<strong>at</strong>ion in general evolves with time even when the true non-conditional<br />
correl<strong>at</strong>ion remains constant. In fact, Malevergne & Sorn<strong>et</strong>te (2002a)<br />
have shown th<strong>at</strong> any approach based on conditional <strong>de</strong>pen<strong>de</strong>nce measures<br />
implies a spurious change of the intrinsic value of the <strong>de</strong>pen<strong>de</strong>nce,<br />
measured for instance by copulas. Recall th<strong>at</strong> the copula of several random<br />
variables is the (unique) function (for continuous marginals) th<strong>at</strong> compl<strong>et</strong>ely<br />
embodies the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween these variables, irrespective of<br />
their marginal behaviour (see Nelsen, 1998, for a m<strong>at</strong>hem<strong>at</strong>ical <strong>de</strong>scription<br />
of the notion of copula).<br />
In view of these limit<strong>at</strong>ions of the standard st<strong>at</strong>istical tools, it is n<strong>at</strong>ural<br />
to turn to extreme value theory. In the univari<strong>at</strong>e case, extreme value theory<br />
is very useful and provi<strong>de</strong>s many tools for investig<strong>at</strong>ing the extreme<br />
Portfolio tail risk l<br />
tails of distributions of ass<strong>et</strong>s’ r<strong>et</strong>urns. These new <strong>de</strong>velopments rest on<br />
the existence of a few fundamental results on extremes, such as the Gne<strong>de</strong>nko-Pickands-Balkema-<strong>de</strong><br />
Haan theorem, which gives a general expression<br />
for the conditional distribution of exceedance over a large<br />
threshold. In this framework, the study of large and extreme co-movements<br />
requires the multivari<strong>at</strong>e extreme values theory, which, in contrast with the<br />
univari<strong>at</strong>e case, cannot be used to constrain accur<strong>at</strong>ely the distribution of<br />
large co-movements since the class of limiting extreme-value distributions<br />
is too broad.<br />
In the spirit of the mean-variance portfolio or of utility theory, which<br />
establish an investment <strong>de</strong>cision on a unique risk measure, we use the coefficient<br />
of tail <strong>de</strong>pen<strong>de</strong>nce, which, to our knowledge, was first introduced<br />
in a financial context by Embrechts, McNeil & Straumann (2002). The coefficient<br />
of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s Xi and Xj is a very n<strong>at</strong>ural and<br />
easily comprehensible measure of extreme co-movements. It is <strong>de</strong>fined as<br />
the probability th<strong>at</strong> the ass<strong>et</strong> Xi incurs a large loss (or gain) assuming th<strong>at</strong><br />
the ass<strong>et</strong> Xj also un<strong>de</strong>rgoes a large loss (or gain) <strong>at</strong> the same probability<br />
level, in the limit where this probability level explores the extreme tails of<br />
the distribution of r<strong>et</strong>urns of the two ass<strong>et</strong>s. M<strong>at</strong>hem<strong>at</strong>ically speaking, the<br />
coefficient of lower tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two ass<strong>et</strong>s Xi and Xj , <strong>de</strong>noted<br />
by λ _<br />
ij , is <strong>de</strong>fined by:<br />
−<br />
λij u→0<br />
−1 −1<br />
{ Xi Fi u Xj Fj u }<br />
= lim Pr < ( ) < ( )<br />
where F i –1 (u) and Fj –1 (u) represent the quantiles of ass<strong>et</strong>s Xi and X j <strong>at</strong> the<br />
level u. Similarly, the coefficient of upper tail <strong>de</strong>pen<strong>de</strong>nce is:<br />
{ }<br />
+<br />
−1 −1<br />
= > ( ) > ( )<br />
Cutting edge<br />
Minimising extremes<br />
Portfolio diversific<strong>at</strong>ion often breaks down in stressed mark<strong>et</strong> environments, but the comovement<br />
of ass<strong>et</strong> prices in a tail risk regime may be mo<strong>de</strong>lled using a coefficient of tail<br />
<strong>de</strong>pen<strong>de</strong>nce. Here, Yannick Malevergne and Didier Sorn<strong>et</strong>te show how such coefficients can<br />
be estim<strong>at</strong>ed analytically using the param<strong>et</strong>ers of factor mo<strong>de</strong>ls, while avoiding the problem<br />
of un<strong>de</strong>r-sampling of extreme values<br />
(2)<br />
λ _<br />
ij (respectively λ+ ij ) is of concern to investors with long (respectively<br />
short) positions. We refer to Coles, Heffernan & Tawn (1999) and references<br />
therein for a survey of the properties of the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce.<br />
L<strong>et</strong> us stress th<strong>at</strong> the use of quantiles in the <strong>de</strong>finition of λ _<br />
ij<br />
and λ + ij makes them in<strong>de</strong>pen<strong>de</strong>nt of the marginal distribution of the ass<strong>et</strong><br />
r<strong>et</strong>urns. As a consequence, the tail <strong>de</strong>pen<strong>de</strong>nce param<strong>et</strong>ers are intrinsic<br />
<strong>de</strong>pen<strong>de</strong>nce measures. The obvious gain is an ‘orthogonal’ <strong>de</strong>composition<br />
of the risks into (1) individual risks carried by the marginal distribution<br />
of each ass<strong>et</strong> and (2) their collective risk <strong>de</strong>scribed by their<br />
<strong>de</strong>pen<strong>de</strong>nce structure or copula.<br />
Being a probability, the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce varies b<strong>et</strong>ween<br />
zero and one. A large value of λ _<br />
ij means th<strong>at</strong> large losses are more likely<br />
to occur tog<strong>et</strong>her. Then, large risks cannot be diversified away and the ass<strong>et</strong>s<br />
crash tog<strong>et</strong>her. This investor and portfolio manager nightmare is further<br />
amplified in real life situ<strong>at</strong>ions by the limited liquidity of mark<strong>et</strong>s.<br />
When λ _<br />
λij lim Pr Xi Fi u Xj Fj u<br />
u→1<br />
ij vanishes, these ass<strong>et</strong>s are said to be asymptotically in<strong>de</strong>pen<strong>de</strong>nt,<br />
but this term hi<strong>de</strong>s the subtl<strong>et</strong>y th<strong>at</strong> the ass<strong>et</strong>s can still present a non-zero<br />
<strong>de</strong>pen<strong>de</strong>nce in their tails. For instance, two ass<strong>et</strong>s with a bivari<strong>at</strong>e normal<br />
distribution with correl<strong>at</strong>ion coefficient less than one can be shown to have<br />
a vanishing coefficient of tail <strong>de</strong>pen<strong>de</strong>nce. Nevertheless, unless their correl<strong>at</strong>ion<br />
coefficient is zero, these ass<strong>et</strong>s are never in<strong>de</strong>pen<strong>de</strong>nt. Thus, asymptotic<br />
in<strong>de</strong>pen<strong>de</strong>nce must be un<strong>de</strong>rstood as the weakest <strong>de</strong>pen<strong>de</strong>nce<br />
(1)<br />
WWW.RISK.NET ● NOVEMBER 2002 RISK 129
12.2. Minimiser l’impact <strong>de</strong>s grands co-mouvements 383<br />
Cutting edge l<br />
Portfolio tail risk<br />
1. Tail <strong>de</strong>pen<strong>de</strong>nce versus correl<strong>at</strong>ion<br />
λ<br />
λ<br />
1.0<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
0<br />
–1.0 –0.8 –0.6 –0.4 –0.2 0<br />
ρ<br />
0.2 0.4 0.6 0.8 1.0<br />
1.0<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
Tail in<strong>de</strong>x ν = 3<br />
Tail in<strong>de</strong>x ν = 10<br />
0<br />
–1.0 –0.8 –0.6 –0.4 –0.2 0<br />
ρ<br />
0.2 0.4 0.6 0.8 1.0<br />
Evolution as a function of the correl<strong>at</strong>ion coefficient ρ of the coefficient<br />
of tail <strong>de</strong>pen<strong>de</strong>nce for an elliptical bivari<strong>at</strong>e Stu<strong>de</strong>nt distribution (solid<br />
line) and for the additive factor mo<strong>de</strong>l with Stu<strong>de</strong>nt factor and noise<br />
(dashed line)<br />
th<strong>at</strong> can be quantified by the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce (for other d<strong>et</strong>ails,<br />
the rea<strong>de</strong>r is referred to Ledford & Tawn, 1998).<br />
For practical implement<strong>at</strong>ions, a direct applic<strong>at</strong>ion of the <strong>de</strong>finitions (1)<br />
and (2) fails to provi<strong>de</strong> reasonable estim<strong>at</strong>ions due to the double curse of<br />
dimensionality and un<strong>de</strong>r-sampling of extreme values, so th<strong>at</strong> a fully nonparam<strong>et</strong>ric<br />
approach is not reliable. It turns out to be possible to circumvent<br />
this fundamental difficulty by consi<strong>de</strong>ring the general class of factor<br />
mo<strong>de</strong>ls, which are among the most wi<strong>de</strong>spread and vers<strong>at</strong>ile mo<strong>de</strong>ls in finance.<br />
They come in two classes: multiplic<strong>at</strong>ive and additive factor mo<strong>de</strong>ls<br />
respectively. The multiplic<strong>at</strong>ive factor mo<strong>de</strong>ls are generally used to<br />
mo<strong>de</strong>l ass<strong>et</strong> fluctu<strong>at</strong>ions due to an un<strong>de</strong>rlying stochastic vol<strong>at</strong>ility (see, for<br />
example, Hull & White, 1987, and Taylor, 1994, for a survey of the properties<br />
of these mo<strong>de</strong>ls). The additive factor mo<strong>de</strong>ls are ma<strong>de</strong> to rel<strong>at</strong>e ass<strong>et</strong><br />
fluctu<strong>at</strong>ions to mark<strong>et</strong> fluctu<strong>at</strong>ions, as in the capital ass<strong>et</strong> pricing mo<strong>de</strong>l<br />
and its generalis<strong>at</strong>ions (see, for example, Sharpe, 1964, and Rubinstein,<br />
1973), or to any s<strong>et</strong> of common factors as in Ross’ (1976) arbitrage pricing<br />
theory. The coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is known in closed form for both<br />
classes of factor mo<strong>de</strong>ls, which allows, as we shall see, for an efficient empirical<br />
estim<strong>at</strong>ion.<br />
Tail <strong>de</strong>pen<strong>de</strong>nce gener<strong>at</strong>ed by factor mo<strong>de</strong>ls<br />
We first examine multiplic<strong>at</strong>ive factor mo<strong>de</strong>ls, which account for most of<br />
the stylised facts observed on financial time series. A multivari<strong>at</strong>e stochastic<br />
vol<strong>at</strong>ility mo<strong>de</strong>l with a common stochastic vol<strong>at</strong>ility factor can be<br />
written as:<br />
130 RISK NOVEMBER 2002 ● WWW.RISK.NET<br />
where σ is a positive random variable mo<strong>de</strong>lling the vol<strong>at</strong>ility, Y is a Gaussian<br />
random vector, in<strong>de</strong>pen<strong>de</strong>nt of σ, and X is the vector of ass<strong>et</strong> r<strong>et</strong>urns.<br />
In this framework, the multivari<strong>at</strong>e distribution of ass<strong>et</strong> r<strong>et</strong>urns X is an elliptical<br />
multivari<strong>at</strong>e distribution. For instance, if the inverse of the square<br />
of the vol<strong>at</strong>ility 1/σ2 is a constant times a χ2-distributed random variable<br />
with ν <strong>de</strong>grees of freedom, the distribution of ass<strong>et</strong> r<strong>et</strong>urns will be the Stu<strong>de</strong>nt<br />
distribution with ν <strong>de</strong>grees of freedom. When the vol<strong>at</strong>ility follows<br />
Arch or Garch processes, the ass<strong>et</strong> r<strong>et</strong>urns are also elliptically distributed<br />
with f<strong>at</strong>-tailed marginal distributions. Thus, any ass<strong>et</strong> Xi is asymptotically<br />
distributed according to a regularly varying distribution1 : Pr{|Xi | > x} ~ L(x)<br />
× x –ν , where L(⋅) <strong>de</strong>notes a slowly varying function, with the same exponent<br />
ν for all ass<strong>et</strong>s, due to the ellipticity of their multivari<strong>at</strong>e distribution.<br />
Hult & Lindskog (2002) have shown th<strong>at</strong> the necessary and sufficient<br />
condition for any two ass<strong>et</strong>s Xi and Xj with an elliptical multivari<strong>at</strong>e distribution<br />
to have a non-vanishing coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is th<strong>at</strong> their<br />
distribution be regularly varying. Denoting by ρij the correl<strong>at</strong>ion coefficient<br />
b<strong>et</strong>ween the ass<strong>et</strong>s Xi and Xj and by ν the tail in<strong>de</strong>x of their distributions,<br />
they obtain:<br />
π / 2<br />
ν<br />
∫<br />
dt cos t<br />
± ( π/ 2−arcsin ρij<br />
) / 2<br />
ν + 1<br />
λij<br />
= = 2I<br />
1+<br />
ρ ,<br />
π / 2<br />
ij<br />
(4)<br />
ν<br />
dt cos t<br />
2 2<br />
1 ⎛ ⎞<br />
⎝<br />
⎜<br />
2⎠<br />
⎟<br />
where the function:<br />
∫<br />
0<br />
X =σY<br />
1<br />
Ix( z, w)=<br />
Bzw ( , )<br />
( )<br />
x z−1<br />
w−1<br />
dt t 1 − t<br />
0<br />
<strong>de</strong>notes the incompl<strong>et</strong>e b<strong>et</strong>a function. This expression holds for any regularly<br />
varying elliptical distribution, irrespective of the exact shape of the<br />
distribution. Only the tail in<strong>de</strong>x is important in the d<strong>et</strong>ermin<strong>at</strong>ion of the<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce because λ ± ij probes the extreme end of the<br />
tail of the distributions, which all have, roughly speaking, the same behaviour<br />
for regularly varying distributions. In contrast, when the marginal<br />
distributions <strong>de</strong>cay faster than any power law, such as the Gaussian distribution,<br />
the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce is zero.<br />
L<strong>et</strong> us now turn to the second class of additive factor mo<strong>de</strong>ls, whose<br />
introduction in finance goes back <strong>at</strong> least to the arbitrage pricing theory<br />
(Ross, 1976). They are now wi<strong>de</strong>ly used in many branches of finance, including<br />
to mo<strong>de</strong>l stock r<strong>et</strong>urns, interest r<strong>at</strong>es and credit risks. Here, we<br />
shall only consi<strong>de</strong>r the effect of a single factor, which may represent the<br />
mark<strong>et</strong>, for example. This factor will be <strong>de</strong>noted by Y and its cumul<strong>at</strong>ive<br />
distribution by FY . As previously, the vector X is the vector of ass<strong>et</strong> r<strong>et</strong>urns<br />
and ε will <strong>de</strong>note the vector of idiosyncr<strong>at</strong>ic noises assumed in<strong>de</strong>pen<strong>de</strong>nt2 of Y. β is the vector whose components are the regression coefficients of<br />
the Xi on the factor Y. Thus, the factor mo<strong>de</strong>l reads:<br />
(6)<br />
In contrast with multiplic<strong>at</strong>ive factor mo<strong>de</strong>ls, the multivari<strong>at</strong>e distribution<br />
of X cannot be obtained in an analytical form, in the general case. In<br />
the particular situ<strong>at</strong>ion when Y and ε are normally distributed, the multivari<strong>at</strong>e<br />
distribution of X is also normal but this case is not very interesting.<br />
In a sense, additive factor mo<strong>de</strong>ls are richer than the multiplic<strong>at</strong>ive ones,<br />
since they give birth to a larger s<strong>et</strong> of distributions of ass<strong>et</strong> r<strong>et</strong>urns.<br />
Notwithstanding these difficulties, it turns out to be possible to obtain<br />
the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce for any pair of ass<strong>et</strong>s Xi and Xj . In a first<br />
step, l<strong>et</strong> us consi<strong>de</strong>r the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce λ ± i b<strong>et</strong>ween any<br />
ass<strong>et</strong> Xi and the factor Y itself. Malevergne & Sorn<strong>et</strong>te (2002b) have shown<br />
th<strong>at</strong> λ ± X = βY+ ε<br />
i is also i<strong>de</strong>ntically zero for all rapidly varying factors, th<strong>at</strong> is, for all<br />
factors whose distribution <strong>de</strong>cays faster than any power law, such as the<br />
Gaussian, exponential or gamma laws. When the factor Y has a distribution<br />
th<strong>at</strong> is regularly varying with tail in<strong>de</strong>x ν, we have:<br />
1 See Bingham, Goldie & Teugel (1987) for d<strong>et</strong>ails on regular vari<strong>at</strong>ions<br />
2 In fact, ε and Y can be weakly <strong>de</strong>pen<strong>de</strong>nt (see Malevergne & Sorn<strong>et</strong>te, 2002b, for d<strong>et</strong>ails)<br />
∫<br />
(3)<br />
(5)
384 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
where:<br />
λ<br />
+<br />
i =<br />
l =<br />
A similar expression holds for λ – i , which is obtained by simply replacing<br />
the limit u → 1 by u → 0 in the <strong>de</strong>finition of l. λ ± i is non-zero as long<br />
as l remains finite, th<strong>at</strong> is, when the tail of the distribution of the factor is<br />
not thinner than the tail of the idiosyncr<strong>at</strong>ic noise εi . Therefore, two conditions<br />
must hold for the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce to be non-zero:<br />
the factor must be intrinsically ‘wild’ (to use the terminology of Man<strong>de</strong>lbrot,<br />
1997) so th<strong>at</strong> its distribution is regularly varying; and<br />
the factor must be sufficiently ‘wild’ in its intrinsic variability, so th<strong>at</strong> its<br />
influence is not domin<strong>at</strong>ed by the idiosyncr<strong>at</strong>ic component of the ass<strong>et</strong>.<br />
Then, the amplitu<strong>de</strong> of λ ± i is d<strong>et</strong>ermined by the tra<strong>de</strong>-off b<strong>et</strong>ween the rel<strong>at</strong>ive<br />
tail behaviours of the factor and the idiosyncr<strong>at</strong>ic noise.<br />
As an example, l<strong>et</strong> us consi<strong>de</strong>r th<strong>at</strong> the factor and the idiosyncr<strong>at</strong>ic noise<br />
follow Stu<strong>de</strong>nt distribution with νY and ν <strong>de</strong>grees of freedom and scale<br />
εi<br />
factor σY and σ respectively. Expression (7) leads to:<br />
εi<br />
λi = 0 if νY > νεi<br />
λi =<br />
1<br />
1<br />
σεi<br />
ν<br />
if νY = νε= ν<br />
i<br />
+ ( βσ i Y)<br />
−1<br />
lim<br />
u→1<br />
Y<br />
ν<br />
l { 1 β } i<br />
( )<br />
( )<br />
FXu −1<br />
F u<br />
λ = 1 if ν < ν<br />
i Y<br />
The tail <strong>de</strong>pen<strong>de</strong>nce <strong>de</strong>creases when the idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility increases<br />
rel<strong>at</strong>ive to the factor vol<strong>at</strong>ility. Therefore, λi <strong>de</strong>creases in periods<br />
of high idiosyncr<strong>at</strong>ic vol<strong>at</strong>ility and increases in periods of high mark<strong>et</strong><br />
vol<strong>at</strong>ility. From the viewpoint of the tail <strong>de</strong>pen<strong>de</strong>nce, the vol<strong>at</strong>ility of an<br />
ass<strong>et</strong> is not relevant. Wh<strong>at</strong> is governing extreme co-movement is the rel<strong>at</strong>ive<br />
weights of the different components of the vol<strong>at</strong>ility of the ass<strong>et</strong>.<br />
Figure 1 compares the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce as a function of<br />
the correl<strong>at</strong>ion coefficient for the bivari<strong>at</strong>e Stu<strong>de</strong>nt distribution (expression<br />
(4)) and for the factor mo<strong>de</strong>l with the factor and the idiosyncr<strong>at</strong>ic noise<br />
following Stu<strong>de</strong>nt distributions (equ<strong>at</strong>ion (8)). Contrary to the coefficient<br />
of tail <strong>de</strong>pen<strong>de</strong>nce of the Stu<strong>de</strong>nt factor mo<strong>de</strong>l, the tail <strong>de</strong>pen<strong>de</strong>nce of the<br />
(elliptical) Stu<strong>de</strong>nt distribution does not vanish for neg<strong>at</strong>ive correl<strong>at</strong>ion coefficients.<br />
For large values of the correl<strong>at</strong>ion coefficient, the former is always<br />
larger than the l<strong>at</strong>ter.<br />
Once the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s and the<br />
common factor are known, the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween any<br />
two ass<strong>et</strong>s Xi and Xj with a common factor Y is simply equal to the weakest<br />
tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s and their common factor:<br />
λ min λ , λ<br />
(9)<br />
= { }<br />
ij i j<br />
This result is very intuitive: since the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the two ass<strong>et</strong>s<br />
is due to their common factor, this <strong>de</strong>pen<strong>de</strong>nce cannot be stronger than<br />
the weakest <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween each of the ass<strong>et</strong>s and the factor.<br />
Practical implement<strong>at</strong>ion and consequences<br />
The two m<strong>at</strong>hem<strong>at</strong>ical results (4) and (7) have a very important practical<br />
effect for estim<strong>at</strong>ing the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce. As we have already<br />
pointed out, its direct estim<strong>at</strong>ion is essentially impossible since, by <strong>de</strong>finition,<br />
the number of observ<strong>at</strong>ions goes to zero as the probability level of<br />
the quantile goes to zero (or one). In contrast, the formulas (4) and (7–9)<br />
tell us th<strong>at</strong> one has just to estim<strong>at</strong>e a tail in<strong>de</strong>x and a correl<strong>at</strong>ion coefficient.<br />
These estim<strong>at</strong>ions can be reasonably accur<strong>at</strong>e because they make<br />
use of a significant part of the d<strong>at</strong>a beyond the few extremes targ<strong>et</strong>ed by<br />
λ. Moreover, equ<strong>at</strong>ion (7) does not explicitly assume a power law behaviour,<br />
but only a regularly varying behaviour, which is far more general. In<br />
1<br />
max ,<br />
εi<br />
(7)<br />
(8)<br />
A. Coefficients of tail <strong>de</strong>pen<strong>de</strong>nce<br />
Lower tail <strong>de</strong>pen<strong>de</strong>nce Upper tail <strong>de</strong>pen<strong>de</strong>nce<br />
Bristol-Myers Squibb 0.16 (0.03) 0.14 (0.01)<br />
Chevron 0.05 (0.01) 0.03 (0.01)<br />
Hewl<strong>et</strong>t-Packard 0.13 (0.01) 0.12 (0.01)<br />
Coca-Cola 0.12 (0.01) 0.09 (0.01)<br />
Minnesota Mining & MFG 0.07 (0.01) 0.06 (0.01)<br />
Philip Morris 0.04 (0.01) 0.04 (0.01)<br />
Procter & Gamble 0.12 (0.02) 0.09 (0.01)<br />
Pharmacia 0.06 (0.01) 0.04 (0.01)<br />
Schering-Plough 0.12 (0.01) 0.11 (0.01)<br />
Texaco 0.04 (0.01) 0.03 (0.01)<br />
Texas Instruments 0.17 (0.02) 0.12 (0.01)<br />
Walgreen 0.11 (0.01) 0.09 (0.01)<br />
This table presents the coefficients of lower and upper tail <strong>de</strong>pen<strong>de</strong>nce with the S&P<br />
500 in<strong>de</strong>x for a s<strong>et</strong> of 12 major stocks tra<strong>de</strong>d on the New York Stock Exchange from<br />
January 1991 to December 2000. The numbers in brack<strong>et</strong>s give the estim<strong>at</strong>ed standard<br />
<strong>de</strong>vi<strong>at</strong>ion of the empirical coefficients of tail <strong>de</strong>pen<strong>de</strong>nce<br />
2. Quantile r<strong>at</strong>io<br />
I = X k,N /Y k,N<br />
2.2<br />
2.0<br />
1.8<br />
1.6<br />
1.4<br />
1.2<br />
1.0<br />
0.8<br />
0 0.05 0.10 0.15 0.20 0.25<br />
k/N<br />
^<br />
Empirical estim<strong>at</strong>e l of the quantile r<strong>at</strong>io l in (7) versus the empirical<br />
^<br />
quantile k/N. We observe a very good stability of l for quantiles<br />
ranging b<strong>et</strong>ween 0.005 and 0.05<br />
such a case, the empirical quantile r<strong>at</strong>io l in (7) turns out to be stable<br />
enough for its accur<strong>at</strong>e non-param<strong>et</strong>ric estim<strong>at</strong>ion, as shown in figure 2.<br />
As an example, table A presents the results obtained both for the upper<br />
and lower coefficients of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween several major stocks<br />
and the mark<strong>et</strong> factor represented here by the S&P 500 in<strong>de</strong>x, over the<br />
past <strong>de</strong>ca<strong>de</strong>. The estim<strong>at</strong>ion has been performed un<strong>de</strong>r the assumption<br />
th<strong>at</strong> equ<strong>at</strong>ion (6) holds, r<strong>at</strong>her than un<strong>de</strong>r the ellipticality assumption yielding<br />
equ<strong>at</strong>ion (4). In the present context of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween stocks<br />
and an in<strong>de</strong>x (not b<strong>et</strong>ween two stocks), favouring the factor mo<strong>de</strong>l is very<br />
reasonable since, according to the financial theory, the mark<strong>et</strong>’s r<strong>et</strong>urn is<br />
well known to be the most important explan<strong>at</strong>ory factor for each individual<br />
ass<strong>et</strong> r<strong>et</strong>urn. 3 The technical aspects of the m<strong>et</strong>hod are given in the Appendix.<br />
The coefficient of tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween any two ass<strong>et</strong>s is easily<br />
<strong>de</strong>rived from (9). It is interesting to observe th<strong>at</strong> the coefficients of tail <strong>de</strong>pen<strong>de</strong>nce<br />
seem almost i<strong>de</strong>ntical in the lower and the upper tail. Non<strong>et</strong>heless,<br />
the coefficient of lower tail <strong>de</strong>pen<strong>de</strong>nce is always slightly larger than<br />
the upper one, showing th<strong>at</strong> large losses are more likely to occur tog<strong>et</strong>her<br />
than large gains.<br />
Two clusters of ass<strong>et</strong>s stand out: those with a tail <strong>de</strong>pen<strong>de</strong>nce of about<br />
3 In a situ<strong>at</strong>ion where the common factor cannot be easily i<strong>de</strong>ntified or estim<strong>at</strong>ed, the<br />
ellipticality assumption may provi<strong>de</strong> a useful altern<strong>at</strong>ive<br />
WWW.RISK.NET ● NOVEMBER 2002 RISK 131
12.2. Minimiser l’impact <strong>de</strong>s grands co-mouvements 385<br />
Cutting edge l<br />
Portfolio tail risk<br />
3. Portfolios versus mark<strong>et</strong><br />
Portfolio daily r<strong>et</strong>urn<br />
0.10<br />
0.08<br />
0.06<br />
0.04<br />
0.02<br />
0<br />
0.02<br />
0.04<br />
0.06<br />
0.08<br />
Portfolio 1<br />
Linear regression of<br />
the portfolio 1 on the<br />
S&P 500 in<strong>de</strong>x<br />
Portfolio 2<br />
Linear regression of<br />
the portfolio 2 on the<br />
S&P 500 in<strong>de</strong>x<br />
0.10<br />
0.08 0.06 0.04 0.02 0 0.02 0.04 0.06<br />
S&P 500 daily r<strong>et</strong>urn<br />
Daily r<strong>et</strong>urns of two equally weighted portfolios P 1 (ma<strong>de</strong> of four stocks<br />
with small λ ≤ 0.06) and P 2 (ma<strong>de</strong> of four stocks with large λ ≥ 0.12) as a<br />
function of the daily r<strong>et</strong>urns of the S&P 500 from Jan 1991–Dec 2000<br />
10% (or more) and those with a tail <strong>de</strong>pen<strong>de</strong>nce of about 5%. Since the<br />
estim<strong>at</strong>ion of the tail <strong>de</strong>pen<strong>de</strong>nce coefficients has been performed un<strong>de</strong>r<br />
the assumption th<strong>at</strong> equ<strong>at</strong>ion (6) holds, it is interesting to compare the<br />
values with the tail <strong>de</strong>pen<strong>de</strong>nce coefficient un<strong>de</strong>r the assumption of joint<br />
ellipticality leading to (4). To g<strong>et</strong> a reliable correl<strong>at</strong>ion coefficient ρ, we<br />
calcul<strong>at</strong>e the Kendall’s tau coefficient τ, use the rel<strong>at</strong>ion r = sin(πτ/2) and<br />
<strong>de</strong>rive λ from (4), assuming ν = 3 or 4. For ν = 3 (respectively ν = 4), all<br />
λ’s are in the range 0.25–0.30 (respectively 0.20–0.25). Thus, assuming<br />
joint ellipticality, the tail-<strong>de</strong>pen<strong>de</strong>nce coefficients b<strong>et</strong>ween stocks and the<br />
in<strong>de</strong>x are much more homogeneous than found with the factor mo<strong>de</strong>l.<br />
As we show in figure 3, it is clear th<strong>at</strong> the portfolio with stocks with small<br />
tail-<strong>de</strong>pen<strong>de</strong>nce coefficients with the in<strong>de</strong>x (measured with the factor<br />
mo<strong>de</strong>l) exhibits significantly less <strong>de</strong>pen<strong>de</strong>nce than the portfolio constructed<br />
with stocks with large λ’s (measured with the factor mo<strong>de</strong>l). This<br />
132 RISK NOVEMBER 2002 ● WWW.RISK.NET<br />
should not be observed if all λ’s are the same as predicted un<strong>de</strong>r the assumption<br />
of joint ellipticality.<br />
We now explore some consequences of the existence of stocks with<br />
drastically different tail-<strong>de</strong>pen<strong>de</strong>nce coefficients with the in<strong>de</strong>x. These<br />
stocks offer the interesting possibility of <strong>de</strong>vising a pru<strong>de</strong>ntial portfolio th<strong>at</strong><br />
can be significantly less sensitive to the large mark<strong>et</strong> moves. Figure 3 compares<br />
the daily r<strong>et</strong>urns of the S&P 500 in<strong>de</strong>x with those of two portfolios<br />
P 1 and P 2 . P 1 comprises the four stocks (Chevron, Philip Morris, Pharmacia<br />
and Texaco) with the smallest λ’s while P 2 comprises the four stocks<br />
(Bristol-Meyer Squibb, Hewl<strong>et</strong>t-Packard, Schering-Plough and Texas Instruments)<br />
with the largest λ’s. For each s<strong>et</strong> of stocks, we have constructed<br />
two portfolios, one in which each stock has the same weight 1/4 and<br />
the other with ass<strong>et</strong> weights chosen to minimise the variance of the resulting<br />
portfolio. We find th<strong>at</strong> the results are almost the same b<strong>et</strong>ween the<br />
equally weighted and minimum-variance portfolios. This makes sense since<br />
the tail-<strong>de</strong>pen<strong>de</strong>nce coefficient of a bivari<strong>at</strong>e random vector does not <strong>de</strong>pend<br />
on the variances of the components, which only account for price<br />
moves of mo<strong>de</strong>r<strong>at</strong>e amplitu<strong>de</strong>s.<br />
Figure 3 shows the results for the equally weighted portfolios gener<strong>at</strong>ed<br />
from the two groups of ass<strong>et</strong>s. Observe th<strong>at</strong> only one large drop occurs<br />
simultaneously for P 1 and for the S&P 500 in<strong>de</strong>x, in contrast with P 2 ,<br />
for which several large drops are associ<strong>at</strong>ed with the largest drops of the<br />
in<strong>de</strong>x and only a few occur <strong>de</strong>synchronised. The figure clearly shows an<br />
almost circular sc<strong>at</strong>ter plot for the large moves of P 1 and the in<strong>de</strong>x compared<br />
with a r<strong>at</strong>her narrow ellipse, whose long axis is approxim<strong>at</strong>ely along<br />
the first diagonal, for the the large r<strong>et</strong>urns of P 2 and the in<strong>de</strong>x, illustr<strong>at</strong>ing<br />
th<strong>at</strong> the small tail <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the in<strong>de</strong>x and the four stocks in<br />
P 1 autom<strong>at</strong>ically implies th<strong>at</strong> their mutual tail <strong>de</strong>pen<strong>de</strong>nce is also very small,<br />
according to (9). As a consequence, P 1 offers a b<strong>et</strong>ter diversific<strong>at</strong>ion with<br />
respect to large drops than P 2 . This effect, already quite significant for such<br />
small portfolios, should be overwhelming for large ones. The most interesting<br />
result stressed in figure 3 is th<strong>at</strong> optimising for minimum tail <strong>de</strong>pen<strong>de</strong>nce<br />
autom<strong>at</strong>ically diversifies away the large risks.<br />
These advantages of portfolio P 1 with small tail <strong>de</strong>pen<strong>de</strong>nce compared<br />
with portfolio P 2 with large tail <strong>de</strong>pen<strong>de</strong>nce with the S&P 500<br />
in<strong>de</strong>x come <strong>at</strong> almost no cost in terms of the daily Sharpe r<strong>at</strong>io, which<br />
is equal respectively to 0.058 and 0.061 for the equally weighted and<br />
minimum variance P 1 and to 0.069 and 0.071 for the equally weighted<br />
and minimum variance P 2 .<br />
The straight lines represent the linear regression of the two portfolios’
386 12. Gestion <strong>de</strong>s risques grands <strong>et</strong> extrêmes<br />
r<strong>et</strong>urns on the in<strong>de</strong>x r<strong>et</strong>urns, which shows th<strong>at</strong> there is significantly less<br />
linear correl<strong>at</strong>ion b<strong>et</strong>ween P 1 and the in<strong>de</strong>x (correl<strong>at</strong>ion coefficient of 0.52<br />
for both the equally weighted and the minimum variance P 1 ) compared<br />
with P 2 and the in<strong>de</strong>x (correl<strong>at</strong>ion coefficient of 0.73 for the equally weighted<br />
P 2 and of 0.70 for the minimum variance P 2 ). Theor<strong>et</strong>ically, it is possible<br />
to construct two random variables with small correl<strong>at</strong>ion coefficient<br />
and large λ and vice versa. Recall th<strong>at</strong> the correl<strong>at</strong>ion coefficient and the<br />
tail-<strong>de</strong>pen<strong>de</strong>nce coefficient are two opposite end-members of <strong>de</strong>pen<strong>de</strong>nce<br />
measures. The correl<strong>at</strong>ion coefficient quantifies the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
rel<strong>at</strong>ively small moves while the tail-<strong>de</strong>pen<strong>de</strong>nce coefficient measures the<br />
<strong>de</strong>pen<strong>de</strong>nce during extreme events. The finding th<strong>at</strong> P 1 comes with both<br />
the smallest correl<strong>at</strong>ion and the smallest tail-<strong>de</strong>pen<strong>de</strong>nce coefficients suggests<br />
th<strong>at</strong> they are not in<strong>de</strong>pen<strong>de</strong>nt properties of ass<strong>et</strong>s. This intuition is<br />
in fact explained and encompassed by the factor mo<strong>de</strong>l since the larger<br />
β is, the larger the correl<strong>at</strong>ion coefficient and the larger the tail <strong>de</strong>pen<strong>de</strong>nce.<br />
Diversifying away extreme shocks may provi<strong>de</strong> a useful diversific<strong>at</strong>ion<br />
tool for less extreme <strong>de</strong>pen<strong>de</strong>nces, thus improving the potential<br />
usefulness of a str<strong>at</strong>egy of portfolio management based on tail <strong>de</strong>pen<strong>de</strong>nce<br />
proposed here.<br />
As a final remark, the almost i<strong>de</strong>ntical values of the coefficients of tail<br />
<strong>de</strong>pen<strong>de</strong>nce for neg<strong>at</strong>ive and positive tails show th<strong>at</strong> ass<strong>et</strong>s th<strong>at</strong> are the<br />
most likely to suffer from the large losses of the mark<strong>et</strong> factor are also<br />
those th<strong>at</strong> are the most likely to take advantage of its large gains. This<br />
has the following consequence: minimising the large concomitant losses<br />
b<strong>et</strong>ween the stocks and the mark<strong>et</strong> means renouncing the potential<br />
concomitant large gains. This point is well exemplified by our two portfolios<br />
(see figure 3): P 2 obviously un<strong>de</strong>rwent severe neg<strong>at</strong>ive co-movements<br />
but it also enjoyed large gains with the large positive movements<br />
of the in<strong>de</strong>x. In contrast, P 1 is almost compl<strong>et</strong>ely <strong>de</strong>coupled from the<br />
large neg<strong>at</strong>ive movements of the mark<strong>et</strong> but is also insensitive to the large<br />
positive movements of the in<strong>de</strong>x. Thus, a good dynamic str<strong>at</strong>egy seems<br />
to be: invest in P 1 during bearish or trend-less mark<strong>et</strong> phases and prefer<br />
P 2 in a bullish mark<strong>et</strong>. ■<br />
Yannick Malevergne is a PhD stu<strong>de</strong>nt <strong>at</strong> the University of Nice-Sophia<br />
Antipolis and <strong>at</strong> the <strong>ISFA</strong> Actuarial School – University of Lyon. Didier<br />
Bingham N, C Goldie and J Teugel, 1987<br />
Regular vari<strong>at</strong>ion<br />
Cambridge University Press, Cambridge<br />
Boyer B, M Gibson and M Laur<strong>et</strong>an, 1997<br />
Pitfalls in tests for changes in correl<strong>at</strong>ions<br />
Intern<strong>at</strong>ional Finance Discussion Paper 597, Board of the Governors of the Fe<strong>de</strong>ral<br />
Reserve System<br />
Coles S, J Heffernan and J Tawn, 1999<br />
Depen<strong>de</strong>nce measures for extreme value analysis<br />
Extremes 2, pages 339–365<br />
Embrechts P, A McNeil and D Straumann, 1999<br />
Correl<strong>at</strong>ion: pitfalls and altern<strong>at</strong>ives<br />
Risk May, pages 69–71<br />
Embrechts P, A McNeil and D Straumann, 2002<br />
Correl<strong>at</strong>ion and <strong>de</strong>pen<strong>de</strong>nce in risk management: properties and pitfalls<br />
In Risk Management: Value <strong>at</strong> Risk and Beyond, edited by M Dempster, pages<br />
176–223, Cambridge University Press, Cambridge<br />
Hull J and A White, 1987<br />
The option pricing on ass<strong>et</strong>s with stochastic vol<strong>at</strong>ilities<br />
Journal of Finance 42, pages 281–300<br />
Hult H and F Lindskog, 2002<br />
Multivari<strong>at</strong>e extremes, aggreg<strong>at</strong>ion and <strong>de</strong>pen<strong>de</strong>nce in elliptical distributions<br />
Forthcoming in Advances in Applied Probability 34(3)<br />
Ledford A and J Tawn, 1998<br />
Concomitant tail behaviour for extremes<br />
Advances in Applied Probability 30, pages 197–215<br />
REFERENCES<br />
Appendix: empirical estim<strong>at</strong>ion of the<br />
coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
We show how to estim<strong>at</strong>e the coefficient of tail <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween an ass<strong>et</strong> X and the mark<strong>et</strong> factor Y rel<strong>at</strong>ed by the rel<strong>at</strong>ion<br />
(6) where ε is an idiosyncr<strong>at</strong>ic noise uncorrel<strong>at</strong>ed with X.<br />
Given a sample of N realis<strong>at</strong>ions {X 1 , X 2 , ... , X N } and<br />
{Y 1 , Y 2 , ... , Y N } of X and Y, we first estim<strong>at</strong>e the coefficient β using<br />
the ordinary least square estim<strong>at</strong>or. L<strong>et</strong> β ^ <strong>de</strong>note its estim<strong>at</strong>e. Then,<br />
using Hill’s estim<strong>at</strong>or, we obtain the tail in<strong>de</strong>x ν^ of the factor Y:<br />
⎡ k 1<br />
ˆ = ⎢ ∑ logY −logY<br />
⎣⎢<br />
k j=<br />
1<br />
νk j, N k, N<br />
where Y 1, N ≥ Y 2, N ≥ ... ≥ Y N, N are the or<strong>de</strong>r st<strong>at</strong>istics of the N realis<strong>at</strong>ions<br />
of Y. The constant l is non-param<strong>et</strong>rically estim<strong>at</strong>ed with<br />
the formula:<br />
FXu X<br />
l = lim ≃<br />
u→1<br />
−1<br />
F u Y<br />
for k = o(N), which means th<strong>at</strong> k must remain very small with<br />
respect to N but large enough to ensure an accur<strong>at</strong>e d<strong>et</strong>ermin<strong>at</strong>ion<br />
of l. Figure 2 presents l ^ as a function of k/N.<br />
Finally, using equ<strong>at</strong>ion (7), the estim<strong>at</strong>ed λ ^ is:<br />
ˆ<br />
max , ˆ<br />
+ 1<br />
λ =<br />
νˆ<br />
{ 1<br />
βˆ<br />
} l<br />
Sorn<strong>et</strong>te is a CNRS research director <strong>at</strong> the University of Nice-Sophia<br />
Antipolis and professor of geophysics <strong>at</strong> the University of California <strong>at</strong><br />
Los Angeles. They acknowledge helpful discussions with Jean-Paul<br />
Laurent. This work was partially supported by the James S McDonnell<br />
Found<strong>at</strong>ion twenty-first century scientist award/studying complex<br />
system. e-mail: Yannick.Malevergne@unice.fr and sorn<strong>et</strong>te@unice.fr<br />
Malevergne Y and D Sorn<strong>et</strong>te, 2002a<br />
Investig<strong>at</strong>ing extreme <strong>de</strong>pen<strong>de</strong>nces: concepts and tools<br />
Working paper, available <strong>at</strong> http://papers.ssrn.com/sol3/papers.cfm?abstract_id =<br />
303465<br />
Malevergne Y and D Sorn<strong>et</strong>te, 2002b<br />
Tail <strong>de</strong>pen<strong>de</strong>nce of factor mo<strong>de</strong>ls<br />
Working paper, available <strong>at</strong> http://papers.ssrn.com/sol3/papers.cfm?abstract_id =<br />
301266<br />
Man<strong>de</strong>lbrot B, 1997<br />
Fractals and scaling in finance: discontinuity, concentr<strong>at</strong>ion<br />
Springer-Verlag, New York<br />
Nelsen R, 1998<br />
An introduction to copulas<br />
Lectures Notes in St<strong>at</strong>istics 139, Springer Verlag, New York<br />
Ross S, 1976<br />
The arbitrage theory of capital ass<strong>et</strong> pricing<br />
Journal of Economic Theory 17, pages 254–286<br />
Rubinstein M, 1973<br />
The fundamental theorem of param<strong>et</strong>er-preference security valu<strong>at</strong>ion<br />
Journal of Financial and Quantit<strong>at</strong>ive Analysis 8, pages 61–69<br />
Sharpe W, 1964<br />
Capital ass<strong>et</strong> pricing: a theory of mark<strong>et</strong> equilibrium un<strong>de</strong>r conditions of risk<br />
Journal of Finance 19, pages 425–442<br />
Taylor S, 1994<br />
Mo<strong>de</strong>ling stochastic vol<strong>at</strong>ility<br />
M<strong>at</strong>hem<strong>at</strong>ical Finance 4, pages 183–204<br />
−1<br />
Y<br />
( )<br />
( )<br />
kN ,<br />
kN ,<br />
⎤<br />
⎥<br />
⎦⎥<br />
WWW.RISK.NET ● NOVEMBER 2002 RISK 133
Chapitre 13<br />
Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong><br />
capital économique : l’exemple <strong>de</strong> la<br />
Value-<strong>at</strong>-Risk <strong>et</strong> <strong>de</strong> l’Expected-Shortfall<br />
A l’ai<strong>de</strong> d’une famille <strong>de</strong> distributions <strong>de</strong> Weibull modifiées, englobant à la fois <strong>de</strong>s distributions super<br />
<strong>et</strong> sous-exponentielles (dont l’interêt a été souligné au chapitre 3), nous paramétrons les distributions<br />
<strong>de</strong> ren<strong>de</strong>ments <strong>de</strong>s actifs financiers. L’hypothèse <strong>de</strong> copule gaussienne est utilisée pour modéliser la<br />
dépendance entre les actifs. Cela nous perm<strong>et</strong> d’obtenir les expressions analytiques <strong>de</strong>s queues <strong>de</strong> la<br />
distribution P (S) du ren<strong>de</strong>ment S d’un <strong>portefeuille</strong> composé <strong>de</strong> tels actifs. Nous montrons que les queues<br />
<strong>de</strong> la distribution P (S) <strong>de</strong>meurent asymptotiquement <strong>de</strong>s distributions <strong>de</strong> Weibull modifiées avec un<br />
facteur d’échelle χ fonction du poids <strong>de</strong>s actifs dans le <strong>portefeuille</strong> <strong>et</strong> dont l’expression différe selon<br />
le comportement super ou sous-exponentiel <strong>de</strong>s actifs. Nos traitons alors en détails le problème <strong>de</strong> la<br />
minimis<strong>at</strong>ion du risque pour <strong>de</strong> tels <strong>portefeuille</strong>s, les mesures <strong>de</strong> risque considérées étant la Value-<strong>at</strong>-<br />
Risk <strong>et</strong> l’Expected-Shortfall, que nous montrons être asymptotiquement équivalentes dans le cadre <strong>de</strong> la<br />
représentaion adoptée.<br />
387
388 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique
VaR-Efficient Portfolios for a Class of Super- and<br />
Sub-Exponentially Decaying Ass<strong>et</strong>s R<strong>et</strong>urn Distributions<br />
Y. Malevergne 1,2 and D. Sorn<strong>et</strong>te 1,3<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée CNRS UMR 6622<br />
Université <strong>de</strong> Nice-Sophia Antipolis, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095, USA<br />
email: Yannick.Malevergne@unice.fr and sorn<strong>et</strong>te@unice.fr<br />
fax: (33) 4 92 07 67 54<br />
Abstract<br />
Using a family of modified Weibull distributions, encompassing both sub-exponentials and super-exponentials,<br />
to param<strong>et</strong>erize the marginal distributions of ass<strong>et</strong> r<strong>et</strong>urns and their multivari<strong>at</strong>e generaliz<strong>at</strong>ions<br />
with Gaussian copulas, we offer exact formulas for the tails of the distribution P (S) of r<strong>et</strong>urns S of a<br />
portfolio of arbitrary composition of these ass<strong>et</strong>s. We find th<strong>at</strong> the tail of P (S) is also asymptotically<br />
a modified Weibull distribution with a characteristic scale χ function of the ass<strong>et</strong> weights with different<br />
functional forms <strong>de</strong>pending on the super- or sub-exponential behavior of the marginals and on the<br />
strength of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s. We then tre<strong>at</strong> in d<strong>et</strong>ails the problem of risk minimiz<strong>at</strong>ion<br />
using the Value-<strong>at</strong>-Risk and Expected-Shortfall which are shown to be (asymptotically) equivalent in this<br />
framework.<br />
Introduction<br />
In recent years, the Value-<strong>at</strong>-Risk has become one of the most popular risk assessment tool (Duffie and<br />
Pan 1997, Jorion 1997). The inf<strong>at</strong>u<strong>at</strong>ion for this particular risk measure probably comes from a vari<strong>et</strong>y of<br />
factors, the most prominent ones being its conceptual simplicity and relevance in addressing the ubiquitous<br />
large risks often ina<strong>de</strong>qu<strong>at</strong>ely accounted for by the standard vol<strong>at</strong>ility, and from its prominent role in the<br />
recommend<strong>at</strong>ions of the intern<strong>at</strong>ional banking authorities (Basle Commitee on Banking Supervision 1996,<br />
2001). Moreover, down-si<strong>de</strong> risk measures such as the Value-<strong>at</strong>-risk seem more in accordance with observed<br />
behavior of economic agents. For instance, according to prospect theory (Kahneman and Tversky 1979), the<br />
perception of downward mark<strong>et</strong> movements is not the same as upward movements. This may be reflected<br />
in the so-called leverage effect, first discussed by (Black 1976), who observed th<strong>at</strong> the vol<strong>at</strong>ility of a stock<br />
tends to increase when its price drops (see (Fouque <strong>et</strong> al. 2000, Campbell, Lo and McKinley 1997, Bekaert<br />
and Wu 2000, Bouchaud <strong>et</strong> al. 2001) for reviews and recent works). Thus, it should be more n<strong>at</strong>ural to<br />
consi<strong>de</strong>r down-si<strong>de</strong> risk measures like the VaR than the variance traditionally used in portfolio management<br />
(Markowitz 1959) which does not differenti<strong>at</strong>e b<strong>et</strong>ween positive and neg<strong>at</strong>ive change in future wealth.<br />
1<br />
389
390 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
However, the choice of the Value-<strong>at</strong>-Risk has recently been criticized (Szergö 1999, Danielsson <strong>et</strong> al. 2001)<br />
due to its lack of coherence in the sense of Artzner <strong>et</strong> al. (1999), among other reasons. This <strong>de</strong>ficiency<br />
leads to several theor<strong>et</strong>ical and practical problems. In<strong>de</strong>ed, other than the class of elliptical distributions,<br />
the VaR is not sub-additive (Embrechts <strong>et</strong> al. 2002a), and may lead to inefficient risk diversific<strong>at</strong>ion policies<br />
and to severe problems in the practical implement<strong>at</strong>ion of portfolio optimiz<strong>at</strong>ion algorithms (see (Chabaane<br />
<strong>et</strong> al. 2002) for a discussion). Altern<strong>at</strong>ive have been proposed in terms of Conditional-VaR or Expected-<br />
Shortfall (Artzner <strong>et</strong> al. 1999, Acerbi and Tasche 2002, for instance), which enjoy the property of subadditivity.<br />
This ensures th<strong>at</strong> they yield coherent portfolio alloc<strong>at</strong>ions which can be obtained by the simple<br />
linear optimiz<strong>at</strong>ion algorithm proposed by Rockafellar and Uryasev (2000).<br />
From a practical standpoint, the estim<strong>at</strong>ion of the VaR of a portfolio is a strenuous task, requiring large comput<strong>at</strong>ional<br />
time leading som<strong>et</strong>imes to disappointing results lacking accuracy and stability. As a consequence,<br />
many approxim<strong>at</strong>ion m<strong>et</strong>hods have been proposed (Tasche and Tilib<strong>et</strong>ti 2001, Embrechts <strong>et</strong> al. 2002b, for<br />
instance). Empirical mo<strong>de</strong>ls constitute another wi<strong>de</strong>ly used approach, since they provi<strong>de</strong> a good tra<strong>de</strong>-off<br />
b<strong>et</strong>ween speed and accuracy.<br />
From a general point of view, the param<strong>et</strong>ric d<strong>et</strong>ermin<strong>at</strong>ion of the risks and r<strong>et</strong>urns associ<strong>at</strong>ed with a given<br />
portfolio constituted of N ass<strong>et</strong>s is compl<strong>et</strong>ely embed<strong>de</strong>d in the knowledge of their multivari<strong>at</strong>e distribution<br />
of r<strong>et</strong>urns. In<strong>de</strong>ed, the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween random variables is compl<strong>et</strong>ely <strong>de</strong>scribed by their joint<br />
distribution. This remark entails the two major problems of portfolio theory: 1) the d<strong>et</strong>ermin<strong>at</strong>ion of the<br />
multivari<strong>at</strong>e distribution function of ass<strong>et</strong> r<strong>et</strong>urns; 2) the <strong>de</strong>riv<strong>at</strong>ion from it of a useful measure of portfolio<br />
risks, in the goal of analyzing and optimizing portfolios. These objective can be easily reached if one can<br />
<strong>de</strong>rive an analytical expression of the portfolio r<strong>et</strong>urns distribution from the multivari<strong>at</strong>e distribution of ass<strong>et</strong><br />
r<strong>et</strong>urns.<br />
In the standard Gaussian framework, the multivari<strong>at</strong>e distribution takes the form of an exponential of minus<br />
a quadr<strong>at</strong>ic form X ′ Ω −1 X, where X is the uni-column of ass<strong>et</strong> r<strong>et</strong>urns and Ω is their covariance m<strong>at</strong>rix.<br />
The beauty and simplicity of the Gaussian case is th<strong>at</strong> the essentially impossible task of d<strong>et</strong>ermining a<br />
large multidimensional function is collapsed onto the very much simpler one of calcul<strong>at</strong>ing the N(N +<br />
1)/2 elements of the symm<strong>et</strong>ric covariance m<strong>at</strong>rix. And, by the st<strong>at</strong>ibility of the Gaussian distribution,<br />
the risk is then uniquely and compl<strong>et</strong>ely embodied by the variance of the portfolio r<strong>et</strong>urn, which is easily<br />
d<strong>et</strong>ermined from the covariance m<strong>at</strong>rix. This is the basis of Markowitz (1959)’s portfolio theory and of the<br />
CAPM (Sharpe 1964, Lintner 1965, Mossin 1966). The same phenomenon occurs in the stable Par<strong>et</strong>ian<br />
portfolio analysis <strong>de</strong>rived by (Fama 1965) and generalized to separ<strong>at</strong>e positive and neg<strong>at</strong>ive power law tails<br />
(Bouchaud <strong>et</strong> al. 1998). The stability of the distribution of r<strong>et</strong>urns is essentiel to bypass the difficult problem<br />
of d<strong>et</strong>ermining the <strong>de</strong>cision rules (utility function) of the economic agents since all the risk measures are<br />
equivalent to a single param<strong>et</strong>er (the variance in the case of a Gaussian universe).<br />
However, it is well-known th<strong>at</strong> the empirical distributions of r<strong>et</strong>urns are neither Gaussian nor Lévy Stable<br />
(Lux 1996, Gopikrishnan <strong>et</strong> al. 1998, Gouriéroux and Jasiak 1998) and the <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween ass<strong>et</strong>s are<br />
only imperfectly accounted for by the covariance m<strong>at</strong>rix (Litterman and Winkelmann 1998). It is thus <strong>de</strong>sirable<br />
to find altern<strong>at</strong>ive param<strong>et</strong>eriz<strong>at</strong>ions of multivari<strong>at</strong>e distributions of r<strong>et</strong>urns which provi<strong>de</strong> reasonably<br />
good approxim<strong>at</strong>ions of the ass<strong>et</strong> r<strong>et</strong>urns distribution and which enjoy asymptotic stability properties in the<br />
tails so as to be relevant for the VaR.<br />
To this aim, section 1 presents a specific param<strong>et</strong>eriz<strong>at</strong>ion of the marginal distributions in terms of so-called<br />
modified Weibull distributions introduced by Sorn<strong>et</strong>te <strong>et</strong> al. (2000b), which are essentially exponential of<br />
minus a power law. This family of distributions contains both sub-exponential and super-exponentials,<br />
including the Gaussian law as a special case. It is shown th<strong>at</strong> this param<strong>et</strong>eriz<strong>at</strong>ion is relevant for mo<strong>de</strong>ling<br />
the distribution of ass<strong>et</strong> r<strong>et</strong>urns in both an unconditional and a conditional framework. The <strong>de</strong>pen<strong>de</strong>nce<br />
structure b<strong>et</strong>ween the ass<strong>et</strong> is <strong>de</strong>scribed by a Gaussian copula which allows us to <strong>de</strong>scribe several <strong>de</strong>grees of<br />
2
<strong>de</strong>pen<strong>de</strong>nce: from in<strong>de</strong>pen<strong>de</strong>nce to comonotonicity. The relevance of the Gaussian copula has been put in<br />
light by several recent studies (Sorn<strong>et</strong>te <strong>et</strong> al. 2000a, Sorn<strong>et</strong>te <strong>et</strong> al. 2000b, Malevergne and Sorn<strong>et</strong>te 2001,<br />
Malevergne and Sorn<strong>et</strong>te 2002c).<br />
In section 2, we use the multivari<strong>at</strong>e construction based on (i) the modified Weibull marginal distributions<br />
and (ii) the Gaussian copula to <strong>de</strong>rive the asymptotic analytical form of the tail of the distribution of r<strong>et</strong>urns<br />
of a portfolio composed of an arbitrary combin<strong>at</strong>ion of these ass<strong>et</strong>s. In the case where individual ass<strong>et</strong><br />
r<strong>et</strong>urns have modified-Weibull distributions, we show th<strong>at</strong> the tail of the distribution of portfolio r<strong>et</strong>urns S<br />
is asymptotically of the same form but with a characteristic scale χ function of the ass<strong>et</strong> weights taking<br />
different functional forms <strong>de</strong>pending on the super- or sub-exponential behavior of the marginals and on the<br />
strength of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s. Thus, this particular class of modified-Weibull distributions<br />
enjoys (asymptotically) the same stability properties as the Gaussian or Lévy distributions. The <strong>de</strong>pen<strong>de</strong>nce<br />
properties are shown to be embodied in the N(N + 1)/2 elements of a non-linear covariance m<strong>at</strong>rix and the<br />
individual risk of each ass<strong>et</strong>s are quantified by the sub- or super-exponential behavior of the marginals.<br />
Section 3 then uses this non-Gaussian nonlinear <strong>de</strong>pen<strong>de</strong>nce framework to estim<strong>at</strong>e the Value-<strong>at</strong>-Risk (VaR)<br />
and the Expected-Shortfall. As in the Gaussian framework, the VaR and the Expected-Shortfall are (asymptotically)<br />
controlled only by the non-linear covariance m<strong>at</strong>rix, leading to their equivalence. More generally,<br />
any risk measure based on the (sufficiently far) tail of the distribution of the portfolio r<strong>et</strong>urns are equivalent<br />
since they can be expressed as a function of the non-linear covariance m<strong>at</strong>rix and the weights of the ass<strong>et</strong>s<br />
only.<br />
Section 4 uses this s<strong>et</strong> of results to offer an approach to portfolio optimiz<strong>at</strong>ion based on the asymptotic<br />
form of the tail of the distribution of portfolio r<strong>et</strong>urns. When possible, we give the analytical formulas of<br />
the explicit composition of the optimal portfolio or suggest the use of reliable algorithms when numerical<br />
calcul<strong>at</strong>ion is nee<strong>de</strong>d.<br />
Section 5 conclu<strong>de</strong>s.<br />
Before proceeding with the present<strong>at</strong>ion of our results, we s<strong>et</strong> the not<strong>at</strong>ions to <strong>de</strong>rive the basic problem<br />
addressed in this paper, namely to study the distribution of the sum of weighted random variables with given<br />
marginal distributions and <strong>de</strong>pen<strong>de</strong>nce. Consi<strong>de</strong>r a portfolio with ni shares of ass<strong>et</strong> i of price pi(0) <strong>at</strong> time<br />
t = 0 whose initial wealth is<br />
N<br />
W (0) = nipi(0) . (1)<br />
A time τ l<strong>at</strong>er, the wealth has become W (τ) = N<br />
i=1 nipi(τ) and the wealth vari<strong>at</strong>ion is<br />
where<br />
δτ W ≡ W (τ) − W (0) =<br />
N<br />
i=1<br />
wi =<br />
i=1<br />
nipi(0) pi(τ) − pi(0)<br />
pi(0)<br />
nipi(0)<br />
N<br />
j=1 njpj(0)<br />
= W (0)<br />
391<br />
N<br />
wixi(t, τ), (2)<br />
is the fraction in capital invested in the ith ass<strong>et</strong> <strong>at</strong> time 0 and the r<strong>et</strong>urn xi(t, τ) b<strong>et</strong>ween time t − τ and t of<br />
ass<strong>et</strong> i is <strong>de</strong>fined as:<br />
xi(t, τ) = pi(t) − pi(t − τ)<br />
.<br />
pi(t − τ)<br />
(4)<br />
Using the <strong>de</strong>finition (4), this justifies us to write the r<strong>et</strong>urn Sτ of the portfolio over a time interval τ as the<br />
3<br />
i=1<br />
(3)
392 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
weighted sum of the r<strong>et</strong>urns ri(τ) of the ass<strong>et</strong>s i = 1, ..., N over the time interval τ<br />
Sτ = δτ W<br />
W (0) =<br />
N<br />
wi xi(τ) . (5)<br />
In the sequel, we shall thus consi<strong>de</strong>r the ass<strong>et</strong> r<strong>et</strong>urns Xi as the fundamental variables and study their<br />
aggreg<strong>at</strong>ion properties, namely how the distribution of portfolio r<strong>et</strong>urn equal to their weighted sum <strong>de</strong>rives<br />
for their multivariable distribution. We shall consi<strong>de</strong>r a single time scale τ which can be chosen arbitrarily,<br />
say equal to one day. We shall thus drop the <strong>de</strong>pen<strong>de</strong>nce on τ, un<strong>de</strong>rstanding implicitly th<strong>at</strong> all our results<br />
hold for r<strong>et</strong>urns estim<strong>at</strong>ed over time step τ.<br />
1 Definitions and important concepts<br />
1.1 The modified Weibull distributions<br />
We will consi<strong>de</strong>r a class of distributions with f<strong>at</strong> tails but <strong>de</strong>caying faster than any power law. Such possible<br />
behavior for ass<strong>et</strong>s r<strong>et</strong>urns distributions have been suggested to be relevant by several empirical works<br />
(Mantegna and Stanley 1995, Gouriéroux and Jasiak 1998, Malevergne <strong>et</strong> al. 2002) and has also been asserted<br />
to provi<strong>de</strong> a convenient and flexible param<strong>et</strong>eriz<strong>at</strong>ion of many phenomena found in n<strong>at</strong>ure and in<br />
the social sciences (Lahèrre and Sorn<strong>et</strong>te 1998). In all the following, we will use the param<strong>et</strong>eriz<strong>at</strong>ion<br />
introduced by Sorn<strong>et</strong>te <strong>et</strong> al. (2000b) and <strong>de</strong>fine the modified-Weibull distributions:<br />
DEFINITION 1 (MODIFIED WEIBULL DISTRIBUTION)<br />
A random variable X will be said to follow a modified Weibull distribution with exponent c and scale<br />
param<strong>et</strong>er χ, <strong>de</strong>noted in the sequel X ∼ W(c, χ), if and only if the random variable<br />
follows a Normal distribution.<br />
i=1<br />
Y = sgn(X) √ 2<br />
c<br />
|X| 2<br />
χ<br />
These so-called modified-Weibull distributions can be seen to be general forms of the extreme tails of product<br />
of random variables (Frisch and Sorn<strong>et</strong>te 1997), and using the theorem of change of variable, we can<br />
assert th<strong>at</strong> the <strong>de</strong>nsity of such distributions is<br />
where c and χ are the two key param<strong>et</strong>ers.<br />
p(x) = 1<br />
2 √ c<br />
π χ c |x|<br />
2<br />
c<br />
2 −1 e −|x|<br />
χc<br />
, (7)<br />
These expressions are close to the Weibull distribution, with the addition of a power law prefactor to the<br />
exponential such th<strong>at</strong> the Gaussian law is r<strong>et</strong>rieved for c = 2. Following Sorn<strong>et</strong>te <strong>et</strong> al. (2000b), Sorn<strong>et</strong>te<br />
<strong>et</strong> al. (2000a) and An<strong>de</strong>rsen and Sorn<strong>et</strong>te (2001), we call (7) the modified Weibull distribution. For c < 1,<br />
the pdf is a str<strong>et</strong>ched exponential, which belongs to the class of sub-exponential. The exponent c d<strong>et</strong>ermines<br />
the shape of the distribution, f<strong>at</strong>ter than an exponential if c < 1. The param<strong>et</strong>er χ controls the scale or<br />
characteristic width of the distribution. It plays a role analogous to the standard <strong>de</strong>vi<strong>at</strong>ion of the Gaussian<br />
law.<br />
The interest of these family of distributions for financial purposes have also been recently un<strong>de</strong>rlined by<br />
Brummelhuis and Guégan (2000) and Brummelhuis <strong>et</strong> al. (2002). In<strong>de</strong>ed these authors have shown th<strong>at</strong><br />
4<br />
(6)
given a series of r<strong>et</strong>urn {rt}t following a GARCH(1,1) process, the large <strong>de</strong>vi<strong>at</strong>ions of the r<strong>et</strong>urns rt+k<br />
and of the aggreg<strong>at</strong>ed r<strong>et</strong>urns rt + · · · + rt+k conditional on the r<strong>et</strong>urn <strong>at</strong> time t are distributed according<br />
to a modified-Weibull distribution, where the exponent c is rel<strong>at</strong>ed to the number of step forward k by the<br />
formula c = 2/k .<br />
A more general param<strong>et</strong>eriz<strong>at</strong>ion taking into account a possible asymm<strong>et</strong>ry b<strong>et</strong>ween neg<strong>at</strong>ive and positive<br />
values (thus leading to possible non-zero mean) is<br />
p(x) =<br />
p(x) =<br />
1<br />
2 √ π<br />
1<br />
2 √ π<br />
c+<br />
c +<br />
2<br />
+<br />
χ<br />
c−<br />
c− 2<br />
−<br />
χ<br />
|x| c +<br />
2 −1 e −|x| +<br />
χ +c<br />
|x| c− 2 −1 e −|x| −<br />
χ−c 393<br />
if x ≥ 0 (8)<br />
if x < 0 . (9)<br />
In wh<strong>at</strong> follows, we will assume th<strong>at</strong> the marginal probability distributions of r<strong>et</strong>urns follow modified<br />
Weibull distributions. Figure 1 shows the (neg<strong>at</strong>ive) “Gaussianized” r<strong>et</strong>urns Y <strong>de</strong>fined in (6) of the Standard<br />
and Poor’s 500 in<strong>de</strong>x versus the raw r<strong>et</strong>urns X over the time interval from January 03, 1995 to December<br />
29, 2000. With such a represent<strong>at</strong>ion, the modified-Weibull distributions are qualified by a power law of<br />
exponent c/2, by <strong>de</strong>finition 1. The double logarithmic scales of figure 1 clearly shows a straight line over an<br />
exten<strong>de</strong>d range of d<strong>at</strong>a, qualifying a power law rel<strong>at</strong>ionship. An accur<strong>at</strong>e d<strong>et</strong>ermin<strong>at</strong>ion of the param<strong>et</strong>ers<br />
(χ, c) can be performed by maximum likelihood estim<strong>at</strong>ion (Sorn<strong>et</strong>te 2000, pp 160-162). However, note<br />
th<strong>at</strong>, in the tail, the six most extreme points significantly <strong>de</strong>vi<strong>at</strong>e from the modified-Weibull <strong>de</strong>scription.<br />
Such an anomalous behavior of the most extreme r<strong>et</strong>urns can be probably be associ<strong>at</strong>ed with the notion<br />
of “outliers” introduced by Johansen and Sorn<strong>et</strong>te (1998, 2002) and associ<strong>at</strong>ed with behavioral and crowd<br />
phenomena during turbulent mark<strong>et</strong> phases.<br />
The modified Weibull distributions <strong>de</strong>fined here are of interest for financial purposes and specifically for<br />
portfolio and risk management, since they offer a flexible param<strong>et</strong>ric represent<strong>at</strong>ion of ass<strong>et</strong> r<strong>et</strong>urns distribution<br />
either in a conditional or an unconditional framework, <strong>de</strong>pending on the standpoint prefered by<br />
manager. The rest of the paper uses this family of distributions.<br />
1.2 Tail equivalence for distribution functions<br />
An interesting fe<strong>at</strong>ure of the modified Weibull distributions, as we will see in the next section, is to enjoy the<br />
property of asymptotic stability. Asymptotic stability means th<strong>at</strong>, in the regime of large <strong>de</strong>vi<strong>at</strong>ions, a sum<br />
of in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically distributed modified Weibull variables follows the same modified Weibull<br />
distribution, up to a rescaling.<br />
DEFINITION 2 (TAIL EQUIVALENCE)<br />
L<strong>et</strong> X and Y be two random variables with distribution function F and G respectively.<br />
X and Y are said to be equivalent in the upper tail if and only if there exists λ+ ∈ (0, ∞) such th<strong>at</strong><br />
1 − F (x)<br />
lim<br />
x→+∞ 1 − G(x) = λ+. (10)<br />
Similarly, X and Y are said equivalent in the lower tail if and only if there exists λ− ∈ (0, ∞) such th<strong>at</strong><br />
<br />
F (x)<br />
lim<br />
x→−∞ G(x) = λ−. (11)<br />
5
394 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
Applying l’Hospital’s rule, this gives immedi<strong>at</strong>ely the following corollary:<br />
COROLLARY 1<br />
L<strong>et</strong> X and Y be two random variables with <strong>de</strong>nsities functions f and g respectively. X and Y are equivalent<br />
in the upper (lower) tail if and only if<br />
<br />
1.3 The Gaussian copula<br />
f(x)<br />
lim<br />
x→±∞ g(x) = λ±, λ± ∈ (0, ∞). (12)<br />
We recall only the basic properties about copulas and refer the interested rea<strong>de</strong>r to (Nelsen 1998), for instance,<br />
for more inform<strong>at</strong>ion. L<strong>et</strong> us first give the <strong>de</strong>finition of a copula of n random variables.<br />
DEFINITION 3 (COPULA)<br />
A function C : [0, 1] n −→ [0, 1] is a n-copula if it enjoys the following properties :<br />
• ∀u ∈ [0, 1], C(1, · · · , 1, u, 1 · · · , 1) = u ,<br />
• ∀ui ∈ [0, 1], C(u1, · · · , un) = 0 if <strong>at</strong> least one of the ui equals zero ,<br />
• C is groun<strong>de</strong>d and n-increasing, i.e., the C-volume of every boxes whose vertices lie in [0, 1] n is<br />
positive. <br />
The fact th<strong>at</strong> such copulas can be very useful for representing multivari<strong>at</strong>e distributions with arbitrary<br />
marginals is seen from the following result.<br />
THEOREM 1 (SKLAR’S THEOREM)<br />
Given an n-dimensional distribution function F with continuous marginal distributions F1, · · · , Fn, there<br />
exists a unique n-copula C : [0, 1] n −→ [0, 1] such th<strong>at</strong> :<br />
<br />
F (x1, · · · , xn) = C(F1(x1), · · · , Fn(xn)) . (13)<br />
This theorem provi<strong>de</strong>s both a param<strong>et</strong>eriz<strong>at</strong>ion of multivari<strong>at</strong>e distributions and a construction scheme for<br />
copulas. In<strong>de</strong>ed, given a multivari<strong>at</strong>e distribution F with margins F1, · · · , Fn, the function<br />
C(u1, · · · , un) = F F −1<br />
1 (u1), · · · , F −1<br />
n (un) <br />
(14)<br />
is autom<strong>at</strong>ically a n-copula. Applying this theorem to the multivari<strong>at</strong>e Gaussian distribution, we can <strong>de</strong>rive<br />
the so-called Gaussian copula.<br />
DEFINITION 4 (GAUSSIAN COPULA)<br />
L<strong>et</strong> Φ <strong>de</strong>note the standard Normal distribution and ΦV,n the n-dimensional Gaussian distribution with cor-<br />
rel<strong>at</strong>ion m<strong>at</strong>rix V. Then, the Gaussian n-copula with correl<strong>at</strong>ion m<strong>at</strong>rix V is<br />
−1<br />
CV (u1, · · · , un) = ΦV,n Φ (u1), · · · , Φ −1 (un) , (15)<br />
whose <strong>de</strong>nsity<br />
cV (u1, · · · , un) = ∂CV (u1, · · · , un)<br />
∂u1 · · · ∂un<br />
6<br />
(16)
eads<br />
cV (u1, · · · , un) =<br />
<br />
1<br />
√ exp −<br />
d<strong>et</strong> V 1<br />
2 yt (u) (V−1 <br />
− Id)y (u)<br />
with yk(u) = Φ −1 (uk). Note th<strong>at</strong> theorem 1 and equ<strong>at</strong>ion (14) ensure th<strong>at</strong> CV (u1, · · · , un) in equ<strong>at</strong>ion (15)<br />
is a copula. <br />
It can be shown th<strong>at</strong> the Gaussian copula n<strong>at</strong>urally arises when one tries to d<strong>et</strong>ermine the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween<br />
random variables using the principle of entropy maximiz<strong>at</strong>ion (Rao 1973, Sorn<strong>et</strong>te <strong>et</strong> al. 2000b, for<br />
instance). Its pertinence and limit<strong>at</strong>ions for mo<strong>de</strong>ling the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s r<strong>et</strong>urns has been tested<br />
by Malevergne and Sorn<strong>et</strong>te (2001), who show th<strong>at</strong> in most cases, this <strong>de</strong>scription of the <strong>de</strong>pen<strong>de</strong>nce can<br />
be consi<strong>de</strong>red s<strong>at</strong>isfying, specially for stocks, provi<strong>de</strong>d th<strong>at</strong> one does not consi<strong>de</strong>r too extreme realiz<strong>at</strong>ions<br />
(Malevergne and Sorn<strong>et</strong>te 2002a, Malevergne and Sorn<strong>et</strong>te 2002b, Mashal and Zeevi 2002).<br />
2 Portfolio wealth distribution for several <strong>de</strong>pen<strong>de</strong>nce structures<br />
2.1 Portfolio wealth distribution for in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s<br />
L<strong>et</strong> us first consi<strong>de</strong>r the case of a portfolio ma<strong>de</strong> of in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s. This limiting (and unrealistic) case<br />
is a m<strong>at</strong>hem<strong>at</strong>ical i<strong>de</strong>aliz<strong>at</strong>ion which provi<strong>de</strong>s a first n<strong>at</strong>ural benchmark of the class of portolio r<strong>et</strong>urn distributions<br />
to be expected. Moreover, it is generally the only case for which the calcul<strong>at</strong>ions are analytically<br />
tractable. For such in<strong>de</strong>pen<strong>de</strong>pent ass<strong>et</strong>s distributed with the modified Weibull distributions, the following<br />
results prove the asymptotic stability of this s<strong>et</strong> of distributions:<br />
THEOREM 2 (TAIL EQUIVALENCE FOR I.I.D MODIFIED WEIBULL RANDOM VARIABLES)<br />
L<strong>et</strong> X1, X2, · · · , XN be N in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically W(c, χ)-distributed random variables. Then, the<br />
variable<br />
SN = X1 + X2 + · · · + XN<br />
(18)<br />
is equivalent in the lower and upper tail to Z ∼ W(c, ˆχ), with<br />
<br />
395<br />
(17)<br />
ˆχ = N c−1<br />
c χ, c > 1, (19)<br />
ˆχ = χ, c ≤ 1. (20)<br />
This theorem is a direct consequence of the theorem st<strong>at</strong>ed below and is based on the result given by Frisch<br />
and Sorn<strong>et</strong>te (1997) for c > 1 and on general properties of sub-exponential distributions when c ≤ 1.<br />
THEOREM 3 (TAIL EQUIVALENCE FOR WEIGHTED SUMS OF INDEPENDENT VARIABLES)<br />
L<strong>et</strong> X1, X2, · · · , XN be N in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically W(c, χ)-distributed random variables. L<strong>et</strong> w1,<br />
w2, · · · , wN be N non-random real coefficients. Then, the variable<br />
SN = w1X1 + w2X2 + · · · + wNXN<br />
is equivalent in the upper and the lower tail to Z ∼ W(c, ˆχ) with<br />
ˆχ =<br />
<br />
N<br />
|wi| c<br />
<br />
c−1<br />
c−1<br />
c<br />
· χ, c > 1, (22)<br />
<br />
i=1<br />
(21)<br />
ˆχ = max<br />
i {|w1|, |w2|, · · · , |wN|}, c ≤ 1. (23)<br />
7
396 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
The proof of this theorem is given in appendix A.<br />
COROLLARY 2<br />
L<strong>et</strong> X1, X2, · · · , XN be N in<strong>de</strong>pen<strong>de</strong>nt random variables such th<strong>at</strong> Xi ∼ W(c, χi). L<strong>et</strong> w1, w2, · · · , wN<br />
be N non-random real coefficients. Then, the variable<br />
SN = w1X1 + w2X2 + · · · + wNXN<br />
is equivalent in the upper and the lower tail to Z ∼ W(c, ˆχ) with<br />
<br />
ˆχ =<br />
N<br />
i=1<br />
|wiχi| c<br />
<br />
c−1<br />
c−1<br />
c<br />
(24)<br />
, c > 1, (25)<br />
ˆχ = max<br />
i {|w1χ1|, |w2χ2|, · · · , |wNχN|}, c ≤ 1. (26)<br />
The proof of the corollary is a straightforward applic<strong>at</strong>ion of theorem 3. In<strong>de</strong>ed, l<strong>et</strong> Y1, Y2, · · · , YN be N<br />
in<strong>de</strong>pen<strong>de</strong>nt and i<strong>de</strong>ntically W(c, 1)-distributed random variables. Then,<br />
which yields<br />
(X1, X2, · · · , XN) d = (χ1Y1, χ2Y2, · · · , χNYN), (27)<br />
SN d = w1χ1 · Y1 + w2χ2 · Y2 + · · · + wNχN · YN . (28)<br />
Thus, applying theorem 3 to the i.i.d variables Yi’s with weights wiχi leads to corollary 2.<br />
2.2 Portfolio wealth distribution for comonotonic ass<strong>et</strong>s<br />
The case of comonotonic ass<strong>et</strong>s is of interest as the limiting case of the strongest possible <strong>de</strong>pen<strong>de</strong>nce<br />
b<strong>et</strong>ween random variables. By <strong>de</strong>finition,<br />
DEFINITION 5 (COMONOTONICITY)<br />
the variables X1, X2, · · · , XN are comonotonic if and only if there exits a random variable U and non<strong>de</strong>creasing<br />
functions f1, f2, · · · , fN such th<strong>at</strong><br />
<br />
(X1, X2, · · · , XN) d = (f1(U), f2(U), · · · , fN(U)). (29)<br />
In terms of copulas, the comonotonicity can be expressed by the following form of the copula<br />
C(u1, u2, · · · , uN) = min(u1, u2, · · · , uN) . (30)<br />
This expression is known as the Fréch<strong>et</strong>-Hoeffding upper bound for copulas (Nelsen 1998, for instance). It<br />
would be appealing to think th<strong>at</strong> estim<strong>at</strong>ing the Value-<strong>at</strong>-Risk un<strong>de</strong>r the comonotonicity assumption could<br />
provi<strong>de</strong> an upper bound for the Value-<strong>at</strong>-Risk. However, it turns out to be wrong, due –as we shall see in the<br />
sequel– to the lack of coherence (in the sense of Artzner <strong>et</strong> al. (1999)) of the Value-<strong>at</strong>-Risk, in the general<br />
case. Notwithstanding, an upper and lower bound can always be <strong>de</strong>rived for the Value-<strong>at</strong>-Risk (Embrechts<br />
<strong>et</strong> al. 2002b). But in the present situ<strong>at</strong>ion, where we are only interested in the class of modified Weibull<br />
distributions with a Gaussian copula, the VaR <strong>de</strong>rived un<strong>de</strong>r the comonoticity assumption will actually<br />
represent the upper bound (<strong>at</strong> least for the VaR calcul<strong>at</strong>ed <strong>at</strong> sufficiently hight confi<strong>de</strong>nce levels).<br />
8
THEOREM 4 (TAIL EQUIVALENCE FOR A SUM OF COMONOTONIC RANDOM VARIABLES)<br />
L<strong>et</strong> X1, X2, · · · , XN be N comonotonic random variables such th<strong>at</strong> Xi ∼ W(c, χi). L<strong>et</strong> w1, w2, · · · , wN<br />
be N non-random real coefficients. Then, the variable<br />
SN = w1X1 + w2X2 + · · · + wNXN<br />
is equivalent in the upper and the lower tail to Z ∼ W(c, ˆχ) with<br />
ˆχ = <br />
wiχi . (32)<br />
<br />
The proof is obvious since, un<strong>de</strong>r the assumption of comonotonicity, the portfolio wealth S is given by<br />
S = <br />
and for modified Weibull distributions, we have<br />
i<br />
i<br />
wi · Xi d =<br />
fi(·) = sgn(·) χi<br />
397<br />
(31)<br />
N<br />
wi · fi(U), (33)<br />
i=1<br />
2/ci | · |<br />
√2 , (34)<br />
in the symm<strong>et</strong>ric case while U is a Gaussian random variable. If, in addition, we assume th<strong>at</strong> all ass<strong>et</strong>s have<br />
the same exponent ci = c, it is clear th<strong>at</strong> S ∼ W(c, ˆχ) with<br />
ˆχ = <br />
wiχi. (35)<br />
i<br />
It is important to note th<strong>at</strong> this rel<strong>at</strong>ion is exact and not asymptotic as in the case of in<strong>de</strong>pen<strong>de</strong>nt variables.<br />
When the exponents ci’s are different from an ass<strong>et</strong> to another, a similar result holds, since we can still write<br />
the inverse cumul<strong>at</strong>ive function of S as<br />
F −1<br />
S (p) =<br />
N<br />
i=1<br />
wiF −1<br />
(p), p ∈ (0, 1), (36)<br />
Xi<br />
which is the property of additive comonotonicity of the Value-<strong>at</strong>-Risk1 . L<strong>et</strong> us then sort the Xi’s such th<strong>at</strong><br />
c1 = c2 = · · · = cp < cp+1 ≤ · · · ≤ cN. We immedi<strong>at</strong>ely obtain th<strong>at</strong> S is equivalent in the tail to<br />
Z ∼ W(c1, ˆχ), where<br />
p<br />
ˆχ = wiχi. (37)<br />
i=1<br />
In such a case, only the ass<strong>et</strong>s with the f<strong>at</strong>est tails contributes to the behavior of the sum in the large <strong>de</strong>vi<strong>at</strong>ion<br />
regime.<br />
1 This rel<strong>at</strong>ion shows th<strong>at</strong>, in general, the VaR calcul<strong>at</strong>ed for comonotonic ass<strong>et</strong>s does not provi<strong>de</strong> an upper bound of the VaR,<br />
wh<strong>at</strong>ever the <strong>de</strong>pen<strong>de</strong>nce structure the portfolio may be. In<strong>de</strong>ed, in such a case, we have VaR(X1 + X2) = VaR(X1) + VaR(X2)<br />
while, by lack of coherence, we may have VaR(X1 + X2) ≥ VaR(X1) + VaR(X2) for some <strong>de</strong>pen<strong>de</strong>nce structure b<strong>et</strong>ween X1<br />
and X2.<br />
9
398 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
2.3 Portfolio wealth un<strong>de</strong>r the Gaussian copula hypothesis<br />
2.3.1 Deriv<strong>at</strong>ion of the multivari<strong>at</strong>e distribution with a Gaussian copula and modified Weibull margins<br />
An advantage of the class of modified Weibull distributions (7) is th<strong>at</strong> the transform<strong>at</strong>ion into a Gaussian,<br />
and thus the calcul<strong>at</strong>ion of the vector y introduced in <strong>de</strong>finition 1, is particularly simple. It takes the form<br />
yk = sgn(xk) √ <br />
|xk|<br />
2<br />
χk<br />
c k 2<br />
, (38)<br />
where yk is normally distributed . These variables Yi then allow us to obtain the covariance m<strong>at</strong>rix V of the<br />
Gaussian copula :<br />
<br />
|xi|<br />
Vij = 2 · E sgn(xixj)<br />
χi<br />
c i<br />
2 |xj|<br />
χj<br />
cj <br />
2<br />
, (39)<br />
which always exists and can be efficiently estim<strong>at</strong>ed. The multivari<strong>at</strong>e <strong>de</strong>nsity P (x) is thus given by:<br />
P (x1, · · · , xN) = cV (x1, x2, · · · , xN)<br />
=<br />
1<br />
2 N π N/2√ V<br />
N<br />
pi(xi) (40)<br />
i=1<br />
N ci|xi| c/2−1<br />
⎡<br />
exp ⎣− <br />
i=1<br />
χ c/2<br />
i<br />
i,j<br />
V −1<br />
ij<br />
Obviously, similar transforms hold, mut<strong>at</strong>is mutandis, for the asymm<strong>et</strong>ric case (8,9).<br />
⎤<br />
c/2 c/2 |xi| |xj|<br />
⎦ . (41)<br />
2.3.2 Asymptotic distribution of a sum of modified Weibull variables with the same exponent c > 1<br />
We now consi<strong>de</strong>r a portfolio ma<strong>de</strong> of <strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s with pdf given by equ<strong>at</strong>ion (41) or its asymm<strong>et</strong>ric<br />
generaliz<strong>at</strong>ion. For such distributions of ass<strong>et</strong> r<strong>et</strong>urns, we obtain the following result<br />
THEOREM 5 (TAIL EQUIVALENCE FOR A SUM OF DEPENDENT RANDOM VARIABLES)<br />
L<strong>et</strong> X1, X2, · · · , XN be N random variables with a <strong>de</strong>pen<strong>de</strong>nce structure <strong>de</strong>scribed by the Gaussian copula<br />
with correl<strong>at</strong>ion m<strong>at</strong>rix V and such th<strong>at</strong> each Xi ∼ W(c, χi). L<strong>et</strong> w1, w2, · · · , wN be N (positive) nonrandom<br />
real coefficients. Then, the variable<br />
SN = w1X1 + w2X2 + · · · + wNXN<br />
is equivalent in the upper and the lower tail to Z ∼ W(c, ˆχ) with<br />
ˆχ =<br />
<br />
where the σi’s are the unique (positive) solution of<br />
<br />
<br />
i<br />
i<br />
wiχiσi<br />
c−1<br />
c<br />
χi<br />
χj<br />
(42)<br />
, (43)<br />
V −1<br />
ik σi c/2 = wkχk σ 1−c/2<br />
k , ∀k . (44)<br />
10
In<strong>de</strong>pen<strong>de</strong>nt Ass<strong>et</strong>s<br />
Comonotonic Ass<strong>et</strong>s<br />
N c<br />
i=1 |wiχi| c−1<br />
ˆχ λ−<br />
c−1<br />
c<br />
, c > 1<br />
<br />
N−1<br />
c 2<br />
2(c−1)<br />
max{|w1χ1|, · · · , |wNχN|}, c ≤ 1 Card {|wiχi| = maxj{ |wjχj|}}<br />
N<br />
i=1 wiχi<br />
Gaussian copula ( <br />
c−1<br />
i wiχiσi) c , c > 1 see appendix B<br />
Table 1: Summary of the various scale factors obtained for different distribution of ass<strong>et</strong> r<strong>et</strong>urns.<br />
The proof of this theorem follows the same lines as the proof of theorem 3. We thus only provi<strong>de</strong> a heuristic<br />
<strong>de</strong>riv<strong>at</strong>ion of this result in appendix B. Equ<strong>at</strong>ion (44) is equivalent to<br />
<br />
Vik wkχk σk 1−c/2 = σi c/2 , ∀i . (45)<br />
k<br />
which seems more <strong>at</strong>tractive since it does not require the inversion of the correl<strong>at</strong>ion m<strong>at</strong>rix. In the special<br />
case where V is the i<strong>de</strong>ntity m<strong>at</strong>rix, the variables Xi’s are in<strong>de</strong>pen<strong>de</strong>nt so th<strong>at</strong> equ<strong>at</strong>ion (43) must yield the<br />
same result as equ<strong>at</strong>ion (22). This results from the expression of σk = (wkχk) 1<br />
c−1 valid in the in<strong>de</strong>pen<strong>de</strong>nt<br />
case. Moreover, in the limit where all entries of V equal one, we r<strong>et</strong>rieve the case of comonotonic ass<strong>et</strong>s.<br />
Obviously, V−1 does not exist for comonotonic ass<strong>et</strong>s and the <strong>de</strong>riv<strong>at</strong>ion given in appendix B does not hold,<br />
but equ<strong>at</strong>ion (45) remains well-<strong>de</strong>fined and still has a unique solution σk = ( wkχk) 1<br />
c−1 which yields the<br />
scale factor given in theorem 4.<br />
2.4 Summary<br />
In the previous sections, we have shown th<strong>at</strong> the wealth distribution FS(x) of a portfolio ma<strong>de</strong> of ass<strong>et</strong>s with<br />
modified Weibull distributions with the same exponent c remains equivalent in the tail to a modified Weibull<br />
distribution W(c, ˆχ). Specifically,<br />
FS(x) ∼ λ− FZ(x) , (46)<br />
when x → −∞, and where Z ∼ W(c, ˆχ). Expression (46) <strong>de</strong>fines the proportionality factor or weight λ−<br />
of the neg<strong>at</strong>ive tail of the portfolio wealth distribution FS(x). Table 1 summarizes the value of the scale<br />
param<strong>et</strong>er ˆχ for the different types of <strong>de</strong>pen<strong>de</strong>nce we have studied. In addition, we give the value of the<br />
coefficient λ−, which may also <strong>de</strong>pend on the weights of the ass<strong>et</strong>s in the portfolio in the case of <strong>de</strong>pen<strong>de</strong>nt<br />
ass<strong>et</strong>s.<br />
11<br />
1<br />
399
400 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
3 Value-<strong>at</strong>-Risk<br />
3.1 Calcul<strong>at</strong>ion of the VaR<br />
We consi<strong>de</strong>r a portfolio ma<strong>de</strong> of N ass<strong>et</strong>s with all the same exponent c and scale param<strong>et</strong>ers χi, i ∈<br />
{1, 2, · · · , N}. The weight of the i th ass<strong>et</strong> in the portfolio is <strong>de</strong>noted by wi. By <strong>de</strong>finition, the Value<strong>at</strong>-Risk<br />
<strong>at</strong> the loss probability α, <strong>de</strong>noted by VaRα, is given , for a continuous distribution of profit and loss,<br />
by<br />
Pr{W (τ) − W (0) < −VaRα} = α, (47)<br />
which can be rewritten as<br />
<br />
Pr S < − VaRα<br />
<br />
= α. (48)<br />
W (0)<br />
In this expression, we have assumed th<strong>at</strong> all the wealth is invested in risky ass<strong>et</strong>s and th<strong>at</strong> the risk-free<br />
interest r<strong>at</strong>e equals zero, but it is easy to reintroduce it, if necessary. It just leads to discount VaRα by the<br />
discount factor 1/(1 + µ0), where µ0 <strong>de</strong>notes the risk-free interest r<strong>at</strong>e.<br />
Now, using the fact th<strong>at</strong> FS(x) ∼ λ− FZ(x), when x → −∞, and where Z ∼ W(c, ˆχ), we have<br />
<br />
1<br />
Pr S < − VaRα<br />
<br />
√2<br />
c/2<br />
VaRα<br />
1 − Φ<br />
, (49)<br />
W (0)<br />
W (0) ˆχ<br />
λ−<br />
as VaRα goes to infinity, which allows us to obtain a closed expression for the asymptotic Value-<strong>at</strong>-Risk<br />
with a loss probability α:<br />
VaRα W (0) ˆχ<br />
21/c <br />
Φ −1<br />
<br />
1 − α<br />
2/c , (50)<br />
λ−<br />
ξ(α) 2/c W (0) · ˆχ, (51)<br />
where the function Φ(·) <strong>de</strong>notes the cumul<strong>at</strong>ive Normal distribution function and<br />
ξ(α) ≡ 1<br />
2 Φ−1<br />
<br />
1 − α<br />
<br />
. (52)<br />
λ−<br />
In the case where a fraction w0 of the total wealth is invested in the risk-free ass<strong>et</strong> with interest r<strong>at</strong>e µ0, the<br />
previous equ<strong>at</strong>ion simply becomes<br />
VaRα ξ(α) 2/c (1 − w0) · W (0) · ˆχ − w0W (0)µ0. (53)<br />
Due to the convexity of the scale param<strong>et</strong>er ˆχ, the VaR is itself convex and therefore sub-additive. Thus, for<br />
this s<strong>et</strong> of distributions, the VaR becomes coherent when the consi<strong>de</strong>red quantiles are sufficiently small.<br />
The Expected-Shortfall ESα, which gives the average loss beyond the VaR <strong>at</strong> probability level α, is also<br />
very easily computable:<br />
α<br />
ESα = 1<br />
VaRu du (54)<br />
α 0<br />
= ζ(α)(1 − w0) · W (0) · ˆχ − w0W (0)µ0, (55)<br />
where ζ(α) = 1<br />
α<br />
α 0 ξ(u)2/c du . Thus, the Value-<strong>at</strong>-Risk, the Expected-Shortfall and in fact any downsi<strong>de</strong><br />
risk measure involving only the far tail of the distribution of r<strong>et</strong>urns are entirely controlled by the scale<br />
param<strong>et</strong>er ˆχ. We see th<strong>at</strong> our s<strong>et</strong> of multivari<strong>at</strong>e modified Weibull distributions enjoy, in the tail, exactly the<br />
same properties as the Gaussian distributions, for which, all the risk measures are controlled by the standard<br />
<strong>de</strong>vi<strong>at</strong>ion.<br />
12
3.2 Typical recurrence time of large losses<br />
L<strong>et</strong> us transl<strong>at</strong>e these formulas in intuitive form. For this, we <strong>de</strong>fine a Value-<strong>at</strong>-Risk VaR ∗ which is such th<strong>at</strong><br />
its typical frequency is 1/T0. T0 is by <strong>de</strong>finition the typical recurrence time of a loss larger than VaR ∗ . In<br />
our present example, we take T0 equals 1 year for example, i.e., VaR ∗ is the typical annual shock or crash.<br />
Expression (49) then allows us to predict the recurrence time T of a loss of amplitu<strong>de</strong> VaR equal to β times<br />
this reference value VaR ∗ :<br />
401<br />
<br />
T<br />
ln (β<br />
T0<br />
c ∗ c VaR<br />
− 1)<br />
+ O(ln β) . (56)<br />
W (0) ˆχ<br />
Figure 2 shows ln T versus β. Observe th<strong>at</strong> T increases all the more slowly with β, the smaller is the<br />
T0<br />
exponent c. This quantifies our expect<strong>at</strong>ion th<strong>at</strong> large losses occur more frequently for the “wil<strong>de</strong>r” subexponential<br />
distributions than for super-exponential ones.<br />
4 Optimal portfolios<br />
In this section, we present our results on the problem of the efficient portfolio alloc<strong>at</strong>ion for ass<strong>et</strong> distributed<br />
according to modified Weibull distributions with the different <strong>de</strong>pen<strong>de</strong>nce structures studied in the previous<br />
sections. We focus on the case when all ass<strong>et</strong> modified Weibull distributions have the same exponent c, as<br />
it provi<strong>de</strong>s the richest and more varied situ<strong>at</strong>ion. When this is not the case and the ass<strong>et</strong>s have different<br />
exponents ci, i = 1, ..., N, the asymptotic tail of the portfolio r<strong>et</strong>urn distribution is domin<strong>at</strong>ed by the ass<strong>et</strong><br />
with the heaviest tail. The largest risks of the portfolio are thus controlled by the single most risky ass<strong>et</strong><br />
characterized by the smallest exponent c. Such extreme risk cannot be diversified away. In such a case, for<br />
a risk-averse investor, the best str<strong>at</strong>egy focused on minimizing the extreme risks consists in holding only the<br />
ass<strong>et</strong> with the thinnest tail, i.e., with the largest exponent c.<br />
4.1 Portfolios with minimum risk<br />
L<strong>et</strong> us consi<strong>de</strong>r first the problem of finding the composition of the portfolio with minimum risks, where<br />
the risks are measured by the Value-<strong>at</strong>-Risk. We consi<strong>de</strong>r th<strong>at</strong> short sales are not allowed, th<strong>at</strong> the risk free<br />
interest r<strong>at</strong>e equals zero and th<strong>at</strong> all the wealth is invested in stocks. This last condition is in<strong>de</strong>ed the only<br />
interesting one since allowing to invest in a risk-free ass<strong>et</strong> would autom<strong>at</strong>ically give the trivial solution in<br />
which the minimum risk portfolio is compl<strong>et</strong>ely invested in the risk-free ass<strong>et</strong>.<br />
The problem to solve reads:<br />
VaR ∗ α = min VaRα = ξ(α) 2/c W (0) · min ˆχ (57)<br />
N<br />
i=1 wi = 1 (58)<br />
wi ≥ 0 ∀i. (59)<br />
In some cases (see table 1), the prefactor ξ(α) <strong>de</strong>fined in (52) also <strong>de</strong>pends on the weight wi’s through λ−<br />
<strong>de</strong>fined in (46). But, its contribution remains subdominant for the large losses. This allows to restrict the<br />
minimiz<strong>at</strong>ion to ˆχ instead of ξ(α) 2/c · ˆχ.<br />
13
402 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
4.1.1 Case of in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s<br />
“Super-exponential” portfolio (c > 1)<br />
Consi<strong>de</strong>r ass<strong>et</strong>s distributed according to modified Weibull distributions with the same exponent c > 1.<br />
The Value-<strong>at</strong>-Risk is given by<br />
VaRα = ξ(α) 2/c W (0) ·<br />
N<br />
i=1<br />
|wiχi| c<br />
<br />
c−1<br />
c−1<br />
2<br />
Introducing the Lagrange multiplier λ, the first or<strong>de</strong>r condition yields<br />
∂ ˆχ<br />
∂wi<br />
and the composition of the minimal risk portfolio is<br />
=<br />
λ<br />
ξ(α) W (0)<br />
wi ∗ = χ−c<br />
i <br />
j χ−c<br />
j<br />
which s<strong>at</strong>istifies the positivity of the Hessian m<strong>at</strong>rix Hjk = ∂2 ˆχ<br />
∂wj∂wk {w∗ i }<br />
The minimal risk portfolio is such th<strong>at</strong><br />
VaR ∗ α = ξ(α)2/c W (0)<br />
i χ−c<br />
1<br />
c<br />
j<br />
, (60)<br />
∀i, (61)<br />
<br />
<br />
, µ ∗ <br />
=<br />
i χ−c<br />
i µi<br />
<br />
j χ−c<br />
j<br />
where µi is the r<strong>et</strong>urn of ass<strong>et</strong> i and µ ∗ is the r<strong>et</strong>urn of the minimum risk portfolio.<br />
(second or<strong>de</strong>r condition).<br />
(62)<br />
, (63)<br />
sub-exponential portfolio (c ≤ 1)<br />
Consi<strong>de</strong>r ass<strong>et</strong>s distributed according to modified Weibull distributions with the same exponent c < 1.<br />
The Value-<strong>at</strong>-Risk is now given by<br />
VaRα = ξ(α) c/2 W (0) · max{|w1χ1|, · · · , |wNχN|}. (64)<br />
Since the weights wi are positive, the modulus appearing in the argument of the max() function can be<br />
removed. It is easy to see th<strong>at</strong> the minimum of VaRα is obtained when all the wiχi’s are equal, provi<strong>de</strong>d<br />
th<strong>at</strong> the constraint wi = 1 can be s<strong>at</strong>isfied. In<strong>de</strong>ed, l<strong>et</strong> us start with the situ<strong>at</strong>ion where<br />
w1χ1 = w2χ2 = · · · = wNχN . (65)<br />
L<strong>et</strong> us <strong>de</strong>crease the weight w1. Then, w1χ1 <strong>de</strong>creases with respect to the initial maximum situ<strong>at</strong>ion (65) but,<br />
in or<strong>de</strong>r to s<strong>at</strong>isfy the constraint <br />
i wi = 1, <strong>at</strong> least one of the other weights wj, j ≥ 2 has to increase, so<br />
th<strong>at</strong> wjχj increases, leading to a maximum for the s<strong>et</strong> of the wiχi’s gre<strong>at</strong>er than in the initial situ<strong>at</strong>ion where<br />
(65) holds. Therefore,<br />
and the constraint <br />
i wi = 1 yields<br />
w ∗ i = A<br />
A =<br />
χi<br />
, ∀i, (66)<br />
1<br />
<br />
i χ−1<br />
i<br />
14<br />
, (67)
and finally<br />
w ∗ i = χ−1<br />
i<br />
<br />
j χ−1<br />
j<br />
, VaR ∗ α = ξ(α)c/2 W (0)<br />
, µ ∗ <br />
=<br />
i χ−1<br />
i<br />
i χ−1<br />
i µi<br />
<br />
j χ−1<br />
j<br />
403<br />
. (68)<br />
The composition of the optimal portfolio is continuous in c <strong>at</strong> the value c = 1. This is the consequence<br />
of the continuity as a function of c <strong>at</strong> c = 1 of the scale factor ˆχ for a sum of in<strong>de</strong>pen<strong>de</strong>nt variables. In<br />
this regime c ≤ 1, the Value-<strong>at</strong>-Risk increases as c <strong>de</strong>creases only through its <strong>de</strong>pen<strong>de</strong>nce on the prefactor<br />
ξ(α) 2/c since the scale factor ˆχ remains constant.<br />
4.1.2 Case of comonotonic ass<strong>et</strong>s<br />
For comonotonic ass<strong>et</strong>s, the Value-<strong>at</strong>-Risk is<br />
VaRα = ξ(α) c/2 W (0) · <br />
which leads to a very simple linear optimiz<strong>at</strong>ion problem. In<strong>de</strong>ed, <strong>de</strong>noting χ1 = min{χ1, χ2, · · · , χN},<br />
we have <br />
wi = χ1, (70)<br />
i<br />
wiχi ≥ χ1<br />
i<br />
i<br />
wiχi<br />
which proves th<strong>at</strong> the composition of the optimal portfolio is w ∗ 1 = 1, w∗ i<br />
= 0 i ≥ 2 leading to<br />
(69)<br />
VaR ∗ α = ξ(α) c/2 W (0)χ1, µ ∗ = µ1. (71)<br />
This result is not surprising since all ass<strong>et</strong>s move tog<strong>et</strong>her. Thus, the portfolio with minimum Value-<strong>at</strong>-Risk<br />
is obtained when only the less risky ass<strong>et</strong>, i.e., with the smallest scale factor χi, is held. In the case where<br />
there is a <strong>de</strong>generacy in the smallest χ of or<strong>de</strong>r p (χ1 = χ2 = ... = χp = min{χ1, χ2, · · · , χN}), the<br />
optimal choice lead to invest all the wealth in the ass<strong>et</strong> with the larger expected r<strong>et</strong>urn µj, j ∈ {1, · · · , p}.<br />
However, in an efficient mark<strong>et</strong> with r<strong>at</strong>ional agents, such an opportunity should not exist since the same<br />
risk embodied by χ1 = χ2 = ... = χp should be remuner<strong>at</strong>ed by the same r<strong>et</strong>urn µ1 = µ2 = ... = µp.<br />
4.1.3 Case of ass<strong>et</strong>s with a Gaussian copula<br />
In this situ<strong>at</strong>ion, we cannot solve the problem analytically. We can only assert th<strong>at</strong> the miminiz<strong>at</strong>ion problem<br />
has a unique solution, since the function VaRα({wi}) is convex. In or<strong>de</strong>r to obtain the composition of the<br />
optimal portfolio, we need to perform the following numerical analysis.<br />
It is first nee<strong>de</strong>d to solve the s<strong>et</strong> of equ<strong>at</strong>ions −1<br />
i Vij σc/2<br />
1−c/2<br />
i = wjχjσj or the equivalent s<strong>et</strong> of equ<strong>at</strong>ions<br />
given<br />
<br />
by (45), which can be performed by Newton’s algorithm. Then one have the minimize the quantity<br />
wiχiσi({wi}). To this aim, one can use the gradient algorithm, which requires the calcul<strong>at</strong>ion of the<br />
<strong>de</strong>riv<strong>at</strong>ives of the σi’s with respect to the wk’s. These quantities are easily obtained by solving the linear s<strong>et</strong><br />
of equ<strong>at</strong>ions<br />
c <br />
· V<br />
2 −1<br />
c<br />
2<br />
ij σ −1<br />
i σ c<br />
2 −1<br />
<br />
∂σi c<br />
<br />
1 ∂σj<br />
j + − 1 wjχj = χj · δjk. (72)<br />
∂wk 2 σj ∂wk<br />
i<br />
Then, the analytical solution for in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s or comonotonic ass<strong>et</strong>s can be used to initialize the<br />
minimiz<strong>at</strong>ion algorithm with respect to the weights of the ass<strong>et</strong>s in the portfolio.<br />
15
404 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
4.2 VaR-efficient portfolios<br />
We are now interested in portfolios with minimum Value-<strong>at</strong>-risk, but with a given expected r<strong>et</strong>urn µ =<br />
<br />
i wiµi. We will first consi<strong>de</strong>r the case where, as previously, all the wealth is invested in risky ass<strong>et</strong>s and<br />
we then will discuss the consequences of the introduction of a risk-free ass<strong>et</strong> in the portfolio.<br />
4.2.1 Portfolios without risky ass<strong>et</strong><br />
When the investors have to select risky ass<strong>et</strong>s only, they have to solve the following minimiz<strong>at</strong>ion problem:<br />
VaR ∗ α = min VaRα = ξ(α) W (0) · min ˆχ (73)<br />
N<br />
i=1 wiµi = µ (74)<br />
N<br />
i=1 wi = 1 (75)<br />
wi ≥ 0 ∀i. (76)<br />
In contrast with the research of the minimum risk portfolios where analytical results have been <strong>de</strong>rived, we<br />
need here to use numerical m<strong>et</strong>hods in every situ<strong>at</strong>ion. In the case of super-exponential portfolios, with or<br />
without <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s, the gradient m<strong>et</strong>hod provi<strong>de</strong>s a fast and easy to implement algorithm,<br />
while for sub-exponential portfolios or portfolios ma<strong>de</strong> of comonotonic ass<strong>et</strong>s, one has to use the simplex<br />
m<strong>et</strong>hod since the minimiz<strong>at</strong>ion problem is then linear.<br />
Thus, although not as convenient to handle as analytical results, these optimiz<strong>at</strong>ion problems remain easy to<br />
manage and fast to compute even for large portfolios.<br />
4.2.2 Portfolios with risky ass<strong>et</strong><br />
When a risk-free ass<strong>et</strong> is introduced in the portfolio, the expression of the Value-<strong>at</strong>-Risk is given by equ<strong>at</strong>ion<br />
(53), the minimiz<strong>at</strong>ion problem becomes<br />
VaR ∗ α = min ξ(α) 2/c (1 − w0) · W (0) · ˆχ − w0W (0)µ0<br />
(77)<br />
N<br />
i=1 wiµi = µ (78)<br />
N<br />
i=1 wi = 1 (79)<br />
wi ≥ 0 ∀i. (80)<br />
When the risk-free interest r<strong>at</strong>e µ0 is non zero, we have to use the same numerical m<strong>et</strong>hods as above to<br />
solve the problem. However, if we assume th<strong>at</strong> µ0 = 0, the problem becomes amenable analytically. Its<br />
Lagrangian reads<br />
L = ξ(α) 2/c ⎛<br />
(1 − w0) · W (0) · ˆχ − λ1 ⎝ <br />
⎞ <br />
N<br />
<br />
wiµi − µ ⎠ − λ2 wi − 1 , (81)<br />
= ξ(α) 2/c<br />
⎛<br />
⎝ <br />
j=0<br />
which allows us to show th<strong>at</strong> the weights of the optimal portfolio are<br />
w ∗ i = (1 − w0) ·<br />
wi<br />
ˆwi<br />
N<br />
j=1 ˆwj<br />
i=0<br />
⎞<br />
⎛<br />
⎠ · W (0) · ˆχ − λ1 ⎝ <br />
⎞<br />
wiµi − µ ⎠ , (82)<br />
i=0<br />
i=0<br />
and VaR ∗ α = ξ(α)2/c (1 − w0) · W (0)<br />
2<br />
16<br />
· µ, (83)
where the ˆwi’s are solution of the s<strong>et</strong> of equ<strong>at</strong>ions<br />
ˆχ +<br />
N<br />
i=1<br />
ˆwi<br />
<br />
∂ ˆχ<br />
∂ ˆwi<br />
405<br />
= µi . (84)<br />
Expression (83) shows th<strong>at</strong> the efficient frontier is simply a straight line and th<strong>at</strong> any efficient portfolio is<br />
the sum of two portfolios: a “riskless portfolio” in which a fraction w0 of the initial wealth is invested and<br />
a portfolio with the remaining (1 − w0) of the initial wealth invested in risky ass<strong>et</strong>s. This provi<strong>de</strong>s another<br />
example of the two funds separ<strong>at</strong>ion theorem. A CAPM then holds, since equ<strong>at</strong>ion (84) tog<strong>et</strong>her with the<br />
mark<strong>et</strong> equilibrium assumption yields the proportionality b<strong>et</strong>ween any stock r<strong>et</strong>urn and the mark<strong>et</strong> r<strong>et</strong>urn.<br />
However, these three properties are rigorously established only for a zero risk-free interest r<strong>at</strong>e and may not<br />
remain necessarily true as soon as the risk-free interest r<strong>at</strong>e becomes non zero.<br />
Finally, for practical purpose, the s<strong>et</strong> of weights w∗ i ’s obtained un<strong>de</strong>r the assumption of zero risk-free interest<br />
r<strong>at</strong>e µ0, can be used to initialize the optimiz<strong>at</strong>ion algorithms when µ0 does not vanish.<br />
5 Conclusion<br />
The aim of this work has been to show th<strong>at</strong> the key properties of Gaussian ass<strong>et</strong> distributions of stability<br />
un<strong>de</strong>r convolution, of the equivalence b<strong>et</strong>ween all down-si<strong>de</strong> riks measures, of coherence and of simple<br />
use also hold for a general family of distributions embodying both sub-exponential and super-exponential<br />
behaviors, when restricted to their tail. We then used these results to compute the Value-<strong>at</strong>-Risk (VaR) and<br />
to obtain efficient porfolios in the risk-r<strong>et</strong>urn sense, where the risk is characterized by the Value-<strong>at</strong>-Risk.<br />
Specifically, we have studied a family of modified Weibull distributions to param<strong>et</strong>erize the marginal distributions<br />
of ass<strong>et</strong> r<strong>et</strong>urns, exten<strong>de</strong>d to their multivari<strong>at</strong>e distribution with Gaussian copulas. The relevance to<br />
finance of the family of modified Weibull distributions has been proved in both a context of conditional and<br />
unconditional portfolio management. We have <strong>de</strong>rived exact formulas for the tails of the distribution P (S)<br />
of r<strong>et</strong>urns S of a portfolio of arbitrary composition of these ass<strong>et</strong>s. We find th<strong>at</strong> the tail of P (S) is also<br />
asymptotically a modified Weibull distribution with a characteristic scale χ function of the ass<strong>et</strong> weights<br />
with different functional forms <strong>de</strong>pending on the super- or sub-exponential behavior of the marginals and<br />
on the strength of the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong>s. The <strong>de</strong>riv<strong>at</strong>ion of the portfolio distribution has shown<br />
the asymptotic stability of this family of distribution with the important economic consequence th<strong>at</strong> any<br />
down-si<strong>de</strong> risk measure based upon the tail of the ass<strong>et</strong> r<strong>et</strong>urns distribution are equivalent, in so far as they<br />
all <strong>de</strong>pends on the scale factor χ and keep the same functional form wh<strong>at</strong>ever the number of ass<strong>et</strong>s in the<br />
portfolio may be. Our analytical study of the properties of the VaR has shown the VaR to be coherent. This<br />
justifies the use of the VaR as a coherent risk measure for the class of modified Weibull distributions and<br />
ensures th<strong>at</strong> portfolio optimiz<strong>at</strong>ion problems are always well-conditioned even when not fully analytically<br />
solvable. The Value-<strong>at</strong>-Risk and the Expected-Shortfall have also been shown to be (asymptotically) equivalent<br />
in this framework. In fine, using the large class of modified Weibull distributions, we have provi<strong>de</strong>d<br />
a simple and fast m<strong>et</strong>hod for calcul<strong>at</strong>ing large down-si<strong>de</strong> risks, exemplified by the Value-<strong>at</strong>-Risk, for ass<strong>et</strong>s<br />
with distributions of r<strong>et</strong>urns which fit quite reasonably the empirical distributions.<br />
17
406 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
A Proof of theorem 3 : Tail equivalence for weighted sums of modified<br />
Weibull variables<br />
A.1 Super-exponential case: c > 1<br />
L<strong>et</strong> X1, X2, · · · , XN be N i.i.d random variables with <strong>de</strong>nsity p(·). L<strong>et</strong> us <strong>de</strong>note by f(·) and g(·) two<br />
positive functions such th<strong>at</strong> p(·) = g(·) · e −f(·) . L<strong>et</strong> w1, w2, · · · , wN be N real non-random coefficients,<br />
and S = N<br />
i=1 wixi.<br />
L<strong>et</strong> X = {x ∈ R N , N<br />
i=1 wixi = S}. The <strong>de</strong>nsity of the variable S is given by<br />
<br />
PS(S) =<br />
We will assume the following conditions on the function f<br />
X<br />
dx e −N<br />
i=1 [f(xi)−ln g(xi)] , (85)<br />
1. f(·) is three times continuously differentiable and four times differentiable,<br />
2. f (2) (x) > 0, for |x| large enough,<br />
3. limx→±∞ f (3) (x)<br />
(f (2) (x)) 2 = 0,<br />
4. f (3) is asymptotically monotonous,<br />
5. there is a constant β > 1 such th<strong>at</strong> f (3) (β·x)<br />
f (3) (x)<br />
remains boun<strong>de</strong>d as x goes to infinity,<br />
6. g(·) is ultim<strong>at</strong>ely a monotonous function, regularly varying <strong>at</strong> infinity with indice ν.<br />
L<strong>et</strong> us start with the <strong>de</strong>monstr<strong>at</strong>ion of several propositions.<br />
PROPOSITION 1<br />
un<strong>de</strong>r hypothesis 3, we have<br />
<br />
Proof<br />
lim<br />
x→±∞ |x| · f ′′ (x) = 0. (86)<br />
d 1<br />
Hypothesis 3 can be rewritten as lim<br />
x→±∞ dx f (2) = 0, so th<strong>at</strong><br />
(x)<br />
<br />
<br />
∀ɛ > 0, ∃Aɛ/x > Aɛ =⇒ <br />
d 1<br />
dx<br />
f (2) <br />
<br />
<br />
(x) ≤ ɛ. (87)<br />
Now, since f ′′ is differentiable, 1/f ′′ is also differentiable, and by the mean value theorem, we have<br />
for some ξ ∈ (x, y).<br />
<br />
<br />
<br />
1<br />
f<br />
′′ 1<br />
−<br />
(x) f ′′ <br />
<br />
<br />
(y) = |x − y| · <br />
d 1<br />
dξ<br />
f ′′ <br />
<br />
<br />
(ξ) <br />
18<br />
(88)
Choosing x > y > Aɛ, and applying equ<strong>at</strong>ion (87) tog<strong>et</strong>her with (88) yields<br />
<br />
<br />
<br />
1<br />
f<br />
′′ 1<br />
−<br />
(x) f ′′ <br />
<br />
<br />
(y) ≤ ɛ · |x − y|. (89)<br />
Now, dividing by x and l<strong>et</strong>ting x go to infinity gives<br />
<br />
<br />
lim <br />
1<br />
x→∞ x<br />
· f ′′ <br />
<br />
<br />
(x) ≤ ɛ, (90)<br />
which conclu<strong>de</strong>s the proof. <br />
PROPOSITION 2<br />
Un<strong>de</strong>r assumption 3, we have<br />
<br />
Proof<br />
407<br />
lim<br />
x→±∞ f ′ (x) = +∞. (91)<br />
According to assumption 3 and proposition 1, lim<br />
x→±∞ x · f ′′ (x) = ∞, which means<br />
This thus gives<br />
∀α > 0, ∃Aα/x > Aɛ =⇒ x · f ′′ (x) ≥ α. (92)<br />
∀x ≥ aα, x · f ′′ (x) ≥ α ⇐⇒ f ′′ (x) ≥ α<br />
x<br />
=⇒<br />
x<br />
Aα<br />
f ′′ (t) dt ≥ α ·<br />
x<br />
Aα<br />
dt<br />
t<br />
(93)<br />
(94)<br />
=⇒ f ′ (x) ≥ α · ln x − α · ln Aα + f ′ (Aα). (95)<br />
The right-hand-si<strong>de</strong> of this last equ<strong>at</strong>ion goes to infinity as x goes to infinity, which conclu<strong>de</strong>s the proof. <br />
PROPOSITION 3<br />
Un<strong>de</strong>r assumptions 3 and 6, the function g(·) s<strong>at</strong>isfies<br />
uniformly in h, for any positive constant C. <br />
Proof For g non-<strong>de</strong>creasing, we have<br />
∀|h| ≤ C<br />
f ′′ (x) ,<br />
∀|h| ≤ C<br />
f ′′ g(x + h)<br />
, lim = 1, (96)<br />
(x) x→±∞ g(x)<br />
g<br />
<br />
x<br />
<br />
1 − C<br />
x·f ′′ <br />
(x) g(x + h)<br />
≤<br />
g(x)<br />
g(x) ≤<br />
<br />
g x<br />
1 + C<br />
x·f ′′ (x)<br />
g(x)<br />
<br />
. (97)<br />
If g is non-increasing, the same inequalities hold with the left and right terms exchanged. Therefore, the<br />
final conclusion is easily shown to be in<strong>de</strong>pen<strong>de</strong>nt of the monotocity property of g. From assumption 3 and<br />
proposition 1, we have<br />
∀α > 0, ∃Aα/x > Aɛ =⇒ x · f ′′ (x) ≥ α. (98)<br />
Thus, for all x larger than Aα and all |h| ≤ C/f ′′ (x)<br />
g x 1 − C<br />
<br />
α g(x + h)<br />
≤<br />
g(x) g(x) ≤ g x 1 + C<br />
<br />
α<br />
g(x)<br />
19<br />
(99)
408 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
Now, l<strong>et</strong>ting x go to infinity,<br />
for all α as large as we want, which conclu<strong>de</strong>s the proof. <br />
<br />
1 − C<br />
ν g(x + h)<br />
≤ lim<br />
α x→∞ g(x) ≤<br />
<br />
1 + C<br />
ν , (100)<br />
α<br />
PROPOSITION 4<br />
Un<strong>de</strong>r assumptions 1, 3 and 4 we have, for any positive constant C:<br />
<br />
Proof<br />
L<strong>et</strong> us first remark th<strong>at</strong><br />
∀|h| ≤ C<br />
f ′′ , lim<br />
(x) x→±∞<br />
<br />
sup <br />
ξ∈[x,x+h] f (3) (ξ) <br />
f ′′ (x) 2<br />
<br />
sup <br />
ξ∈[x,x+h] f (3) (ξ) <br />
f ′′ (x) 2 = 0. (101)<br />
= sup <br />
<br />
ξ∈[x,x+h] f (3) (ξ) f<br />
<br />
f (3) (x) <br />
·<br />
(3) (x) <br />
f ′′ . (102)<br />
(x) 2<br />
The rightmost factor in the right-hand-si<strong>de</strong> of the equ<strong>at</strong>ion above goes to zero as x goes to infinity by assumption<br />
3. Therefore, we just have to show th<strong>at</strong> the leftmost factor in the right-hand-si<strong>de</strong> remains boun<strong>de</strong>d<br />
as x goes to infinity to prove Proposition 4.<br />
Applying assumption 4 according to which f (3) is asymptotically monotonous, we have<br />
<br />
sup f ξ∈[x,x+h]<br />
(3) (ξ) <br />
<br />
f (3) (x) <br />
≤<br />
<br />
<br />
f (3)<br />
<br />
x + C<br />
f ′′ =<br />
<br />
<br />
(x)<br />
<br />
f (3) (x) <br />
<br />
<br />
f<br />
(103)<br />
(3)<br />
<br />
x 1 + C<br />
x·f ′′ <br />
<br />
(x)<br />
<br />
f (3) (x) ≤<br />
<br />
,<br />
<br />
f (3)<br />
(104)<br />
x 1 + C<br />
<br />
<br />
α f (3) (x) <br />
, (105)<br />
for every x larger than some positive constant Aα by assumption 3 and proposition 1. Now, for α large<br />
enough, 1 + C<br />
α is less than β (assumption 5) which shows th<strong>at</strong> supξ∈[x,x+h]|f (3) (ξ)|<br />
|f (3) (x)|<br />
remains boun<strong>de</strong>d for large<br />
x, which conclu<strong>de</strong> the proof. <br />
We can now show th<strong>at</strong> un<strong>de</strong>r the assumptions st<strong>at</strong>ed above, the leading or<strong>de</strong>r expansion of PS(S) for large<br />
S and finite N > 1 is obtained by a generaliz<strong>at</strong>ion of Laplace’s m<strong>et</strong>hod which here amounts to remark th<strong>at</strong><br />
the s<strong>et</strong> of x∗ i ’s th<strong>at</strong> maximize the integrand in (85) are solution of<br />
f ′ i(x ∗ i ) = σ(S)wi , (106)<br />
where σ(S) is nothing but a Lagrange multiplier introduced to minimize the expression N<br />
i=1 fi(xi) un<strong>de</strong>r<br />
the constraint N<br />
i=1 wixi = S. This constraint shows th<strong>at</strong> <strong>at</strong> least one xi, for instance x1, goes to infinity<br />
as S → ∞. Since f ′ (x1) is an increasing function by assumption 2 which goes to infinity as x1 → +∞<br />
(proposition 2), expression (106) shows th<strong>at</strong> σ(S) goes to infinity with S, as long as the weight of the ass<strong>et</strong><br />
1 is not zero. Putting the divergence of σ(S) with S in expression (106) for i = 2, ..., N ensures th<strong>at</strong> each<br />
x∗ i increases when S increases and goes to infinity when S goes to infinity.<br />
20
Expanding fi(xi) around x ∗ i yields<br />
where the s<strong>et</strong> of hi = xi − x∗ i obey the condition<br />
409<br />
f(xi) = f(x ∗ i ) + f ′ (x ∗ x∗ i +hi<br />
i ) · hi +<br />
x∗ t<br />
dt<br />
i<br />
x∗ du f<br />
i<br />
′′ (u) (107)<br />
N<br />
wihi = 0 . (108)<br />
i=1<br />
Summing (106) in the presence of rel<strong>at</strong>ion (108), we obtain<br />
N<br />
N<br />
f(xi) = f(x ∗ N<br />
i ) +<br />
i=1<br />
i=1<br />
i=1<br />
i=1<br />
i=1<br />
x ∗ i +hi<br />
Thus exp(− f(xi)) can be rewritten as follows :<br />
<br />
N<br />
<br />
N<br />
exp − f(xi) = exp f(x ∗ N<br />
i ) +<br />
x ∗ i<br />
i=1<br />
t<br />
dt<br />
x∗ du f<br />
i<br />
′′ (u) . (109)<br />
x ∗ i +hi<br />
x ∗ i<br />
t<br />
dt<br />
x∗ du f<br />
i<br />
′′ <br />
(u) . (110)<br />
L<strong>et</strong> us now <strong>de</strong>fine the compact s<strong>et</strong> AC = {h ∈ RN , N i=1 f ′′ (x∗ i )2 · h2 i ≤ C2 } for any given positive<br />
constant C and the s<strong>et</strong> H = {h ∈ RN , N i=1 wihi = 0}. We can thus write<br />
PS(S) =<br />
=<br />
<br />
<br />
H<br />
dh e −N<br />
i=1 [f(xi)−ln g(xi)] , (111)<br />
AC∩H<br />
dh e −N<br />
i=1 [f(xi)−ln<br />
<br />
g(xi)]<br />
+<br />
AC∩H<br />
dh e −N<br />
i=1 [f(xi)−ln g(xi)]<br />
, (112)<br />
We are now going to analyze in turn these two terms in the right-hand-si<strong>de</strong> of (112).<br />
First term of the right-hand-si<strong>de</strong> of (112).<br />
L<strong>et</strong> us start with the first term. We are going to show th<strong>at</strong><br />
lim<br />
S→∞<br />
<br />
i=1x<br />
AC∩H dh e−N<br />
∗ i +hi x∗ dtt<br />
x<br />
i<br />
∗ du f<br />
i<br />
′′ (u)−ln g(xi)<br />
(2π) N−1 i<br />
2 g(x∗ i ) N w<br />
i=1<br />
2 f<br />
iN<br />
j=1 ′′<br />
j (x∗ j )<br />
f ′′<br />
i (x∗ i )<br />
= 1, for some positive C. (113)<br />
In or<strong>de</strong>r to prove this assertion, we will first consi<strong>de</strong>r the leftmost factor of the right-hand-si<strong>de</strong> of (112):<br />
g(xi) e −f(xi) = g(x ∗ i + hi) e −x ∗<br />
= g(x ∗ 1<br />
−<br />
i + hi) e<br />
Since for all ξ ∈ R, e −|ξ| ≤ e −ξ ≤ e |ξ| , we have<br />
e −x ∗ x∗ i<br />
i +h i<br />
dtt<br />
x∗ du [f<br />
i<br />
′′ (u)−f ′′ (x∗ i )]<br />
≤ e −x ∗ x∗ i<br />
i +h i<br />
x ∗ i<br />
i +h i<br />
dtt<br />
x∗ du f<br />
i<br />
′′ (u)<br />
, (114)<br />
2f ′′ (x∗ i )h2 i e −x ∗ i +hi x∗ i<br />
dtt<br />
x ∗ i<br />
21<br />
du [f ′′ (u)−f ′′ (x∗ i )] ≤ ex ∗ i +hi x∗ i<br />
dtt<br />
x∗ du [f<br />
i<br />
′′ (u)−f ′′ (x∗ i )] . (115)<br />
dtt<br />
x∗ du [f<br />
i<br />
′′ (u)−f ′′ (x∗ i )]<br />
,<br />
(116)
410 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
and<br />
e −x ∗ x∗ i<br />
i +h i<br />
dtt<br />
x ∗ i<br />
du |f ′′ (u)−f ′′ (x ∗ i )| ≤ e −x ∗<br />
x ∗ i<br />
i +h i<br />
since wh<strong>at</strong>ever the sign of hi, the quantity x ∗ i +hi<br />
x ∗ i<br />
dtt<br />
x ∗ i<br />
du [f ′′ (u)−f ′′ (x∗ i )] ≤ ex ∗ i +hi x∗ i<br />
dt t<br />
x∗ du |f<br />
i<br />
′′ (u) − f ′′ (x∗ i<br />
dtt<br />
x∗ du |f<br />
i<br />
′′ (u)−f ′′ (x∗ i )| ,<br />
(117)<br />
)| remains always positive.<br />
But, |u − x∗ i | ≤ |hi| ≤ C<br />
f ′′ (x∗ i ), which leads, by the mean value theorem and assumption 1, to<br />
which yields<br />
Thus<br />
|f ′′ (u) − f ′′ (x ∗ i )| ≤ sup<br />
ξ∈(x∗ i ,x∗ i +hi)<br />
|f (3) (ξ)| · |u − x ∗ i |, (118)<br />
0 ≤ <br />
1<br />
− 2i<br />
e sup |f (3) (ξ)|<br />
i<br />
≤ sup<br />
ξ∈(x∗ i ,x∗ i +hi)<br />
|f (3) (ξ)| C<br />
f ′′ (x∗ i ),<br />
≤ sup |f<br />
ξ∈Gi<br />
(3) (ξ)| C<br />
f ′′ (x∗ i ), where Gi =<br />
x ∗<br />
i +hi<br />
x ∗ i<br />
t<br />
dt<br />
x∗ du |f<br />
i<br />
′′ (u) − f ′′ (x ∗ i )| ≤ 1 <br />
2<br />
i<br />
C<br />
f ′′ (x∗ i ) h2 i<br />
≤ e −x ∗ i +hi x∗ i<br />
dtt<br />
x ∗ i<br />
(119)<br />
<br />
x ∗ i − C<br />
f ′′ (x∗ i ), x∗i + C<br />
f ′′ (x∗ i )<br />
<br />
, (120)<br />
sup<br />
ξ∈Gi<br />
|f (3) (ξ)|<br />
du [f ′′ (u)−f ′′ (x∗ i )] 1<br />
2i<br />
≤ e<br />
sup |f (3) (ξ)|<br />
C<br />
f ′′ (x ∗ i )h2 i . (121)<br />
C<br />
f ′′ (x ∗ i ) h2 i , (122)<br />
where sup ξ∈Gi |f (3) (ξ)|, have been <strong>de</strong>noted by sup |f (3) (ξ)| in the previous expression, in or<strong>de</strong>r not to<br />
cumber the not<strong>at</strong>ions.<br />
By proposition 4, we know th<strong>at</strong> for all h ∈ AC and all ɛi > 0<br />
<br />
<br />
sup<br />
|f<br />
<br />
<br />
(3) (ξ)<br />
f ′′ (x∗ i )<br />
<br />
<br />
<br />
<br />
≤ ɛi, for x ∗ i large enough, (123)<br />
so th<strong>at</strong><br />
∀ɛ ′ C·ɛ′<br />
i −<br />
> 0 and ∀h ∈ AC, e 2 h2 i ≤ e −x ∗ x∗ i<br />
for |x| large enough.<br />
i +h i<br />
Moreover, from proposition 3, we have for all ɛi > 0 and x∗ i large enough:<br />
so, for all ɛ ′′ > 0<br />
dtt<br />
x∗ du [f<br />
i<br />
′′ (u)−f ′′ (x∗ i )] ≤ e C·ɛ′<br />
i 2 h2 i , (124)<br />
∀h ∈ AC, (1 − ɛi) ν ≤ g(x∗i + hi)<br />
g(x∗ i ) ≤ (1 + ɛi) ν , (125)<br />
∀h ∈ AC, (1 − ɛ ′′ ) Nν ≤ <br />
Then for all ɛ > 0 and |x| large enough, this yields :<br />
i<br />
g(x ∗ i<br />
+ hi)<br />
g(x∗ i ) ≤ (1 + ɛ′′ ) Nν . (126)<br />
(1 − ɛ) Nν 1<br />
−<br />
e 2i (f ′′ (x∗ i )+C·ɛ)·h2 i ≤ g(xi)<br />
g(x∗ i )e−f(xi) Nν −<br />
≤ (1 + ɛ) e 1<br />
2i (f ′′ (x∗ i )−C·ɛ)·h2 i , (127)<br />
i<br />
22
for all h ∈ AC. Thus, integr<strong>at</strong>ing over all the h ∈ AC ∩ H and by continuity of the mapping<br />
<br />
G(Y) = dh g(h, Y) (128)<br />
AC∩H<br />
1<br />
− where g(h, Y) = e 2Yi·h 2 i , we can conclu<strong>de</strong> th<strong>at</strong>,<br />
Now, we remark th<strong>at</strong><br />
<br />
with<br />
H<br />
<br />
AC∩H<br />
<br />
g(x∗ i ) <br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i =<br />
and <br />
dh e<br />
AC∩H<br />
<br />
H<br />
<br />
g(xi) e −f(xi)<br />
AC∩H<br />
AC∩H<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i =<br />
dh e− 1<br />
2f ′′ (x ∗ i )h2 i<br />
411<br />
S→∞<br />
−−−→ 1. (129)<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 <br />
i +<br />
AC∩H<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i , (130)<br />
1<br />
− 2f ′′ (x∗ i )h2 i ∼ O<br />
(2π) N−1<br />
2<br />
<br />
N w<br />
i=1<br />
2 iN<br />
j=1 f ′′<br />
j (x∗ j )<br />
f ′′<br />
i (x∗ i )<br />
, (131)<br />
α<br />
−<br />
e f ′′ (x∗ <br />
) , α > 0, (132)<br />
where x ∗ = max{x ∗ i } (note th<strong>at</strong> 1/f ′′ (x) → ∞ with x by Proposition 1). In<strong>de</strong>ed, we clearly have<br />
<br />
AC∩H<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i ≤<br />
=<br />
<br />
AC<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i , (133)<br />
(2π) N/2<br />
f ′′ (x ∗ i )<br />
<br />
BC<br />
du <br />
i<br />
1 u<br />
− 2<br />
e 2 i<br />
f ′′ (x∗ i )<br />
<br />
2π f ′′ (x∗ i )<br />
, (134)<br />
where we have performed the change of variable ui = f ′′ (x∗ i ) · hi and <strong>de</strong>noted by BC the s<strong>et</strong> {h ∈<br />
RN , u2 i ≤ C2 }. Now, l<strong>et</strong> x∗ max = max{x∗ i } and x∗min = min{x∗i }. Expression (134) then gives<br />
<br />
AC∩H<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i ≤<br />
SN−1<br />
<br />
(2π) N/2<br />
f ′′ (x∗ min )N/2<br />
du<br />
BC<br />
e−<br />
1<br />
2 f ′′ (x∗ max )u 2 i<br />
(2π f ′′ (x∗ min<br />
f<br />
= SN−1<br />
′′ (x∗ max) N/2<br />
f ′′ (x∗ <br />
N<br />
Γ<br />
min<br />
)N 2 ,<br />
C2 2 f ′′ (x∗ max)<br />
N<br />
2 −1<br />
f ′′ (x ∗ max) N/2<br />
f ′′ (x ∗ min )N<br />
C 2<br />
2 f ′′ (x ∗ max)<br />
))N/2 , (135)<br />
<br />
, (136)<br />
· e −<br />
C 2<br />
2 f ′′ (x∗ max ) , (137)<br />
which <strong>de</strong>cays exponentially fast for large S (or large x ∗ max) as long as f ′′ goes to zero <strong>at</strong> infinity, i.e, for any<br />
function f which goes to infinity not faster than x 2 . So, finally<br />
<br />
AC∩H<br />
1<br />
−<br />
dh e 2f ′′ (x∗ i )h2 i =<br />
which conclu<strong>de</strong>s the proof of equ<strong>at</strong>ion (113).<br />
(2π) N−1<br />
2<br />
<br />
N w<br />
i=1<br />
2 iN<br />
j=1 f ′′<br />
j (x∗ j )<br />
f ′′<br />
i (x∗ i )<br />
23<br />
<br />
+ O e −<br />
α<br />
f ′′ (x∗ max )<br />
<br />
, (138)
412 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
Second term of the right-hand-si<strong>de</strong> of (112).<br />
We now have to show th<strong>at</strong> <br />
dh e<br />
AC∩H<br />
−f(x ∗ i +hi)−g(x∗ i +hi)<br />
(139)<br />
can be neglected. This is obvious since, by assumption 2 and 6, the function f(x) − ln g(x) remains convex<br />
for x large enough, which ensures th<strong>at</strong> f(x) − ln g(x) ≥ C1|x| for some positive constant C1 and x large<br />
enough. Thus, choosing the constant C in AC large enough, we have<br />
<br />
dh e −N<br />
i=1 f(xi)−ln<br />
<br />
g(xi)<br />
≤ dh e −C1N<br />
i=1 |x∗ i +hi|<br />
∼ O<br />
AC∩H<br />
AC∩H<br />
Thus, for S large enough, the <strong>de</strong>nsity PS(S) is asymptotically equal to<br />
PS(S) = <br />
i<br />
g(x ∗ i )<br />
In the case of the modified Weibull variables, we have<br />
f(x) =<br />
(2π) N−1<br />
2<br />
<br />
N w<br />
i=1<br />
2 iN<br />
j=1 f ′′<br />
j (x∗ j )<br />
f ′′<br />
i (x∗ i )<br />
|x|<br />
χ<br />
<br />
α′<br />
−<br />
e f ′′ (x∗ <br />
) . (140)<br />
. (141)<br />
c<br />
, (142)<br />
and<br />
c<br />
g(x) =<br />
2 √ c<br />
· |x| 2<br />
πχc/2 −1 , (143)<br />
which s<strong>at</strong>isfy our assumptions if and only if c > 1. In such a case, we obtain<br />
x ∗ i = w<br />
1<br />
c−1<br />
i<br />
c<br />
c−1<br />
wi<br />
which, after some simple algebraic manipul<strong>at</strong>ions, yield<br />
N−1<br />
c<br />
2 c<br />
P (S) ∼<br />
2(c − 1) 2 √ π<br />
with<br />
as announced in theorem 3.<br />
A.2 Sub-exponential case: c ≤ 1<br />
<br />
ˆχ = w<br />
c<br />
c−1<br />
i<br />
· S, (144)<br />
1 c<br />
|S| 2<br />
ˆχ c/2 −1 e −|S|<br />
ˆχc<br />
c−1<br />
c<br />
(145)<br />
· χ. (146)<br />
L<strong>et</strong> X1, X2, · · · , XN be N i.i.d sub-exponential modified Weibull random variables W(c, χ), with distribution<br />
function F . L<strong>et</strong> us <strong>de</strong>note by GS the distribution function of the variable<br />
where w1, w2, · · · , wN are real non-random coefficients.<br />
SN = w1X1 + w2X2 + · · · + wNXN, (147)<br />
L<strong>et</strong> w ∗ = max{|w1|, |w2|, · · · , |wN|}. Then, theorem 5.5 (b) of Goldie and Klüppelberg (1998) st<strong>at</strong>es th<strong>at</strong><br />
GS(x/w<br />
lim<br />
x→∞<br />
∗ )<br />
F (x) = Card {i ∈ {1, 2, , N} : |wi| = w ∗ }. (148)<br />
By <strong>de</strong>finition 2, this allows us to conclu<strong>de</strong> th<strong>at</strong> SN is equivalent in the upper tail to Z ∼ W(c, w ∗ χ).<br />
A similar calcul<strong>at</strong>ion yields an analogous result for the lower tail.<br />
24
B Asymptotic distribution of the sum of Weibull variables with a Gaussian<br />
copula.<br />
We assume th<strong>at</strong> the marginal distributions are given by the modified Weibull distributions:<br />
Pi(xi) = 1<br />
2 √ π<br />
c<br />
χ c/2<br />
i<br />
|xi| c/2−1 e −|x i |<br />
χ ic<br />
413<br />
(149)<br />
and th<strong>at</strong> the χi’s are all equal to one, in or<strong>de</strong>r not to cumber the not<strong>at</strong>ion. As in the proof of corollary 2, it<br />
will be sufficient to replace wi by wiχi to reintroduce the scale factors.<br />
Un<strong>de</strong>r the Gaussian copula assumption, we obtain the following form for the multivari<strong>at</strong>e distribution :<br />
c<br />
P (x1, · · · , xN) =<br />
N<br />
2Nπ N/2√ N<br />
x<br />
d<strong>et</strong> V<br />
c/2−1<br />
⎡<br />
i exp ⎣− <br />
V −1<br />
ij xc/2<br />
i xc/2<br />
⎤<br />
⎦<br />
j . (150)<br />
L<strong>et</strong><br />
i=1<br />
f(x1, · · · , xN) = <br />
i,j<br />
i,j<br />
V −1<br />
ij xi c/2 xj c/2 . (151)<br />
We have to minimize f un<strong>de</strong>r the constraint wixi = S. As for the in<strong>de</strong>pen<strong>de</strong>nt case, we introduce a<br />
Lagrange multiplier λ which leads to<br />
c <br />
j<br />
V −1<br />
jk x∗ j c/2 x ∗ k c/2−1 = λwk . (152)<br />
The left-hand-si<strong>de</strong> of this equ<strong>at</strong>ion is a homogeneous function of <strong>de</strong>gree c − 1 in the x∗ i ’s, thus necessarily<br />
where the σi’s are solution of <br />
j<br />
x ∗ i =<br />
1<br />
λ c−1<br />
· σi, (153)<br />
c<br />
V −1<br />
jk σj c/2 σk c/2−1 = wk . (154)<br />
The s<strong>et</strong> of equ<strong>at</strong>ions (154) has a unique solution due to the convexity of the minimiz<strong>at</strong>ion problem. This<br />
s<strong>et</strong> of equ<strong>at</strong>ions can be easily solved by a numerical m<strong>et</strong>hod like Newton’s algorithm. It is convenient to<br />
simplify the problem and avoid the inversion of the m<strong>at</strong>rix V , by rewritting (154) as<br />
Using the constraint wix ∗ i<br />
so th<strong>at</strong><br />
<br />
k<br />
= S, we obtain<br />
Vjk wk σk 1−c/2 = σ c/2<br />
j . (155)<br />
1<br />
λ c−1<br />
c<br />
x ∗ i =<br />
=<br />
S<br />
, (156)<br />
wiσi<br />
σi<br />
· S. (157)<br />
wiσi<br />
25
414 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
Thus<br />
f(x1, · · · , xN) = f(x ∗ 1, · · · , x ∗ N) + ∂f<br />
=<br />
Sc ( 1<br />
+<br />
wiσi) c−1 2<br />
i<br />
ij<br />
∂xi<br />
hi + 1 <br />
2<br />
ij<br />
∂2f hihj + · · · (158)<br />
∂xi∂xj<br />
∂2f hihj + · · · , (159)<br />
∂xi∂xj<br />
where, as in the previous section, hi = xi − x ∗ i and the <strong>de</strong>riv<strong>at</strong>ives of f are expressed <strong>at</strong> x∗ 1 , ..., x∗ N .<br />
It is easy to check th<strong>at</strong> the nth-or<strong>de</strong>r <strong>de</strong>riv<strong>at</strong>ive of f with respect to the xi’s evalu<strong>at</strong>ed <strong>at</strong> {x∗ i } is proportional<br />
to Sc−n . In the sequel, we will use the following not<strong>at</strong>ion :<br />
∂n <br />
f <br />
<br />
∂xi1 · · · ∂xin<br />
We can write :<br />
f(x1, · · · , xN) =<br />
up to the fourth or<strong>de</strong>r. This leads to<br />
P (S) ∝ e −<br />
Sc ( Sc−2<br />
+<br />
wiσi) c−1 2<br />
( w<br />
S c<br />
iσi ) c−1<br />
<br />
{x ∗ i }<br />
<br />
ij<br />
= M (n)<br />
i1···in Sc−n . (160)<br />
M (2)<br />
ij hihj + Sc−3<br />
6<br />
<br />
ijk<br />
Sc−2ij<br />
(2)<br />
− M<br />
dh1 · · · dhNe 2<br />
ij hihj<br />
<br />
δ wihi<br />
×<br />
⎡<br />
⎣1 + Sc−3<br />
6<br />
Using the rel<strong>at</strong>ion δ ( wihi) = dk<br />
2π e−ikj wjhj , we obtain :<br />
P (S) ∝ e −<br />
or in vectorial not<strong>at</strong>ion :<br />
( w<br />
S c<br />
iσi ) c−1<br />
P (S) ∝ e −<br />
( w<br />
dk<br />
2π<br />
<br />
S c<br />
iσi ) c−1<br />
×<br />
<br />
M (3)<br />
ijk<br />
M (3)<br />
ijk hihjhk + · · · (161)<br />
ijk hihjhk + · · ·<br />
×<br />
⎤<br />
Sc−2ij<br />
(2)<br />
− M<br />
dh1 · · · dhNe 2<br />
ij hihj−ikj wjhj ×<br />
×<br />
<br />
dk<br />
2π<br />
⎡<br />
⎡<br />
⎣1 + Sc−3<br />
6<br />
⎣1 + Sc−3<br />
6<br />
L<strong>et</strong> us perform the following standard change of variables :<br />
<br />
M (3)<br />
ijk<br />
ijk hihjhk + · · ·<br />
Sc−2<br />
−<br />
dh e 2 htM (2) h−ikwth ×<br />
⎤<br />
<br />
M (3)<br />
(M (2)−1 exists since f is assumed convex and thus M (2) positive) :<br />
S c−2<br />
ijk<br />
ijk hihjhk + · · ·<br />
⎦ . (162)<br />
⎤<br />
⎦ , (163)<br />
⎦ . (164)<br />
h = h ′ − ik<br />
S c−2 M(2)−1 w . (165)<br />
2 ht M (2) h + ikw t h = Sc−2<br />
2 h′t M (2) h ′ + k2<br />
2S c−2 wt M (2)−1 w . (166)<br />
26
This yields<br />
<br />
×<br />
h+<br />
Sc−2<br />
− 2 dh e ik<br />
Sc−2 M(2)−1 w wt<br />
M (2)h+<br />
ik<br />
Sc−2 M(2)−1<br />
S c ( wiσi ) c−1<br />
P (S) ∝ e −<br />
⎡<br />
⎣1 + Sc−3 <br />
M<br />
6<br />
(3)<br />
<br />
dk k2<br />
e− 2S<br />
2π c−2 wtM (2)−1w ×<br />
ijk<br />
ijk hihjhk + · · ·<br />
⎤<br />
415<br />
⎦ . (167)<br />
Denoting by 〈·〉h the average with respect to the Gaussian distribution of h and by 〈·〉k the average with<br />
respect to the Gaussian distribution of k, we have :<br />
<br />
×<br />
⎡<br />
⎣1 + Sc−3<br />
6<br />
<br />
ijk<br />
P (S) ∝<br />
M (3)<br />
ijk 〈〈hihjhk〉h〉k + Sc−4<br />
24<br />
d<strong>et</strong> M (2)−1<br />
w t M (2)−1 w (2πS2−c ) N−1<br />
2 e −<br />
S c ( wiσi ) c−1 ×<br />
<br />
M (4)<br />
ijkl 〈〈hihjhkhl〉h〉k<br />
⎤<br />
+ · · · ⎦ . (168)<br />
We now invoke Wick’s theorem 2 , which st<strong>at</strong>es th<strong>at</strong> each term 〈〈hi · · · hp〉h〉k can be expressed as a product<br />
of pairwise correl<strong>at</strong>ion coefficients. Evalu<strong>at</strong>ing the average with respect to the symm<strong>et</strong>ric distribution of k,<br />
it is obvious th<strong>at</strong> odd-or<strong>de</strong>r terms will vanish and th<strong>at</strong> the count of powers of S involved in each even-or<strong>de</strong>r<br />
term shows th<strong>at</strong> all are sub-dominant. So, up to the leading or<strong>de</strong>r :<br />
<br />
P (S) ∝<br />
The m<strong>at</strong>rix M (2) can be calcul<strong>at</strong>ed, which yields<br />
M (2)<br />
kl<br />
and shows th<strong>at</strong> <br />
=<br />
=<br />
ijkl<br />
d<strong>et</strong> M (2)−1<br />
w t M (2)−1 w (2πS2−c ) N−1<br />
2 e −<br />
S c ( wiσi ) c−1 . (169)<br />
1<br />
( wiσi) c−2<br />
<br />
c<br />
<br />
wk<br />
c − 1 δkl +<br />
2 σk<br />
c2 c<br />
−1 2 V<br />
2 kl σ −1<br />
k σ c<br />
2 −1<br />
<br />
l , (170)<br />
1<br />
( wiσi) c−2 ˜ Mkl, (171)<br />
d<strong>et</strong> M (2)−1<br />
w t M (2)−1 w =<br />
c<br />
(N−1)( 2<br />
wiσi<br />
−1)<br />
The inverse m<strong>at</strong>rix ˜ M −1 s<strong>at</strong>isfies <br />
l ˜ Mkl · ( ˜ M −1 )lj = δkj which can be rewritten:<br />
or equivalently<br />
<br />
c<br />
<br />
wk<br />
c − 1 (<br />
2 σk<br />
˜ M −1 )kj + c2<br />
2<br />
<br />
c<br />
<br />
c − 1 wk (<br />
2 ˜ M −1 )kj + c2<br />
2<br />
<br />
V −1<br />
kl · ( ˜ M −1 )ljσ c<br />
k<br />
l<br />
<br />
V −1<br />
kl · ( ˜ M −1 )ljσ c<br />
2<br />
k<br />
l<br />
<br />
d<strong>et</strong> ˜M −1<br />
wt ˜M −1 . (172)<br />
w<br />
2 −1<br />
σ c<br />
2 −1<br />
l = δkj (173)<br />
σ c<br />
2 −1<br />
l = δkj · σk (174)<br />
2 See for instance (Brézin <strong>et</strong> al. 1976) for a general introduction, (Sorn<strong>et</strong>te 1998) for an early applic<strong>at</strong>ion to the portfolio problem<br />
and (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b) for a system<strong>at</strong>ic utiliz<strong>at</strong>ion with the help of diagrams.<br />
27
416 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
which gives<br />
<br />
c<br />
c − 1<br />
2<br />
<br />
j,k<br />
wk ( ˜ M −1 )kj wj + c2<br />
2<br />
<br />
V −1<br />
kl · ( ˜ M −1 )ljσ c c<br />
2<br />
k σl j,k,l<br />
2 −1<br />
wj = <br />
δkj · σkwj . (175)<br />
Summing the rightmost factor of the left-hand-si<strong>de</strong> over k, and accounting for equ<strong>at</strong>ion (154) leads to<br />
so th<strong>at</strong><br />
Moreover<br />
<br />
c<br />
c − 1<br />
2<br />
c N<br />
<br />
2 N π N/2√ d<strong>et</strong> V<br />
j,k<br />
wk ( ˜ M −1 )kj wj + c2<br />
2<br />
N<br />
w t ˜M −1 w =<br />
<br />
j,l<br />
1<br />
c(c − 1)<br />
c N<br />
x<br />
i=1<br />
∗ i c/2−1 =<br />
2Nπ N/2√d<strong>et</strong> V<br />
Thus, putting tog<strong>et</strong>her equ<strong>at</strong>ions (169), (172), (177) and (178) yields<br />
<br />
P (S) c(c − 1) d<strong>et</strong> ˜ M −1<br />
1<br />
·<br />
d<strong>et</strong> V<br />
2 √ π<br />
with<br />
ˆχ =<br />
c N−1 σ c/2−1<br />
i<br />
2 (N−1)/2<br />
wiχiσi<br />
28<br />
j,k<br />
wl ( ˜ M −1 )ljwj = <br />
j<br />
σjwj . (176)<br />
<br />
wjσj . (177)<br />
j<br />
c<br />
2 σ −1<br />
i<br />
( c<br />
wiσi)<br />
N( 2<br />
−1) SN( c<br />
2 −1) . (178)<br />
c<br />
ˆχ c/2 |S|c/2−1e −|S|<br />
ˆχc<br />
, (179)<br />
c−1<br />
c<br />
. (180)
References<br />
Acerbi, A. and D. Tasche, 2002, On the coherence of expected-shortfall, Journal of Banking and Finance<br />
26, 1487-1503.<br />
An<strong>de</strong>rsen, J.V., and D. Sorn<strong>et</strong>te, 2001, Have your cake and e<strong>at</strong> it too: increasing r<strong>et</strong>urns while lowering large<br />
risks! Journal of risk finance 2, 70-82.<br />
Artzner, P., F. Delbaen, J.M. Eber and D. He<strong>at</strong>h, 1999, Coherent measures of risk, M<strong>at</strong>hem<strong>at</strong>ical Finance 9,<br />
203-228.<br />
Basle Commitee on Banking Supervision, 1996, Amen<strong>de</strong>ment to the capital accord to incorpor<strong>at</strong>e mark<strong>et</strong><br />
risks.<br />
Basle Commitee on Banking Supervision, 2001, The new Basel capital accord.<br />
Bekaert, G. and G.J. Wu, 2000, Asymm<strong>et</strong>ric vol<strong>at</strong>ility and risk in equity mark<strong>et</strong>s, Review of Financial<br />
Studies 13, 1-42.<br />
Black, F., 1976, in Proceedings of the 1976 American St<strong>at</strong>istical Associ<strong>at</strong>ion, Business and Economical<br />
St<strong>at</strong>istics Section (American St<strong>at</strong>istical Associ<strong>at</strong>ion, Alexandria, VA), p. 177.<br />
Bouchaud, J.P., M<strong>at</strong>acz, A. and Potters, M., 2001, Leverage effect in financial mark<strong>et</strong>s: The r<strong>et</strong>ar<strong>de</strong>d vol<strong>at</strong>ility<br />
mo<strong>de</strong>l, Physical Review E 61, 228701.<br />
Bouchaud, J.-P., D. Sorn<strong>et</strong>te, C. Walter and J.-P. Aguilar, 1998, Taming large events: Optimal portfolio<br />
theory for strongly fluctu<strong>at</strong>ing ass<strong>et</strong>s, Intern<strong>at</strong>ional Journal of Theor<strong>et</strong>ical and Applied Finance 1,<br />
25-41.<br />
Brézin, E., J.C. Le Guillou <strong>et</strong> J. Zinn-Justin, 1976, Field theor<strong>et</strong>ical approach to critical phenomena, in<br />
C. Domb and M. Green, vol.6, The renormaliz<strong>at</strong>ion group and its applic<strong>at</strong>ions (Aca<strong>de</strong>mic Press,<br />
London), pp. 125-247.<br />
Brummelhuis, R.G.M. and D. Guégan, 2000, Extreme values of conditional distributions of GARCH(1,1)<br />
processes, Working Paper, University of Reims<br />
Brummelhuis, R.G.M., D. Guégan and S. LaDouc<strong>et</strong>te, in prepar<strong>at</strong>ion.<br />
Campbell, J.Y., A. W. Lo, and A. C. McKinley, 1997, The Econom<strong>et</strong>rics of Financial Mark<strong>et</strong>s (Princ<strong>et</strong>on<br />
University Press, Princ<strong>et</strong>on, NJ),<br />
Chabaane, A., E. Duclos, J.P. Laurent, Y. Malevergne and F. Turpin, 2002, Looking for efficient portfolios:<br />
An empirical investig<strong>at</strong>ion, Working Paper.<br />
Danielsson, J., P. Embrechts, C. Goodhart, C. Ke<strong>at</strong>ing, F. Muennich, O. Renault and H.-S. Shin, 2001, An<br />
aca<strong>de</strong>mic response to Basel II, Working Paper, FMG and ESRC, London.<br />
Duffie, D. and J. Pan, 1997, An Overview of Value <strong>at</strong> Risk, Journal of Deriv<strong>at</strong>ives 4, 7-49.<br />
Embrechts, P., C. Kluppelberg and T. Mikosh, 1997, Mo<strong>de</strong>lling extremal events (Springel-Verlag, Applic<strong>at</strong>ions<br />
of M<strong>at</strong>hem<strong>at</strong>ics 33).<br />
Embrechts, P., A. McNeil and D. Straumann, 2002a, Correl<strong>at</strong>ion and Depen<strong>de</strong>nce in risk management:<br />
properties and pitfalls, In: Risk Management: Value <strong>at</strong> Risk and Beyond, ed. M.A.H. Dempster,<br />
Cambridge University Press, Cambridge, pp. 176-223.<br />
29<br />
417
418 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
Embrecht, P., A. Höing and A. Juri, 2002b, Using copulae to bound the Value-<strong>at</strong>-Risk for functions of<br />
<strong>de</strong>pen<strong>de</strong>nt risks, Working Paper, Risklab.<br />
Fama, E.F., 1965, Portfolio analysis in a stable Par<strong>et</strong>ian mark<strong>et</strong>, Management Science 11, 404-419.<br />
Fouque, J.-P., G. Papanicolaou, and R. Sircar, 2000, Deriv<strong>at</strong>ives in Financial Mark<strong>et</strong>s with Stochastic<br />
Vol<strong>at</strong>ility (Cambridge University Press, Cambridge, UK).<br />
Frisch, U. and D.Sorn<strong>et</strong>te, 1997, Extreme Devi<strong>at</strong>ions and Applic<strong>at</strong>ions, Journal <strong>de</strong>. Physique I France 7,<br />
1155-1171.<br />
Goldie, C. M. and C. Klüppelberg, 1998, Subexponential distributions, in A Practical Gui<strong>de</strong> to Heavy Tails:<br />
St<strong>at</strong>istical Techniques and Applic<strong>at</strong>ions, ed. R. L. Adler, R. Feldman, M. S. Taqqu, pp. 435-459 (<br />
Birkhuser, Boston)<br />
Gopikrishnan, P., M. Meyer, L.A. Nunes Amaral and H.E. Stanley, 1998, Inverse cubic law for the distribution<br />
of stock price vari<strong>at</strong>ions, European Physical Journal B 3, 139-140.<br />
Gouriéroux, C. and J. Jasiak, 1998, Trunc<strong>at</strong>ed maximum likelyhood, goodness of fit tests and tail analysis,<br />
Working paper.<br />
Johansen, A. and D. Sorn<strong>et</strong>te, 1998, Stock mark<strong>et</strong> crashes are outliers, European Physical Journal B 1,<br />
141-144.<br />
Johansen, A. and D. Sorn<strong>et</strong>te, 1998, Large mark<strong>et</strong> price drawdowns are outliers, Journal of Risk 4, 69-110.<br />
Jorion, P., 1997, Value-<strong>at</strong>-Risk: The New Benchmark for Controlling Deriv<strong>at</strong>ives Risk (Irwin Publishing,<br />
Chicago, IL).<br />
Kahneman, D. and A. Tversky, 1979, Prospect theort: An analysis of <strong>de</strong>cision un<strong>de</strong>r risk, Econom<strong>et</strong>rica 47,<br />
263-291.<br />
Laherrère, J. and D. Sorn<strong>et</strong>te, 1998, Str<strong>et</strong>ched exponential distributions in n<strong>at</strong>ure and economy : ”f<strong>at</strong> tails”<br />
with characteristic scales, European Physical Journal B 2, 525-539.<br />
Lintner, J., 1965, The valu<strong>at</strong>ion of risk ass<strong>et</strong>s and the selection of risky investments in stock portfolios and<br />
the capital bug<strong>et</strong>s, Review of Economics and St<strong>at</strong>itics 41, 13–37.<br />
Litterman, R. and K. Winkelmann, 1998, Estim<strong>at</strong>ing covariance m<strong>at</strong>rices (Risk Management Series, Goldman<br />
Sachs).<br />
Lux, T., 1996, The stable Par<strong>et</strong>ian hypothesis and the frequency of large r<strong>et</strong>urns: an examin<strong>at</strong>ion of major<br />
German stocks, Applied Financial Economics 6, 463-475.<br />
Malevergne, Y., V. Pisarenko and D. Sorn<strong>et</strong>te, 2002, Empirical distributions of log-r<strong>et</strong>urns: Exponential or<br />
power-like?, Working paper.<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2001, Testing the Gaussian copulas hypothesis for mo<strong>de</strong>ling financial ass<strong>et</strong>s<br />
<strong>de</strong>pen<strong>de</strong>nce, Working paper, cond-m<strong>at</strong>/0111310.<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2002a, Minimising extremes, Risk 15(11), 129-133.<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2002b, How to account for extreme co-movements b<strong>et</strong>ween individual<br />
stocks and the mark<strong>et</strong>, Working paper.<br />
30
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2002c, Multi-Moments M<strong>et</strong>hod for Portfolio Management: Generalized<br />
Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>l in Homogeneous and H<strong>et</strong>erogeneous mark<strong>et</strong>s, working paper http :<br />
//papers.ssrn.com/paper.taf?abstract id = 319544<br />
Mantegna, R.N. and H.E. Stanley, 1995, Scaling bahavior of an economic in<strong>de</strong>x, N<strong>at</strong>ure 376, 46-55.<br />
Markovitz, H., 1959, Portfolio selection : Efficient diversific<strong>at</strong>ion of investments (John Wiley and Sons, New<br />
York).<br />
Mashal, R. and A. Zeevi, 2002, Beyond Correl<strong>at</strong>ion: Extreme Co-movements B<strong>et</strong>ween Financial Ass<strong>et</strong>s,<br />
Working paper, Columbia University, preprint <strong>at</strong> www.columbia.edu\ ∼rm586<br />
Mossin, J., 1966, Equilibrium in capital ass<strong>et</strong> mark<strong>et</strong>, Econom<strong>et</strong>rica 35, 768–783.<br />
Nelsen, R.B., 1998, An Introduction to Copulas. Lectures Notes in st<strong>at</strong>istic, 139 (Springer Verlag, New<br />
York).<br />
Rao, C.R., 1973, Linear st<strong>at</strong>istical inference and its applic<strong>at</strong>ions, 2d ed. (New York Willey).<br />
Rockafeller, R.T and S. Uryasev, 2000, Optimiz<strong>at</strong>ion of the conditional value-<strong>at</strong>-risk, journal of Risk 2,<br />
21-41.<br />
Sharpe, W., 1964, Capital ass<strong>et</strong>s prices: a theory of mark<strong>et</strong> equilibrium un<strong>de</strong>r conditions of risk, Journal of<br />
Finance 19, 425–442.<br />
Sorn<strong>et</strong>te, D., 1998, Large <strong>de</strong>vi<strong>at</strong>ions and portfolio optimiz<strong>at</strong>ion, Physica A 256, 251-283.<br />
Sorn<strong>et</strong>te, D., 2000, Critical Phenomena in N<strong>at</strong>ural Sciences, Chaos, Fractals, Self-organiz<strong>at</strong>ion and Disor<strong>de</strong>r:<br />
Concepts and Tools (Springer Series in Synerg<strong>et</strong>ics).<br />
Sorn<strong>et</strong>te, D., J. V. An<strong>de</strong>rsen and P. Simon<strong>et</strong>ti, 2000a, Portfolio Theory for “F<strong>at</strong> Tails”, Intern<strong>at</strong>ional Journal<br />
of Theor<strong>et</strong>ical and Applied Finance 3, 523-535.<br />
Sorn<strong>et</strong>te, D., P. Simon<strong>et</strong>ti, J.V. An<strong>de</strong>rsen, 2000b, φ q -field theory for portfolio optimiz<strong>at</strong>ion : ”f<strong>at</strong>-tails” and<br />
non-linear correl<strong>at</strong>ions, Physics Reports 335, 19-92.<br />
Szergö, G., 1999, A critique to Basel regul<strong>at</strong>ion, or how to enhance (im)moral hazards, Proceedings of the<br />
Intern<strong>at</strong>ional Conference on Risk Management and Regul<strong>at</strong>ion in Banking, Bank of Israel, Kluwer.<br />
Tasche, D. and L. Tilib<strong>et</strong>ti, 2001, Approxim<strong>at</strong>ions for the Value-<strong>at</strong>-Risk approach to risk-r<strong>et</strong>urn, Working<br />
Paper.<br />
31<br />
419
420 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique<br />
Gaussian R<strong>et</strong>urns<br />
10 1<br />
10 0<br />
10 −1<br />
10 −4<br />
10 −3<br />
Standard & Poor’s 500<br />
Raw R<strong>et</strong>urns<br />
c/2 = 0.73<br />
Figure 1: Graph of the Gaussianized Standard & Poor’s 500 in<strong>de</strong>x r<strong>et</strong>urns versus its raw r<strong>et</strong>urns, during the<br />
time interval from January 03, 1995 to December 29, 2000 for the neg<strong>at</strong>ive tail of the distribution.<br />
32<br />
10 −2<br />
10 −1
ln(T/T 0 )<br />
<br />
T<br />
Figure 2: Logarithm ln of the r<strong>at</strong>io of the recurrence time T to a reference time T0 for the recurrence<br />
T0<br />
of a given loss V aR as a function of β <strong>de</strong>fined by β = VaR<br />
VaR∗ . VaR∗ (resp. VaR) is the Value-<strong>at</strong>-Risk over a<br />
time interval T0 (resp. T ).<br />
α<br />
33<br />
c>1<br />
c=1<br />
c
422 13. Gestion <strong>de</strong> <strong>portefeuille</strong> sous contraintes <strong>de</strong> capital économique
Chapitre 14<br />
Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong><br />
équilibre <strong>de</strong> marché<br />
Nous présentons un nouvel ensemble <strong>de</strong> mesures <strong>de</strong> risque qualifiées <strong>de</strong> consistantes, en termes <strong>de</strong>s<br />
semi-invariants <strong>de</strong> la distribution <strong>de</strong> ren<strong>de</strong>ment d’un <strong>portefeuille</strong>, tels que les moments centrés ou les<br />
cumulants, <strong>et</strong> qui accor<strong>de</strong>nt plus <strong>de</strong> poids aux événements rares. Nous dérivons les frontières efficientes<br />
basées sur c<strong>et</strong> ensemble <strong>de</strong> mesures <strong>de</strong> risques <strong>et</strong> présentons une version généralisée du CAPM dans le<br />
cas où le marché est composé d’agents formant <strong>de</strong>s anticip<strong>at</strong>ions homogènes ou hétérogènes.<br />
Utilisant alors une famille <strong>de</strong> distributions <strong>de</strong> Weibull modifiées, qui englobe aussi bien <strong>de</strong>s lois superexponentielles<br />
que sous-exponentielles, pour paramétrer les distributions marginales du ren<strong>de</strong>ment <strong>de</strong>s<br />
actifs <strong>et</strong> une copule gaussienne pour représenter la dépendance entre les actifs, nous dérivons les expressions<br />
analytiques <strong>de</strong>s moments <strong>et</strong> cumulants <strong>de</strong> la distribution <strong>de</strong> ren<strong>de</strong>ment du <strong>portefeuille</strong> pour<br />
une composition arbitraire <strong>de</strong> ses actifs. Faisant usage <strong>de</strong> fonctions hypergéométriques, nous sommes en<br />
particulier capables d’étendre certains résult<strong>at</strong>s antérieurs au cas où l’exposant <strong>de</strong> Weibull est différent<br />
d’un actif à l’autre.<br />
A l’ai<strong>de</strong> <strong>de</strong> c<strong>et</strong>te représent<strong>at</strong>ion paramétrique, nous traitons en détail le problème <strong>de</strong> la minimis<strong>at</strong>ion<br />
du risque mesuré par les cumulants pour un <strong>portefeuille</strong> <strong>de</strong> <strong>de</strong>ux actifs <strong>et</strong> comparons nos prédictions<br />
théoriques aux estim<strong>at</strong>ions empiriques directes. Ces formules nous perm<strong>et</strong>tent en outre <strong>de</strong> déterminer<br />
analytiquement dans quelles conditions il est possible <strong>de</strong> construire un <strong>portefeuille</strong> présentant à la fois<br />
un ren<strong>de</strong>ment plus élevé <strong>et</strong> moins <strong>de</strong> grands risques.<br />
423
424 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché
Multi-Moments M<strong>et</strong>hod for Portfolio Management:<br />
Generalized Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>l<br />
in Homogeneous and H<strong>et</strong>erogeneous mark<strong>et</strong>s ∗<br />
Y. Malevergne 1,2 and D. Sorn<strong>et</strong>te 1,3<br />
1 Labor<strong>at</strong>oire <strong>de</strong> Physique <strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée<br />
CNRS UMR6622 and Université <strong>de</strong> Nice-Sophia Antipolis<br />
Parc Valrose, 06108 Nice Ce<strong>de</strong>x 2, France<br />
2 Institut <strong>de</strong> Science Financière <strong>et</strong> d’Assurances - Université Lyon I<br />
43, Bd du 11 Novembre 1918, 69622 Villeurbanne Ce<strong>de</strong>x<br />
3 Institute of Geophysics and Plan<strong>et</strong>ary Physics and Department of Earth and Space Science<br />
University of California, Los Angeles, California 90095<br />
e-mails: Yannick.Malevergne@unice.fr and sorn<strong>et</strong>te@unice.fr<br />
Abstract<br />
We use a new s<strong>et</strong> of consistent measures of risks, in terms of the semi-invariants of pdf’s, such th<strong>at</strong> the<br />
centered moments and the cumulants of the portfolio distribution of r<strong>et</strong>urns th<strong>at</strong> put more emphasis on<br />
the tail the distributions. We <strong>de</strong>rive generalized efficient frontiers, based on these novel measures of<br />
risks and present the generalized CAPM, both in the cases of homogeneous and h<strong>et</strong>erogeneous mark<strong>et</strong>s.<br />
Then, using a family of modified Weibull distributions, encompassing both sub-exponentials and superexponentials,<br />
to param<strong>et</strong>erize the marginal distributions of ass<strong>et</strong> r<strong>et</strong>urns and their n<strong>at</strong>ural multivari<strong>at</strong>e<br />
generaliz<strong>at</strong>ions, we offer exact formulas for the moments and cumulants of the distribution of r<strong>et</strong>urns of<br />
a portfolio ma<strong>de</strong> of an arbitrary composition of these ass<strong>et</strong>s. Using combin<strong>at</strong>orial and hypergeom<strong>et</strong>ric<br />
functions, we are in particular able to extend previous results to the case where the exponents of the<br />
Weibull distributions are different from ass<strong>et</strong> to ass<strong>et</strong> and in the presence of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s.<br />
In this param<strong>et</strong>eriz<strong>at</strong>ion, we tre<strong>at</strong> in d<strong>et</strong>ails the problem of risk minimiz<strong>at</strong>ion using the cumulants as<br />
measures of risks for a portfolio ma<strong>de</strong> of two ass<strong>et</strong>s and compare the theor<strong>et</strong>ical predictions with direct<br />
empirical d<strong>at</strong>a. Our exten<strong>de</strong>d formulas enable us to d<strong>et</strong>ermine analytically the conditions un<strong>de</strong>r which<br />
it is possible to “have your cake and e<strong>at</strong> it too”, i.e., to construct a portfolio with both larger r<strong>et</strong>urn and<br />
smaller “large risks”.<br />
1 Introduction<br />
The Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>l (CAPM) is still the most wi<strong>de</strong>ly used approach to rel<strong>at</strong>ive ass<strong>et</strong> evalu<strong>at</strong>ion,<br />
although its empirical roots are been found weaker and weaker in recent years. This ass<strong>et</strong> valu<strong>at</strong>ion mo<strong>de</strong>l<br />
<strong>de</strong>scribing the rel<strong>at</strong>ionship b<strong>et</strong>ween expected risk and expected r<strong>et</strong>urn for mark<strong>et</strong>able ass<strong>et</strong>s is strongly<br />
∗ We acknowledge helpful discussions and exchanges with J.V. An<strong>de</strong>rsen, J.P. Laurent and V. Pisarenko. We are gr<strong>at</strong>eful to<br />
participants of the workshop on “Multi-moment Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>ls and Rel<strong>at</strong>ed Topics”, ESCP-EAP European School<br />
of Management, Paris, April,19, 2002, and in particular to Philippe Spieser, for their comments. This work was partially supported<br />
by the James S. Mc Donnell Found<strong>at</strong>ion 21st century scientist award/studying complex system.<br />
1<br />
425
426 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
entangled with the Mean-Variance Portfolio Mo<strong>de</strong>l. In<strong>de</strong>ed both of them fundamentally rely on the <strong>de</strong>scription<br />
of the probability distribution function (pdf) of ass<strong>et</strong> r<strong>et</strong>urns in terms of Gaussian functions. The<br />
Mean-Variance <strong>de</strong>scription is thus <strong>at</strong> the basis of Markovitz’s portfolio theory (Markovitz 1959) and of the<br />
CAPM (see for instance (Merton 1990)).<br />
Otherwise, the d<strong>et</strong>ermin<strong>at</strong>ion of the risks and r<strong>et</strong>urns associ<strong>at</strong>ed with a given portfolio constituted of N<br />
ass<strong>et</strong>s is compl<strong>et</strong>ely embed<strong>de</strong>d in the knowledge of their multivari<strong>at</strong>e distribution of r<strong>et</strong>urns. In<strong>de</strong>ed, the<br />
<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween random variables is compl<strong>et</strong>ely <strong>de</strong>scribed by their joint distribution. This remark<br />
entails the two major problems of portfolio theory: 1) d<strong>et</strong>ermine the multivari<strong>at</strong>e distribution function of<br />
ass<strong>et</strong> r<strong>et</strong>urns; 2) <strong>de</strong>rive from it useful measures of portfolio risks and use them to analyze and optimize<br />
portfolios.<br />
The variance (or vol<strong>at</strong>ility) of portfolio r<strong>et</strong>urns provi<strong>de</strong>s the simplest way to quantify its fluctu<strong>at</strong>ions and<br />
is <strong>at</strong> the fund<strong>at</strong>ion of the (Markovitz 1959)’s portfolio selection theory. Non<strong>et</strong>heless, the variance of a<br />
portfolio offers only a limited quantific<strong>at</strong>ion of incurred risks (in terms of fluctu<strong>at</strong>ions), as the empirical<br />
distributions of r<strong>et</strong>urns have “f<strong>at</strong> tails” (Lux 1996, Gopikrishnan <strong>et</strong> al. 1998, among many others) and the<br />
<strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween ass<strong>et</strong>s are only imperfectly accounted for by the covariance m<strong>at</strong>rix (Litterman and<br />
Winkelmann 1998). It is thus essential to extend portfolio theory and the CAPM to tackle these empirical<br />
facts.<br />
The Value-<strong>at</strong>-Risk (Jorion 1997) and many other measures of risks (Artzner <strong>et</strong> al. 1997, Sorn<strong>et</strong>te 1998,<br />
Artzner <strong>et</strong> al. 1999, Bouchaud <strong>et</strong> al. 1998, Sorn<strong>et</strong>te <strong>et</strong> al. 2000b) have then been <strong>de</strong>veloped to account for the<br />
larger moves allowed by non-Gaussian distributions and non-linear correl<strong>at</strong>ions but they mainly allow for the<br />
assessment of down-si<strong>de</strong> risks. Here, we consi<strong>de</strong>r both-si<strong>de</strong> risk and <strong>de</strong>fine general measures of fluctu<strong>at</strong>ions.<br />
It is the first goal of this article. In<strong>de</strong>ed, consi<strong>de</strong>ring the minimum s<strong>et</strong> of properties a fluctu<strong>at</strong>ion measure<br />
must fulfil, we characterize these measures. In particular, we show th<strong>at</strong> any absolute central moments and<br />
some cumulants s<strong>at</strong>isfy these requirement as well as do any combin<strong>at</strong>ion of these quantities. Moreover,<br />
the weights involved in these combin<strong>at</strong>ions can be interpr<strong>et</strong>ed in terms of the portfolio manager’s aversion<br />
against large fluctu<strong>at</strong>ions.<br />
Once the <strong>de</strong>finition of the fluctu<strong>at</strong>ion measures have been s<strong>et</strong>, it is possible to classify the ass<strong>et</strong>s and portfolios<br />
using for instance a risk adjustment m<strong>et</strong>hod (Sharpe 1994, Dowd 2000) and to <strong>de</strong>velop a portfolio<br />
selection and optimiz<strong>at</strong>ion approach. It is the second goal of this article.<br />
Then a new mo<strong>de</strong>l of mark<strong>et</strong> equilibrium can be <strong>de</strong>rived, which generalizes the usual Capital Ass<strong>et</strong> Pricing<br />
Mo<strong>de</strong>l (CAPM). This is the third goal of our paper. This improvement is necessary since, although the use<br />
of the CAPM is still wi<strong>de</strong>ly spread, its empirical justific<strong>at</strong>ion has been found less and less convincing in the<br />
past years (Lim 1989, Harvey and Siddique 2000).<br />
The last goal of this article is to present an efficient param<strong>et</strong>ric m<strong>et</strong>hod allowing for the estim<strong>at</strong>ion of the<br />
centered moments and cumulants, based upon a maximum entropy principle. This param<strong>et</strong>eriz<strong>at</strong>ion of<br />
the problem is necessary in or<strong>de</strong>r to obtain accur<strong>at</strong>e estim<strong>at</strong>es of the high or<strong>de</strong>r moment-based quantities<br />
involved the portfolio optimiz<strong>at</strong>ion problem with our generalized measures of fluctu<strong>at</strong>ions.<br />
The paper is organized as follows.<br />
Section 2 presents a new s<strong>et</strong> of consistent measures of risks, in terms of the semi-invariants of pdf’s, such as<br />
the centered moments and the cumulants of the portfolio distribution of r<strong>et</strong>urns, for example.<br />
Section 3 <strong>de</strong>rives the generalized efficient frontiers, based on these novel measures of risks. Both cases with<br />
and without risk-free ass<strong>et</strong> are analyzed.<br />
Section 4 offers a generaliz<strong>at</strong>ion of the Sharpe r<strong>at</strong>io and thus provi<strong>de</strong>s new tools to classify ass<strong>et</strong>s with<br />
respect to their risk adjusted performance. In particular, we show th<strong>at</strong> this classific<strong>at</strong>ion may <strong>de</strong>pend on the<br />
2
choosen risk measure.<br />
Section 5 presents the generalized CAPM based on these new measures of risks, both in the cases of homogeneous<br />
and h<strong>et</strong>erogeneous agents.<br />
Section 6 introduces a novel general param<strong>et</strong>eriz<strong>at</strong>ion of the multivari<strong>at</strong>e distribution of r<strong>et</strong>urns based on two<br />
steps: (i) the projection of the empirical marginal distributions onto Gaussian laws via nonlinear mappings;<br />
(ii) the use of an entropy maximiz<strong>at</strong>ion to construct the corresponding most parsimonious represent<strong>at</strong>ion of<br />
the multivari<strong>at</strong>e distribution.<br />
Section 7 offers a specific param<strong>et</strong>eriz<strong>at</strong>ion of marginal distributions in terms of so-called modified Weibull<br />
distributions, which are essentially exponential of minus a power law. Notwithstanding their possible f<strong>at</strong>-tail<br />
n<strong>at</strong>ure, all their moments and cumulants are finite and can be calcul<strong>at</strong>ed. We present empirical calibr<strong>at</strong>ion<br />
of the two key param<strong>et</strong>ers of the modified Weibull distribution, namely the exponent c and the characteristic<br />
scale χ.<br />
Section 8 provi<strong>de</strong>s the analytical expressions of the cumulants of the distribution of portfolio r<strong>et</strong>urns for the<br />
param<strong>et</strong>eriz<strong>at</strong>ion of marginal distributions in terms of so-called modified Weibull distributions, introduced<br />
in section 6. Empirical tests comparing the direct numerical evalu<strong>at</strong>ion of the cumulants of financial time<br />
series to the values predicted from our analytical formulas find a good consistency.<br />
Section 9 uses these two s<strong>et</strong>s of results to illustr<strong>at</strong>e how portfolio optimiz<strong>at</strong>ion works in this context. The<br />
main novel result is an analytical un<strong>de</strong>rstanding of the conditions un<strong>de</strong>r which it is possible to simultaneously<br />
increase the portfolio r<strong>et</strong>urn and <strong>de</strong>creases its large risks quantified by large-or<strong>de</strong>r cumulants. It thus appears<br />
th<strong>at</strong> the multidimensional n<strong>at</strong>ure of risks allows one to break the stalem<strong>at</strong>e of no b<strong>et</strong>ter r<strong>et</strong>urn without more<br />
risks, for some special kind of r<strong>at</strong>ional agents.<br />
Section 10 conclu<strong>de</strong>s.<br />
Before proceeding with the present<strong>at</strong>ion of our results, we s<strong>et</strong> the not<strong>at</strong>ions to <strong>de</strong>rive the basic problem<br />
addressed in this paper, namely to study the distribution of the sum of weighted random variables with<br />
arbitrary marginal distributions and <strong>de</strong>pen<strong>de</strong>nce. Consi<strong>de</strong>r a portfolio with ni shares of ass<strong>et</strong> i of price pi(0)<br />
<strong>at</strong> time t = 0 whose initial wealth is<br />
N<br />
W (0) = nipi(0) . (1)<br />
A time τ l<strong>at</strong>er, the wealth has become W (τ) = N<br />
i=1 nipi(τ) and the wealth vari<strong>at</strong>ion is<br />
where<br />
δτ W ≡ W (τ) − W (0) =<br />
N<br />
i=1<br />
wi =<br />
i=1<br />
nipi(0) pi(τ) − pi(0)<br />
pi(0)<br />
nipi(0)<br />
N<br />
j=1 njpj(0)<br />
= W (0)<br />
427<br />
N<br />
wiri(t, τ), (2)<br />
is the fraction in capital invested in the ith ass<strong>et</strong> <strong>at</strong> time 0 and the r<strong>et</strong>urn ri(t, τ) b<strong>et</strong>ween time t − τ and t of<br />
ass<strong>et</strong> i is <strong>de</strong>fined as:<br />
ri(t, τ) = pi(t) − pi(t − τ)<br />
.<br />
pi(t − τ)<br />
(4)<br />
Using the <strong>de</strong>finition (4), this justifies us to write the r<strong>et</strong>urn Sτ of the portfolio over a time interval τ as the<br />
weighted sum of the r<strong>et</strong>urns ri(τ) of the ass<strong>et</strong>s i = 1, ..., N over the time interval τ<br />
Sτ = δτ W<br />
W (0) =<br />
3<br />
i=1<br />
(3)<br />
N<br />
wi ri(τ) . (5)<br />
i=1
428 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
In the sequel, we shall thus consi<strong>de</strong>r ass<strong>et</strong> r<strong>et</strong>urns as the fundamental variables (<strong>de</strong>noted xi or Xi in the<br />
sequel) and study their aggreg<strong>at</strong>ion properties, namely how the distribution of portfolio r<strong>et</strong>urn equal to their<br />
weighted sum <strong>de</strong>rives for their multivariable distribution. We shall consi<strong>de</strong>r a single time scale τ which<br />
can be chosen arbitrarily, say equal to one day. We shall thus drop the <strong>de</strong>pen<strong>de</strong>nce on τ, un<strong>de</strong>rstanding<br />
implicitely th<strong>at</strong> all our results hold for r<strong>et</strong>urns estim<strong>at</strong>ed over the time step τ.<br />
2 Measuring large risks of a portfolio<br />
The question on how to assess risk is recurrent in finance (and in many other fields) and has not y<strong>et</strong> received<br />
a general solution. Since the middle of the twenti<strong>et</strong>h century, several p<strong>at</strong>hs have been explored. The pioneering<br />
work by (Von Neuman and Morgenstern 1947) has given birth to the m<strong>at</strong>hem<strong>at</strong>ical <strong>de</strong>finition of the<br />
expected utility function which provi<strong>de</strong>s interesting insights on the behavior of a r<strong>at</strong>ional economic agent<br />
and formalized the concept of risk aversion. Based upon the properties of the utility function, (Rothschild<br />
and Stiglitz 1970) and (Rothschild and Stiglitz 1971) have <strong>at</strong>tempted to <strong>de</strong>fine the notion of increasing risks.<br />
But, as revealed by (Allais 1953, Allais 1990), empiric investig<strong>at</strong>ions has proven th<strong>at</strong> the postul<strong>at</strong>es chosen<br />
by (Von Neuman and Morgenstern 1947) are actually often viol<strong>at</strong>ed. Many generaliz<strong>at</strong>ions have been<br />
proposed for curing the so-called Allais’ Paradox, but up to now, no generally accepted procedure has been<br />
found in this way.<br />
Recently, a theory due to (Artzner <strong>et</strong> al. 1997, Artzner <strong>et</strong> al. 1999) and its generaliz<strong>at</strong>ion by (Föllmer and<br />
Schied 2002a, Föllmer and Schied 2002b), have appeared. Based on a series of postul<strong>at</strong>es th<strong>at</strong> are quite<br />
n<strong>at</strong>ural, this theory allows one to build coherent (convex) measures of risks. In fact, this theory seems welladapted<br />
to the assessment of the nee<strong>de</strong>d economic capital, th<strong>at</strong> is, of the fraction of capital a company must<br />
keep as risk-free ass<strong>et</strong>s in or<strong>de</strong>r to face its commitments and thus avoid ruin. However, for the purpose of<br />
quantifying the fluctu<strong>at</strong>ions of the ass<strong>et</strong> r<strong>et</strong>urns and of <strong>de</strong>veloping a theory of portfolios, this approach does<br />
not seem to be oper<strong>at</strong>ional. Here, we shall r<strong>at</strong>her revisit (Markovitz 1959)’s approach to investig<strong>at</strong>e how<br />
its extension to higher-or<strong>de</strong>r moments or cumulants, and any combin<strong>at</strong>ion of these quantities, can be used<br />
oper<strong>at</strong>ionally to account for large risks.<br />
2.1 Why do higher moments allow to assess larger risks?<br />
In principle, the compl<strong>et</strong>e <strong>de</strong>scription of the fluctu<strong>at</strong>ions of an ass<strong>et</strong> <strong>at</strong> a given time scale is given by the<br />
knowledge of the probability distribution function (pdf) of its r<strong>et</strong>urns. The pdf encompasses all the risk<br />
dimensions associ<strong>at</strong>ed with this ass<strong>et</strong>. Unfortun<strong>at</strong>ely, it is impossible to classify or or<strong>de</strong>r the risks <strong>de</strong>scribed<br />
by the entire pdf, except in special cases where the concept of stochastic dominance applies. Therefore, the<br />
whole pdf can not provi<strong>de</strong> an a<strong>de</strong>qu<strong>at</strong>e measure of risk, embodied by a single variable. In or<strong>de</strong>r to perform<br />
a selection among a bask<strong>et</strong> of ass<strong>et</strong>s and construct optimal portfolios, one needs measures given as real<br />
numbers, not functions, which can be or<strong>de</strong>red according to the n<strong>at</strong>ural or<strong>de</strong>ring of real numbers on the line.<br />
In this vein, (Markovitz 1959) has proposed to summarize the risk of an ass<strong>et</strong> by the variance of its pdf of<br />
r<strong>et</strong>urns (or equivalently by the corresponding standard <strong>de</strong>vi<strong>at</strong>ion). It is clear th<strong>at</strong> this <strong>de</strong>scription of risks is<br />
fully s<strong>at</strong>isfying only for ass<strong>et</strong>s with Gaussian pdf’s. In any other case, the variance generally provi<strong>de</strong>s a very<br />
poor estim<strong>at</strong>e of the real risk. In<strong>de</strong>ed, it is a well-established empirical fact th<strong>at</strong> the pdf’s of ass<strong>et</strong> r<strong>et</strong>urns<br />
has f<strong>at</strong> tails (Lux 1996, Pagan 1996, Gopikrishnan <strong>et</strong> al. 1998), so th<strong>at</strong> the Gaussian approxim<strong>at</strong>ion un<strong>de</strong>restim<strong>at</strong>es<br />
significantly the large prices movements frequently observed on stock mark<strong>et</strong>s. Consequently, the<br />
variance can not be taken as a suitable measure of risks, since it only accounts for the smallest contributions<br />
to the fluctu<strong>at</strong>ions of the ass<strong>et</strong>s r<strong>et</strong>urns.<br />
4
The variance of the r<strong>et</strong>urn X of an ass<strong>et</strong> involves its second moment E[X2 <br />
] and, more precisely, is equal<br />
to its second centered moment (or moment about the mean) E (X − E[X]) 2<br />
. Thus, the weight of a given<br />
fluctu<strong>at</strong>ion X entering in the <strong>de</strong>finition of the variance of the r<strong>et</strong>urns is proportional to its square. Due to the<br />
<strong>de</strong>cay of the pdf of X for large X boun<strong>de</strong>d from above by ∼ 1/|X| 1+α with α > 2, the largest fluctu<strong>at</strong>ions<br />
do not contribute significantly to this expect<strong>at</strong>ion. To increase their contributions, and in this way to account<br />
for the largest fluctu<strong>at</strong>ions, it is n<strong>at</strong>ural to invoke higher or<strong>de</strong>r moments of or<strong>de</strong>r n > 2. The large n is, the<br />
larger is the contribution of the rare and large r<strong>et</strong>urns in the tail of the pdf. This phenomenon is <strong>de</strong>monstr<strong>at</strong>ed<br />
in figure 1, where we can observe the evolution of the quantity x n · P (x) for n = 1, 2 and 4, where P (x), in<br />
this example, is the standard exponential distribution e −x . The expect<strong>at</strong>ion E[X n ] is then simply represented<br />
geom<strong>et</strong>rically as equal to the area below the curve x n ·P (x). These curves provi<strong>de</strong> an intuitive illustr<strong>at</strong>ion of<br />
the fact th<strong>at</strong> the main contributions to the moment E[X n ] of or<strong>de</strong>r n come from values of X in the vicinity<br />
of the maximum of x n · P (x) which increases fast with the or<strong>de</strong>r n of the moment we consi<strong>de</strong>r, all the more<br />
so, the f<strong>at</strong>ter is the tail of the pdf of the r<strong>et</strong>urns X. For the exponential distribution chosen to construct figure<br />
1, the value of x corresponding to the maximum of x n · P (x) is exactly equal to n. Thus, increasing the<br />
or<strong>de</strong>r of the moment allows one to sample larger fluctu<strong>at</strong>ions of the ass<strong>et</strong> prices.<br />
2.2 Quantifying the fluctu<strong>at</strong>ions of an ass<strong>et</strong><br />
L<strong>et</strong> us now examine wh<strong>at</strong> should be the properties th<strong>at</strong> coherent measures of risks adapted to the portfolio<br />
problem must s<strong>at</strong>isfy in or<strong>de</strong>r to best quantify the ass<strong>et</strong> price fluctu<strong>at</strong>ions. L<strong>et</strong> us consi<strong>de</strong>r an ass<strong>et</strong> <strong>de</strong>noted<br />
X, and l<strong>et</strong> G be the s<strong>et</strong> of all the risky ass<strong>et</strong>s available on the mark<strong>et</strong>. Its profit and loss distribution is the<br />
distribution of δX = X(τ) − X(0), while the r<strong>et</strong>urn distribution is given by the distribution of X(τ)−X(0)<br />
X(0) .<br />
The risk measures will be <strong>de</strong>fined for the profit and loss distributions and then shown to be equivalent to<br />
another <strong>de</strong>finition applied to the r<strong>et</strong>urn distribution.<br />
Our first requirement is th<strong>at</strong> the risk measure ρ(·), which is a functional on G, should always remain positive<br />
AXIOM 1 ∀X ∈ G, ρ(δX) ≥ 0 ,<br />
where the equality holds if and only if X is certain. L<strong>et</strong> us now add to this ass<strong>et</strong> a given amount a invested<br />
in the risk free-ass<strong>et</strong> whose r<strong>et</strong>urn is µ0 (with therefore no randomness in its price trajectory) and <strong>de</strong>fine the<br />
new ass<strong>et</strong> Y = X +a. Since a is non-random, the fluctu<strong>at</strong>ions of X and Y are the same. Thus, it is <strong>de</strong>sirable<br />
th<strong>at</strong> ρ enjoys the property of transl<strong>at</strong>ional invariance, wh<strong>at</strong>ever the ass<strong>et</strong> X and the non-random coefficient<br />
a may be:<br />
AXIOM 2 ∀X ∈ G, ∀a ∈ R, ρ(δX + µ · a) = ρ(δX).<br />
We also require th<strong>at</strong> our risk measure increases with the quantity of ass<strong>et</strong>s held in the portfolio. A priori,<br />
one should expect th<strong>at</strong> the risk of a position is proportional to its size. In<strong>de</strong>ed, the fluctu<strong>at</strong>ions associ<strong>at</strong>ed<br />
with the variable 2 · X are n<strong>at</strong>urally twice larger as the fluctu<strong>at</strong>ions of X. This is true as long as we can<br />
consi<strong>de</strong>r th<strong>at</strong> a large position can be liquid<strong>at</strong>ed as easily as a smaller one. This is obviously not true, due<br />
to the limited liquidity of real mark<strong>et</strong>s. Thus, a large position in a given ass<strong>et</strong> is more risky than the sum<br />
of the risks associ<strong>at</strong>ed with the many smaller positions which add up to the large position. To account for<br />
this point, we assume th<strong>at</strong> ρ <strong>de</strong>pends on the size of the position in the same manner for all ass<strong>et</strong>s. This<br />
assumption is slightly restrictive but not unrealistic for companies with comparable properties in terms of<br />
mark<strong>et</strong> capitaliz<strong>at</strong>ion or sector of activity. This requirement reads<br />
AXIOM 3 ∀X ∈ G, ∀λ ∈ R+, ρ(λ · δX) = f(λ) · ρ(δX),<br />
5<br />
429
430 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
where the function f : R+ −→ R+ is increasing and convex to account for liquidity risk. In fact, it is<br />
straightforward to show 1 th<strong>at</strong> the only functions st<strong>at</strong>istying this axiom are the fonctions fα(λ) = λ α with<br />
α ≥ 1, so th<strong>at</strong> axiom 3 can be reformul<strong>at</strong>ed in terms of positive homogeneity of <strong>de</strong>gree α:<br />
AXIOM 4<br />
∀X ∈ G, ∀λ ∈ R+, ρ(λ · δX) = λ α · ρ(δX). (6)<br />
Note th<strong>at</strong> the case of liquid mark<strong>et</strong>s is recovered by α = 1 for which the risk is directly proportionnal to the<br />
size of the position.<br />
These axioms, which <strong>de</strong>fine our risk measures for profit and loss can easily be exten<strong>de</strong>d to the r<strong>et</strong>urns of<br />
the ass<strong>et</strong>s. In<strong>de</strong>ed, the r<strong>et</strong>urn is nothing but the profit and loss divi<strong>de</strong>d by the initial value X(0) of the ass<strong>et</strong>.<br />
One can thus easily check th<strong>at</strong> the risk <strong>de</strong>fined on the profit and loss distribution is X(0) α times the risk<br />
<strong>de</strong>fined on the r<strong>et</strong>urn distribution. In the sequel, we will only consi<strong>de</strong>r this l<strong>at</strong>er <strong>de</strong>finition, and, to simplify<br />
the not<strong>at</strong>ions since we will only consi<strong>de</strong>r the r<strong>et</strong>urns and not the profit and loss, the not<strong>at</strong>ion X will be used<br />
to <strong>de</strong>note the ass<strong>et</strong> and its r<strong>et</strong>urn as well.<br />
We can remark th<strong>at</strong> the risk measures ρ enjoying the two properties <strong>de</strong>fined by the axioms 2 and 4 are known<br />
as the semi-invariants of the distribution of the profit and loss / r<strong>et</strong>urns of X (see (Stuart and Ord 1994, p<br />
86-87)). Among the large familly of semi-invariants, we can cite the well-known centered moments and<br />
cumulants of X.<br />
2.3 Examples<br />
The s<strong>et</strong> of risk measures obeying axioms 1-4 is huge since it inclu<strong>de</strong>s all the homogeneous functionals of<br />
(X − E[X]), for instance. The centered moments (or moments about the mean) and the cumulants are two<br />
well-known classes of semi-invariants. Then, a given value of α can be seen as nothing but a specific choice<br />
of the or<strong>de</strong>r n of the centered moments or of the cumulants. In this case, our risk measure <strong>de</strong>fined via these<br />
semi-invariants fulfills the two following conditions:<br />
ρ(X + µ) = ρ(X), (7)<br />
ρ(λ · X) = λ n · ρ(X). (8)<br />
In or<strong>de</strong>r to s<strong>at</strong>isfy the positivity condition (axiom 1), we need to restrict the s<strong>et</strong> of values taken by n. By<br />
construction, the centered moments of even or<strong>de</strong>r are always positive while the odd or<strong>de</strong>r centered moments<br />
can be neg<strong>at</strong>ive. Thus, only the even or<strong>de</strong>r centered moments are acceptable risk measures. The situ<strong>at</strong>ion<br />
is not so clear for the cumulants, since the even or<strong>de</strong>r cumulants, as well as the odd or<strong>de</strong>r ones, can be<br />
neg<strong>at</strong>ive. In full generality, only the centered moments provi<strong>de</strong> reasonable risk measures s<strong>at</strong>ifying our<br />
axioms. However, for a large class of distributions, even or<strong>de</strong>r cumulants remain positive, especially for<br />
f<strong>at</strong> tail distributions (eventhough there are simple but somewh<strong>at</strong> artificial counter-examples). Therefore,<br />
cumulants of even or<strong>de</strong>r can be useful risk measures when restricted to these distributions.<br />
In<strong>de</strong>ed, the cumulants enjoy a property which can be consi<strong>de</strong>red as a n<strong>at</strong>ural requirement for a risk measure.<br />
It can be <strong>de</strong>sirable th<strong>at</strong> the risk associ<strong>at</strong>ed with a portfolio ma<strong>de</strong> of in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s is exactly the sum<br />
of the risk associ<strong>at</strong>ed with each individual ass<strong>et</strong>. Thus, given N in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s {X1, · · · , XN}, and the<br />
portfolio SN = X1 + · · · + XN, we wish to have<br />
ρn(SN) = ρn(X1) + · + ρn(XN) . (9)<br />
1 using the trick ρ(λ1λ2 · δX) = f(λ1) · ρ(λ2 · δX) = f(λ1) · f(λ2) · ρ(δX) = f(λ1 · λ2) · ρ(δX) leading to f(λ1 · λ2) =<br />
f(λ1) · f(λ2). The unique increasing convex solution of this functional equ<strong>at</strong>ion is fα(λ) = λ α with α ≥ 1.<br />
6
This property is verified for all cumulants while is not true for centered moments. In addition, as seen from<br />
their <strong>de</strong>finition in terms of the characteristic function (63), cumulants of or<strong>de</strong>r larger than 2 quantify <strong>de</strong>vi<strong>at</strong>ion<br />
from the Gaussian law, and thus large risks beyond the variance (equal to the second-or<strong>de</strong>r cumulant).<br />
Thus, centered moments of even or<strong>de</strong>rs possess all the minimal properties required for a suitable portfolio<br />
risk measure. Cumulants fulfill these requirement only for well behaved distributions, but have an additional<br />
advantage compared to the centered moments, th<strong>at</strong> is, they fulfill the condition (9). For these reasons, we<br />
shall consi<strong>de</strong>r below both the centered moments and the cumulants.<br />
In fact, we can be more general. In<strong>de</strong>ed, as we have written, the centered moments or the cumulants of or<strong>de</strong>r<br />
n are homogeneous functions of or<strong>de</strong>r n, and due to the positivity requirement, we have to restrict ourselves<br />
to even or<strong>de</strong>r centered moments and cumulants. Thus, only homogeneous functions of or<strong>de</strong>r 2n can be<br />
consi<strong>de</strong>red. Actually, this restrictive constraint can be relaxed by recalling th<strong>at</strong>, given any homogeneous<br />
function f(·) of or<strong>de</strong>r p, the function f(·) q is also homogeneous of or<strong>de</strong>r p · q. This allows us to <strong>de</strong>couple<br />
the or<strong>de</strong>r of the moments to consi<strong>de</strong>r, which quantifies the impact of the large fluctu<strong>at</strong>ions, from the influence<br />
of the size of the positions held, measured by the <strong>de</strong>gres of homogeneity of ρ. Thus, consi<strong>de</strong>ring any even<br />
or<strong>de</strong>r centered moments, we can build a risk measure ρ(X) = E (X − E[X]) 2n α/2n which account for<br />
the fluctu<strong>at</strong>ions measured by the centered moment of or<strong>de</strong>r 2n but with a <strong>de</strong>gree of homogeneity equal to α.<br />
A further generaliz<strong>at</strong>ion is possible to odd-or<strong>de</strong>r moments. In<strong>de</strong>ed, the absolute centered moments s<strong>at</strong>isfy<br />
our three axioms for any odd or even or<strong>de</strong>r. We can go one step further and use non-integer or<strong>de</strong>r absolute<br />
centered moments, and <strong>de</strong>fine the more general risk measure<br />
where γ <strong>de</strong>notes any positve real number.<br />
431<br />
ρ(X) = E [|X − E[X]| γ ] α/γ , (10)<br />
These s<strong>et</strong> of risk measures are very interesting since, due to the Minkowsky inegality, they are convex for<br />
any α and γ larger than 1 :<br />
ρ(u · X + (1 − u) · Y ) ≤ u · ρ(X) + (1 − u) · ρ(Y ), (11)<br />
which ensures th<strong>at</strong> aggreg<strong>at</strong>ing two risky ass<strong>et</strong>s lead to diversify their risk. In fact, in the special case γ = 1,<br />
these measures enjoy the stronger sub-additivity property.<br />
Finally, we should stress th<strong>at</strong> any discr<strong>et</strong>e or continuous (positive) sum of these risk measures, with the same<br />
<strong>de</strong>gree of homogeneity is again a risk measure. This allows us to <strong>de</strong>fine “spectral measures of fluctu<strong>at</strong>ions”<br />
in the same spirit as in (Acerbi 2002):<br />
<br />
ρ(X) = dγ φ(γ) E [(X − E[X]) γ ] α/γ , (12)<br />
where φ is a positive real valued function <strong>de</strong>fined on any subinterval of [1, ∞) such th<strong>at</strong> the integral in<br />
(12) remains finite. It is interesting to restrict oneself to the functions φ whose integral sums up to one:<br />
dγ φ(γ) = 1, which is always possible, up to a renormaliz<strong>at</strong>ion. In<strong>de</strong>ed, in such a case, φ(γ) represents<br />
the rel<strong>at</strong>ive weight <strong>at</strong>tributed to the fluctu<strong>at</strong>ions measured by a given moment or<strong>de</strong>r. Thus, the function φ<br />
can be consi<strong>de</strong>red as a measure of the risk aversion of the risk manager with respect to the large fluctu<strong>at</strong>ions.<br />
L<strong>et</strong> us stress th<strong>at</strong> the variance, originally used in (Markovitz 1959)’s portfolio theory, is nothing but the<br />
second centered moment, also equal to the second or<strong>de</strong>r cumulant (the three first cumulants and centered<br />
moments are equal). Therefore, a portfolio theory based on the centered moments or on the cumulants<br />
autom<strong>at</strong>ically contain (Markovitz 1959)’s theory as a special case, and thus offers a n<strong>at</strong>ural generaliz<strong>at</strong>ion<br />
emcompassing large risks of this masterpiece of the financial science. It also embodies several other generaliz<strong>at</strong>ions<br />
where homogeneous measures of risks are consi<strong>de</strong>red, a for instance in (Hwang and S<strong>at</strong>chell 1999).<br />
7
432 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
3 The generalized efficient frontier and some of its properties<br />
We now address the problem of the portfolio selection and optimiz<strong>at</strong>ion, based on the risk measures introduced<br />
in the previous section. As we have already seen, there is a large choice of relevant risk measures<br />
from which the portfolio manager is free to choose as a function of his own aversion to small versus large<br />
risks. A strong risk aversion to large risks will lead him to choose a risk measure which puts the emphasis<br />
on the large fluctu<strong>at</strong>ions. The simplest examples of such risk measures are provi<strong>de</strong>d by the high-or<strong>de</strong>r centered<br />
moments or cumulants. Obviously, the utility function of the fund manager plays a central role in his<br />
choice of the risk measure. The rel<strong>at</strong>ion b<strong>et</strong>ween the central moments and the utility function has already<br />
been un<strong>de</strong>rlined by several authors such as (Rubinstein 1973) or (Jurczenko and Maill<strong>et</strong> 2002), who have<br />
shown th<strong>at</strong> an economic agent with a quartic utility function is n<strong>at</strong>urally sensitive to the first four moments<br />
of his expected wealth distribution. But, as stressed before, we do not wish to consi<strong>de</strong>r the expected utility<br />
formalism since our goal, in this paper, is not to study the un<strong>de</strong>rlying behavior leading to the choice of any<br />
risk measure.<br />
The choice of the risk measure also <strong>de</strong>pends upon the time horizon of investment. In<strong>de</strong>ed, as the time<br />
scale increases, the distribution of ass<strong>et</strong> r<strong>et</strong>urns progressively converges to the Gaussian pdf, so th<strong>at</strong> only<br />
the variance remains relevant for very long term investment horizons. However, for shorter time horizons,<br />
say, for portfolio rebalanced <strong>at</strong> a weekly, daily or intra-day time scales, choosing a risk measure putting the<br />
emphasis on the large fluctu<strong>at</strong>ions, such as the centered moments µ6 or µ8 or the cumulants C6 or C8 (or of<br />
larger or<strong>de</strong>rs), may be necessary to account for the “wild” price fluctu<strong>at</strong>ions usually observed for such short<br />
time scales.<br />
Our present approach uses a single time scale over which the r<strong>et</strong>urns are estim<strong>at</strong>ed, and is thus restricted<br />
to portfolio selection with a fixed investment horizon. Extensions to a portofolio analysis and optimiz<strong>at</strong>ion<br />
in terms of high-or<strong>de</strong>r moments and cumulants performed simultaneously over different time scales can be<br />
found in (Muzy <strong>et</strong> al. 2001).<br />
3.1 Efficient frontier without risk-free ass<strong>et</strong><br />
L<strong>et</strong> us consi<strong>de</strong>r N risky ass<strong>et</strong>s, <strong>de</strong>noted by X1, · · · , XN. Our goal is to find the best possible alloc<strong>at</strong>ion, given<br />
a s<strong>et</strong> of constraints.The portfolio optimiz<strong>at</strong>ion generalizing the approach of (Sorn<strong>et</strong>te <strong>et</strong> al. 2000a, An<strong>de</strong>rsen<br />
and Sorn<strong>et</strong>te 2001) corresponds to accounting for large fluctu<strong>at</strong>ions of the ass<strong>et</strong>s through the risk measures<br />
introduced above in the presence of a constraint on the r<strong>et</strong>urn as well as the “no-short sells” constraint:<br />
⎧<br />
⎪⎨<br />
⎪⎩<br />
inf wi∈[0,1] ρα({wi})<br />
<br />
i≥1 wi = 1<br />
<br />
i≥1 wiµ(i) = µ ,<br />
wi ≥ 0, ∀i > 0,<br />
where wi is the weight of Xi and µ(i) its expected r<strong>et</strong>urn. In all the sequel, the subscript α in ρα will refer<br />
to the <strong>de</strong>gree of homogeneity of the risk measure.<br />
This problem cannot be solved analytically (except in the Markovitz’s case where the risk measure is given<br />
by the variance). We need to perform numerical calcul<strong>at</strong>ions to obtain the shape of the efficient frontier.<br />
Non<strong>et</strong>heless, when the ρα’s <strong>de</strong>notes the centered moments or any convex risk measure, we can assert th<strong>at</strong><br />
this optimiz<strong>at</strong>ion problem is a convex optimiz<strong>at</strong>ion problem and th<strong>at</strong> it admits one and only one solution<br />
which can be easily d<strong>et</strong>ermined by standard numerical relax<strong>at</strong>ion or gradient m<strong>et</strong>hods.<br />
As an example, we have represented In figure 2, the mean-ρα efficient frontier for a portfolio ma<strong>de</strong> of seventeen<br />
ass<strong>et</strong>s (see appendix A for d<strong>et</strong>ails) in the plane (µ, ρ 1/α<br />
α ), where ρα represents the centered moments<br />
8<br />
(13)
µn=α of or<strong>de</strong>r n = 2, 4, 6 and 8. The efficient frontier is concave, as expected from the n<strong>at</strong>ure of the optimiz<strong>at</strong>ion<br />
problem (13). For a given value of the expected r<strong>et</strong>urn µ, we observe th<strong>at</strong> the amount of risk<br />
measured by µ 1/n<br />
n increases with n, so th<strong>at</strong> there is an additional price to pay for earning more: not only<br />
the µ2-risk increases, as usual according to Markowitz’s theory, but the large risks increases faster, the more<br />
so, the larger n is. This means th<strong>at</strong>, in this example, the large risks increases more rapidly than the small<br />
risks, as the required r<strong>et</strong>urn increases. This is an important empirical result th<strong>at</strong> has obvious implic<strong>at</strong>ions for<br />
portfolio selection and risk assessment. For instance, l<strong>et</strong> us consi<strong>de</strong>r an efficient portfolio whose expected<br />
(daily) r<strong>et</strong>urn equals 0.12%, which gives an annualized r<strong>et</strong>urn equal to 30%. We can see in table 1 th<strong>at</strong> the<br />
typical fluctu<strong>at</strong>ions around the expected r<strong>et</strong>urn are about twice larger when measured by µ6 compared with<br />
µ2 and th<strong>at</strong> they are 1.5 larger when measured with µ8 compared with µ4.<br />
3.2 Efficient frontier with a risk-free ass<strong>et</strong><br />
L<strong>et</strong> us now assume the existence of a risk-free ass<strong>et</strong> X0. The optimiz<strong>at</strong>ion problem with the same s<strong>et</strong> of<br />
constraints as previoulsy can be written as:<br />
⎧<br />
⎪⎨<br />
⎪⎩<br />
inf wi∈[0,1] ρα({wi})<br />
<br />
i≥0 wi = 1<br />
<br />
i≥0 wiµ(i) = µ ,<br />
wi ≥ 0, ∀i > 0,<br />
This optimiz<strong>at</strong>ion problem can be solved exactly. In<strong>de</strong>ed, due to existence of a risk-free ass<strong>et</strong>, the normaliz<strong>at</strong>ion<br />
condition wi = 1 is not-constraining since one can always adjust, by lending or borrowing money,<br />
the fraction w0 to a value s<strong>at</strong>isfying the normaliz<strong>at</strong>ion condition. Thus, as shown in appendix B, the efficient<br />
frontier is a straight line in the plane (µ, ρα 1/α ), with positive slope and whose intercept is given by the<br />
value of the risk-free interest r<strong>at</strong>e:<br />
µ = µ0 + ξ · ρα 1/α , (15)<br />
where ξ is a coefficient given explicitely below. This result is very n<strong>at</strong>ural when ρα <strong>de</strong>notes the variance,<br />
since it is then nothing but (Markovitz 1959)’s result. But in addition, it shows th<strong>at</strong> the mean-variance result<br />
can be generalized to every mean-ρα optimal portfolios.<br />
We present in figure 3 the results given by numerical simul<strong>at</strong>ions. The s<strong>et</strong> of ass<strong>et</strong>s is the same as before and<br />
the risk-free interest r<strong>at</strong>e has been s<strong>et</strong> to 5% a year. The optimiz<strong>at</strong>ion procedure has been performed using<br />
a gen<strong>et</strong>ic algorithm on the risk measure given by the centered moments µ2, µ4, µ6 and µ8. As expected,<br />
we observe three increasing straight lines, whose slopes monotonically <strong>de</strong>cay with the or<strong>de</strong>r of the centered<br />
moment un<strong>de</strong>r consi<strong>de</strong>r<strong>at</strong>ion. Below, we will discuss this property in gre<strong>at</strong>er d<strong>et</strong>ail.<br />
3.3 Two funds separ<strong>at</strong>ion theorem<br />
The two funds separ<strong>at</strong>ion theorem is a well-known result associ<strong>at</strong>ed with the mean-variance efficient portfolios.<br />
It results from the concavity of the Markovitz’s efficient frontier for portfolios ma<strong>de</strong> of risky ass<strong>et</strong>s<br />
only. It st<strong>at</strong>es th<strong>at</strong>, if the investors can choose b<strong>et</strong>ween a s<strong>et</strong> of risky ass<strong>et</strong>s and a risk-free ass<strong>et</strong>, they invest<br />
a fraction w0 of their wealth in the risk-free ass<strong>et</strong> and the fraction 1 − w0 in a portfolio composed only with<br />
risky ass<strong>et</strong>s. This risky portofolio is the same for all the investors and the fraction w0 of wealth invested<br />
in the risk-free ass<strong>et</strong> <strong>de</strong>pends on the risk aversion of the investor or on the amount of economic capital an<br />
institution must keep asi<strong>de</strong> due to the legal requirements insuring its solvency <strong>at</strong> a given confi<strong>de</strong>nce level.<br />
We shall see th<strong>at</strong> this result can be generalized to any mean-ρα efficient portfolio.<br />
9<br />
433<br />
(14)
434 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
In<strong>de</strong>ed, it can be shown (see appendix B) th<strong>at</strong> the weights of the optimal portfolios th<strong>at</strong> are solutions of (14)<br />
are given by:<br />
w ∗ 0 = w0, (16)<br />
w ∗ i = (1 − w0) · ˜wi, i ≥ 1, (17)<br />
where the ˜wi’s are constants such th<strong>at</strong> ˜wi = 1 and whose expressions are given appendix B. Thus,<br />
<strong>de</strong>noting by Π the portfolio only ma<strong>de</strong> of risky ass<strong>et</strong>s whose weights are the ˜wi’s, the optimal portfolios are<br />
the linear combin<strong>at</strong>ion of the risk-free ass<strong>et</strong>, with weight w0, and of the portfolio Π, with weigth 1 − w0.<br />
This result generalizes the mean-variance two fund theorem to any mean-ρα efficient portfolio.<br />
To check numerically this prediction, figure 4 represents the five largest weights of ass<strong>et</strong>s in the portfolios<br />
previously investig<strong>at</strong>ed as a function of the weight of the risk-free ass<strong>et</strong>, for the four risk measures given<br />
by the centered moments µ2, µ4, µ6 and µ8. One can observe <strong>de</strong>caying straight lines th<strong>at</strong> intercept the<br />
horizontal axis <strong>at</strong> w0 = 1, as predicted by equ<strong>at</strong>ions (16-17).<br />
In figure 2, the straight lines representing the efficient portfolios with a risk-free ass<strong>et</strong> are also represented.<br />
They are tangent to the efficient frontiers without risk-free ass<strong>et</strong>. This is n<strong>at</strong>ural since the efficient portfolios<br />
with the risk-free ass<strong>et</strong> are the weighted sum of the risk-free ass<strong>et</strong> and the optimal portfolio Π only ma<strong>de</strong><br />
of risky ass<strong>et</strong>s. Since Π also belongs to the efficient frontier without risk-free ass<strong>et</strong>, the optimum is reached<br />
when the straight line <strong>de</strong>scribing the efficient frontier with a risk-free ass<strong>et</strong> and the (concave) curve of the<br />
efficient frontier without risk-free ass<strong>et</strong> are tangent.<br />
3.4 Influence of the risk-free interest r<strong>at</strong>e<br />
Figure 3 has shown th<strong>at</strong> the slope of the efficient frontier (with a risk-free ass<strong>et</strong>) <strong>de</strong>creases when the or<strong>de</strong>r<br />
n of the centered moment used to measure risks increases. This is an important qualit<strong>at</strong>ive properties of the<br />
risk measures offered by the centered moments, as this means th<strong>at</strong> higher and higher large risks are sampled<br />
un<strong>de</strong>r increasing imposed r<strong>et</strong>urn.<br />
Is it possible th<strong>at</strong> the largest risks captured by the high-or<strong>de</strong>r centered moments could increase <strong>at</strong> a slower<br />
r<strong>at</strong>e than the small risks embodied in the small-or<strong>de</strong>r centered cumulants? For instance, is it possible for<br />
the slope of the mean-µ6 efficient frontier to be larger than the slope of the mean-µ4 frontier? This is an<br />
important question as it conditions the rel<strong>at</strong>ive costs in terms of the panel of risks un<strong>de</strong>r increasing specified<br />
r<strong>et</strong>urns. To address this question, consi<strong>de</strong>r figure 2. Changing the value of the risk-free interest r<strong>at</strong>e amounts<br />
to move the intercept of the straight lines along the ordin<strong>at</strong>e axis so as to keep them tangent to the efficient<br />
frontiers without risk-free ass<strong>et</strong>. Therefore, it is easy to see th<strong>at</strong>, in the situ<strong>at</strong>ion <strong>de</strong>picted in figure 2, the<br />
slope of the four straight lines will always <strong>de</strong>cay with the or<strong>de</strong>r of the centered moment.<br />
In or<strong>de</strong>r to observe an inversion in the or<strong>de</strong>r of the slopes, it is necessary and sufficient th<strong>at</strong> the efficient<br />
frontiers without risk-free ass<strong>et</strong> cross each other. This assertion is proved by visual inspection of figure<br />
5. Can we observe such crossing of efficient frontiers? In the most general case of risk measure, nothing<br />
forbids this occurence. Non<strong>et</strong>heless, we think th<strong>at</strong> this kind of behavior is not realistic in a financial context<br />
since, as said above, it would mean th<strong>at</strong> the large risks could increase <strong>at</strong> a slower r<strong>at</strong>e than the small risks,<br />
implying an irr<strong>at</strong>ional behavior of the economic agents.<br />
4 Classific<strong>at</strong>ion of the ass<strong>et</strong>s and of portfolios<br />
L<strong>et</strong> us consi<strong>de</strong>r two ass<strong>et</strong>s or portfolios X1 and X2 with different expected r<strong>et</strong>urns µ(1), µ(2) and different<br />
levels of risk measured by ρα(X1) and ρα(X2). An important question is then to be able to compare<br />
10
these two ass<strong>et</strong>s or portfolios. The most general way to perform such a comparison is to refer to <strong>de</strong>cision<br />
theory and to calcul<strong>at</strong>e the utility of each of them. But, as already said, the utility function of an agent is<br />
generally not known, so th<strong>at</strong> other approaches have to be <strong>de</strong>veloped. The simplest solution is to consi<strong>de</strong>r<br />
th<strong>at</strong> the couple (expected r<strong>et</strong>urn, risk measure) fully characterizes the behavior of the economic agent and<br />
thus provi<strong>de</strong>s a sufficiently good approxim<strong>at</strong>ion for her utility function.<br />
In the (Markovitz 1959)’s world for instance, the preferences of the agents are summarized by the two first<br />
moments of the distribution of ass<strong>et</strong>s r<strong>et</strong>urns. Thus, as shown by (Sharpe 1966, Sharpe 1994) a simple way<br />
to synth<strong>et</strong>ize these two param<strong>et</strong>ers, in or<strong>de</strong>r to g<strong>et</strong> a measure of the performance of the ass<strong>et</strong>s or portfolios,<br />
is to build the r<strong>at</strong>io of the expected r<strong>et</strong>urn µ (minus the risk free interest r<strong>at</strong>e) over the standard <strong>de</strong>vi<strong>at</strong>ion σ:<br />
S =<br />
435<br />
µ − µ0<br />
, (18)<br />
σ<br />
which is the so-called Sharpe r<strong>at</strong>io and simply represents the amount of expected r<strong>et</strong>urn per unit of risk,<br />
measured by the standard <strong>de</strong>vi<strong>at</strong>ion. It is an increasing function of the expected r<strong>et</strong>urn and a <strong>de</strong>creasing<br />
function of the level of risk, which is n<strong>at</strong>ural for risk-averse or pru<strong>de</strong>ntial agent.<br />
4.1 The risk-adjustment approach<br />
This approach can be generalized to any type of risk measures (see (Dowd 2000), for instance) and thus<br />
allows for the comparison of ass<strong>et</strong>s whose risks are not well accounted for by the variance (or the standard<br />
<strong>de</strong>vi<strong>at</strong>ion). In<strong>de</strong>ed, instead of consi<strong>de</strong>ring the variance, which only accounts for the small risks, one can build<br />
the r<strong>at</strong>io of the expected r<strong>et</strong>urn over any risk measure. In fact, looking <strong>at</strong> the equ<strong>at</strong>ion (113) in appendix B,<br />
the expression<br />
µ − µ0<br />
, (19)<br />
ρα(X) 1/α<br />
n<strong>at</strong>urally arises and is constant for every efficient portfolios. In this expression, α <strong>de</strong>notes the coefficient<br />
of homogeneity of the risk measure. It is nothing but a simple generalis<strong>at</strong>ion of the usual Sharpe r<strong>at</strong>io.<br />
In<strong>de</strong>ed, when ρα is given by the variance σ 2 , the expression above recovers the Sharpe r<strong>at</strong>io. Thus, once<br />
the portfolio manager has chosen his measure of fluctu<strong>at</strong>ions ρα, he can build a consistent risk-adjusted<br />
performance measure, as shown by (19).<br />
As just said, these generalized Sharpe r<strong>at</strong>ios are constant for every efficient portfolios. In fact, they are not<br />
only constant but also maximum for every efficient portfolios, so th<strong>at</strong> looking for the portfolio with maximum<br />
generalized Sharpe r<strong>at</strong>io yields the same optimal portfolios as those found with the whole optimiz<strong>at</strong>ion<br />
program solved in the previous section.<br />
As an illutr<strong>at</strong>ion, table 2 gives the risk-adjusted performance of the s<strong>et</strong> of seventeen ass<strong>et</strong>s already studied,<br />
for several risk measures. We have consi<strong>de</strong>red the three first even or<strong>de</strong>r centered moments (columns 2 to 4)<br />
and the three first even or<strong>de</strong>r cumulants (columns 2, 5 and 6) as fluctu<strong>at</strong>ion measures. Obviously the second<br />
or<strong>de</strong>r centered moment and the second or<strong>de</strong>r cumulant are the same, and give again the usual Sharpe r<strong>at</strong>io<br />
(18). The ass<strong>et</strong>s have been sorted with respect to their Sharpe R<strong>at</strong>io.<br />
The first point to note is th<strong>at</strong> the rank of an ass<strong>et</strong> in terms of risk-adjusted perfomance strongly <strong>de</strong>pends on<br />
the risk measure un<strong>de</strong>r consi<strong>de</strong>r<strong>at</strong>ion. The case of MCI Worldcom is very striking in this respect. In<strong>de</strong>ed,<br />
according to the usual Sharpe r<strong>at</strong>io, it appears in the 12 th position with a value larger than 0.04 while<br />
according to the other measures it is the last ass<strong>et</strong> of our selection with a value lower than 0.02.<br />
The second interesting point is th<strong>at</strong>, for a given ass<strong>et</strong>, the generalize Sharpe r<strong>at</strong>io is always a <strong>de</strong>creasing<br />
function of the or<strong>de</strong>r of the consi<strong>de</strong>red centered moment. This is not particular to our s<strong>et</strong> of ass<strong>et</strong>s since we<br />
11
436 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
can prove th<strong>at</strong><br />
so th<strong>at</strong><br />
∀p > q,<br />
(E [|X| p ]) 1/p ≥ (E [|X| q ]) 1/q , (20)<br />
µ − µ0<br />
(E [|X| p ≤<br />
1/p<br />
])<br />
µ − µ0<br />
(E [|X| q . (21)<br />
1/q<br />
])<br />
On the contrary, when the cumulants are used as risk measures, the generalized Sharpe r<strong>at</strong>ios are not monotonically<br />
<strong>de</strong>creasing, as exhibited by Procter & Gamble for instance. This can be surprising in view of our<br />
previous remark th<strong>at</strong> the larger is the or<strong>de</strong>r of the moments involved in a risk measure, the larger are the fluctu<strong>at</strong>ions<br />
it is accounting for. Extrapol<strong>at</strong>ing this property to cumulants, it would mean th<strong>at</strong> Procter & Gamble<br />
presents less large risks according to C6 than according to C4, while according to the centered moments, the<br />
reverse evolution is observed.<br />
Thus, the question of the coherence of the cumulants as measures of fluctu<strong>at</strong>ions may arise. And if we accept<br />
th<strong>at</strong> such measures are coherent, wh<strong>at</strong> are the implic<strong>at</strong>ions on the preferences of the agents employing such<br />
measures ? To answer this question, it is inform<strong>at</strong>ive to express the cumulants as a function of the moments.<br />
For instance, l<strong>et</strong> us consi<strong>de</strong>r the fourth or<strong>de</strong>r cumulant<br />
C4 = µ4 − 3 · µ2 2 , (22)<br />
= µ4 − 3 · C2 2 . (23)<br />
An agent assessing the fluctu<strong>at</strong>ions of an ass<strong>et</strong> with respect to C4 presents aversion for the fluctu<strong>at</strong>ions<br />
quantified by the fourth central moment µ4 – since C4 increases with µ4 – but is <strong>at</strong>tracted by the fluctu<strong>at</strong>ions<br />
measured by the variance - since C4 <strong>de</strong>creases with µ2. This behavior is not irr<strong>at</strong>ional since it remains<br />
globally risk-averse. In<strong>de</strong>ed, it <strong>de</strong>picts an agent which tries to avoid the larger risks but is ready to accept<br />
the smallest ones.<br />
This kind of behavior is characteristic of any agent using the cumulants as risk measures. It thus allows us to<br />
un<strong>de</strong>rstand why Procter & Gamble is more <strong>at</strong>tractive for an agent sentitive to C6 than for an agent sentitive<br />
to C4. From the expression of C6, we remark th<strong>at</strong> the agent sensitive to this cumulant is risk-averse with<br />
respect to the fluctu<strong>at</strong>ions mesured by µ6 and µ2 but is risk-seeker with respect to the fluctu<strong>at</strong>ions mesured<br />
by µ4 and µ3. Then, is this particular case, the l<strong>at</strong>er ones compens<strong>at</strong>e the former ones.<br />
It also allows us to un<strong>de</strong>rstand from a behavioral stand-point why it is possible to “have your cake and e<strong>at</strong><br />
it too” in the sense of (An<strong>de</strong>rsen and Sorn<strong>et</strong>te 2001), th<strong>at</strong> is, why, when the cumulants are choosen as risk<br />
measures, it may be possible to increase the expected r<strong>et</strong>urn of a portfolio while lowering its large risks, or in<br />
other words, why its generalized Sharpe r<strong>at</strong>io may increase when one consi<strong>de</strong>r larger cumulants to measure<br />
its risks. We will discuus this point again in section 9.<br />
4.2 Marginal risk of an ass<strong>et</strong> within a portofolio<br />
Another important question th<strong>at</strong> arises is the contribution of a given ass<strong>et</strong> to the risk of the whole portfolio.<br />
In<strong>de</strong>ed, it is crucial to know wh<strong>et</strong>her the risk is homogeneously shared by all the ass<strong>et</strong>s of the portfolio or if<br />
it is only held by a few of them. The quality of the diversific<strong>at</strong>ion is then <strong>at</strong> stake. Moreover, this also allows<br />
for the sensitivity analysis of the risk of the portfolio with respect to small changes in its composition 2 ,<br />
which is of practical interest since it can prevent us from recalcul<strong>at</strong>ing the whole risk of the portfolio after a<br />
small re-adjustment of its composition.<br />
2 see (Gouriéroux <strong>et</strong> al. 2000, Scaill<strong>et</strong> 2000) for a sensitivity analysis of the Value-<strong>at</strong>-Risk and the expected shortfall.<br />
12
Due to the homogeneity property of the fluctu<strong>at</strong>ion measures and to Euler’s theorem for homogeneous<br />
functions, we can write th<strong>at</strong><br />
ρ({w1, · · · , wN}) = 1<br />
N<br />
wi ·<br />
α<br />
∂ρ<br />
, (24)<br />
∂wi<br />
provi<strong>de</strong>d the risk measure ρ is differentiable which will be assumed in all the sequel. In this expression, the<br />
coefficient α again <strong>de</strong>notes the <strong>de</strong>gree of homogeneity of the risk measure ρ<br />
This rel<strong>at</strong>ion simply shows th<strong>at</strong> the amount of risk brought by one unit of the ass<strong>et</strong> i in the portfolio is given<br />
by the first <strong>de</strong>riv<strong>at</strong>ive of the risk of the portfolio with respect to the weight wi ot this ass<strong>et</strong>. Thus, α −1 · ∂ρ<br />
∂wi<br />
represents the marginal amount of risk of ass<strong>et</strong> i in the portfolio. It is then easy to check th<strong>at</strong>, in a portfolio<br />
with minimum risk, irrespective of the expected r<strong>et</strong>urn, the weight of each ass<strong>et</strong> is such th<strong>at</strong> the marginal<br />
risks of the ass<strong>et</strong>s in the portfolio are equal.<br />
5 A new equilibrum mo<strong>de</strong>l for ass<strong>et</strong> prices<br />
Using the portfolio selection m<strong>et</strong>hod explained in the two previous sections, we now present an equilibrium<br />
mo<strong>de</strong>l generalizing the original Capital Ass<strong>et</strong> Pricing Mo<strong>de</strong>l <strong>de</strong>veloped by (Sharpe 1964, Lintner 1965,<br />
Mossin 1966). Many generaliz<strong>at</strong>ions have already been proposed to account for the f<strong>at</strong>-tailness of the ass<strong>et</strong>s<br />
r<strong>et</strong>urn distributions, which led to the multi-moments CAPM. For instance (Rubinstein 1973) and (Krauss<br />
and Lintzenberger 1976) or (Lim 1989) and (Harvey and Siddique 2000) have un<strong>de</strong>rlined and tested the role<br />
of the asymm<strong>et</strong>ry in the risk premium by accounting for the skewness of the distribution of r<strong>et</strong>urns. More<br />
recently, (Fang and Lai 1997) and (Hwang and S<strong>at</strong>chell 1999) have introduced a four-moments CAPM to<br />
take into account the l<strong>et</strong>pokurtic behavior of the ass<strong>et</strong>s r<strong>et</strong>urn distributions. Many other extentions have<br />
been presented such as the VaR-CAPM (see (Alexan<strong>de</strong>r and Baptista 2002)) or the Distributional-CAPM<br />
by (Polimenis 2002). All these generaliz<strong>at</strong>ion become more and more complic<strong>at</strong>ed and not do not provi<strong>de</strong><br />
necessarily more accur<strong>at</strong>e prediction of the expected r<strong>et</strong>urns.<br />
Here, we will assume th<strong>at</strong> the relevant risk measure is given by any measure of fluctu<strong>at</strong>ions previously<br />
presented th<strong>at</strong> obey the axioms I-IV of section 2. We will also relax the usual assumption of an homogeneous<br />
mark<strong>et</strong> to give to the economic agents the choice of their own risk measure: some of them may choose a<br />
risk measure which put the emphasis on the small fluctu<strong>at</strong>ions while others may prefer those which account<br />
for the large ones. We will show th<strong>at</strong>, in such an h<strong>et</strong>erogeneous mark<strong>et</strong>, an equilibrium can still be reached<br />
and th<strong>at</strong> the excess r<strong>et</strong>urns of individual stocks remain proportional to the mark<strong>et</strong> excess r<strong>et</strong>urn.<br />
For this, we need the following assumptions about the mark<strong>et</strong>:<br />
• H1: We consi<strong>de</strong>r a one-period mark<strong>et</strong>, such th<strong>at</strong> all the positions held <strong>at</strong> the begining of a period are<br />
cleared <strong>at</strong> the end of the same period.<br />
• H2: The mark<strong>et</strong> is perfect, i.e., there are no transaction cost or taxes, the mark<strong>et</strong> is efficient and the<br />
investors can lend and borrow <strong>at</strong> the same risk-free r<strong>at</strong>e µ0.<br />
We will now add another assumption th<strong>at</strong> specifies the behavior of the agents acting on the mark<strong>et</strong>, which<br />
will lead us to make the distinction b<strong>et</strong>ween homogeneous and h<strong>et</strong>erogeneous mark<strong>et</strong>s.<br />
13<br />
i1<br />
437
438 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
5.1 Equilibrium in a homogeneous mark<strong>et</strong><br />
The mark<strong>et</strong> is said to be homogeneous if all the agents acting on this mark<strong>et</strong> aim <strong>at</strong> fulfilling the same<br />
objective. This means th<strong>at</strong>:<br />
• H3-1: all the agents want to maximize the expected r<strong>et</strong>urn of their portfolio <strong>at</strong> the end of the period<br />
un<strong>de</strong>r a given constraint of measured risk, using the same measure of risks ρα for all of them.<br />
In the special case where ρα <strong>de</strong>notes the variance, all the agents follow a Markovitz’s optimiz<strong>at</strong>ion procedure,<br />
which leads to the CAPM equilibrium, as proved by (Sharpe 1964). When ρα represents the centered<br />
moments, we will be led to the mark<strong>et</strong> equilibrium <strong>de</strong>scribed by (Rubinstein 1973). Thus, this approach<br />
allows for a generaliz<strong>at</strong>ion of the most popular ass<strong>et</strong> pricing in equilibirum mark<strong>et</strong> mo<strong>de</strong>ls.<br />
When all the agents have the same risk function ρα, wh<strong>at</strong>ever α may be, we can assert th<strong>at</strong> they have all a<br />
fraction of their capital invested in the same portfolio Π, whose composition is given in appendix B, and the<br />
remaining in the risk-free ass<strong>et</strong>. The amount of capital invested in the risky fund only <strong>de</strong>pends on their risk<br />
aversion or on the legal margin requirement they have to fulfil.<br />
L<strong>et</strong> us now assume th<strong>at</strong> the mark<strong>et</strong> is <strong>at</strong> equilibrium, i.e., supply equals <strong>de</strong>mand. In such a case, since the<br />
optimal portfolios can be any linear combin<strong>at</strong>ions of the risk-free ass<strong>et</strong> and of the risky portfolio Π, it is<br />
straightforward to show (see appendix C) th<strong>at</strong> the mark<strong>et</strong> portfolio, ma<strong>de</strong> of all tra<strong>de</strong>d ass<strong>et</strong>s in proportion<br />
of their mark<strong>et</strong> capitaliz<strong>at</strong>ion, is nothing but the risky portfolio Π. Thus, as shown in appendix D, we can<br />
st<strong>at</strong>e th<strong>at</strong>, wh<strong>at</strong>ever the risk measure ρα chosen by the agents to perform their optimiz<strong>at</strong>ion, the excess r<strong>et</strong>urn<br />
of any ass<strong>et</strong> over the risk-free interest r<strong>at</strong>e is proportional to the excess r<strong>et</strong>urn of the mark<strong>et</strong> portfolio Π over<br />
the risk-free interest r<strong>at</strong>e:<br />
µ(i) − µ0 = β i α · (µΠ − µ0), (25)<br />
where<br />
β i α = ·<br />
<br />
∂ ln<br />
1<br />
ρα α<br />
∂wi<br />
<br />
<br />
<br />
<br />
<br />
w ∗ 1 ,···,w ∗ N<br />
, (26)<br />
where w ∗ 1 , · · · , w∗ N are <strong>de</strong>fined in appendix D. When ρα <strong>de</strong>notes the variance, we recover the usual β i given<br />
by the mean-variance approach:<br />
β i = Cov(Xi, Π)<br />
. (27)<br />
Var(Π)<br />
Thus, the rel<strong>at</strong>ions (25) and (26) generalize the usual CAPM formula, showing th<strong>at</strong> the specific choice of<br />
the risk measure is not very important, as long as it follows the axioms I-IV characterizing the fluctu<strong>at</strong>ions<br />
of the distribution of ass<strong>et</strong> r<strong>et</strong>urns.<br />
5.2 Equilibrium in a h<strong>et</strong>erogeneous mark<strong>et</strong><br />
Does this result hold in the more realistic situ<strong>at</strong>ion of an h<strong>et</strong>erogeneous mark<strong>et</strong>? A mark<strong>et</strong> will be said to be<br />
h<strong>et</strong>erogeneous if the agents seek to fulfill different objectives. We thus consi<strong>de</strong>r the following assumption:<br />
• H3-2: There exists N agents. Each agent n is characterized by her choice of a risk measure ρα(n) so<br />
th<strong>at</strong> she invests only in the mean-ρα(n) efficient portfolios.<br />
According to this hypothesis, an agent n invests a fraction of her wealth in the risk-free ass<strong>et</strong> and the<br />
remaining in Πn, the mean-ρα(n) efficient portfolio, only ma<strong>de</strong> of risky ass<strong>et</strong>s. The fraction of wealth<br />
14
invested in the risky fund <strong>de</strong>pends on the risk aversion of each agents, which may vary from an agent to<br />
another one.<br />
The composition of the mark<strong>et</strong> portfolio for such a h<strong>et</strong>erogenous mark<strong>et</strong> is <strong>de</strong>rived in appendix C. We find<br />
th<strong>at</strong> the mark<strong>et</strong> portfolio Π is nothing but the weighted sum of the mean-ρα(n) optimal portfolio Πn:<br />
Π =<br />
439<br />
N<br />
γnΠn, (28)<br />
where γn is the fraction of the total wealth invested in the fund Πn by the n th agent.<br />
n=1<br />
Appendix D <strong>de</strong>monstr<strong>at</strong>es th<strong>at</strong>, for every ass<strong>et</strong> i and for any mean-ρα(n) efficient portfolio Πn, for all n, the<br />
following equ<strong>at</strong>ion holds<br />
µ(i) − µ0 = β i n · (µΠn − µ0) . (29)<br />
Multiplying these equ<strong>at</strong>ions by γn/β i n, we g<strong>et</strong><br />
γn<br />
βi n<br />
for all n, and summing over the different agents, we obtain<br />
<br />
<br />
γn<br />
<br />
· (µ(i) − µ0) =<br />
so th<strong>at</strong><br />
with<br />
n<br />
β i n<br />
· (µ(i) − µ0) = γn · (µΠn − µ0), (30)<br />
n<br />
γn · µΠn<br />
<br />
− µ0, (31)<br />
µ(i) − µ0 = β i · (µΠ − µ0), (32)<br />
β i =<br />
<br />
n<br />
γn<br />
βi n<br />
−1<br />
. (33)<br />
This allows us to conclu<strong>de</strong> th<strong>at</strong>, even in a h<strong>et</strong>erogeneous mark<strong>et</strong>, the expected excess r<strong>et</strong>urn of each individual<br />
stock is directly proportionnal to the expected excess r<strong>et</strong>urn of the mark<strong>et</strong> portfolio, showing th<strong>at</strong><br />
the homogeneity of the mark<strong>et</strong> is not a key property necessary for observing a linear rel<strong>at</strong>ionship b<strong>et</strong>ween<br />
individual excess ass<strong>et</strong> r<strong>et</strong>urns and the mark<strong>et</strong> excess r<strong>et</strong>urn.<br />
6 Estim<strong>at</strong>ion of the joint probability distribution of r<strong>et</strong>urns of several ass<strong>et</strong>s<br />
A priori, one of the main practical advantage of (Markovitz 1959)’s m<strong>et</strong>hod and its generaliz<strong>at</strong>ion presented<br />
above is th<strong>at</strong> one does not need the multivari<strong>at</strong>e probability distribution function of the ass<strong>et</strong>s r<strong>et</strong>urns, as the<br />
analysis solely relies on the coherent measures ρ(X) <strong>de</strong>fined in section 2, such as the centered moments<br />
or the cumulants of all or<strong>de</strong>rs th<strong>at</strong> can in principle be estim<strong>at</strong>ed empirically. Unfortun<strong>at</strong>ely, this apparent<br />
advantage maybe an illusion. In<strong>de</strong>ed, as un<strong>de</strong>rlined by (Stuart and Ord 1994) for instance, the error of<br />
the empirically estim<strong>at</strong>ed moment of or<strong>de</strong>r n is proportional to the moment of or<strong>de</strong>r 2n, so th<strong>at</strong> the error<br />
becomes quickly of the same or<strong>de</strong>r as the estim<strong>at</strong>ed moment itself. Thus, above n = 6 (or may be n = 8) it is<br />
not reasonable to estim<strong>at</strong>e the moments and/or cumulants directly. Thus, the knowledge of the multivari<strong>at</strong>e<br />
distribution of ass<strong>et</strong>s r<strong>et</strong>urns remains necessary. In addition, there is a current of thoughts th<strong>at</strong> provi<strong>de</strong>s<br />
evi<strong>de</strong>nce th<strong>at</strong> marginal distributions of r<strong>et</strong>urns may be regularly varying with in<strong>de</strong>x µ in the range 3-4<br />
(Lux 1996, Pagan 1996, Gopikrishnan <strong>et</strong> al. 1998), suggesting the non-existence of asymptotically <strong>de</strong>fined<br />
moments and cumulants of or<strong>de</strong>r equal to or larger than µ.<br />
15
440 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
In the standard Gaussian framework, the multivari<strong>at</strong>e distribution takes the form of an exponential of minus<br />
a quadr<strong>at</strong>ic form X ′ Ω −1 X, where X is the unicolumn of ass<strong>et</strong> r<strong>et</strong>urns and Ω is their covariance m<strong>at</strong>rix. The<br />
beauty and simplicity of the Gaussian case is th<strong>at</strong> the essentially impossible task of d<strong>et</strong>ermining a large multidimensional<br />
function is reduced into the very much simpler one of calcul<strong>at</strong>ing the N(N + 1)/2 elements<br />
of the symm<strong>et</strong>ric covariance m<strong>at</strong>rix. Risk is then uniquely and compl<strong>et</strong>ely embodied by the variance of the<br />
portfolio r<strong>et</strong>urn, which is easily d<strong>et</strong>ermined from the covariance m<strong>at</strong>rix. This is the basis of Markovitz’s<br />
portfolio theory (Markovitz 1959) and of the CAPM (see for instance (Merton 1990)).<br />
However, as is well-known, the variance (vol<strong>at</strong>ility) of portfolio r<strong>et</strong>urns provi<strong>de</strong>s <strong>at</strong> best a limited quantific<strong>at</strong>ion<br />
of incurred risks, as the empirical distributions of r<strong>et</strong>urns have “f<strong>at</strong> tails” (Lux 1996, Gopikrishnan <strong>et</strong><br />
al. 1998) and the <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween ass<strong>et</strong>s are only imperfectly accounted for by the covariance m<strong>at</strong>rix<br />
(Litterman and Winkelmann 1998).<br />
In this section, we present a novel approach based on (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b) to <strong>at</strong>tack this problem in terms<br />
of the param<strong>et</strong>eriz<strong>at</strong>ion of the multivari<strong>at</strong>e distribution of r<strong>et</strong>urns involving two steps: (i) the projection of<br />
the empirical marginal distributions onto Gaussian laws via nonlinear mappings; (ii) the use of an entropy<br />
maximiz<strong>at</strong>ion to construct the corresponding most parsimonious represent<strong>at</strong>ion of the multivari<strong>at</strong>e distribution.<br />
6.1 A brief exposition and justific<strong>at</strong>ion of the m<strong>et</strong>hod<br />
We will use the m<strong>et</strong>hod of d<strong>et</strong>ermin<strong>at</strong>ion of multivari<strong>at</strong>e distributions introduced by (Karlen 1998) and<br />
(Sorn<strong>et</strong>te <strong>et</strong> al. 2000b). This m<strong>et</strong>hod consists in two steps: (i) transform each r<strong>et</strong>urn x into a Gaussian<br />
variable y by a nonlinear monotonous increasing mapping; (ii) use the principle of entropy maximiz<strong>at</strong>ion to<br />
construct the corresponding multivari<strong>at</strong>e distribution of the transformed variables y.<br />
The first concern to address before going any further is wh<strong>et</strong>her the nonlinear transform<strong>at</strong>ion, which is in<br />
principle different for each ass<strong>et</strong> r<strong>et</strong>urn, conserves the structure of the <strong>de</strong>pen<strong>de</strong>nce. In wh<strong>at</strong> sense is the<br />
<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the transformed variables y the same as the <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the ass<strong>et</strong> r<strong>et</strong>urns x? It<br />
turns out th<strong>at</strong> the notion of “copulas” provi<strong>de</strong>s a general and rigorous answer which justifies the procedure<br />
of (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b).<br />
For compl<strong>et</strong>eness and use l<strong>at</strong>er on, we briefly recall the <strong>de</strong>finition of a copula (for further d<strong>et</strong>ails about<br />
the concept of copula see (Nelsen 1998)). A function C : [0, 1] n −→ [0, 1] is a n-copula if it enjoys the<br />
following properties :<br />
• ∀u ∈ [0, 1], C(1, · · · , 1, u, 1 · · · , 1) = u ,<br />
• ∀ui ∈ [0, 1], C(u1, · · · , un) = 0 if <strong>at</strong> least one of the ui equals zero ,<br />
• C is groun<strong>de</strong>d and n-increasing, i.e., the C-volume of every boxes whose vertices lie in [0, 1] n is<br />
positive.<br />
Skar’s Theorem then st<strong>at</strong>es th<strong>at</strong>, given an n-dimensional distribution function F with continuous marginal<br />
distributions F1, · · · , Fn, there exists a unique n-copula C : [0, 1] n −→ [0, 1] such th<strong>at</strong> :<br />
F (x1, · · · , xn) = C(F1(x1), · · · , Fn(xn)) . (34)<br />
This elegant result shows th<strong>at</strong> the study of the <strong>de</strong>pen<strong>de</strong>nce of random variables can be performed in<strong>de</strong>pen<strong>de</strong>ntly<br />
of the behavior of the marginal distributions. Moreover, the following result shows th<strong>at</strong> copulas<br />
are intrinsic measures of <strong>de</strong>pen<strong>de</strong>nce. Consi<strong>de</strong>r n continuous random variables X1, · · · , Xn with copula<br />
16
C. Then, if g1(X1), · · · , gn(Xn) are strictly increasing on the ranges of X1, · · · , Xn, the random variables<br />
Y1 = g1(X1), · · · , Yn = gn(Xn) have exactly the same copula C (Lindskog 2000). The copula is thus<br />
invariant un<strong>de</strong>r strictly increasing tranform<strong>at</strong>ion of the variables. This provi<strong>de</strong>s a powerful way of studying<br />
scale-invariant measures of associ<strong>at</strong>ions. It is also a n<strong>at</strong>ural starting point for construction of multivari<strong>at</strong>e<br />
distributions and provi<strong>de</strong>s the theor<strong>et</strong>ical justific<strong>at</strong>ion of the m<strong>et</strong>hod of d<strong>et</strong>ermin<strong>at</strong>ion of mutivari<strong>at</strong>e distributions<br />
th<strong>at</strong> we will use in the sequel.<br />
6.2 Transform<strong>at</strong>ion of an arbitrary random variable into a Gaussian variable<br />
L<strong>et</strong> us consi<strong>de</strong>r the r<strong>et</strong>urn X, taken as a random variable characterized by the probability <strong>de</strong>nsity p(x). The<br />
transform<strong>at</strong>ion y(x) which obtains a standard normal variable y from x is d<strong>et</strong>ermined by the conserv<strong>at</strong>ion<br />
of probability:<br />
Integr<strong>at</strong>ing this equ<strong>at</strong>ion from −∞ and x, we obtain:<br />
441<br />
p(x)dx = 1 y2<br />
− √ e 2 dy . (35)<br />
2π<br />
F (x) = 1<br />
2<br />
where F (x) is the cumul<strong>at</strong>ive distribution of X:<br />
F (x) =<br />
This leads to the following transform<strong>at</strong>ion y(x):<br />
<br />
1 + erf<br />
<br />
y√2<br />
, (36)<br />
x<br />
dx<br />
−∞<br />
′ p(x ′ ) . (37)<br />
y = √ 2 erf −1 (2F (x) − 1) , (38)<br />
which is obvously an increasing function of X as required for the applic<strong>at</strong>ion of the invariance property of<br />
the copula st<strong>at</strong>ed in the previous section. An illustr<strong>at</strong>ion of the nonlinear transform<strong>at</strong>ion (38) is shown in<br />
figure 6. Note th<strong>at</strong> it does not require any special hypothesis on the probability <strong>de</strong>nsity X, apart from being<br />
non-<strong>de</strong>gener<strong>at</strong>e.<br />
In the case where the pdf of X has only one maximum, we may use a simpler expression equivalent to (38).<br />
Such a pdf can be written un<strong>de</strong>r the so-called Von Mises param<strong>et</strong>riz<strong>at</strong>ion (Embrechts <strong>et</strong> al. 1997) :<br />
p(x) = C f ′ (x) 1<br />
− e 2<br />
|f(x)| f(x) , (39)<br />
where C is a constant of normaliz<strong>at</strong>ion. For f(x)/x 2 → 0 when |x| → +∞, the pdf has a “f<strong>at</strong> tail,” i.e., it<br />
<strong>de</strong>cays slower than a Gaussian <strong>at</strong> large |x|.<br />
L<strong>et</strong> us now <strong>de</strong>fine the change of variable<br />
Using the rel<strong>at</strong>ionship p(y) = p(x) dx<br />
dy , we g<strong>et</strong>:<br />
y = sgn(x) |f(x)| . (40)<br />
p(y) = 1 y2<br />
− √ e 2 . (41)<br />
2π<br />
It is important to stress the presence of the sign function sgn(x) in equ<strong>at</strong>ion (40), which is essential in or<strong>de</strong>r<br />
to correctly quantify <strong>de</strong>pen<strong>de</strong>nces b<strong>et</strong>ween random variables. This transform<strong>at</strong>ion (40) is equivalent to (38)<br />
but of a simpler implement<strong>at</strong>ion and will be used in the sequel.<br />
17
442 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
6.3 D<strong>et</strong>ermin<strong>at</strong>ion of the joint distribution : maximum entropy and Gaussian copula<br />
L<strong>et</strong> us now consi<strong>de</strong>r N random variables Xi with marginal distributions pi(xi). Using the transform<strong>at</strong>ion<br />
(38), we <strong>de</strong>fine N standard normal variables Yi. If these variables were in<strong>de</strong>pen<strong>de</strong>nt, their joint distribution<br />
would simply be the product of the marginal distributions. In many situ<strong>at</strong>ions, the variables are not<br />
in<strong>de</strong>pen<strong>de</strong>nt and it is necessary to study their <strong>de</strong>pen<strong>de</strong>nce.<br />
The simplest approach is to construct their covariance m<strong>at</strong>rix. Applied to the variables Yi, we are certain th<strong>at</strong><br />
the covariance m<strong>at</strong>rix exists and is well-<strong>de</strong>fined since their marginal distributions are Gaussian. In contrast,<br />
this is not ensured for the variables Xi. In<strong>de</strong>ed, in many situ<strong>at</strong>ions in n<strong>at</strong>ure, in economy, finance and in<br />
social sciences, pdf’s are found to have power law tails ∼ A<br />
x1+µ for large |x|. If µ ≤ 2, the variance and the<br />
covariances can not be <strong>de</strong>fined. If 2 < µ ≤ 4, the variance and the covariances exit in principle but their<br />
sample estim<strong>at</strong>ors converge poorly.<br />
We thus <strong>de</strong>fine the covariance m<strong>at</strong>rix:<br />
V = E[yy t ] , (42)<br />
where y is the vector of variables Yi and the oper<strong>at</strong>or E[·] represents the m<strong>at</strong>hem<strong>at</strong>ical expect<strong>at</strong>ion. A<br />
classical result of inform<strong>at</strong>ion theory (Rao 1973) tells us th<strong>at</strong>, given the covariance m<strong>at</strong>rix V , the best joint<br />
distribution (in the sense of entropy maximiz<strong>at</strong>ion) of the N variables Yi is the multivari<strong>at</strong>e Gaussian:<br />
1<br />
P (y) =<br />
(2π) N/2d<strong>et</strong>(V ) exp<br />
<br />
− 1<br />
2 ytV −1 <br />
y . (43)<br />
In<strong>de</strong>ed, this distribution implies the minimum additional inform<strong>at</strong>ion or assumption, given the covariance<br />
m<strong>at</strong>rix.<br />
Using the joint distribution of the variables Yi, we obtain the joint distribution of the variables Xi:<br />
<br />
∂yi <br />
P (x) = P (y) <br />
∂xj<br />
, (44)<br />
<br />
<br />
where ∂yi<br />
<br />
<br />
is the Jacobian of the transform<strong>at</strong>ion. Since<br />
we g<strong>et</strong><br />
∂xj<br />
This finally yields<br />
P (x) =<br />
∂yi<br />
∂xj<br />
= √ 2πpj(xj)e 1<br />
2 y2 i δij , (45)<br />
<br />
∂yi <br />
N<br />
<br />
∂xj<br />
= (2π)N/2<br />
i=1<br />
pi(xi)e 1<br />
2 y2 i . (46)<br />
<br />
1<br />
exp −<br />
d<strong>et</strong>(V ) 1<br />
2 yt (x) (V −1 <br />
N<br />
− I)y (x) pi(xi) . (47)<br />
i=1<br />
As expected, if the variables are in<strong>de</strong>pen<strong>de</strong>nt, V = I, and P (x) becomes the product of the marginal<br />
distributions of the variables Xi.<br />
L<strong>et</strong> F (x) <strong>de</strong>note the cumul<strong>at</strong>ive distribution function of the vector x and Fi(xi), i = 1, ..., N the N corresponding<br />
marginal distributions. The copula C is then such th<strong>at</strong><br />
F (x1, · · · , xn) = C(F1(x1), · · · , Fn(xn)) . (48)<br />
18
Differenti<strong>at</strong>ing with respect to x1, · · · , xN leads to<br />
where<br />
is the <strong>de</strong>nsity of the copula C.<br />
P (x1, · · · , xn) = ∂F (x1, · · · , xn)<br />
∂x1 · · · ∂xn<br />
= c(F1(x1), · · · , Fn(xn))<br />
c(u1, · · · , uN) = ∂C(u1, · · · , uN)<br />
∂u1 · · · ∂uN<br />
443<br />
N<br />
pi(xi) , (49)<br />
Comparing (50) with (47), the <strong>de</strong>nsity of the copula is given in the present case by<br />
1<br />
c(u1, · · · , uN) = exp<br />
d<strong>et</strong>(V )<br />
<br />
, (51)<br />
i=1<br />
<br />
− 1<br />
2 yt (u) (V −1 − I)y (u)<br />
which is the “Gaussian copula” with covariance m<strong>at</strong>rix V. This result clarifies and justifies the m<strong>et</strong>hod<br />
of (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b) by showing th<strong>at</strong> it essentially amounts to assume arbitrary marginal distributions<br />
with Gaussian copulas. Note th<strong>at</strong> the Gaussian copula results directly from the transform<strong>at</strong>ion to Gaussian<br />
marginals tog<strong>et</strong>her with the choice of maximizing the Shannon entropy un<strong>de</strong>r the constraint of a fixed covariance<br />
m<strong>at</strong>rix. Un<strong>de</strong>r differents constraint, we would have found another maximum entropy copula. This<br />
is not unexpected in analogy with the standard result th<strong>at</strong> the Gaussian law is maximizing the Shannon entropy<br />
<strong>at</strong> fixed given variance. If we were to extend this formul<strong>at</strong>ion by consi<strong>de</strong>ring more general expressions<br />
of the entropy, such th<strong>at</strong> Tsallis entropy (Tsallis 1998), we would have found other copulas.<br />
6.4 Empirical test of the Gaussian copula assumption<br />
We now present some tests of the hypothesis of Gaussian copulas b<strong>et</strong>ween r<strong>et</strong>urns of financial ass<strong>et</strong>s. This<br />
present<strong>at</strong>ion is only for illustr<strong>at</strong>ion purposes, since testing the gaussian copula hypothesis is a <strong>de</strong>lic<strong>at</strong>e task<br />
which has been addressed elsewhere (see (Malevergne and Sorn<strong>et</strong>te 2001)). Here, as an example, we propose<br />
two simple standard m<strong>et</strong>hods.<br />
The first one consists in using the property th<strong>at</strong> Gaussian variables are stable in distribution un<strong>de</strong>r addition.<br />
Thus, a (quantile-quantile or Q − Q) plot of the cumul<strong>at</strong>ive distribution of the sum y1 + · · · + yp versus<br />
the cumul<strong>at</strong>ive Normal distribution with the same estim<strong>at</strong>ed variance should give a straight line in or<strong>de</strong>r to<br />
qualify a multivari<strong>at</strong>e Gaussian distribution (for the transformed y variables). Such tests on empirical d<strong>at</strong>a<br />
are presented in figures 7-9.<br />
The second test amounts to estim<strong>at</strong>ing the covariance m<strong>at</strong>rix V of the sample we consi<strong>de</strong>r. This step is<br />
simple since, for fast <strong>de</strong>caying pdf’s, robust estim<strong>at</strong>ors of the covariance m<strong>at</strong>rix are available. We can then<br />
estim<strong>at</strong>e the distribution of the variable z 2 = y t V −1 y. It is well known th<strong>at</strong> z 2 follows a χ 2 distribution<br />
if y is a Gaussian random vector. Again, the empirical cumul<strong>at</strong>ive distribution of z 2 versus the χ 2 cumul<strong>at</strong>ive<br />
distribution should give a straight line in or<strong>de</strong>r to qualify a multivari<strong>at</strong>e Gaussian distribution (for the<br />
transformed y variables). Such tests on empirical d<strong>at</strong>a are presented in figures 10-12.<br />
First, one can observe th<strong>at</strong> the Gaussian copula hypothesis appears b<strong>et</strong>ter for stocks than for currencies.<br />
As discussed in (Malevergne and Sorn<strong>et</strong>te 2001), this result is quite general. A plausible explan<strong>at</strong>ion lies<br />
in the stronger <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween the currencies compared with th<strong>at</strong> b<strong>et</strong>ween stocks, which is due to<br />
the mon<strong>et</strong>ary policies limiting the fluctu<strong>at</strong>ions b<strong>et</strong>ween the currencies of a group of countries, such as was<br />
the case in the European Mon<strong>et</strong>ary System before the unique Euro currency. Note also th<strong>at</strong> the test of<br />
aggreg<strong>at</strong>ion seems system<strong>at</strong>ically more in favor of the Gaussian copula hypothesis than is the χ 2 test, maybe<br />
due to its smaller sensitivity. Non<strong>et</strong>heless, the very good performance of the Gaussian hypothesis un<strong>de</strong>r the<br />
19<br />
(50)
444 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
aggreg<strong>at</strong>ion test bears good news for a porfolio theory based on it, since by <strong>de</strong>finition a portfolio corresponds<br />
to ass<strong>et</strong> aggreg<strong>at</strong>ion. Even if sums of the transformed r<strong>et</strong>urns are not equivalent to sums of r<strong>et</strong>urns (as we<br />
shall see in the sequel), such sums qualify the collective behavior whose properties are controlled by the<br />
copula.<br />
Notwithstanding some <strong>de</strong>vi<strong>at</strong>ions from linearity in figures 7-12, it appears th<strong>at</strong>, for our purpose of <strong>de</strong>veloping<br />
a generalized portfolio theory, the Gaussian copula hypothesis is a good approxim<strong>at</strong>ion. A more system<strong>at</strong>ic<br />
test of this goodness of fit requires the quantific<strong>at</strong>ion of a confi<strong>de</strong>nce level, for instance using the Kolmogorov<br />
test, th<strong>at</strong> would allow us to accept or reject the Gaussian copula hypothesis. Such a test has been performed<br />
in (Malevergne and Sorn<strong>et</strong>te 2001), where it is shown th<strong>at</strong> this test is sensitive enough only in the bulk of the<br />
distribution, and th<strong>at</strong> an An<strong>de</strong>rson-Darling test is preferable for the tails of the distributions. Non<strong>et</strong>heless,<br />
the quantit<strong>at</strong>ive conclusions of these tests are i<strong>de</strong>ntical to the qualit<strong>at</strong>ive results presented here. Some other<br />
tests would be useful, such as the multivari<strong>at</strong>e Gaussianity test presented by (Richardson and Smith 1993).<br />
7 Choice of an exponential family to param<strong>et</strong>erize the marginal distributions<br />
7.1 The modified Weibull distributions<br />
We now apply these constructions to a class of distributions with f<strong>at</strong> tails, th<strong>at</strong> have been found to provi<strong>de</strong><br />
a convenient and flexible param<strong>et</strong>eriz<strong>at</strong>ion of many phenomena found in n<strong>at</strong>ure and in the social sciences<br />
(Laherrère and Sorn<strong>et</strong>te 1998). These so-called str<strong>et</strong>ched exponential distributions can be seen to be general<br />
forms of the extreme tails of product of random variables (Frisch and Sorn<strong>et</strong>te 1997).<br />
Following (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b), we postul<strong>at</strong>e the following marginal probability distributions of r<strong>et</strong>urns:<br />
p(x) = 1<br />
2 √ c<br />
π χ c |x|<br />
2<br />
c<br />
2 −1 e −|x|<br />
χc<br />
, (52)<br />
where c and χ are the two key param<strong>et</strong>ers. A more general param<strong>et</strong>eriz<strong>at</strong>ion taking into account a possible<br />
asymm<strong>et</strong>ry b<strong>et</strong>ween neg<strong>at</strong>ive and positive r<strong>et</strong>urns (thus leading to possible non-zero average r<strong>et</strong>urn) is<br />
p(x) = Q √<br />
π<br />
χ<br />
p(x) =<br />
1 − Q<br />
√ π<br />
c+<br />
c +<br />
2<br />
+<br />
|x| c +<br />
2 −1 e −|x| +<br />
χ +c<br />
c−<br />
c− 2<br />
−<br />
χ<br />
|x| c− 2 −1 e −|x| −<br />
χ−c if x ≥ 0 (53)<br />
if x < 0 , (54)<br />
where Q (respectively 1 − Q) is the fraction of positive (respectively neg<strong>at</strong>ive) r<strong>et</strong>urns. In the sequel, we<br />
will only consi<strong>de</strong>r the case Q = 1<br />
2 , which is the only analytically tractable case. Thus the pdf’s asymm<strong>et</strong>ry<br />
will be only accounted for by the exponents c+, c− and the scale factors χ+, χ−.<br />
We can note th<strong>at</strong> these expressions are close to the Weibull distribution, with the addition of a power law<br />
prefactor to the exponential such th<strong>at</strong> the Gaussian law is r<strong>et</strong>rieved for c = 2. Following (Sorn<strong>et</strong>te <strong>et</strong><br />
al. 2000b, Sorn<strong>et</strong>te <strong>et</strong> al. 2000a, An<strong>de</strong>rsen and Sorn<strong>et</strong>te 2001), we call (52) the modified Weibull distribution.<br />
For c < 1, the pdf is a str<strong>et</strong>ched exponential, also called sub-exponential. The exponent c d<strong>et</strong>ermines the<br />
shape of the distribution, which is f<strong>at</strong>ter than an exponential if c < 1. The param<strong>et</strong>er χ controls the scale or<br />
characteristic width of the distribution. It plays a role analogous to the standard <strong>de</strong>vi<strong>at</strong>ion of the Gaussian<br />
law. See chapter 6 of (?) for a recent review on maximum likelihood and other estim<strong>at</strong>ors of such generalized<br />
Weibull distributions.<br />
20
7.2 Transform<strong>at</strong>ion of the modified Weibull pdf into a Gaussian Law<br />
One advantage of the class of distributions (52) is th<strong>at</strong> the transform<strong>at</strong>ion into a Gaussian is particularly<br />
simple. In<strong>de</strong>ed, the expression (52) is of the form (39) with<br />
c |x|<br />
f(x) = 2 . (55)<br />
χ<br />
Applying the change of variable (40) which reads<br />
leads autom<strong>at</strong>ically to a Gaussian distribution.<br />
yi = sgn(xi) √ <br />
|xi|<br />
2<br />
χi<br />
c i<br />
2<br />
These variables Yi then allow us to obtain the covariance m<strong>at</strong>rix V :<br />
Vij = 2<br />
T<br />
2 N π N/2√ V exp<br />
T<br />
<br />
|xi|<br />
sgn(xixj)<br />
n=1<br />
i,j<br />
V −1<br />
ij<br />
χi<br />
χi<br />
c i<br />
2 |xj|<br />
χj<br />
445<br />
, (56)<br />
χj<br />
c j<br />
2<br />
and thus the multivari<strong>at</strong>e distributions P (y) and P (x) :<br />
P (x1, · · · , xN) =<br />
1<br />
⎡<br />
⎣− <br />
⎤<br />
c/2 c/2 |xi| |xj|<br />
⎦<br />
, (57)<br />
<br />
N<br />
ci|xi| c/2−1<br />
i=1<br />
χ c/2<br />
i<br />
e −|x i |<br />
χc <br />
.<br />
(58)<br />
Similar transforms hold, mut<strong>at</strong>is mutandis, for the asymm<strong>et</strong>ric case. In<strong>de</strong>ed, for asymm<strong>et</strong>ric ass<strong>et</strong>s of interest<br />
for financial risk managers, the equ<strong>at</strong>ions (53) and (54) yields the following change of variable:<br />
yi = √ 2<br />
<br />
xi<br />
c+ i<br />
2<br />
χ + i<br />
yi = − √ <br />
|xi|<br />
2<br />
χ − i<br />
c− i 2<br />
and xi ≥ 0, (59)<br />
and xi < 0 . (60)<br />
This allows us to <strong>de</strong>fine the correl<strong>at</strong>ion m<strong>at</strong>rix V and to obtain the multivari<strong>at</strong>e distribution P (x), generalizing<br />
equ<strong>at</strong>ion (58) for asymm<strong>et</strong>ric ass<strong>et</strong>s. Since this expression is r<strong>at</strong>her cumbersome and nothing but a<br />
straightforward generaliz<strong>at</strong>ion of (58), we do not write it here.<br />
7.3 Empirical tests and estim<strong>at</strong>ed param<strong>et</strong>ers<br />
In or<strong>de</strong>r to test the validity of our assumption, we have studied a large bask<strong>et</strong> of financial ass<strong>et</strong>s including<br />
currencies and stocks. As an example, we present in figures 13 to 17 typical log-log plot of the transformed<br />
r<strong>et</strong>urn variable Y versus the r<strong>et</strong>urn variable X for a certain number of ass<strong>et</strong>s. If our assumption was right,<br />
we should observe a single straight line whose slope is given by c/2. In contrast, we observe in general<br />
two approxim<strong>at</strong>ely linear regimes separ<strong>at</strong>ed by a cross-over. This means th<strong>at</strong> the marginal distribution of<br />
r<strong>et</strong>urns can be approxim<strong>at</strong>ed by two modified Weibull distributions, one for small r<strong>et</strong>urns which is close to a<br />
Gaussian law and one for large r<strong>et</strong>urns with a f<strong>at</strong> tail. Each regime is <strong>de</strong>picted by its corresponding straight<br />
line in the graphs. The exponents c and the scale factors χ for the different ass<strong>et</strong>s we have studied are given<br />
in tables 3 for currencies and 4 for stocks. The coefficients within brack<strong>et</strong>s are the coefficients estim<strong>at</strong>ed for<br />
small r<strong>et</strong>urns while the non-brack<strong>et</strong>ed coefficients correspond to the second f<strong>at</strong> tail regime.<br />
21
446 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
The first point to note is the difference b<strong>et</strong>ween currencies and stocks. For small as well as for large r<strong>et</strong>urns,<br />
the exponents c− and c+ for currencies (excepted Poland and Thailand) are all close to each other. Additional<br />
tests are required to establish wh<strong>et</strong>her their rel<strong>at</strong>ively small differences are st<strong>at</strong>istically significant. Similarly,<br />
the scale factors are also comparable. In contrast, many stocks exhibit a large asymm<strong>et</strong>ric behavior for large<br />
r<strong>et</strong>urns with c+ −c− 0.5 in about one-half of the investig<strong>at</strong>ed stocks. This means th<strong>at</strong> the tails of the large<br />
neg<strong>at</strong>ive r<strong>et</strong>urns (“crashes”) are often much f<strong>at</strong>ter than those of the large positive r<strong>et</strong>urns (“rallies”).<br />
The second important point is th<strong>at</strong>, for small r<strong>et</strong>urns, many stocks have an exponent 〈c+〉 ≈ 〈c−〉 2 and<br />
thus have a behavior not far from a pure Gaussian in the bulk of the distribution, while the average exponent<br />
for currencies is about 1.5 in the same “small r<strong>et</strong>urn” regime. Therefore, even for small r<strong>et</strong>urns, currencies<br />
exhibit a strong <strong>de</strong>parture from Gaussian behavior.<br />
In conclusion, this empirical study shows th<strong>at</strong> the modified Weibull param<strong>et</strong>eriz<strong>at</strong>ion, although not exact on<br />
the entire range of vari<strong>at</strong>ion of the r<strong>et</strong>urns X, remains consistent within each of the two regimes of small<br />
versus large r<strong>et</strong>urns, with a sharp transition b<strong>et</strong>ween them. It seems especially relevant in the tails of the<br />
r<strong>et</strong>urn distributions, on which we shall focus our <strong>at</strong>tention next.<br />
8 Cumulant expansion of the portfolio r<strong>et</strong>urn distribution<br />
8.1 link b<strong>et</strong>ween moments and cumulants<br />
Before <strong>de</strong>riving the main result of this section, we recall a standard rel<strong>at</strong>ion b<strong>et</strong>ween moments and cumulants<br />
th<strong>at</strong> we need below.<br />
The moments Mn of the distribution P are <strong>de</strong>fined by<br />
ˆP (k) =<br />
+∞<br />
n=0<br />
(ik) n<br />
where ˆ P is the characteristic function, i.e., the Fourier transform of P :<br />
Similarly, the cumulants Cn are given by<br />
Differenti<strong>at</strong>ing n times the equ<strong>at</strong>ion<br />
ln<br />
ˆP (k) =<br />
ˆP (k) = exp<br />
+∞<br />
n=0<br />
(ik) n<br />
+∞<br />
−∞<br />
n! Mn<br />
+∞<br />
n! Mn , (61)<br />
dS P (S)e ikS . (62)<br />
n=1<br />
<br />
=<br />
(ik) n<br />
n! Cn<br />
+∞<br />
n=1<br />
<br />
(ik) n<br />
. (63)<br />
n! Cn , (64)<br />
we obtain the following recurrence rel<strong>at</strong>ions b<strong>et</strong>ween the moments and the cumulants :<br />
Mn =<br />
n−1 <br />
<br />
n − 1<br />
MpCn−p ,<br />
p<br />
(65)<br />
p=0<br />
n−1 <br />
<br />
n − 1<br />
Cn = Mn −<br />
CpMn−p . (66)<br />
n − p<br />
p=1<br />
22
In the sequel, we will first evalu<strong>at</strong>e the moments, which turns out to be easier, and then using eq (66) we<br />
will be able to calcul<strong>at</strong>e the cumulants.<br />
8.2 Symm<strong>et</strong>ric ass<strong>et</strong>s<br />
We start with the expression of the distribution of the weighted sum of N ass<strong>et</strong>s :<br />
<br />
PS(s) =<br />
R N<br />
447<br />
N<br />
dx P (x)δ( wixi − s) , (67)<br />
where δ(·) is the Dirac distribution. Using the change of variable (40), allowing us to go from the ass<strong>et</strong><br />
r<strong>et</strong>urns Xi’s to the transformed r<strong>et</strong>urns Yi’s, we g<strong>et</strong><br />
1<br />
PS(s) =<br />
(2π) N/2 <br />
d<strong>et</strong>(V )<br />
R N<br />
i=1<br />
1<br />
−<br />
dy e 2 ytV −1y δ(<br />
Taking its Fourier transform ˆ PS(k) = dsPS(s)eiks , we obtain<br />
ˆPS(k)<br />
1<br />
=<br />
(2π) N/2 <br />
d<strong>et</strong>(V )<br />
where ˆ PS is the characteristic function of PS.<br />
R N<br />
N<br />
wisgn(yi)f −1 (y 2 i ) − s) . (68)<br />
i=1<br />
1<br />
−<br />
dy e 2 ytV −1y+ikN i=1 wisgn(yi)f −1 (y2 i ) , (69)<br />
In the particular case of interest here where the marginal distributions of the variables Xi’s are the modified<br />
Weibull pdf,<br />
f −1 (yi) = χi| yi<br />
√2 | qi (70)<br />
with<br />
the equ<strong>at</strong>ion (69) becomes<br />
ˆPS(k)<br />
1<br />
=<br />
(2π) N/2 <br />
d<strong>et</strong>(V )<br />
R N<br />
qi = 2/ci , (71)<br />
1<br />
−<br />
dy e 2 ytV −1y+ikN i=1 wisgn(yi)χi| y √2 i | qi . (72)<br />
The task in front of us is to evalu<strong>at</strong>e this expression through the d<strong>et</strong>ermin<strong>at</strong>ion of the moments and/or<br />
cumulants.<br />
8.2.1 Case of in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s<br />
In this case, the cumulants can be obtained explicitely (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b). In<strong>de</strong>ed, the expression (72)<br />
can be expressed as a product of integrals of the form<br />
We obtain<br />
+∞<br />
0<br />
C2n =<br />
u2<br />
− 2 du e +ikwiχiu i<br />
√2q<br />
. (73)<br />
N<br />
c(n, qi)(χiwi) 2n , (74)<br />
i=1<br />
23
448 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
and<br />
⎧<br />
⎨n−2<br />
<br />
c(n, qi) = (2n)! (−1)<br />
⎩<br />
n Γ qi(n − p) + 1<br />
<br />
2<br />
(2n − 2p)!π1/2 p=0<br />
Γ qi + 1<br />
2<br />
2!π 1/2<br />
<br />
p<br />
− (−1)n<br />
n<br />
Γ qi + 1<br />
2<br />
2!π<br />
<br />
1/2<br />
⎫<br />
n⎬ .<br />
⎭<br />
(75)<br />
Note th<strong>at</strong> the coefficient c(n, qi) is the cumulant of or<strong>de</strong>r n of the marginal distribution (52) with c = 2/qi<br />
and χ = 1. The equ<strong>at</strong>ion (74) expresses simply the fact th<strong>at</strong> the cumulants of the sum of in<strong>de</strong>pen<strong>de</strong>nt variables<br />
is the sum of the cumulants of each variable. The odd-or<strong>de</strong>r cumulants are zero due to the symm<strong>et</strong>ry<br />
of the distributions.<br />
8.2.2 Case of <strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s<br />
Here, we restrict our exposition to the case of two random variables. The case with N arbitrary can be<br />
tre<strong>at</strong>ed in a similar way but involves r<strong>at</strong>her complex formulas. The equ<strong>at</strong>ion (72) reads<br />
ˆPS(k)<br />
1<br />
=<br />
2π 1 − ρ2 <br />
<br />
dy1dy2 exp − 1<br />
2 ytV −1 <br />
q1<br />
y1 <br />
y + ik χ1w1sgn(y1) <br />
√<br />
<br />
2<br />
+<br />
q2<br />
y2 <br />
+χ2w2sgn(y2) <br />
√<br />
<br />
2<br />
<br />
<br />
, (76)<br />
and we can show (see appendix E) th<strong>at</strong> the moments read<br />
n<br />
<br />
n<br />
Mn =<br />
p<br />
with<br />
γq1q2<br />
γq1q2 (2n, 2p) = χ2p<br />
1 χ2(n−p) 2<br />
(2n, 2p + 1) = 2χ2p+1<br />
1<br />
p=0<br />
Γ q1p + 1<br />
2<br />
χ 2(n−p)−1<br />
2<br />
, −q2(n − p) + q2 + 1<br />
2<br />
where 2F1 is an hypergeom<strong>et</strong>ric function.<br />
<br />
w p<br />
1wn−pγq1q2 2 (n, p) , (77)<br />
<br />
Γ q2(n − p) + 1<br />
<br />
2<br />
2F1 −q1p, −q2(n − p);<br />
π<br />
1<br />
<br />
; ρ2 , (78)<br />
2<br />
<br />
Γ q2(n − p) + 1 − q2<br />
<br />
2 ρ 2F1 −q1p −<br />
π<br />
q1 − 1<br />
,<br />
2<br />
; 3<br />
<br />
; ρ2 , (79)<br />
2<br />
Γ q1p + 1 + q1<br />
2<br />
These two rel<strong>at</strong>ions allow us to calcul<strong>at</strong>e the moments and cumulants for any possible values of q1 = 2/c1<br />
and q2 = 2/c2. If one of the qi’s is an integer, a simplific<strong>at</strong>ion occurs and the coefficients γ(n, p) reduce to<br />
polynomials. In the simpler case where all the qi’s are odd integer the expression of moments becomes :<br />
with<br />
Mn =<br />
n<br />
p=0<br />
<br />
n<br />
(w1χ1)<br />
p<br />
p (w2χ2) n−p<br />
min{q1p,q2(n−p)} <br />
a (2n)<br />
2p = (2p)!<br />
a (2n)<br />
2p+1<br />
s=0<br />
<br />
2(n − p)<br />
(2(n − p) − 1)!! =<br />
2n<br />
ρ s s! a (q1p)<br />
s a (q2(n−p))<br />
s , (80)<br />
(2n)!<br />
2n−p , (81)<br />
(n − p)!<br />
= 0 , (82)<br />
a (2n+1)<br />
2p = 0 , (83)<br />
a (2n+1)<br />
<br />
2(n − p)<br />
(2n + 1)!<br />
2p+1 = (2p + 1)!<br />
(2(n − p) − 1)!! =<br />
2n + 1<br />
2n−p . (84)<br />
(n − p)!<br />
24
8.3 Non-symm<strong>et</strong>ric ass<strong>et</strong>s<br />
In the case of asymm<strong>et</strong>ric ass<strong>et</strong>s, we have to consi<strong>de</strong>r the formula (53-54), and using the same not<strong>at</strong>ion as in<br />
the previous section, the moments are again given by (77) with the coefficient γ(n, p) now equal to :<br />
<br />
Γ<br />
4π<br />
− <br />
q1 p<br />
+2Γ + 1 Γ<br />
2<br />
<br />
Γ<br />
4π<br />
− <br />
q1 p<br />
−2Γ + 1 Γ<br />
2<br />
<br />
Γ<br />
4π<br />
+ <br />
q 1 p<br />
−2Γ + 1 Γ<br />
2<br />
<br />
Γ<br />
4π<br />
+ <br />
q 1 p<br />
+2Γ + 1 Γ<br />
2<br />
γ(n, p) = (−1)n (χ − 1 )p (χ − 2 )n−p<br />
+ (−1)p (χ − 1 )p (χ + 2 )n−p<br />
+ (−1)n−p (χ + 1 )p (χ − 2 )n−p<br />
+ (χ+ 1 )p (χ + 2 )n−p<br />
− − <br />
q1 p + 1 q2 (n − p) + 1<br />
Γ<br />
2F1<br />
2<br />
2<br />
− <br />
q2 (n − p)<br />
+ 1 ρ 2F1 −<br />
2<br />
q− 1 p − 1<br />
2<br />
− + <br />
q1 p + 1 q 2 (n − p) + 1<br />
Γ<br />
2F1<br />
2<br />
2<br />
+ <br />
q 2 (n − p)<br />
+ 1 ρ 2F1 −<br />
2<br />
q− 1 p − 1<br />
2<br />
+ − <br />
q 1 p + 1 q2 (n − p) + 1<br />
Γ<br />
2F1<br />
2<br />
2<br />
− <br />
q2 (n − p)<br />
+ 1 ρ 2F1 −<br />
2<br />
q+ 1 p − 1<br />
2<br />
+ + <br />
q 1 p + 1 q 2 (n − p) + 1<br />
Γ<br />
2F1<br />
2<br />
2<br />
+ <br />
q 2 (n − p)<br />
+ 1 ρ 2F1 −<br />
2<br />
q+ 1<br />
<br />
− q− 1 p<br />
2 , −q− 2<br />
, − q− 2<br />
<br />
− q− 1 p<br />
2 , −q+ 2<br />
, − q+ 2<br />
<br />
− q+ 1 p<br />
2 , −q− 2<br />
, − q− 2<br />
<br />
− q+ 1 p<br />
2 , −q+ 2<br />
p − 1<br />
, −<br />
2<br />
q+ 2<br />
(n − p)<br />
2<br />
449<br />
; 1<br />
<br />
; ρ2 +<br />
2<br />
(n − p) − 1<br />
;<br />
2<br />
3<br />
<br />
; ρ2 +<br />
2<br />
(n − p)<br />
;<br />
2<br />
1<br />
<br />
; ρ2 +<br />
2<br />
(n − p) − 1<br />
;<br />
2<br />
3<br />
<br />
; ρ2 +<br />
2<br />
(n − p)<br />
;<br />
2<br />
1<br />
<br />
; ρ2 +<br />
2<br />
(n − p) − 1<br />
;<br />
2<br />
3<br />
<br />
; ρ2 +<br />
2<br />
(n − p)<br />
;<br />
2<br />
1<br />
<br />
; ρ2 +<br />
2<br />
(n − p) − 1<br />
;<br />
2<br />
3<br />
<br />
; ρ2 .<br />
2<br />
(85)<br />
This formula is obtained in the same way as for the formulas given in the symm<strong>et</strong>ric case. We r<strong>et</strong>rieve the<br />
formula (78) as it should if the coefficients with in<strong>de</strong>x ’+’ are equal to the coefficients with in<strong>de</strong>x ’-’.<br />
8.4 Empirical tests<br />
Extensive tests have been performed for currencies un<strong>de</strong>r the assumption th<strong>at</strong> the distributions of ass<strong>et</strong><br />
r<strong>et</strong>urns are symm<strong>et</strong>ric (Sorn<strong>et</strong>te <strong>et</strong> al. 2000b).<br />
As an exemple, l<strong>et</strong> us consi<strong>de</strong>r the Swiss franc and the Japanese Yen against the US dollar. The calibr<strong>at</strong>ion<br />
of the modified Weibull distribution to the tail of the empirical histogram of daily r<strong>et</strong>urns give (qCHF =<br />
1.75, cCHF = 1.14, χCHF = 2.13) and (qJP Y = 2.50, cJP Y = 0.8, χJP Y = 1.25) and their correl<strong>at</strong>ion<br />
coefficient is ρ = 0.43.<br />
Figure 18 plots the excess kurtosis of the sum wCHF xCHF + wJP Y xJP Y as a function of wCHF , with<br />
the constraint wCHF + wJP Y = 1. The thick solid line is d<strong>et</strong>ermined empirically, by direct calcul<strong>at</strong>ion of<br />
the kurtosis from the d<strong>at</strong>a. The thin solid line is the theor<strong>et</strong>ical prediction using our theor<strong>et</strong>ical formulas<br />
with the empirically d<strong>et</strong>ermined exponents c and characteristic scales χ given above. While there is a nonnegligible<br />
difference, the empirical and theor<strong>et</strong>ical excess kurtosis have essentially the same behavior with<br />
their minimum reached almost <strong>at</strong> the same value of wCHF .<br />
Three origins of the discrepancy b<strong>et</strong>ween theory and empirical d<strong>at</strong>a can be invoked. First, as already pointed<br />
out in the preceding section, the modified Weibull distribution with constant exponent and scale param<strong>et</strong>ers<br />
<strong>de</strong>scribes accur<strong>at</strong>ely only the tail of the empirical distributions while, for small r<strong>et</strong>urns, the empirical<br />
distributions are close to a Gaussian law. While putting a strong emphasis on large fluctu<strong>at</strong>ions, cumulants<br />
of or<strong>de</strong>r 4 are still significantly sensitive to the bulk of the distributions. Moreover, the excess kurtosis is<br />
25
450 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
normalized by the square second-or<strong>de</strong>r cumulant, which is almost exclusively sensitive to the bulk of the<br />
distribution. Cumulants of higher or<strong>de</strong>r should thus be b<strong>et</strong>ter <strong>de</strong>scribed by the modified Weibull distribution.<br />
However, a careful comparison b<strong>et</strong>ween theory and d<strong>at</strong>a would then be hin<strong>de</strong>red by the difficulty in estim<strong>at</strong>ing<br />
reliable empirical cumulants of high or<strong>de</strong>r. This estim<strong>at</strong>ion problem is often invoked as a criticism<br />
against using high-or<strong>de</strong>r moments or cumulants. Our approach suggests th<strong>at</strong> this problem can be in large<br />
part circumvented by focusing on the estim<strong>at</strong>ion of a reasonable param<strong>et</strong>ric expression for the probability<br />
<strong>de</strong>nsity or distribution function of the ass<strong>et</strong>s r<strong>et</strong>urns. The second possible origin of the discrepancy b<strong>et</strong>ween<br />
theory and d<strong>at</strong>a is the existence of a weak asymm<strong>et</strong>ry of the empirical distributions, particularly of the Swiss<br />
franc, which has not been taken into account. The figure also suggests th<strong>at</strong> an error in the d<strong>et</strong>ermin<strong>at</strong>ion of<br />
the exponents c can also contribute to the discrepancy.<br />
In or<strong>de</strong>r to investig<strong>at</strong>e the sensitivity with respect to the choice of the param<strong>et</strong>ers q and ρ, we have also<br />
constructed the dashed line corresponding to the theor<strong>et</strong>ical curve with ρ = 0 (instead of ρ = 0.43) and<br />
the dotted line corresponding to the theor<strong>et</strong>ical curve with qCHF = 2 r<strong>at</strong>her than 1.75. Finally, the dasheddotted<br />
line corresponds to the theor<strong>et</strong>ical curve with qCHF = 1.5. We observe th<strong>at</strong> the dashed line remains<br />
r<strong>at</strong>her close to the thin solid line while the dotted line <strong>de</strong>parts significantly when wCHF increases. Therefore,<br />
the most sensitive param<strong>et</strong>er is q, which is n<strong>at</strong>ural because it controls directly the extend of the f<strong>at</strong> tail of the<br />
distributions.<br />
In or<strong>de</strong>r to account for the effect of asymm<strong>et</strong>ry, we have plotted the fourth cumulant of a portfolio composed<br />
of Swiss Francs and British Pounds. On figure 19, the solid line represents the empirical cumulant while<br />
the dashed line shows the theor<strong>et</strong>ical cumulant. The agreement b<strong>et</strong>ween the two curves is b<strong>et</strong>ter than un<strong>de</strong>r<br />
the symm<strong>et</strong>ric asumption. Note once again th<strong>at</strong> an accur<strong>at</strong>e d<strong>et</strong>ermin<strong>at</strong>ion of the param<strong>et</strong>ers is the key point<br />
to obtain a good agreement b<strong>et</strong>ween empirical d<strong>at</strong>a and theor<strong>et</strong>ical prediction. As we can see in figure 19,<br />
the param<strong>at</strong>ers of the Swiss Franc seem well adjusted since the theor<strong>et</strong>ical and empirical cumulants are both<br />
very close when wCHF 1, i.e., when the Swiss Franc is almost the sole ass<strong>et</strong> in the portfolio, while when<br />
wCHF 0, the theor<strong>et</strong>ical cumulant is far from the empirical one, i.e., the param<strong>et</strong>ers of the Bristish Pound<br />
are not sufficiently well-adjusted.<br />
9 Can you have your cake and e<strong>at</strong> it too ?<br />
Now th<strong>at</strong> we have shown how to accur<strong>at</strong>ely estim<strong>at</strong>e the multivari<strong>at</strong>e distribution fonction of the ass<strong>et</strong>s<br />
r<strong>et</strong>urn, l<strong>et</strong> us come back to the portfolio selection problem. In figure 2, we can see th<strong>at</strong> the expected r<strong>et</strong>urn<br />
of the portfolios with minimum risk according to Cn <strong>de</strong>creases when n increases. But, this is not the general<br />
situ<strong>at</strong>ion.<br />
Figure 20 and 21 show the generalized efficient frontiers using C2 (Markovitz case), C4 or C6 as relevant<br />
measures of risks, for two portfolios composed of two stocks : IBM and Hewl<strong>et</strong>t-Packard in the first case<br />
and IBM and Coca-Cola in the second case.<br />
Obviously, given a certain amount of risk, the mean r<strong>et</strong>urn of the portfolio changes when the cumulant<br />
consi<strong>de</strong>red changes. It is interesting to note th<strong>at</strong>, in figure 20, the minimis<strong>at</strong>ion of large risks, i.e., with<br />
respect to C6, increases the average r<strong>et</strong>urn while, in figure 21, the minimis<strong>at</strong>ion of large risks lead to <strong>de</strong>crease<br />
the average r<strong>et</strong>urn.<br />
This allows us to make precise and quantit<strong>at</strong>ive the previously reported empirical observ<strong>at</strong>ion th<strong>at</strong> it is<br />
possible to “have your cake and e<strong>at</strong> it too” (An<strong>de</strong>rsen and Sorn<strong>et</strong>te 2001). We can in<strong>de</strong>ed give a general<br />
criterion to d<strong>et</strong>ermine un<strong>de</strong>r which values of the param<strong>et</strong>ers (exponents c and characteristic scales χ of<br />
the distributions of the ass<strong>et</strong> r<strong>et</strong>urns) the average r<strong>et</strong>urn of the portfolio may increase while the large risks<br />
<strong>de</strong>crease <strong>at</strong> the same time, thus allowing one to gain on both account (of course, the small risks quantified<br />
26
y the variance will then increase). For two in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s, assuming th<strong>at</strong> the cumulants of or<strong>de</strong>r n and<br />
n + k of the portfolio admit a minimum in the interval ]0, 1[, we can show th<strong>at</strong><br />
if and only if<br />
(µ(1) − µ(2)) ·<br />
µ ∗ n < µ ∗ n+k<br />
Cn(1) 1<br />
n−1<br />
−<br />
Cn(2)<br />
451<br />
(86)<br />
1 <br />
Cn+k(1) n+k−1<br />
> 0 , (87)<br />
Cn+k(2)<br />
where µ ∗ n <strong>de</strong>notes the r<strong>et</strong>urn of the portfolio evalu<strong>at</strong>ed with respect to the minimum of the cumulant of or<strong>de</strong>r<br />
n and Cn(i) is the cumulant of or<strong>de</strong>r n for the ass<strong>et</strong> i.<br />
The proof of this result and its generalis<strong>at</strong>ion to N > 2 are given in appendix F. In fact, we have observed<br />
th<strong>at</strong> when the exponent c of the ass<strong>et</strong>s remains sufficiently different, this result still holds in presence of<br />
<strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s. This last empirical observ<strong>at</strong>ion in the presence of <strong>de</strong>pen<strong>de</strong>nce b<strong>et</strong>ween ass<strong>et</strong>s<br />
has not been proved m<strong>at</strong>hem<strong>at</strong>ically. It seems reasonable for ass<strong>et</strong>s with mo<strong>de</strong>r<strong>at</strong>e <strong>de</strong>pen<strong>de</strong>nce while it may<br />
fail when the <strong>de</strong>pen<strong>de</strong>nce becomes too strong as occurs for comonotonic ass<strong>et</strong>s.<br />
For the ass<strong>et</strong>s consi<strong>de</strong>red above, we have found µIBM = 0.13, µHW P = 0.07, µKO = 0.05 and<br />
C2(IBM)<br />
C2(HW P )<br />
C2(IBM)<br />
C2(KO)<br />
= 1.76 ><br />
= 0.96 <<br />
1<br />
C4(IBM) 3<br />
C4(HW P )<br />
1<br />
C4(IBM) 3<br />
C4(KO)<br />
<br />
C6(IBM)<br />
= 1.03 ><br />
C6(HW P )<br />
<br />
C6(IBM)<br />
= 1.01 <<br />
C6(KO)<br />
1<br />
5<br />
1<br />
5<br />
= 0.89 (88)<br />
= 1.06 , (89)<br />
which shows th<strong>at</strong>, for the portfolio IBM / Hewl<strong>et</strong>t-Packard, the efficient r<strong>et</strong>urn is an increasing function of<br />
the or<strong>de</strong>r of the cumulants while, for the portfolio IBM / Coca-Cola, the inverse phenomenon occurs. This<br />
is exactly wh<strong>at</strong> is shown on figures 20 and 21.<br />
The un<strong>de</strong>rlying intuitive mechanism is the following: if a portfolio contains an ass<strong>et</strong> with a r<strong>at</strong>her f<strong>at</strong> tail<br />
(many “large” risks) but narrow waist (few “small” risks) with very little r<strong>et</strong>urn to gain from it, minimizing<br />
the variance C2 of the r<strong>et</strong>urn portfolio will overweight this ass<strong>et</strong> which is wrongly perceived as having little<br />
risk due to its small variance (small waist). In contrast, controlling for the larger risks quantified by C4 or<br />
C6 leads to <strong>de</strong>crease the weight of this ass<strong>et</strong> in the portfolio, and correspondingly to increase the weight<br />
of the more profitable ass<strong>et</strong>s. We thus see th<strong>at</strong> the effect of “both <strong>de</strong>creasing large risks and increasing<br />
profit” appears when the ass<strong>et</strong>(s) with the f<strong>at</strong>ter tails, and therefore the narrower central part, has(ve) the<br />
smaller overall r<strong>et</strong>urn(s). A mean-variance approach will weight them more than <strong>de</strong>emed appropri<strong>at</strong>e from<br />
a pru<strong>de</strong>ntial consi<strong>de</strong>r<strong>at</strong>ion of large risks and consi<strong>de</strong>r<strong>at</strong>ion of profits.<br />
From a behavioral point of view, this phenomenon is very interesting and can probably be linked with the<br />
fact th<strong>at</strong> the main risk measure consi<strong>de</strong>red by the agents is the vol<strong>at</strong>ility (or the variance), so th<strong>at</strong> the other<br />
dimensions of the risk, measured by higher moments, are often neglected. This may som<strong>et</strong>imes offer the<br />
opportunity of increasing the expected r<strong>et</strong>urn while lowering large risks.<br />
10 Conclusion<br />
We have introduced three axioms th<strong>at</strong> <strong>de</strong>fine a consistent s<strong>et</strong> of risk measures, in the spirit of (Artzner <strong>et</strong><br />
al. 1997, Artzner <strong>et</strong> al. 1999). Contrarily to the risk measures of (Artzner <strong>et</strong> al. 1997, Artzner <strong>et</strong> al. 1999),<br />
our consistent risk measures may account for both-si<strong>de</strong> risks and not only for down-si<strong>de</strong> risks. Thus, they<br />
supplement the notion of coherent measures of risk and are well adapted to the problem of portfolio risk<br />
27
452 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
assessment and optimiz<strong>at</strong>ion. We have shown th<strong>at</strong> these risk measures, which contain centered moments<br />
(and cumulants with some restriction) as particular examples, generalize them significantly. We have presented<br />
a generaliz<strong>at</strong>ion of previous generaliz<strong>at</strong>ions of the efficient frontiers and of the CAPM based on these<br />
risk measures in the cases of homogeneous and h<strong>et</strong>erogeneous agents. We have then proposed a simple but<br />
powerful specific von Mises represent<strong>at</strong>ion of multivari<strong>at</strong>e distribution of r<strong>et</strong>urns th<strong>at</strong> allowed us to obtain<br />
new analytical results on and empirical tests of a general framework for a portfolio theory of non-Gaussian<br />
risks with non-linear correl<strong>at</strong>ions. Quantit<strong>at</strong>ive tests have been presented on a bask<strong>et</strong> of seventeen stocks<br />
among the largest capitaliz<strong>at</strong>ion on the NYSE.<br />
This work opens several novel interesting avenues for research. One consists in extending the Gaussian copula<br />
assumption, for instance by using the maximum-entropy principle with non-extensive Tsallis entropies,<br />
known to be the correct m<strong>at</strong>hem<strong>at</strong>ical inform<strong>at</strong>ion-theor<strong>et</strong>ical represent<strong>at</strong>ion of power laws. A second line<br />
of research would be to extend the present framework to encompass simultaneously different time scales τ<br />
in the spirit of (Muzy <strong>et</strong> al. 2001) in the case of a casca<strong>de</strong> mo<strong>de</strong>l of vol<strong>at</strong>ilities.<br />
28
A Description of the d<strong>at</strong>a s<strong>et</strong><br />
We have consi<strong>de</strong>red a s<strong>et</strong> of seventeen ass<strong>et</strong>s tra<strong>de</strong>d on the New York Stock Exchange: Applied M<strong>at</strong>erial,<br />
Coca-Cola, EMC, Exxon-Mobil, General Electric, General Motors, Hewl<strong>et</strong>t Packard, IBM, Intel, MCI<br />
WorldCom, Medtronic, Merck, Pfizer, Procter & Gambel, SBC Communic<strong>at</strong>ion, Texas Instrument, Wall<br />
Mart. These ass<strong>et</strong>s have been choosen since they are among the largest capitaliz<strong>at</strong>ions of the NYSE <strong>at</strong> the<br />
time of writing.<br />
The d<strong>at</strong>as<strong>et</strong> comes from the Center for Research in Security Prices (CRSP) d<strong>at</strong>abase and covers the time<br />
interval from the end of January 1995 to the end of December 2000, which represents exactly 1500 trading<br />
days. The main st<strong>at</strong>istical fe<strong>at</strong>ures of the compagnies composing the d<strong>at</strong>as<strong>et</strong> are presented in the table 5.<br />
Note the high kurtosis of each distribution of r<strong>et</strong>urns as well as the large values of the observed minimum and<br />
maximum r<strong>et</strong>urns compared with the standard <strong>de</strong>vi<strong>at</strong>ions, th<strong>at</strong> clearly un<strong>de</strong>rlines the non-Gaussian behavior<br />
of these ass<strong>et</strong>s.<br />
29<br />
453
454 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
B Generalized efficient frontier and two funds separ<strong>at</strong>ion theorem<br />
L<strong>et</strong> us consi<strong>de</strong>r a s<strong>et</strong> of N risky ass<strong>et</strong>s X1, · · · , XN and a risk-free ass<strong>et</strong> X0. The problem is to find the<br />
optimal alloc<strong>at</strong>ion of these ass<strong>et</strong>s in the following sense:<br />
⎧<br />
⎨<br />
⎩<br />
inf wi∈[0,1] ρα({wi})<br />
<br />
i≥0 wi = 1<br />
<br />
i≥0 wiµ(i) = µ ,<br />
In other words, we search for the portfolio P with minimum risk as measured by any risk measure ρα<br />
obeying axioms I-IV of section 2 for a given amount of expected r<strong>et</strong>urn µ and normalized weights wi.<br />
Short-sells are forbid<strong>de</strong>n except for the risk-free ass<strong>et</strong> which can be lent and borrowed <strong>at</strong> the same interest<br />
r<strong>at</strong>e µ0. Thus, the weights wi’s are assumed positive for all i ≥ 1.<br />
B.1 Case of in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s when the risk is measured by the cumulants<br />
To start with a simple example, l<strong>et</strong> us assume th<strong>at</strong> the risky ass<strong>et</strong>s are in<strong>de</strong>pen<strong>de</strong>nt and th<strong>at</strong> we choose to<br />
measure the risk with the cumulants of their distributions of r<strong>et</strong>urns. The case when the ass<strong>et</strong>s are <strong>de</strong>pen<strong>de</strong>nt<br />
and/or when the risk is measured by any ρα will be consi<strong>de</strong>red l<strong>at</strong>er. Since the ass<strong>et</strong>s are assumed<br />
in<strong>de</strong>pen<strong>de</strong>nt, the cumulant of or<strong>de</strong>r n of the pdf of r<strong>et</strong>urns of the portfolio is simply given by<br />
Cn =<br />
(90)<br />
N<br />
wi n Cn(i), (91)<br />
i=1<br />
where Cn(i) <strong>de</strong>notes the marginal nth or<strong>de</strong>r cumulant of the pdf of r<strong>et</strong>urns of the ass<strong>et</strong> i. In or<strong>de</strong>r to solve<br />
this problem, l<strong>et</strong> us introduce the Lagrangian<br />
<br />
N<br />
<br />
N<br />
<br />
L = Cn − λ1 wi µ(i) − µ − λ2 wi − 1 , (92)<br />
i=0<br />
where λ1 and λ2 are two Lagrange multipliers. Differenti<strong>at</strong>ing with respect to w0 yields<br />
i=0<br />
λ2 = µ0 λ1, (93)<br />
which by substitution in equ<strong>at</strong>ion (92) gives<br />
<br />
N<br />
<br />
L = Cn − λ1 wi (µ(i) − µ0) − (µ − µ0) . (94)<br />
i=1<br />
L<strong>et</strong> us now differenti<strong>at</strong>e L with respect to wi, i ≥ 1, we obtain<br />
so th<strong>at</strong><br />
n w ∗ i n−1 Cn(i) − λ1(µ(i) − µ0) = 0, (95)<br />
w ∗ 1<br />
i = λ1<br />
n−1<br />
Applying the normaliz<strong>at</strong>ion constraint yields<br />
1<br />
w0 + λ1<br />
n−1<br />
N<br />
i=1<br />
1<br />
µ(i) − µ0<br />
n−1<br />
. (96)<br />
n Cn(i)<br />
1<br />
µ(i) − µ0<br />
n−1<br />
n Cn(i)<br />
30<br />
= 1, (97)
thus<br />
and finally<br />
1<br />
λ1<br />
n−1 =<br />
w ∗ i = (1 − w0)<br />
N<br />
i=1<br />
N<br />
i=1<br />
1 − w0<br />
µ(i)−µ0<br />
n Cn(i)<br />
1<br />
µ(i)−µ0 n−1<br />
Cn(i)<br />
µ(i)−µ0<br />
Cn(i)<br />
L<strong>et</strong> us now <strong>de</strong>fine the portfolio Π exclusively ma<strong>de</strong> of risky ass<strong>et</strong>s with weights<br />
˜wi =<br />
N<br />
i=1<br />
1<br />
µ(i)−µ0 n−1<br />
Cn(i)<br />
µ(i)−µ0<br />
Cn(i)<br />
455<br />
1<br />
n−1<br />
, (98)<br />
1<br />
n−1<br />
. (99)<br />
1<br />
n−1<br />
, i ≥ 1. (100)<br />
The optimal portfolio P can be split in two funds : the risk-free ass<strong>et</strong> whose weight is w0 and a risky fund<br />
Π with weight (1 − w0). The expected r<strong>et</strong>urn of the portfolio P is thus<br />
where µΠ <strong>de</strong>notes the expected r<strong>et</strong>urn of portofolio Π:<br />
µΠ =<br />
µ = w0 µ0 + (1 − w0)µΠ, (101)<br />
N<br />
i=1 µ(i)<br />
N<br />
i=1<br />
1<br />
µ(i)−µ0 n−1<br />
Cn(i)<br />
µ(i)−µ0<br />
Cn(i)<br />
1<br />
n−1<br />
The risk associ<strong>at</strong>ed with P and measured by the cumulant Cn of or<strong>de</strong>r n is<br />
Cn = (1 − w0) n<br />
N i=1 Cn(i)<br />
N<br />
i=1<br />
n<br />
µ(i)−µ0 n−1<br />
Cn(i)<br />
µ(i)−µ0<br />
Cn(i)<br />
1<br />
n−1<br />
. (102)<br />
n . (103)<br />
Putting tog<strong>et</strong>her the three last equ<strong>at</strong>ions allows us to obtain the equ<strong>at</strong>ion of the efficient frontier:<br />
µ = µ0 +<br />
which is a straight line in the plane (Cn 1/n , µ).<br />
B.2 General case<br />
(µ(i) − µ0) n<br />
n−1<br />
Cn(i) 1<br />
n−1<br />
n−1<br />
n<br />
1<br />
· Cn n , (104)<br />
L<strong>et</strong> us now consi<strong>de</strong>r the more realistic case when the risky ass<strong>et</strong>s are <strong>de</strong>pen<strong>de</strong>nt and/or when the risk is<br />
measured by any risk measure ρα obeying the axioms I-IV presented in section 2, where α <strong>de</strong>notes the<br />
<strong>de</strong>gres of homogeneity of ρα. Equ<strong>at</strong>ion (94) always holds (with Cn replaced by ρα), and the differenti<strong>at</strong>ion<br />
with respect to wi, i ≥ 1 yields the s<strong>et</strong> of equ<strong>at</strong>ions:<br />
∂ρα<br />
∂wi<br />
(w ∗ 1, · · · , w ∗ N) = λ1 (µ(i) − µ0), i ∈ {1, · · · , N}. (105)<br />
31
456 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
Since ρα(w1, · · · , wN) is a homogeneous function of or<strong>de</strong>r α, its first-or<strong>de</strong>r <strong>de</strong>riv<strong>at</strong>ive with respect to wi is<br />
also a homogeneous function of or<strong>de</strong>r α − 1. Using this homogeneity property allows us to write<br />
∂ρα<br />
∂wi<br />
−1 ∂ρα<br />
λ1 (w<br />
∂wi<br />
∗ 1, · · · , w ∗ N) = (µ(i) − µ0), i ∈ {1, · · · , N}, (106)<br />
1<br />
−<br />
λ1<br />
α−1 w ∗ 1<br />
−<br />
1, · · · , λ1<br />
α−1 w ∗ <br />
N = (µ(i) − µ0), i ∈ {1, · · · , N} . (107)<br />
Denoting by { ˆw1, · · · , ˆwN} the solution of<br />
this shows th<strong>at</strong> the optimal weights are<br />
∂ρα<br />
( ˆw1, · · · , ˆwN) = (µ(i) − µ0), i ∈ {1, · · · , N}, (108)<br />
∂wi<br />
w ∗ i = λ1<br />
1<br />
α−1 ˆwi. (109)<br />
Now, performing the same calcul<strong>at</strong>ion as in the case of in<strong>de</strong>pen<strong>de</strong>nt risky ass<strong>et</strong>s, the efficient portfolio P<br />
can be realized by investing a weight w0 of the initial wealth in the risk-free ass<strong>et</strong> and the weight (1 − w0)<br />
in the risky fund Π, whose weights are given by<br />
˜wi =<br />
Therefore, the expected r<strong>et</strong>urn of every efficient portfolio is<br />
ˆwi<br />
N i=1 ˆwi<br />
. (110)<br />
µ = w0 · µ0 + (1 − w0) · µΠ, (111)<br />
where µΠ <strong>de</strong>notes the expected r<strong>et</strong>urn of the mark<strong>et</strong> portfolio Π, while the risk, measured by ρα is<br />
so th<strong>at</strong><br />
ρα = (1 − w0) α ρα(Π) , (112)<br />
µ = µ0 + µΠ − µ0<br />
ρα(Π) 1/α ρα 1/α . (113)<br />
This expression is the n<strong>at</strong>ural generaliz<strong>at</strong>ion of the rel<strong>at</strong>ion obtained by (Markovitz 1959) for mean-variance<br />
efficient portfolios.<br />
32
C Composition of the mark<strong>et</strong> portfolio<br />
In this appendix, we <strong>de</strong>rive the rel<strong>at</strong>ionship b<strong>et</strong>ween the composition of the mark<strong>et</strong> portfolio and the composition<br />
of the optimal portfolio Π obtained by the minimiz<strong>at</strong>ion of the risks measured by ρα(n).<br />
C.1 Homogeneous case<br />
We first consi<strong>de</strong>r a homogeneous mark<strong>et</strong>, peopled with agents choosing their optimal portfolio with respect<br />
to the same risk measure ρα. A given agent p invests a fraction w0(p) of his wealth W (p) in the risk-free<br />
ass<strong>et</strong> and a fraction 1 − w0(p) in the optimal portfolio Π. Therefore, the total <strong>de</strong>mand Di of ass<strong>et</strong> i is the<br />
sum of the <strong>de</strong>mand Di(p) over all agents p in ass<strong>et</strong> i:<br />
Di = <br />
Di(p) , (114)<br />
p<br />
457<br />
= <br />
W (p) · (1 − w0(p)) · ˜wi , (115)<br />
p<br />
= ˜wi · <br />
W (p) · (1 − w0(p)) , (116)<br />
p<br />
where the ˜wi’s are given by (110). The aggreg<strong>at</strong>ed <strong>de</strong>mand D over all ass<strong>et</strong>s is<br />
D = <br />
Di, (117)<br />
i<br />
= <br />
˜wi · <br />
W (p) · (1 − w0(p)), (118)<br />
i<br />
p<br />
= <br />
W (p) · (1 − w0(p)). (119)<br />
p<br />
By <strong>de</strong>finition, the weight of ass<strong>et</strong> i, <strong>de</strong>noted by wm i , in the mark<strong>et</strong> portfolio equals the r<strong>at</strong>io of its capitaliz<strong>at</strong>ion<br />
(the supply Si of ass<strong>et</strong> i) over the total capitaliz<strong>at</strong>ion of the mark<strong>et</strong> S = Si. At the equilibrium,<br />
<strong>de</strong>mand equals supply, so th<strong>at</strong><br />
w m i = Si Di<br />
=<br />
S D = ˜wi. (121)<br />
Thus, <strong>at</strong> the equilibrium, the optimal portfolio Π is the mark<strong>et</strong> portfolio.<br />
C.2 H<strong>et</strong>erogeneous case<br />
(120)<br />
We now consi<strong>de</strong>r a h<strong>et</strong>erogenous mark<strong>et</strong>, <strong>de</strong>fined such th<strong>at</strong> the agents choose their optimal portfolio with<br />
respect to different risk measures. Some of them choose the usual mean-variance optimal portfolios, others<br />
prefer any mean-ρα efficient portfolio, and so on. L<strong>et</strong> us <strong>de</strong>note by Πn the mean-ρα(n) optimal portfolio<br />
ma<strong>de</strong> only of risky ass<strong>et</strong>s. L<strong>et</strong> φn be the fraction of agents who choose the mean-ρα(n) efficient portfolios.<br />
By normaliz<strong>at</strong>ion, <br />
n φn = 1. The <strong>de</strong>mand Di(n) of ass<strong>et</strong> i from the agents optimizing with respect to<br />
ρα(n) is<br />
Di(n) = <br />
W (p) · (1 − w0(p)) · ˜wi(n), (122)<br />
p∈Sn<br />
= ˜wi(n) <br />
W (p) · (1 − w0(p)), (123)<br />
p∈Sn<br />
33
458 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
where Sn <strong>de</strong>notes the s<strong>et</strong> of agents, among all the agents, who follow the optimiz<strong>at</strong>ion stragtegy with respect<br />
to ρα(n). Thus, the total <strong>de</strong>mand of ass<strong>et</strong> i is<br />
Di = <br />
N φn · Di(n), (124)<br />
n<br />
= N <br />
φn · ˜wi(n) <br />
W (p) · (1 − w0(p)), (125)<br />
n<br />
where N is the total number of agents. This finally yields the total <strong>de</strong>mand D for all ass<strong>et</strong>s and for all agents<br />
D = <br />
Di, (126)<br />
i<br />
p∈Sn<br />
= N <br />
φn · ˜wi(n) <br />
W (p) · (1 − w0(p)), (127)<br />
i<br />
= N <br />
n<br />
n<br />
φn<br />
p∈Sn<br />
<br />
W (p) · (1 − w0(p)), (128)<br />
p∈Sn<br />
since <br />
i ˜wi(n) = 1, for every n. Thus, s<strong>et</strong>ting<br />
<br />
γn = φn p∈Sn <br />
n φn<br />
<br />
p∈Sn<br />
W (p) · (1 − w0(p))<br />
, (129)<br />
W (p) · (1 − w0(p))<br />
the mark<strong>et</strong> portfolio is the weighted sum of the mean-ρα(n) optimal portfolios Πn:<br />
w m i = Si<br />
S<br />
= Di<br />
D<br />
<br />
= γn · ˜wi(n) . (130)<br />
34<br />
n
D Generalized capital ass<strong>et</strong> princing mo<strong>de</strong>l<br />
Our proof of the generalized capital ass<strong>et</strong> princing mo<strong>de</strong>l is similar to the usual <strong>de</strong>montr<strong>at</strong>ion of the CAPM.<br />
L<strong>et</strong> us consi<strong>de</strong>r an efficient portfolio P. It necessarily s<strong>at</strong>isfies equ<strong>at</strong>ion (105) in appendix B :<br />
459<br />
∂ρα<br />
(w<br />
∂wi<br />
∗ 1, · · · , w ∗ N) = λ1 (µ(i) − µ0), i ∈ {1, · · · , N}. (131)<br />
L<strong>et</strong> us now choose any portfolio R ma<strong>de</strong> only of risky ass<strong>et</strong>s and l<strong>et</strong> us <strong>de</strong>note by wi(R) its weights. We<br />
can thus write<br />
N<br />
i=1<br />
wi(R) · ∂ρα<br />
(w<br />
∂wi<br />
∗ 1, · · · , w ∗ N) = λ1<br />
N<br />
wi(R) · (µ(i) − µ0), (132)<br />
i=1<br />
= λ1 (µR − µ0). (133)<br />
We can apply this last rel<strong>at</strong>ion to the mark<strong>et</strong> portfolio Π, because it is only composed of risky ass<strong>et</strong>s (as<br />
proved in appendix B). This leads to wi(R) = w ∗ i and µR = µΠ, so th<strong>at</strong><br />
N<br />
i=1<br />
which, by the homogeneity of the risk measures ρα, yields<br />
Substituting equ<strong>at</strong>ion (131) into (135) allows us to obtain<br />
where<br />
w ∗ i · ∂ρα<br />
(w<br />
∂wi<br />
∗ 1, · · · , w ∗ N) = λ1 (µΠ − µ0), (134)<br />
α · ρα(w ∗ 1, · · · , w ∗ N) = λ1 (µΠ − µ0) . (135)<br />
µj − µ0 = β j α · (µΠ − µ0), (136)<br />
β j α =<br />
<br />
∂<br />
1<br />
ln ρα α<br />
∂wj<br />
<br />
, (137)<br />
calcul<strong>at</strong>ed <strong>at</strong> the point {w∗ 1 , · · · , w∗ N }. Expression (135) with (137) provi<strong>de</strong>s our CAPM, generalized with<br />
respect to the risk measures ρα.<br />
In the case where ρα <strong>de</strong>notes the variance, the second-or<strong>de</strong>r centered moment is equal to the second-or<strong>de</strong>r<br />
cumulant and reads<br />
Since<br />
we find<br />
C2 = w ∗ 1 · Var[X1] + 2w ∗ 1w ∗ 2 · Cov(X1, X2) + w ∗ 2 · Var[X2], (138)<br />
= Var[Π] . (139)<br />
1 ∂C2<br />
·<br />
2 ∂w1<br />
= w ∗ 1 · Var[X1] + w ∗ 2 · Cov(X1, X2) , (140)<br />
= Cov(X1, Π), (141)<br />
β = Cov(X1, XΠ)<br />
Var[XΠ]<br />
which is the standard result of the CAPM <strong>de</strong>rived from the mean-variance theory.<br />
35<br />
, (142)
460 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
E Calcul<strong>at</strong>ion of the moments of the distribution of portfolio r<strong>et</strong>urns<br />
L<strong>et</strong> us start with equ<strong>at</strong>ion (72) in the 2-ass<strong>et</strong> case :<br />
ˆPS(k)<br />
1<br />
=<br />
2π 1 − ρ2 <br />
<br />
dy1dy2 exp − 1<br />
2 ytV −1 <br />
<br />
y1 <br />
y + ik χ1w1sgn(y1) <br />
√<br />
<br />
2<br />
<br />
<br />
y2 <br />
+χ2w2sgn(y2) <br />
√<br />
<br />
2<br />
<br />
Expanding the exponential and using the <strong>de</strong>finition (67) of moments, we g<strong>et</strong><br />
Posing<br />
1<br />
Mn =<br />
2π 1 − ρ2 <br />
dy1 dy2<br />
γq1q2 (n, p) = χ1 p χ2 n−p<br />
this leads to<br />
2π 1 − ρ 2<br />
<br />
n<br />
p=0<br />
<br />
n<br />
χ<br />
p<br />
p<br />
dy1dy2 sgn(y1) p<br />
<br />
y1 <br />
<br />
√<br />
<br />
2<br />
<br />
Mn =<br />
1χn−p 2 wp 1wn−p 2<br />
×sgn(y2) n−p<br />
q1p<br />
<br />
y2 <br />
<br />
√<br />
<br />
2<br />
<br />
sgn(y1) p<br />
<br />
y1 <br />
<br />
√<br />
<br />
2<br />
<br />
q2(n−p)<br />
sgn(y2) n−p<br />
<br />
y2 <br />
<br />
√<br />
<br />
2<br />
<br />
L<strong>et</strong> us <strong>de</strong>fined the auxiliary variables α and β such th<strong>at</strong><br />
α = (V −1 )11 = (V −1 )22 = 1<br />
1−ρ 2 ,<br />
β = −(V −1 )12 = −(V −1 )21 = ρ<br />
1−ρ 2 .<br />
n<br />
p=0<br />
q1<br />
+<br />
q2 <br />
q1p<br />
×<br />
. (143)<br />
1<br />
−<br />
e 2 ytV −1y . (144)<br />
q2(n−p)<br />
1<br />
−<br />
e 2 ytV −1y , (145)<br />
<br />
n<br />
w<br />
p<br />
p<br />
1wn−pγq1q2 2 (n, p) . (146)<br />
(147)<br />
Performing a simple change of variable in (145), we can transform the integr<strong>at</strong>ion such th<strong>at</strong> it is <strong>de</strong>fined<br />
solely within the first quadrant (y1 ≥ 0, y2 ≥ 0), namely<br />
γq1q2 (n, p) = χ1 p n−p 1 + (−1)n<br />
χ2<br />
2π 1 − ρ2 +∞ +∞<br />
dy1 dy2<br />
0<br />
0<br />
q1p q2(n−p) y1 y2<br />
α<br />
− √2 √2 e 2 (y2 1 +y2 2 ) ×<br />
<br />
× e βy1y2<br />
<br />
p −βy1y2<br />
+ (−1) e<br />
. (148)<br />
This equ<strong>at</strong>ion imposes th<strong>at</strong> the coefficients γ vanish if n is odd. This leads to the vanishing of the moments<br />
of odd or<strong>de</strong>rs, as expected for a symm<strong>et</strong>ric distribution. Then, we expand e βy1y2 + (−1) p e −βy1y2 in series.<br />
Permuting the sum sign and the integral allows us to <strong>de</strong>couple the integr<strong>at</strong>ions over the two variables y1 and<br />
y2:<br />
γq1q2 (n, p) = χ1 p n−p 1 + (−1)n<br />
χ2<br />
2π 1 − ρ2 +∞<br />
[1 + (−1)<br />
s=0<br />
p+s ] βs<br />
s!<br />
36<br />
+∞<br />
dy1<br />
0<br />
+∞<br />
× dy2<br />
0<br />
y q1p+s<br />
1<br />
2 q1p 2<br />
y q2(n−p)+s<br />
2<br />
2 q2 (n−p)<br />
2<br />
α<br />
−<br />
e 2 y2 1<br />
<br />
α<br />
−<br />
e 2 y2 1<br />
<br />
×<br />
. (149)
This brings us back to the problem of calcul<strong>at</strong>ing the same type of integrals as in the uncorrel<strong>at</strong>ed case.<br />
Using the expressions of α and β, and taking into account the parity of n and p, we obtain:<br />
γq1q2 (2n, 2p) = χ1 2p χ2 2n−2p (1 − ρ2 1<br />
q1p+q2(n−p)+ ) 2<br />
π<br />
<br />
×Γ q2(n − p) + s + 1<br />
<br />
2<br />
γq1q2 (2n, 2p + 1) = χ1 2p+1 χ2 2n−2p−1 (1 − ρ2 ) q1p+q2(n−p)+ q1−q2 +1<br />
2<br />
<br />
×Γ<br />
q1p + s + 1 + q1<br />
2<br />
<br />
Γ<br />
+∞<br />
s=0<br />
(2ρ) 2s<br />
(2s)! Γ<br />
<br />
q1p + s + 1<br />
<br />
×<br />
2<br />
461<br />
, (150)<br />
+∞<br />
(2ρ)<br />
π<br />
s=0<br />
2s+1<br />
(2s + 1)! ×<br />
q2(n − p) + s + 1 − q2<br />
<br />
. (151)<br />
2<br />
Using the <strong>de</strong>finition of the hypergeom<strong>et</strong>ric functions 2F1 (Abramovitz and Stegun 1972), and the rel<strong>at</strong>ion<br />
(9.131) of (Gradshteyn and Ryzhik 1965), we finally obtain<br />
γq1q2 (2n, 2p) = χ1 2p χ2 2n−2p Γ q1p + 1<br />
<br />
2 Γ q2(n − p) + 1<br />
<br />
2<br />
2F1 −q1p, −q2(n − p);<br />
π<br />
1<br />
<br />
; ρ2(152)<br />
,<br />
2<br />
γq1q2 (2n, 2p + 1) = χ1 2p+1 χ2 2n−2p−1 2Γ q1p + 1 + q1<br />
<br />
2 Γ q2(n − p) + 1 − q2<br />
<br />
2 ρ ×<br />
<br />
π<br />
× 2F1 −q1p − q1 − 1<br />
, −q2(n − p) +<br />
2<br />
q2 + 1<br />
;<br />
2<br />
3<br />
<br />
; ρ2 . (153)<br />
2<br />
In the asymm<strong>et</strong>ric case, a similar calcul<strong>at</strong>ion follows, with the sole difference th<strong>at</strong> the results involves four<br />
terms in the integral (148) instead of two.<br />
37
462 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
F Conditions un<strong>de</strong>r which it is possible to increase the r<strong>et</strong>urn and <strong>de</strong>crease<br />
large risks simultaneously<br />
We consi<strong>de</strong>r N in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s {1 · · · N}, whose r<strong>et</strong>urns are <strong>de</strong>noted by µ(1) · · · µ(N). We aggreg<strong>at</strong>e<br />
these ass<strong>et</strong>s in a portfolio. L<strong>et</strong> w1 · · · wN be their weights. We consi<strong>de</strong>r th<strong>at</strong> short positions are forbid<strong>de</strong>n<br />
and th<strong>at</strong> <br />
i wi = 1. The r<strong>et</strong>urn µ of the portfolio is<br />
µ =<br />
The risk of the portfolio is quantified by the cumulants of the distribution of µ.<br />
N<br />
wiµ(i). (154)<br />
i=1<br />
L<strong>et</strong> us <strong>de</strong>note µ ∗ n the r<strong>et</strong>urn of the portfolio evalu<strong>at</strong>ed for the ass<strong>et</strong> weights which minimize the cumulant of<br />
or<strong>de</strong>r n.<br />
F.1 Case of two ass<strong>et</strong>s<br />
L<strong>et</strong> Cn be the cumulant of or<strong>de</strong>r n for the portfolio. The ass<strong>et</strong>s being in<strong>de</strong>pen<strong>de</strong>nt, we have<br />
Cn = Cn(1)w1 n + Cn(2)w2 n , (155)<br />
= Cn(1)w1 n + Cn(2)(1 − w1) n . (156)<br />
In the following, we will drop the subscript 1 in w1, and only write w. L<strong>et</strong> us evalu<strong>at</strong>e the value w = w ∗ <strong>at</strong><br />
the minimum of Cn, n > 2 :<br />
dCn<br />
dw = 0 ⇐⇒ Cn(1)w n−1 − Cn(2)(1 − w) n−1 = 0, (157)<br />
⇐⇒ Cn(1)<br />
Cn(2) =<br />
1 − w ∗<br />
w ∗<br />
n−1<br />
, (158)<br />
and assuming th<strong>at</strong> Cn(1)/Cn(2) > 0, which is s<strong>at</strong>isfied according to our positivity axiom 1, we obtain<br />
w ∗ =<br />
This leads to the following expression for µ ∗ n :<br />
µ ∗ n =<br />
Cn(2) 1<br />
n−1<br />
Cn(1) 1<br />
n−1 + Cn(2) 1<br />
n−1<br />
µ(1) · Cn(2) 1<br />
n−1 + µ(2) · Cn(1) 1<br />
n−1<br />
Cn(1) 1<br />
n−1 + Cn(2) 1<br />
n−1<br />
Thus, after simple algebraic manipul<strong>at</strong>ions, we find<br />
<br />
⇐⇒ (µ(1) − µ(2)) Cn(1) 1<br />
n−1 Cn+k(2)<br />
µ ∗ n < µ ∗ n+k<br />
1<br />
n+k−1 − Cn(2) 1<br />
. (159)<br />
. (160)<br />
n−1 Cn+k(1)<br />
which conclu<strong>de</strong>s the proof of the result announced in the main body of the text.<br />
38<br />
1<br />
n+k−1<br />
<br />
> 0, (161)
F.2 General case<br />
We now consi<strong>de</strong>r a portfolio with N in<strong>de</strong>pen<strong>de</strong>nt ass<strong>et</strong>s. Assuming th<strong>at</strong> the cumulants Cn(i) have the same<br />
sign for all i (according to axiom 1), we are going to show th<strong>at</strong> the minimum of Cn is obtained for a portfolio<br />
whose weights are given by<br />
and we have<br />
µ ∗ n =<br />
wi =<br />
<br />
N<br />
i=1<br />
In<strong>de</strong>ed, the cumulant of the portfolio is given by<br />
subject to the constraint<br />
Cn =<br />
1<br />
N<br />
j=i Cn(j) n−1<br />
1<br />
N<br />
j=1 Cn(j) n−1<br />
µ(i) N<br />
j=i<br />
1<br />
N<br />
j=1 Cn(j) n−1<br />
N<br />
i=1<br />
Cn(i) w n i<br />
463<br />
, (162)<br />
1<br />
Cn(j) n−1<br />
Introducing a Lagrange multiplier λ, the first or<strong>de</strong>r conditions yields<br />
so th<strong>at</strong><br />
Cn(i) w n−1<br />
i<br />
. (163)<br />
(164)<br />
N<br />
wi = 1. (165)<br />
i=1<br />
− λ = 0, ∀i ∈ {1, · · · , N}, (166)<br />
w n−1<br />
i<br />
λ<br />
= . (167)<br />
Cn(i)<br />
Since all the Cn(i) are positive, we can find a λ such th<strong>at</strong> all the wi are real and positive, which yields the<br />
announced result (162). From here, there is no simple condition th<strong>at</strong> ensures µ ∗ n < µ ∗ n+k . The simplest way<br />
is to calcul<strong>at</strong>e dir<strong>et</strong>ly these quantities using the formula (163).<br />
to compare µ ∗ n and µ ∗ n+k<br />
39
464 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
References<br />
Acerbi, C., 2002, Spectrale measures of risk : a coherent represent<strong>at</strong>ion of subjective risk aversion, Journal<br />
of Banking and Finance.<br />
Alexan<strong>de</strong>r, G.J. and A.M. Baptista, 2002, Economic implic<strong>at</strong>ions of using a mean-VaR mo<strong>de</strong>l for portfolio<br />
selection: A comparison with mean-variance analysis, Journal of Economic Dynamics and Control<br />
26, 1159-1193.<br />
Abramovitz, E. and I.A. Stegun, 1972, Handbook of M<strong>at</strong>hem<strong>at</strong>ical functions (Dover Public<strong>at</strong>ions, New<br />
York).<br />
Allais, M., 1953, Le comportement <strong>de</strong> l’homme r<strong>at</strong>ionel <strong>de</strong>vant le risque, critique <strong>de</strong>s postul<strong>at</strong> <strong>de</strong> l’école<br />
américaine, Econom<strong>et</strong>rica 21, 503-546.<br />
Allais, M., 1990, Allais Paradox, in The new palgrave, Utility and probability, Macmillan, 3-9.<br />
An<strong>de</strong>rsen, J.V., and D. Sorn<strong>et</strong>te, 2001, Have your cake and e<strong>at</strong> it too: increasing r<strong>et</strong>urns while lowering large<br />
risks! Journal of Risk Finance 2, 70-82.<br />
Artzner, P., F. Delbaen, J.M. Eber and D. He<strong>at</strong>h, 1997, Thinking coherently, Risk 10, 68-71.<br />
Artzner, P., F. Delbaen, J.M. Eber and D. He<strong>at</strong>h, 1999, Coherent measures of risk, M<strong>at</strong>h. Finance 9, 203-228.<br />
Bouchaud, J.-P., D. Sorn<strong>et</strong>te, C. Walter and J.-P. Aguilar, 1998, Taming large events: Optimal portfolio<br />
theory for strongly fluctu<strong>at</strong>ing ass<strong>et</strong>s, Intern<strong>at</strong>ional Journal of Theor<strong>et</strong>ical and Applied Finance 1,<br />
25-41.<br />
Dowd, K., 2000, Adjusting for risk: an improved Sharpe r<strong>at</strong>io, Intern<strong>at</strong>ional Review of Economics and<br />
Finance 9, 209-222.<br />
Embrechts, P., C. Kluppelberg and T. Mikosh, 1997, Mo<strong>de</strong>lling extremal events (Springel-Verlag, Applic<strong>at</strong>ions<br />
of M<strong>at</strong>hem<strong>at</strong>ics 33).<br />
Embrecht, P., A. McNeil and D. Straumann, 1998, Correl<strong>at</strong>ion and Depen<strong>de</strong>nce in risk management: properties<br />
and pitfalls, Proceedings of the Risk Management workshop <strong>at</strong> the Newton Institute Cambridge,<br />
Cambridge University Press.<br />
Fang, H. and T. Lai, 1997, Co-kurtosis and capital ass<strong>et</strong> pricing, Financial Review 32, 293-307.<br />
Föllmer, H. and A. Schied, 2002, Convex measures of risk and trading constraints, Finance and Stochastics<br />
6, forthcoming.<br />
Föllmer, H. and A. Schied, 2002, Robust preferences and convex measures of risk, Working paper.<br />
Frisch, U. and D. Sorn<strong>et</strong>te, 1997, Extreme Devi<strong>at</strong>ions and Applic<strong>at</strong>ions, J. Phys. I France 7, 1155-1171.<br />
Gopikrishnan, P., M. Meyer, L.A. Nunes Amaral and H.E. Stanley,1998, Inverse cubic law for the distribution<br />
of stock price vari<strong>at</strong>ions, European Physical Journal B 3, 139-140.<br />
Gouriéroux, C., J.P. Laurent and O. Scaill<strong>et</strong>, 2000, Sensitivity analysis of Values <strong>at</strong> Risk, Journal of Empirical<br />
Finance 7,225-245.<br />
Gradshteyn, I.S. and I. M. Ryzhik, 1965, Table of integrals Series and Products, Aca<strong>de</strong>mic Press.<br />
40
Harvey, C.R. and A. Siddique, 2000, Conditional skewness in ass<strong>et</strong> pricing tests, Journal of Finance 55,<br />
1263-1295.<br />
Hill, B.M., 1975, A Simple General Approach to Inference about the Tail of a Distribution, Annals of<br />
st<strong>at</strong>istics, 3(5), 1163-1174.<br />
Hwang, S. and S. S<strong>at</strong>chell, 1999, Mo<strong>de</strong>lling emerging mark<strong>et</strong> risk premia using higher moments, Intern<strong>at</strong>ional<br />
Journal of Finance and Economics 4, 271-296.<br />
Jorion, P., 1997, Value-<strong>at</strong>-Risk: The New Benchmark for Controlling Deriv<strong>at</strong>ives Risk (Irwin Publishing,<br />
Chicago, IL).<br />
Jurcenko, E. and B. Maill<strong>et</strong>, 2002, The four-moment capital ass<strong>et</strong> pricing mo<strong>de</strong>l: some basic results, Working<br />
Paper.<br />
Karlen, D., 1998, Using projection and correl<strong>at</strong>ion to approxim<strong>at</strong>e probability distributions, Computer in<br />
Physics 12, 380-384.<br />
Krauss, A. and R. Litzenberger, 1976, Skewness preference and the valu<strong>at</strong>ion of risk ass<strong>et</strong>s, Journal of<br />
Finance 31, 1085-1099.<br />
Laherrère, J. and D. Sorn<strong>et</strong>te, 1998, Str<strong>et</strong>ched exponential distributions in n<strong>at</strong>ure and economy : ”f<strong>at</strong> tails”<br />
with characteristic scales, European Physical Journal B 2, 525-539.<br />
Lim, K.G., 1989, A new test for the three-moment capital ass<strong>et</strong> pricing mo<strong>de</strong>l, Journal of Financial and<br />
Quantit<strong>at</strong>ive Analysis 24, 205-216.<br />
Lindskog, F., 2000, Mo<strong>de</strong>lling Depen<strong>de</strong>nce with Copulas,<br />
http : //www.risklab.ch/P apers.html#MT Lindskog<br />
Lintner, J. 1975, The valu<strong>at</strong>ion of risk ass<strong>et</strong>s and the selection of risky investments in stock portfolios and<br />
capital budg<strong>et</strong>s. Review of Economics and St<strong>at</strong>istics 13, 13-37.<br />
Litterman, R. and K. Winkelmann, 1998, Estim<strong>at</strong>ing covariance m<strong>at</strong>rices (Risk Management Series, Goldman<br />
Sachs).<br />
Lux, T., 1996, The stable Par<strong>et</strong>ian hypothesis and the frequency of large r<strong>et</strong>urns: an examin<strong>at</strong>ion of major<br />
German stocks, Applied Financial Economics 6, 463-475.<br />
Malevergne, Y. and D. Sorn<strong>et</strong>te, 2001, Testing the Gaussian copula hypothesis for financial ass<strong>et</strong>s <strong>de</strong>pen<strong>de</strong>nces.<br />
http : //papers.ssrn.com/sol3/papers.cfm?abstract id = 291140<br />
Markovitz, H., 1959, Portfolio selection : Efficient diversific<strong>at</strong>ion of investments (John Wiley and Sons,<br />
New York).<br />
Merton, R.C., 1990, Continuous-time finance, (Blackwell, Cambridge).<br />
Mossin, J., 1966, Equilibrium in a capital mark<strong>et</strong>, Econom<strong>et</strong>rica 34, 768-783.<br />
Muzy, J.-F., D. Sorn<strong>et</strong>te, J. Delour and A. Arneodo, 2001, Multifractal r<strong>et</strong>urns and Hierarchical Portfolio<br />
Theory, Quantit<strong>at</strong>ive Finance 1 (1), 131-148.<br />
Nelsen, R.B. 1998, An Introduction to Copulas. Lectures Notes in st<strong>at</strong>istic, 139, Springer Verlag, New York.<br />
Pagan, A., 1996, The Econom<strong>et</strong>rics of Financial Mark<strong>et</strong>s, Journal of Empirical Finance, 3, 15 - 102.<br />
41<br />
465
466 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
Pickhands, J., 1975, St<strong>at</strong>istical Inference Using Extreme Or<strong>de</strong>r St<strong>at</strong>itstics, Annals of St<strong>at</strong>istics, 3(1), 119-<br />
131.<br />
Polimenis, V., 2002, The distributional CAPM: Connecting risk premia to r<strong>et</strong>urn distributions. Working<br />
Parper<br />
Rao, C.R., 1973, Linear st<strong>at</strong>istical inference and its applic<strong>at</strong>ions, 2d ed. (New York Willey).<br />
Richardson M. and T. Smith, 1993, A test for multivari<strong>at</strong>e normality in stocks, Journal of Business 66,<br />
295-321.<br />
Rothschild, M; and J.E. Stiglitz, 1970, Increasing risk I: A <strong>de</strong>finition, Journal of Economic Theory 2, 225-<br />
243.<br />
Rothschild, M; and J.E. Stiglitz, 1971, Increasing risk II: Its economic consequences, Journal of Economic<br />
Theory 3, 66-84.<br />
Rubinstein, M., 1973, The fundamental theorem of param<strong>et</strong>er-preference security valu<strong>at</strong>ion. Journal of Financial<br />
and Quantit<strong>at</strong>ive Analysis 8, 61-69.<br />
Scaill<strong>et</strong>, O., 2000, Nonparam<strong>et</strong>ric estim<strong>at</strong>ion and sensitivity analysis of expected shortfall, Working paper.<br />
Sharpe, N.F., 1964, Capital ass<strong>et</strong> prices: A theory of mark<strong>et</strong> equilibrium un<strong>de</strong>r conditions of risk, Journal of<br />
Finance, 425-442.<br />
Sharpe, W.F., 1966, Mutual fund performance, Journal of Business 39, 119-138.<br />
Sharpe, W.F., 1994, The Sharpe r<strong>at</strong>io, Journal of Portfolio Management, 49-58.<br />
Sorn<strong>et</strong>te, D., 1998, Large <strong>de</strong>vi<strong>at</strong>ions and portfolio optimiz<strong>at</strong>ion, Physica A 256, 251-283.<br />
orn<strong>et</strong>te, D., 2000, Critical Phenomena in N<strong>at</strong>ural Sciences, Chaos, Fractals, Self-organiz<strong>at</strong>ion and Disor<strong>de</strong>r:<br />
Concepts and Tools, (Springer Series in Synerg<strong>et</strong>ics).<br />
Sorn<strong>et</strong>te, D., J. V. An<strong>de</strong>rsen and P. Simon<strong>et</strong>ti, 2000a, Portfolio Theory for “F<strong>at</strong> Tails”, Intern<strong>at</strong>ional Journal<br />
of Theor<strong>et</strong>ical and Applied Finance 3 (3), 523-535.<br />
Sorn<strong>et</strong>te, D., P. Simon<strong>et</strong>ti, J.V. An<strong>de</strong>rsen, 2000b, φ q -field theory for portfolio optimiz<strong>at</strong>ion : ”f<strong>at</strong>-tails” and<br />
non-linear correl<strong>at</strong>ions, Physics Reports 335 (2), 19-92.<br />
Stuart, A. and J.K. Ord, , Kendall’s advanced theory of st<strong>at</strong>istics, 1994, 6th edition, Edward Arnold London,<br />
Halsted Press, New York.<br />
Tsallis, C., 1998, Possible generaliz<strong>at</strong>ion of Boltzmann-Gibbs st<strong>at</strong>istics, J. St<strong>at</strong>. Phys. 52, 479-487; for<br />
upd<strong>at</strong>ed bibliography on this subject, see http : //tsallis.c<strong>at</strong>.cbpf.br/biblio.htm.<br />
Von Neuman, J. and O. Morgenstern, 1944, Theory of games and economic behavior, Princ<strong>et</strong>own University<br />
Press.<br />
42
µ µ2 1/2 µ4 1/4 µ6 1/6 µ8 1/8<br />
0.10% 0.92% 1.36% 1.79% 2.15%<br />
0.12% 0.96% 1.43% 1.89% 2.28%<br />
0.14% 1.05% 1.56% 2.06% 2.47%<br />
0.16% 1.22% 1.83% 2.42% 2.91%<br />
0.18% 1.47% 2.21% 2.92% 3.55%<br />
0.20% 1.77% 2.65% 3.51% 4.22%<br />
Table 1: This table presents the risk measured by µ 1/n<br />
n<br />
(daily) r<strong>et</strong>urn µ.<br />
µ<br />
µ 1/2<br />
2<br />
µ<br />
µ 1/4<br />
4<br />
467<br />
for n=2,4,6,8, for a given value of the expectedd<br />
µ<br />
µ 1/6<br />
6<br />
µ<br />
C 1/4<br />
4<br />
µ<br />
C 1/6<br />
6<br />
Wall Mart 0.0821 0.0555 0.0424 0.0710 0.0557<br />
EMC 0.0801 0.0552 0.0430 0.0730 0.0612<br />
Intel 0.0737 0.0512 0.0397 0.0694 0.0532<br />
Hewl<strong>et</strong>t Packard 0.0724 0.0472 0.0354 0.0575 0.0439<br />
IBM 0.0705 0.0465 0.0346 0.0574 0.0421<br />
Merck 0.0628 0.0415 0.0292 0.0513 0.0331<br />
Procter & Gamble 0.0590 0.0399 0.0314 0.0510 0.0521<br />
General Motors 0.0586 0.0362 0.0247 0.0418 0.0269<br />
SBC Communic<strong>at</strong>ion 0.0584 0.0386 0.0270 0.0477 0.0302<br />
General Electric 0.0569 0.0334 0.0233 0.0373 0.0258<br />
Applied M<strong>at</strong>erial 0.0525 0.0357 0.0269 0.0462 0.0338<br />
MCI WorldCom 0.0441 0.0173 0.0096 0.0176 0.0098<br />
Medtronic 0.0432 0.0278 0.0202 0.0333 0.0237<br />
Coca-Cola 0.0430 0.0278 0.0207 0.0335 0.0252<br />
Exxon-Mobil 0.0410 0.0256 0.0178 0.0299 0.0197<br />
Texas Instrument 0.0324 0.0224 0.0171 0.0301 0.0218<br />
Pfizer 0.0298 0.0184 0.0131 0.0213 0.0148<br />
Table 2: This table presents the values of the generalized Sharpe r<strong>at</strong>ios for the s<strong>et</strong> of seventeen ass<strong>et</strong>s listed<br />
in the first column. The ass<strong>et</strong>s are ranked with respect to their Sharpe r<strong>at</strong>io, given in the second column. The<br />
third and fourth columns give the generalized Sharpe r<strong>at</strong>io calcul<strong>at</strong>ed with respect to the fourth and sixth<br />
centered moments µ4 and µ6 while the fifth and sixth columns give the generalized Sharpe r<strong>at</strong>io calcul<strong>at</strong>ed<br />
with respect to the fourth and sixth cumulants C4 and C6.<br />
43
468 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
Positive Tail Neg<strong>at</strong>ive Tail<br />
< χ+ > < c+ > χ+ c+ < χ− > < c− > χ− c−<br />
CHF 2.45 1.61 2.33 1.26 2.34 1.53 1.72 0.93<br />
DEM 2.09 1.65 1.74 1.03 2.01 1.58 1.45 0.91<br />
JPY 2.10 1.28 1.30 0.76 1.89 1.47 0.99 0.76<br />
MAL 1.00 1.22 1.25 0.41 1.01 1.25 0.44 0.48<br />
POL 1.55 1.02 1.30 0.73 1.60 2.13 1.25 0.62<br />
THA 0.78 0.75 0.75 0.54 0.82 0.73 0.30 0.38<br />
UKP 1.89 1.52 1.38 0.92 2.00 1.41 1.82 1.09<br />
Table 3: Table of the exponents c and the scale param<strong>et</strong>ers χ for different currencies. The subscript ”+”<br />
or ”-” <strong>de</strong>notes the positive or neg<strong>at</strong>ive part of the distribution of r<strong>et</strong>urns and the terms b<strong>et</strong>ween brack<strong>et</strong>s<br />
refer to param<strong>et</strong>ers estim<strong>at</strong>ed in the bulk of the distribution while naked param<strong>et</strong>ers refer to the tails of the<br />
distribution.<br />
Positive Tail Neg<strong>at</strong>ive Tail<br />
< χ+ > < c+ > χ+ c+ < χ− > < c− > χ− c−<br />
Applied M<strong>at</strong>erial 12.47 1.82 8.75 0.99 11.94 1.66 8.11 0.98<br />
Coca-Cola 5.38 1.88 4.46 1.04 5.06 1.74 2.98 0.78<br />
EMC 13.53 1.63 13.18 1.55 11.44 1.61 3.05 0.57<br />
General Electric 5.21 1.89 1.81 1.28 4.80 1.81 4.31 1.16<br />
General Motors 5.78 1.71 0.63 0.48 5.32 1.89 2.80 0.79<br />
Hewl<strong>et</strong>t Packart 7.51 1.93 4.20 0.84 7.26 1.76 1.66 0.52<br />
IBM 5.46 1.71 3.85 0.87 5.07 1.90 0.18 0.33<br />
Intel 8.93 2.31 2.79 0.64 9.14 1.60 3.56 0.62<br />
MCI WorldCom 9.80 1.74 11.01 1.56 9.09 1.56 2.86 0.58<br />
Medtronic 6.82 1.95 6.09 1.11 6.49 1.54 2.55 0.67<br />
Merck 5.36 1.91 4.56 1.16 5.00 1.73 1.32 0.59<br />
Pfizer 6.41 2.01 5.84 1.27 6.04 1.70 0.26 0.35<br />
Procter & Gambel 4.86 1.83 3.53 0.96 4.55 1.74 2.96 0.82<br />
SBC Communic<strong>at</strong>ion 5.21 1.97 1.26 0.59 4.89 1.59 1.56 0.60<br />
Texas Instrument 9.06 1.78 4.07 0.72 8.24 1.84 2.18 0.54<br />
Wall Mart 7.41 1.83 5.81 1.01 6.80 1.64 3.75 0.78<br />
Table 4: Table of the exponents c and the scale param<strong>et</strong>ers χ for different stocks. The subscript ”+” or ”-”<br />
<strong>de</strong>notes the positive or neg<strong>at</strong>ive part of the distribution and the terms b<strong>et</strong>ween brack<strong>et</strong>s refer to param<strong>et</strong>ers<br />
estim<strong>at</strong>ed in the bulk of the distribution while naked param<strong>et</strong>ers refer to the tails of the distribution.<br />
44
Mean (10 −3 ) Variance (10 −3 ) Skewness Kurtosis min max<br />
Applied M<strong>at</strong>erial 2.11 1.62 0.41 4.68 -14% 21%<br />
Coca-Cola 0.81 0.36 0.13 5.71 -11% 10%<br />
EMC 2.76 1.13 0.23 4.79 -18% 15%<br />
Exxon-Mobil 0.92 0.25 0.30 5.26 -7% 11%<br />
General Electric 1.38 0.30 0.08 4.46 -7% 8%<br />
General Motors 0.64 0.39 0.12 4.35 -11% 8%<br />
Hewl<strong>et</strong>t Packard 1.17 0.81 0.16 6.58 -14% 21%<br />
IBM 1.32 0.54 0.08 8.43 -16% 13%<br />
Intel 1.71 0.85 -0.31 6.88 -22% 14%<br />
MCI WorldCom 0.87 0.85 -0.18 6.88 -20% 13%<br />
Medtronic 1.70 0.55 0.23 5.52 -12% 12%<br />
Merck 1.32 0.35 0.18 5.29 -9% 10%<br />
Pfizer 1.57 0.46 0.01 4.28 -10% 10%<br />
Procter&Gambel 0.90 0.41 -2.57 42.75 -31% 10%<br />
SBC Communic<strong>at</strong>ion 0.86 0.39 0.06 5.86 -13% 9%<br />
Texas Instrument 2.20 1.23 0.50 5.26 -12% 24%<br />
Wall Mart 1.35 0.52 0.16 4.79 -10% 9%<br />
Table 5: This table presents the main st<strong>at</strong>istical fe<strong>at</strong>ures of the daily r<strong>et</strong>urns of the s<strong>et</strong> of seventeen ass<strong>et</strong>s<br />
studied here over the time interval from the end of January 1995 to the end of December 2000.<br />
45<br />
469
470 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
x n P(x)<br />
5<br />
4.5<br />
4<br />
3.5<br />
3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
x P(x)<br />
x 2 P(x)<br />
x 4 P(x)<br />
0<br />
0 1 2 3 4 5<br />
x<br />
6 7 8 9 10<br />
Figure 1: This figure represents the function x n · e −x for n = 1, 2 and 4. It shows the typycal size of the<br />
fluctu<strong>at</strong>ions involved in the moment of or<strong>de</strong>r n.<br />
46
μ (daily r<strong>et</strong>urn)<br />
x 10−3<br />
2.5<br />
2<br />
1.5<br />
1<br />
0.5<br />
Mean−μ 2 Efficient Frontier<br />
Mean−μ 4 Efficient Frontier<br />
Mean−μ 6 Efficient Frontier<br />
Mean−μ 8 Efficient Frontier<br />
0<br />
0 0.01 0.02 0.03<br />
1/n<br />
μ<br />
n<br />
0.04 0.05 0.06<br />
Figure 2: This figure represents the generalized efficient frontier for a portfolio ma<strong>de</strong> of seventeen risky<br />
ass<strong>et</strong>s. The optimiz<strong>at</strong>ion problem is solved numerically, using a gen<strong>et</strong>ic algorithm, with risk measures given<br />
respectively by the centered moments µ2, µ4, µ6 and µ8. The straight lines are the efficient frontiers when<br />
we add to these ass<strong>et</strong>s a risk-free ass<strong>et</strong> whose interest r<strong>at</strong>e is s<strong>et</strong> to 5% a year.<br />
47<br />
471
472 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
μ (daily r<strong>et</strong>urn)<br />
x 10−3<br />
1.5<br />
1<br />
0.5<br />
Mean−μ 2 Efficient Frontier<br />
Mean−μ 4 Efficient Frontier<br />
Mean−μ 6 Efficient Frontier<br />
Mean−μ 8 Efficient Frontier<br />
0<br />
0 0.005 0.01 0.015 0.02 0.025<br />
1/n<br />
μ<br />
n<br />
Figure 3: This figure represents the generalized efficient frontier for a portfolio ma<strong>de</strong> of seventeen risky<br />
ass<strong>et</strong>s and a risk-free ass<strong>et</strong> whose interest r<strong>at</strong>e is s<strong>et</strong> to 5% a year. The optimiz<strong>at</strong>ion problem is solved<br />
numerically, using a gen<strong>et</strong>ic algorithm, with risk measures given by the centered moments µ2, µ4, µ6 and<br />
µ8.<br />
48
w i<br />
w i<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
Mean−μ 2<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
w<br />
0<br />
0.3<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
Mean−μ 6<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
w<br />
0<br />
w i<br />
w i<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
Mean−μ 4<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
w<br />
0<br />
0.3<br />
0.25<br />
0.2<br />
0.15<br />
0.1<br />
0.05<br />
Mean−μ 8<br />
0<br />
0 0.2 0.4 0.6 0.8 1<br />
w<br />
0<br />
Figure 4: Depen<strong>de</strong>nce of the five largest weights of risky ass<strong>et</strong>s in the efficient portfolios found in figure 3<br />
as a function of the weight w0 invested in the risk-free ass<strong>et</strong>, for the four risk measures given by the centered<br />
moments µ2, µ4, µ6 and µ8. The same symbols always represent the same ass<strong>et</strong>.<br />
49<br />
473
474 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
μ<br />
Cn 1/n<br />
Figure 5: The dark and grey thick curves represent two efficient frontiers for a portfolio without risk-free<br />
interest r<strong>at</strong>e obtained with two measures of risks. The dark and grey thin straight lines represent the efficient<br />
frontiers in the presence of a risk-free ass<strong>et</strong>, whose value is given by the intercept of the straight lines with<br />
the ordin<strong>at</strong>e axis. This illustr<strong>at</strong>es the existence of an inversion of the <strong>de</strong>pen<strong>de</strong>nce of the slope of the efficient<br />
frontier with risk-free ass<strong>et</strong> as a function of the or<strong>de</strong>r n of the measures of risks, which can occur only when<br />
the efficient frontiers without risk-free ass<strong>et</strong> cross each other.<br />
50
Figure 6: Schem<strong>at</strong>ic represent<strong>at</strong>ion of the nonlinear mapping Y = u(X) th<strong>at</strong> allows one to transform a variable<br />
X with an arbitrary distribution into a variable Y with a Gaussian distribution. The probability <strong>de</strong>nsities<br />
for X and Y are plotted outsi<strong>de</strong> their respective axes. Consistent with the conserv<strong>at</strong>ion of probability, the<br />
sha<strong>de</strong>d regions have equal area. This conserv<strong>at</strong>ion of probability d<strong>et</strong>ermines the nonlinear mapping.<br />
51<br />
475
476 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
Empirical Cumul<strong>at</strong>ive Distribution<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
CHF−UKP<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Cumul<strong>at</strong>ive Normal Ditribution<br />
Figure 7: Quantile of the normalized sum of the Gaussianized r<strong>et</strong>urns of the Swiss Franc and The British<br />
Pound versus the quantile of the Normal distribution, for the time interval from Jan. 1971 to Oct. 1998.<br />
Different weights in the sum give similar results.<br />
52
Empirical Cumul<strong>at</strong>ive Distribution<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
KO−PG<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Cumul<strong>at</strong>ive Normale Ditribution<br />
Figure 8: Quantile of the normalized sum of the Gaussianized r<strong>et</strong>urns of Coca-Cola and Procter&Gamble<br />
versus the quantile of the Normal distribution, for the time interval from Jan. 1970 to Dec. 2000. Different<br />
weights in the sum give similar results.<br />
53<br />
477
478 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
Empirical Cumul<strong>at</strong>ive Distribution<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
MRK−GE<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1<br />
Cumul<strong>at</strong>ive Normal Ditribution<br />
Figure 9: Quantile of the normalized sum of the Gaussianized r<strong>et</strong>urns of Merk and General Electric versus<br />
the quantile of the Normal distribution, for the time interval from Jan. 1970 to Dec. 2000. Different weights<br />
in the sum give similar results.<br />
54
Z 2<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
CHF−UKP<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5<br />
χ<br />
0.6 0.7 0.8 0.9 1<br />
2<br />
Figure 10: Cumul<strong>at</strong>ive distribution of z 2 = y t V −1 y versus the cumul<strong>at</strong>ive distribution of chi-square (<strong>de</strong>noted<br />
χ 2 ) with two <strong>de</strong>grees of freedom for the couple Swiss Franc / British Pound, for the time interval from<br />
Jan. 1971 to Oct. 1998. This χ 2 should not be confused with the characteristic scale used in the <strong>de</strong>finition<br />
of the modified Weibull distributions.<br />
55<br />
479
480 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
Z 2<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
KO−PG<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5<br />
χ<br />
0.6 0.7 0.8 0.9 1<br />
2<br />
Figure 11: Cumul<strong>at</strong>ive distribution of z 2 = y t V −1 y versus the cumul<strong>at</strong>ive distribution of the chi-square<br />
χ 2 with two <strong>de</strong>grees of freedom for the couple Coca-Cola / Procter&Gamble, for the time interval from Jan.<br />
1970 to Dec. 2000. This χ 2 should not be confused with the characteristic scale used in the <strong>de</strong>finition of the<br />
modified Weibull distributions.<br />
56
Z 2<br />
1<br />
0.9<br />
0.8<br />
0.7<br />
0.6<br />
0.5<br />
0.4<br />
0.3<br />
0.2<br />
0.1<br />
MRK−GE<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5<br />
χ<br />
0.6 0.7 0.8 0.9 1<br />
2<br />
Figure 12: Cumul<strong>at</strong>ive distribution of z 2 = y t V −1 y versus the cumul<strong>at</strong>ive distribution of the chi-square<br />
χ 2 with two <strong>de</strong>grees of freedom for the couple Merk / General Electric, for the time interval from Jan. 1970<br />
to Dec. 2000. This χ 2 should not be confused with the characteristic scale used in the <strong>de</strong>finition of the<br />
modified Weibull distributions.<br />
57<br />
481
482 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
10 0<br />
10 0<br />
10 −1<br />
10 −1<br />
10 −1<br />
10 0<br />
10 0<br />
MAL +<br />
MAL −<br />
Figure 13: Graph of Gaussianized Malaysian Ringgit r<strong>et</strong>urns versus Malaysian Ringgit r<strong>et</strong>urns, for the time<br />
interval from Jan. 1971 to Oct. 1998. The upper graph gives the positive tail and the lower one the neg<strong>at</strong>ive<br />
tail. The two straight lines represent the curves y = √ 〈c±〉<br />
x 2 〈χ±〉 and y = √ c± x 2 χ±<br />
58<br />
10 1<br />
10 1<br />
10 2<br />
10 2
10 0<br />
10 −1<br />
10 0<br />
10 −1<br />
10 −1<br />
10 −1<br />
10 0<br />
10 0<br />
UKP +<br />
UKP −<br />
Figure 14: Graph of Gaussianized British Pound r<strong>et</strong>urns versus British Pound r<strong>et</strong>urns, for the time interval<br />
from Jan. 1971 to Oct. 1998. The upper graph gives the positive tail and the lower one the neg<strong>at</strong>ive tail. The<br />
two straight lines represent the curves y = √ 〈c±〉<br />
x 2 〈χ±〉 and y = √ c± x 2 χ±<br />
59<br />
10 1<br />
10 1<br />
10 2<br />
10 2<br />
483
484 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
10 0<br />
10 −1<br />
10 0<br />
10 −1<br />
10 −1<br />
10 −1<br />
10 0<br />
10 0<br />
GE +<br />
GE −<br />
Figure 15: Graph of Gaussianized General Electric r<strong>et</strong>urns versus General Electric r<strong>et</strong>urns, for the time<br />
interval from Jan. 1970 to Dec. 2000. The upper graph gives the positive tail and the lower one the neg<strong>at</strong>ive<br />
tail. The two straight lines represent the curves y = √ 〈c±〉<br />
x 2 〈χ±〉 and y = √ c± x 2 χ±<br />
60<br />
10 1<br />
10 1<br />
10 2<br />
10 2
10 0<br />
10 −1<br />
10 0<br />
10 −1<br />
10 −1<br />
10 −1<br />
10 0<br />
10 0<br />
IBM +<br />
IBM −<br />
Figure 16: Graph of Gaussianized IBM r<strong>et</strong>urns versus IBM r<strong>et</strong>urns, for the time interval from Jan. 1970 to<br />
Dec. 2000. The upper graph gives the positive tail and the lower one the neg<strong>at</strong>ive tail. The two straight lines<br />
represent the curves y = √ 〈c±〉<br />
x 2 〈χ±〉 and y = √ c± x 2 χ±<br />
61<br />
10 1<br />
10 1<br />
10 2<br />
10 2<br />
485
486 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
10 0<br />
10 −1<br />
10 0<br />
10 −1<br />
10 −1<br />
10 −1<br />
10 0<br />
10 0<br />
WMT +<br />
WMT −<br />
Figure 17: Graph of Gaussianized Wall Mart r<strong>et</strong>urns versus Wall Mart r<strong>et</strong>urns, for the time interval from<br />
Sep. 1972 to Dec. 2000. The upper graph gives the positive tail and the lower one the neg<strong>at</strong>ive tail. The two<br />
straight lines represent the curves y = √ 〈c±〉<br />
x 2 〈χ±〉 and y = √ c± x 2 χ±<br />
62<br />
10 1<br />
10 1<br />
10 2<br />
10 2
κ<br />
22<br />
20<br />
18<br />
16<br />
14<br />
12<br />
10<br />
8<br />
6<br />
4<br />
2<br />
Excess Kurtosis for CHF / JPY<br />
0<br />
0 0.1 0.2 0.3 0.4 0.5<br />
w<br />
CHF<br />
0.6 0.7 0.8 0.9 1<br />
Figure 18: Excess kurtosis of the distribution of the price vari<strong>at</strong>ion wCHF xCHF + wJP Y xJP Y of the<br />
portfolio ma<strong>de</strong> of a fraction wCHF of Swiss franc and a fraction wJP Y = 1 − wCHF of the Japanese Yen<br />
against the US dollar, as a function of wCHF . Thick solid line : empirical curve, thin solid line : theor<strong>et</strong>ical<br />
curve, dashed line : theor<strong>et</strong>ical curve with ρ = 0 (instead of ρ = 0.43), dotted line: theor<strong>et</strong>ical curve with<br />
qCHF = 2 r<strong>at</strong>her than 1.75 and dashed-dotted line: theor<strong>et</strong>ical curve with qCHF = 1.5. The excess kurtosis<br />
has been evalu<strong>at</strong>ed for the time interval from Jan. 1971 to Oct. 1998.<br />
63<br />
487
488 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
C 4<br />
65<br />
60<br />
55<br />
50<br />
45<br />
40<br />
35<br />
30<br />
25<br />
20<br />
Fourth Cumulant for CHF / UKP<br />
15<br />
0 0.1 0.2 0.3 0.4 0.5<br />
w<br />
CHF<br />
0.6 0.7 0.8 0.9 1<br />
Figure 19: Fourth cumulant for a portfolio ma<strong>de</strong> of a fraction wCHF of Swiss Franc and 1−wCHF of British<br />
Pound. The thick solid line represents the empirical cumulant while the dotted line represents the theor<strong>et</strong>ical<br />
cumulant un<strong>de</strong>r the symm<strong>et</strong>ric assumption. The dashed line shows the theor<strong>et</strong>ical cumulant when the slight<br />
asymm<strong>et</strong>ry of the ass<strong>et</strong>s has been taken into account. This cumulant has been evalu<strong>at</strong>ed for the time interval<br />
from Jan. 1971 to Oct. 1998.<br />
64
R<strong>et</strong>urn μ<br />
0.12<br />
0.11<br />
0.1<br />
0.09<br />
0.08<br />
0.07<br />
Efficient Frontier forIBM−HWP<br />
1 1.2 1.4 1.6<br />
min<br />
C /C , n={1,2,3}<br />
2n 2n<br />
1.8 2 2.2<br />
Figure 20: Efficient frontier for a portfolio composed of two stocks: IBM and Hewl<strong>et</strong>t-Packard. The dashed<br />
line represents the efficient frontier with respect to the second cumulant, i.e., the standard Markovitz efficient<br />
frontier, the dash-dotted line represents the efficient frontier with respect to the fourth cumulant and the solid<br />
line is the efficient frontier with respect to the sixth cumulant. The d<strong>at</strong>a s<strong>et</strong> used covers the time interval<br />
from Jan. 1977 to Dec 2000.<br />
65<br />
489
490 14. Gestion <strong>de</strong> Portefeuilles multimoments <strong>et</strong> équilibre <strong>de</strong> marché<br />
R<strong>et</strong>urn μ<br />
0.13<br />
0.12<br />
0.11<br />
0.1<br />
0.09<br />
0.08<br />
0.07<br />
0.06<br />
Efficient Frontier forIBM−KO<br />
0.05<br />
1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45 1.5<br />
min<br />
C /C , n={1,2,3}<br />
2n 2n<br />
Figure 21: Efficient frontier for a portfolio composed of two stocks: IBM and Coca-Cola. The dashed line<br />
represents the efficient frontier with respect to the second cumulant, i.e., the standard Markovitz efficient<br />
frontier, the dash-dotted line represents the efficient frontier with respect to the fourth cumulant and the solid<br />
line the efficient frontier with repect to the sixth cumulant. The d<strong>at</strong>a s<strong>et</strong> used covers the time interval from<br />
Jan. 1970 to Dec 2000.<br />
66
Conclusions <strong>et</strong> Perspectives<br />
L’obj<strong>et</strong> <strong>de</strong> notre étu<strong>de</strong> était <strong>de</strong> contribuer à une meilleure compréhension <strong>de</strong>s risques extrêmes sur les<br />
marchés financiers afin <strong>de</strong> perm<strong>et</strong>tre d’esquisser les contours <strong>de</strong> nouvelles métho<strong>de</strong>s <strong>de</strong> contrôle <strong>de</strong> ce<br />
type <strong>de</strong> risques <strong>et</strong> d’en présenter certaines conséquences pour ce qui est <strong>de</strong> la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>s.<br />
Pour cela, nous avons suivi une approche dont le but était <strong>de</strong> chercher à isoler les différentes sources <strong>de</strong><br />
risques selon que leur origine provienne <strong>de</strong>s fluctu<strong>at</strong>ions individuelles <strong>de</strong>s actifs ou bien <strong>de</strong> leur comportement<br />
collectif. C<strong>et</strong>te distinction nous parait particulièrement importante pour la <strong>gestion</strong> <strong>de</strong>s risques<br />
dans la mesure où il nous semble nécessaire pour mener à bien c<strong>et</strong>te tâche (i) d’être capable d’estimer les<br />
grands risques associés à chaque actif - ce qui ne repose que sur l’étu<strong>de</strong> <strong>de</strong> leur “variabilité” intrinsèque<br />
- <strong>et</strong> (ii) <strong>de</strong> savoir si l’on peut espérer diversifier ces grands risques <strong>et</strong> donc d’en réduire l’impact par la<br />
form<strong>at</strong>ion <strong>de</strong> <strong>portefeuille</strong>s - ce qui fait appel à leurs caractéristiques collectives.<br />
C’est pourquoi nous avons tout d’abord procédé à une étu<strong>de</strong> à la fois empirique <strong>et</strong> théorique <strong>de</strong>s vari<strong>at</strong>ions<br />
<strong>de</strong> cours <strong>de</strong>s actifs financiers, afin d’en déterminer les propriétés <strong>st<strong>at</strong>istique</strong>s <strong>et</strong> d’essayer <strong>de</strong> m<strong>et</strong>tre à<br />
jour certains mécanismes micro-structurels <strong>et</strong> comportementaux perm<strong>et</strong>tant <strong>de</strong> justifier les observ<strong>at</strong>ions<br />
empiriques. L’étu<strong>de</strong> s’est focalisée sur trois points :<br />
– les aspects st<strong>at</strong>iques (monovariés), avec l’étu<strong>de</strong> <strong>de</strong>s distributions marginales <strong>de</strong> ren<strong>de</strong>ments,<br />
– mais aussi les aspects dynamiques <strong>de</strong> par l’étu<strong>de</strong> <strong>de</strong> la modélis<strong>at</strong>ion <strong>de</strong> la dépendance temporelle à<br />
l’ai<strong>de</strong> <strong>de</strong> processus <strong>de</strong> marches alé<strong>at</strong>oires multifractales,<br />
– ainsi que la prise en compte <strong>de</strong> l’interaction entre les actifs au travers <strong>de</strong> l’étu<strong>de</strong> <strong>de</strong> la représent<strong>at</strong>ion<br />
<strong>de</strong> la structure <strong>de</strong> dépendance en terme <strong>de</strong> copule <strong>et</strong> surtout l’estim<strong>at</strong>ion <strong>de</strong>s dépendances extrêmes à<br />
l’ai<strong>de</strong> du coefficient <strong>de</strong> dépendance <strong>de</strong> queue.<br />
Ceci nous a conduit, pour ce qui est du premier point, à nous pencher sur l’hypothèse d’après laquelle<br />
les ren<strong>de</strong>ments sont distribués selon <strong>de</strong>s lois <strong>de</strong> puissances (ou plus généralement <strong>de</strong>s lois régulièrement<br />
variables). Conformément à certains travaux récents, nous avons confirmé qu’il ne pouvait cependant être<br />
exclu que <strong>de</strong>s distributions exponentielles étirées sont à même <strong>de</strong> représenter les distributions <strong>de</strong> ren<strong>de</strong>ments<br />
aussi bien (sinon mieux) que les distributions régulièrement variables généralement utilisées. La<br />
prise en compte <strong>de</strong> c<strong>et</strong>te nouvelle représent<strong>at</strong>ion paramétrique <strong>de</strong>s distributions <strong>de</strong> ren<strong>de</strong>ments est importante<br />
<strong>et</strong> nous semble <strong>de</strong>voir être employée conjointement avec la <strong>de</strong>scription en terme <strong>de</strong> distributions<br />
régulièrement variables (ce que nous avons fait dans la suite <strong>de</strong> notre étu<strong>de</strong>) afin <strong>de</strong> ne pas négliger le<br />
risque <strong>de</strong> modèle inhérent à toute <strong>de</strong>scription paramétrique <strong>de</strong>s données.<br />
Concernant l’étu<strong>de</strong> dynamique <strong>de</strong> l’évolution <strong>de</strong>s cours, nous avons préféré nous tourner vers l’utilis<strong>at</strong>ion<br />
<strong>de</strong> processus <strong>de</strong> marche alé<strong>at</strong>oire multifractale, plutôt que vers les traditionnels processus <strong>de</strong> la famille<br />
ARCH, car ils semblent les seuls à pouvoir rendre compte <strong>de</strong> la manière suivant laquelle la vol<strong>at</strong>ilité<br />
r<strong>et</strong>ourne à la moyenne (relaxe) après un grand choc. En eff<strong>et</strong>, nous avons montré que le processus <strong>de</strong><br />
marche alé<strong>at</strong>oire multifractale prédisait une relax<strong>at</strong>ion hyperbolique dont l’exposant diffère selon que<br />
le choc est endogène ou exogène, ce qui a pu être effectivement observé sur les données. En outre, un<br />
tel comportement ne peut pas être expliqué par les processus <strong>de</strong> type ARCH. Donc, le processus <strong>de</strong><br />
marche alé<strong>at</strong>oire multifractale semble bien adapté à la <strong>de</strong>scription <strong>de</strong> la dynamique <strong>de</strong>s cours <strong>de</strong>s actifs<br />
491
492 Conclusion<br />
financiers, <strong>et</strong> perm<strong>et</strong> une meilleure compréhension <strong>de</strong> l’incorpor<strong>at</strong>ion du flux d’inform<strong>at</strong>ion dans le prix<br />
<strong>de</strong>s actifs, ce qui présente un grand intérêt pour la prédiction <strong>de</strong> la vol<strong>at</strong>ilité, par exemple, mais aussi pour<br />
la <strong>gestion</strong> <strong>de</strong> scenarii perm<strong>et</strong>tant <strong>de</strong> mieux appréhen<strong>de</strong>r l’impact <strong>de</strong> telle ou telle nouvelle sur l’évolution<br />
du prix futur d’un actif <strong>et</strong> notamment la durée <strong>de</strong>s pério<strong>de</strong>s <strong>de</strong> fortes turbulences ou <strong>de</strong> calme anormal<br />
<strong>de</strong>s marchés. Notons toutefois qu’il conviendra d’améliorer c<strong>et</strong>te <strong>de</strong>scription par la prise en compte <strong>de</strong><br />
l’asymétrie entre fluctu<strong>at</strong>ions à la hausse <strong>et</strong> à la baisse (eff<strong>et</strong> <strong>de</strong> levier) qui a été totalement négligée<br />
jusqu’ici.<br />
Pour synthétiser ces <strong>de</strong>ux premiers points, <strong>et</strong> essayer <strong>de</strong> percer certains <strong>de</strong>s mécanismes microscopiques<br />
perm<strong>et</strong>tant <strong>de</strong> les justifier, nous nous sommes intéressés aux modèles d’agents en interaction. Ceci avait<br />
pour but d’intégrer les comportements non totalement r<strong>at</strong>ionnels <strong>de</strong>s acteurs économiques afin <strong>de</strong> comprendre<br />
comment - en dépit <strong>de</strong> c<strong>et</strong>te r<strong>at</strong>ionnalité limitée - peut émerger une structure <strong>de</strong> marché la plupart<br />
du temps efficiente, mais aussi pourquoi <strong>et</strong> comment les marchés s’écartent parfois violemment<br />
<strong>de</strong> c<strong>et</strong>te structure efficiente. Dans c<strong>et</strong>te étu<strong>de</strong>, nous avons mis en exergue l’importance <strong>de</strong>s comportements<br />
mimétiques <strong>et</strong> antagonistes entre agents afin <strong>de</strong> préciser leur rôle dans l’émergence <strong>et</strong> l’explosion<br />
<strong>de</strong>s bulles spécul<strong>at</strong>ives. Là encore, c<strong>et</strong>te approche offre une possibilité <strong>de</strong> simuler certaines phases <strong>de</strong><br />
marchés <strong>et</strong> fournit un moyen <strong>de</strong> mieux anticiper les risques à venir.<br />
Une <strong>de</strong>s limites actuelles <strong>de</strong>s modèles d’agents en interaction est <strong>de</strong> se concentrer uniquement sur la<br />
modélis<strong>at</strong>ion <strong>de</strong> marchés où un seul actif (plus éventuellement un actif sans risque) peut être échangé.<br />
Nous avons, nous aussi, choisi <strong>de</strong> suivre c<strong>et</strong>te voie pour <strong>de</strong>s raisons <strong>de</strong> simplicité. En eff<strong>et</strong>, à partir<br />
du moment où l’on considère un marché doté <strong>de</strong> plusieurs actifs, il convient <strong>de</strong> donner aux agents les<br />
moyens <strong>de</strong> choisir entre les divers actifs <strong>et</strong> donc d’introduire une fonction d’utilité - a priori différente<br />
- pour chaque agent, ce qui pose quelques difficultés, au premier rang <strong>de</strong>squelles, comme souligné au<br />
chapitre 10, le choix <strong>de</strong> c<strong>et</strong>te fonction. Il nous semble clair qu’une <strong>de</strong>s évolutions futures <strong>de</strong>s modèles<br />
d’agents en interaction <strong>de</strong>vra apporter <strong>de</strong>s réponses à ce problème qui à notre avis constitue l’une <strong>de</strong>s<br />
toutes prochaines évolutions à apporter à ce type <strong>de</strong> modélis<strong>at</strong>ion.<br />
Le troisième point que nous avons abordé concernait la <strong>de</strong>scription <strong>de</strong> la dépendance entre les actifs. Nous<br />
avons pour cela commencé par essayer <strong>de</strong> déterminer <strong>de</strong> manière globale la structure <strong>de</strong> dépendance<br />
en tentant d’estimer la copule <strong>de</strong>s ren<strong>de</strong>ments <strong>de</strong>s actifs financiers. Nous avons pu conclure que pour<br />
<strong>de</strong>s ren<strong>de</strong>ments modérés, une copule gaussienne semblait tout à fait à même <strong>de</strong> rendre compte <strong>de</strong> leur<br />
dépendance, mais risquait <strong>de</strong> sous-estimer les dépendances extrêmes. Pour confirmer c<strong>et</strong>te hypothèse,<br />
nous avons voulu estimer le coefficient <strong>de</strong> dépendance <strong>de</strong> queue, qui mesure la propension <strong>de</strong> <strong>de</strong>ux<br />
actifs financiers à subir ensemble <strong>de</strong> grands mouvements. La difficulté à estimer <strong>de</strong> manière directe c<strong>et</strong>te<br />
quantité nous a d’abord conduit à la calculer théoriquement en nous appuyant sur une <strong>de</strong>scription <strong>de</strong>s<br />
fluctu<strong>at</strong>ions <strong>de</strong>s prix <strong>de</strong>s actifs à l’ai<strong>de</strong> d’un modèle à facteur. Puis, <strong>de</strong> la calibr<strong>at</strong>ion <strong>de</strong>s paramètres <strong>de</strong> ce<br />
modèle, nous avons pu déduire la valeur du coefficient <strong>de</strong> dépendance <strong>de</strong> queue pour chaque pair d’actifs.<br />
Ceci a alors confirmé l’inadéqu<strong>at</strong>ion <strong>de</strong> la copule gaussienne à décrire la dépendance entre les gran<strong>de</strong>s<br />
vari<strong>at</strong>ions <strong>de</strong>s cours.<br />
C<strong>et</strong>te étu<strong>de</strong> <strong>de</strong>vra être poursuivie, car nous n’avons pas pu trouver <strong>de</strong> copule réellement s<strong>at</strong>isfaisante<br />
quant à la <strong>de</strong>scription complète <strong>de</strong> la dépendance entre actifs. Si nous sommes désormais convaincus qu’il<br />
existe une dépendance <strong>de</strong> queue entre les actifs, il reste néanmoins à essayer <strong>de</strong> déterminer quelle copule<br />
est la mieux adaptée pour rendre compte <strong>de</strong> la structure <strong>de</strong> dépendance extrême entre les actifs. Ceci<br />
est important car, comme nous l’avons déjà signalé, les paramètres intervenant dans la représent<strong>at</strong>ion<br />
paramétrique <strong>de</strong> la copule perm<strong>et</strong>tent <strong>de</strong> résumer toute la dépendance <strong>de</strong> la même manière que le coefficient<br />
<strong>de</strong> corrél<strong>at</strong>ion capture complètement la dépendance pour une distribution multivariée gaussienne.<br />
Ces paramètres constituent, à notre avis, les “bonnes variables” qu’il convient <strong>de</strong> rechercher afin <strong>de</strong> nous<br />
perm<strong>et</strong>tre <strong>de</strong> modéliser à l’échelle macroscopique le comportement <strong>de</strong>s actifs <strong>et</strong> notamment la rel<strong>at</strong>ion<br />
entre risque <strong>et</strong> ren<strong>de</strong>ment. Que l’on pense au lien existant entre le coefficient <strong>de</strong> corrél<strong>at</strong>ion <strong>et</strong> la variance
Conclusion 493<br />
en univers gaussien d’une part <strong>et</strong> coefficient b<strong>et</strong>a mesurant le risque systém<strong>at</strong>ique associé à chaque actif<br />
selon le CAPM d’autre part.<br />
En outre, notre étu<strong>de</strong> <strong>de</strong> la structure <strong>de</strong> dépendance entre actifs n’a été que <strong>de</strong>scriptive <strong>et</strong> n’en a pas<br />
sondé les causes microscopiques. Cela rejoint en fait ce que nous avons écrit plus haut concernant la<br />
nécessité <strong>de</strong> développer <strong>de</strong>s modèles d’agents considérant <strong>de</strong>s marchés où peuvent être échangés plusieurs<br />
actifs. Ce sera alors un moyen <strong>de</strong> mieux comprendre les raisons <strong>de</strong> l’existence <strong>de</strong> divers facteurs<br />
<strong>et</strong> leur influence sur c<strong>et</strong>te dépendance entre actifs : les facteurs économiques, tels que l’appartenance<br />
à un même secteur d’activité par exemple, sont-ils <strong>de</strong>s éléments primordiaux <strong>de</strong> c<strong>et</strong>te dépendance, ou<br />
bien les agents jouent-ils un rôle prépondérant ? Existe-t-il <strong>de</strong>s phases où la structure <strong>de</strong> dépendance est<br />
dominée par les fondamentaux économiques <strong>et</strong> d’autre part l’activité spécul<strong>at</strong>ive, comme on l’observe<br />
pour le développement <strong>de</strong>s fluctu<strong>at</strong>ions dans le cas marginal ? Toutes ces questions restent ouvertes <strong>et</strong> en<br />
<strong>at</strong>tente <strong>de</strong> réponses.<br />
Après avoir considéré les <strong>de</strong>ux fac<strong>et</strong>tes <strong>de</strong>s grands risques que sont d’une part l’aspect individuel ou marginal<br />
<strong>et</strong> d’autre part l’aspect collectif, nous avons regroupé nos résult<strong>at</strong>s afin d’en déduire les conséquences<br />
pr<strong>at</strong>iques sur la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>s en vue <strong>de</strong> la prévention <strong>de</strong> ces grands risques. Ceci nous a tout<br />
d’abord amené à discuter <strong>de</strong> la manière dont il convient <strong>de</strong> les mesurer. Nous avons pour cela utilisé <strong>de</strong>ux<br />
approches : l’une basée sur la mesure du capital économique <strong>et</strong> l’autre prenant en compte la propension<br />
du <strong>portefeuille</strong> à ne pas trop s’écarter <strong>de</strong>s objectifs <strong>de</strong> rentabilité que l’on en espère. Ces <strong>de</strong>ux voies ont<br />
été explorées <strong>et</strong> nous avons discuté <strong>de</strong> leur mise en œuvre pr<strong>at</strong>ique ainsi que <strong>de</strong> leurs conséquences sur<br />
les prix d’équilibre <strong>de</strong>s actifs. Nous nous sommes bornés à considérer <strong>de</strong>s mesures <strong>de</strong> risques <strong>et</strong> à abor<strong>de</strong>r<br />
le problème <strong>de</strong> l’optimis<strong>at</strong>ion <strong>de</strong> <strong>portefeuille</strong> dans un cadre mono-périodique (ou st<strong>at</strong>ique), alors qu’il<br />
apparaît très clairement que ces <strong>de</strong>ux problèmes doivent se poser en termes dynamiques. De nombreuses<br />
recherches sont actuellement menées pour définir <strong>de</strong> manière axiom<strong>at</strong>ique la notion <strong>de</strong> risque dans un<br />
cadre dynamique ainsi que pour en étudier les conséquences sur la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>s. Nos futures<br />
recherches sur ce suj<strong>et</strong> <strong>de</strong>vront s’inscrire dans ce courant.<br />
Enfin, nous focalisant sur l’impact <strong>de</strong>s risques extrêmes, il est notamment apparu que l’existence d’une<br />
dépendance <strong>de</strong> queue entre les actifs en limitait les possibilités <strong>de</strong> diversific<strong>at</strong>ion. Nous avons alors montrer<br />
comment on parvenait à en minimiser les eff<strong>et</strong>s en ne sélectionnant que les actifs présentant les<br />
plus faibles dépendances <strong>de</strong> queues. Cependant, si les risques extrêmes ne sont pas totalement diversifiables,<br />
nous <strong>de</strong>vrions être en droit <strong>de</strong> nous <strong>at</strong>tendre à ce que le marché les rémunére. Nous n’avons pas<br />
étudié c<strong>et</strong>te question qui nous semble cruciale mais envisageons <strong>de</strong> traiter ce problème lors <strong>de</strong> recherches<br />
ultérieures.<br />
Nous voyons donc que les travaux que nous venons <strong>de</strong> présenter, loin d’être achevés <strong>et</strong> d’apporter <strong>de</strong>s<br />
réponses définitives, posent <strong>de</strong> nombreuses questions <strong>et</strong> ouvrent donc <strong>de</strong> multiples perspectives <strong>de</strong> recherches<br />
futures.
494 Conclusion
Annexe A<br />
Evalu<strong>at</strong>ion <strong>de</strong> la conduite du proj<strong>et</strong> <strong>de</strong><br />
thèse<br />
Ce nouveau chapitre s’inscrit dans le cadre <strong>de</strong> la poursuite d’un proj<strong>et</strong> pilote développé par le ministère<br />
<strong>de</strong> l’éduc<strong>at</strong>ion n<strong>at</strong>ionale, <strong>de</strong> la recherche <strong>et</strong> <strong>de</strong> la technologie (MENRT) afin d’ai<strong>de</strong>r les jeunes docteurs<br />
à s’insérer plus facilement dans la vie active, que ce soit dans le milieu <strong>de</strong> la recherche académique ou<br />
bien au sein d’entreprises privées. L’obj<strong>et</strong> <strong>de</strong> ce chapitre est <strong>de</strong> mener une réflexion sur la conduite <strong>de</strong> la<br />
thèse, non seulement en termes <strong>de</strong> r<strong>et</strong>ombées scientifiques mais également d’un point <strong>de</strong> vue comptable,<br />
avec l’évalu<strong>at</strong>ion du coût du proj<strong>et</strong>, <strong>et</strong> <strong>de</strong> faire le bilan <strong>de</strong>s expériences <strong>et</strong> compétences acquises au cours<br />
<strong>de</strong> ces années.<br />
A.1 Bref résumé du suj<strong>et</strong><br />
L’obj<strong>et</strong> <strong>de</strong> c<strong>et</strong>te thèse était l’étu<strong>de</strong> <strong>de</strong>s risques extrêmes sur les marchés financiers. Pour nous <strong>at</strong>taquer à ce<br />
problème, nous avons tout abord cherché à modéliser les mécanismes perm<strong>et</strong>tant d’expliquer l’apparition<br />
<strong>de</strong> mouvements extrêmes sur ce type <strong>de</strong> marchés afin d’en avoir une meilleure compréhension <strong>et</strong> notamment<br />
d’en isoler certaines causes. A partir <strong>de</strong> là, nous avons, dans un premier temps, pu développer <strong>de</strong><br />
nouvelles mesures <strong>de</strong> risques perm<strong>et</strong>tant <strong>de</strong> mieux rendre compte <strong>de</strong>s gran<strong>de</strong>s fluctu<strong>at</strong>ions observées sur<br />
les cours <strong>de</strong>s actifs financiers. Cela nous a ensuite permis d’élaborer <strong>de</strong> nouvelles <strong>théorie</strong>s <strong>de</strong> <strong>portefeuille</strong>s<br />
visant à se prémunir au mieux contre les risques extrêmes. Enfin, nous en avons déduit d’intéressants<br />
résult<strong>at</strong>s concernant le prix <strong>de</strong>s actifs sur les marchés financiers à l’équilibre.<br />
Notre étu<strong>de</strong> s’est basée sur l’hypothèse que les marchés financiers peuvent être considérés comme <strong>de</strong>s<br />
systèmes complexes auto-organisés : <strong>de</strong> l’interaction entre agents économiques – aux objectifs divers <strong>et</strong><br />
souvent opposés – résultent les propriétés macroscopiques <strong>de</strong>s marchés. Ceci nous a notamment conduit<br />
à abandonner les hypothèses traditionnelles fondées sur l’existence d’agents économiques représent<strong>at</strong>ifs<br />
<strong>et</strong> r<strong>at</strong>ionnels, dont découle le concept <strong>de</strong> marché efficient <strong>et</strong> qui conduit à décrire la dynamique <strong>de</strong>s<br />
cours par un mouvement brownien, dont on sait pertinemment aujourd’hui, qu’il est inadapté à une telle<br />
modélis<strong>at</strong>ion.<br />
495
496 A. Evalu<strong>at</strong>ion <strong>de</strong> la conduite du proj<strong>et</strong> <strong>de</strong> thèse<br />
A.2 Eléments <strong>de</strong> contexte<br />
A.2.1 Choix du suj<strong>et</strong><br />
Après avoir préparé le diplôme d’étu<strong>de</strong>s approfondies en physique théorique <strong>de</strong> l’Ecole Normale Supérieure<br />
<strong>de</strong> Lyon durant l’année scolaire 1999-2000, j’ai décidé <strong>de</strong> m’inscrire en thèse. A l’issue d’un tel DEA, il<br />
est n<strong>at</strong>urel <strong>de</strong> choisir un suj<strong>et</strong> <strong>de</strong> thèse portant sur la physique <strong>st<strong>at</strong>istique</strong> ou la physique <strong>de</strong>s particules.<br />
Cependant, il était bien clair pour moi que <strong>de</strong> tels suj<strong>et</strong>s ne conduisent que rarement à un poste dans la<br />
recherche publique – leur nombre étant extrêmement faible – <strong>et</strong> ne suscitent presque jamais l’intérêt <strong>de</strong>s<br />
acteurs privés, qui jugent ces thèmes <strong>de</strong> recherche trop fondamentaux <strong>et</strong> sans applic<strong>at</strong>ions concrètes. Ceci<br />
risquait donc <strong>de</strong> poser quelques problèmes en vue <strong>de</strong> ma future insertion dans la vie professionnelle <strong>et</strong><br />
laissait présager dès le départ d’une nécessaire réorient<strong>at</strong>ion à l’issue d’une telle thèse. Aussi, je préférais<br />
d’emblée me détourner <strong>de</strong> la voie habituelle.<br />
En fait, le domaine <strong>de</strong> la finance m’<strong>at</strong>tirait <strong>de</strong>puis longtemps, à tel point que j’avais un instant envisagé<br />
<strong>de</strong> suivre un DEA <strong>de</strong> finance plutôt qu’un DEA <strong>de</strong> physique théorique. Mais, sachant que l’étu<strong>de</strong> <strong>de</strong>s<br />
marchés financiers au moyen <strong>de</strong>s outils <strong>de</strong> la physique <strong>st<strong>at</strong>istique</strong> se développait activement en physique,<br />
au sein d’une branche appelée “ éconophysique ”, je m’étais alors résolu à apprendre les bases<br />
<strong>de</strong> la physique théorique dans le but <strong>de</strong> réinvestir ces connaissances dans l’étu<strong>de</strong> <strong>et</strong> la modélis<strong>at</strong>ion <strong>de</strong>s<br />
mécanismes à l’œuvre sur les marchés financiers. C’est donc tout n<strong>at</strong>urellement que m’est venu l’idée <strong>de</strong><br />
préparer une thèse portant sur la problém<strong>at</strong>ique <strong>de</strong>s marchés financiers.<br />
Durant les mois qui ont précédé le début <strong>de</strong> ma thèse j’ai rencontré <strong>et</strong> discuté avec <strong>de</strong> nombreuses<br />
personnes, tant <strong>de</strong>s enseignants <strong>et</strong>/ou chercheurs en physique <strong>et</strong> finance que <strong>de</strong>s professionnels, afin<br />
d’obtenir leur avis <strong>et</strong> conseil sur le choix <strong>de</strong> thèse que je souhaitais réaliser <strong>et</strong> sur la meilleure façon d’y<br />
parvenir. J’avoue qu’à ce moment-là les avis que j’ai reçus étaient partagés <strong>et</strong> que <strong>de</strong> nombreuses mises<br />
en gar<strong>de</strong> sur les difficultés auxquelles je risquais d’être confronté m’ont été faites. Cependant, cela n’a<br />
nullement entamé mon enthousiasme, <strong>et</strong> je dirais même qu’au contraire, cela m’a poussé à relever ce qui<br />
m’apparaissait alors comme un véritable défi.<br />
A.2.2 Choix <strong>de</strong> l’encadrement <strong>et</strong> du labor<strong>at</strong>oire d’accueil<br />
Le thème central <strong>de</strong> ma thèse étant choisi, il me fallait encore trouver un directeur <strong>de</strong> thèse pouvant<br />
m’encadrer sur un tel suj<strong>et</strong> ainsi qu’un labor<strong>at</strong>oire d’accueil. Le choix du directeur <strong>de</strong> thèse fut en fait<br />
assez rapi<strong>de</strong>, car si <strong>de</strong> plus en plus <strong>de</strong> chercheurs s’intéressent à ce domaine encore émergeant qu’est<br />
l’éconophysique, très peu sont à même <strong>de</strong> diriger une thèse sur ce suj<strong>et</strong>. C’est donc fort logiquement que<br />
je pris contact avec Didier Sorn<strong>et</strong>te, directeur <strong>de</strong> recherche au CNRS, affecté au Labor<strong>at</strong>oire <strong>de</strong> Physique<br />
<strong>de</strong> la M<strong>at</strong>ière Con<strong>de</strong>nsée (Université <strong>de</strong> Nice – Sophia Antipolis) où il dirige l’équipe <strong>de</strong> “ physique<br />
pluridisciplinaire ” dont le but est l’étu<strong>de</strong> <strong>de</strong>s phénomènes critiques auto-organisés dans <strong>de</strong>s domaines<br />
aussi divers que la rupture <strong>de</strong> m<strong>at</strong>ériaux, la prédiction <strong>de</strong>s tremblements <strong>de</strong> terre ou l’étu<strong>de</strong> <strong>de</strong>s marchés<br />
financiers. Très vite nous décidâmes <strong>de</strong> nous rencontrer, ce qui me permis <strong>de</strong> découvrir plus en détails<br />
les thèmes <strong>de</strong> recherche abordés au sein <strong>de</strong> c<strong>et</strong>te équipe <strong>et</strong> <strong>de</strong> prévoir le cadre <strong>de</strong> ma thèse. Il fut aussi<br />
décidé que j’effectuerais mon stage <strong>de</strong> DEA au sein <strong>de</strong> l’équipe, ce qui me perm<strong>et</strong>trait d’une part <strong>de</strong> me<br />
familiariser avec les thèmes <strong>et</strong> métho<strong>de</strong>s que j’aurais à employer au cours <strong>de</strong> ma thèse, <strong>et</strong> d’autre part <strong>de</strong><br />
vérifier que le travail qui m’<strong>at</strong>tendait en thèse correspondait bien à mes aspir<strong>at</strong>ions.<br />
Il apparut aussi assez rapi<strong>de</strong>ment qu’il serait très intéressant <strong>de</strong> s’adjoindre les services d’un professeur<br />
<strong>de</strong> finance afin d’avoir <strong>de</strong> réels échanges avec ce milieu, <strong>de</strong> mieux cerner leurs problèmes <strong>et</strong> ainsi ne pas<br />
s’égarer dans <strong>de</strong>s recherches stériles car coupées <strong>de</strong>s réalités du mon<strong>de</strong> financier. De plus, c<strong>et</strong>te colla-
A.2. Eléments <strong>de</strong> contexte 497<br />
bor<strong>at</strong>ion perm<strong>et</strong>tait <strong>de</strong> garantir le caractère pluridisciplinaire <strong>de</strong> la thèse dans laquelle je m’engageais.<br />
C’est pourquoi je suis allé rencontrer Jean-Paul Laurent, professeur à l’Institut <strong>de</strong> Science Financière<br />
<strong>et</strong> d’Assurances (Université Lyon I), afin <strong>de</strong> lui exposer mon proj<strong>et</strong> <strong>et</strong> <strong>de</strong> lui <strong>de</strong>man<strong>de</strong>r <strong>de</strong> bien vouloir<br />
co-diriger ma thèse, ce qu’il accepta tout <strong>de</strong> suite.<br />
A.2.3 Financement du proj<strong>et</strong><br />
Ce proj<strong>et</strong> étant essentiellement théorique, <strong>et</strong> les r<strong>et</strong>ombées pr<strong>at</strong>iques a priori difficilement évaluables, il<br />
semblait délic<strong>at</strong> d’obtenir un financement autre que public. En fait, en tant qu’élève <strong>de</strong> l’Ecole Normale<br />
Supérieure <strong>de</strong> Lyon, je pouvais postuler pour l’obtention d’une bourse spécifique, dite alloc<strong>at</strong>ion couplée,<br />
consistant en une alloc<strong>at</strong>ion <strong>de</strong> recherche du MENRT <strong>et</strong> d’un monitor<strong>at</strong> (à l’université <strong>de</strong> Nice). Le<br />
suj<strong>et</strong> <strong>de</strong> thèse peu orthodoxe que je présentais pouvait me <strong>de</strong>sservir quant à l’obtention <strong>de</strong> c<strong>et</strong>te bourse,<br />
mais finalement, celui-ci fut jugé innovant <strong>et</strong> en plein accord avec la politique ministérielle visant à<br />
promouvoir la pluridisciplinarité dans les thèmes <strong>de</strong> recherche, ce qui me permit d’obtenir l’<strong>at</strong>tribution <strong>de</strong><br />
c<strong>et</strong>te alloc<strong>at</strong>ion couplée. Une telle bourse représente un montant annuel d’environ 24000 euros, charges<br />
p<strong>at</strong>ronales comprises.<br />
A ce montant, s’ajoute un pourcentage du coût salarial <strong>de</strong> mes directeur <strong>et</strong> co-directeur <strong>de</strong> thèse. Ce coût<br />
est difficilement évaluable mais on peut estimer que chacun <strong>de</strong> mes encadrants a consacré entre dix <strong>et</strong><br />
vingt pour-cent <strong>de</strong> son temps à l’encadrement <strong>de</strong> ma thèse ce qui, globalement doit représenter prés <strong>de</strong><br />
15000 euros par an. On peut donc estimer le coût total en ressources humaines à approxim<strong>at</strong>ivement<br />
39000 euros par an.<br />
Du point <strong>de</strong> vue investissement m<strong>at</strong>ériel, le coût se révèle beaucoup plus mo<strong>de</strong>ste. Mon travail <strong>de</strong><br />
thèse étant essentiellement théorique, il n’a nécessité que l’ach<strong>at</strong> d’un ordin<strong>at</strong>eur, quelques logiciels<br />
<strong>et</strong> quelques livres. On peut donc chiffrer c<strong>et</strong>te dépense à 3000 euros.<br />
J’ai aussi eu à effectuer <strong>de</strong> nombreux déplacements soit pour assister à <strong>de</strong>s conférences soit pour rencontrer<br />
mon co-directeur à Lyon ou Paris. Sur toute la durée <strong>de</strong> ma thèse, les sommes engagées se chiffrent<br />
à 4000 euros. J’ai également participé à <strong>de</strong>ux écoles organisées par le département <strong>de</strong> la form<strong>at</strong>ion continue<br />
du CNRS <strong>et</strong> pris en charge par lui. Le coût engagé par le CNRS à ces <strong>de</strong>ux occasions avoisine les<br />
2500 euros.<br />
Enfin, il faut ajouter à cela la fraction qui m’est imputable <strong>de</strong>s frais <strong>de</strong> fonctionnement du labor<strong>at</strong>oire qui<br />
m’a accueilli pendant ces <strong>de</strong>ux ans <strong>et</strong> <strong>de</strong>mi. Ces frais sont difficilement évaluables avec précision, mais<br />
il semble raisonnable <strong>de</strong> les estimer à 5000 euros.<br />
En résumé, les frais engagés sont les suivants :<br />
Ressources humaines (30 mois) 97500 euro<br />
Frais <strong>de</strong> m<strong>at</strong>ériels 3000 euro<br />
Frais <strong>de</strong> déplacement 4000 euro<br />
Form<strong>at</strong>ion continue 2500 euro<br />
Frais <strong>de</strong> fonctionnement 5000 euro<br />
Total 112000 euro<br />
Le total <strong>de</strong>s frais engagés <strong>at</strong>teint donc 112000 euros sur trente mois, soit un coût annuel <strong>de</strong> 44800 euros.<br />
Les frais <strong>de</strong> ressources humaines sont bien évi<strong>de</strong>mment prépondérants, puisqu’ils représentent plus <strong>de</strong><br />
85% du coût total du proj<strong>et</strong>, ce qui est en ligne avec ce qui est généralement observé pour tout suj<strong>et</strong> <strong>de</strong><br />
recherche théorique.
498 A. Evalu<strong>at</strong>ion <strong>de</strong> la conduite du proj<strong>et</strong> <strong>de</strong> thèse<br />
A.3 Evolution du proj<strong>et</strong><br />
A.3.1 Elabor<strong>at</strong>ion du proj<strong>et</strong><br />
Le proj<strong>et</strong> initial visait, tout d’abord, à développer une métho<strong>de</strong> <strong>de</strong> mesure <strong>de</strong>s grands risques financiers en<br />
univers non gaussien <strong>et</strong> à les modéliser à l’ai<strong>de</strong> <strong>de</strong>s outils employés en physique <strong>st<strong>at</strong>istique</strong> <strong>et</strong> <strong>théorie</strong> <strong>de</strong>s<br />
champs, qui semblaient particulièrement bien adaptés à c<strong>et</strong>te problém<strong>at</strong>ique. Nous souhaitions ensuite<br />
pouvoir appliquer ces métho<strong>de</strong>s au problème <strong>de</strong> la <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>s. Donc, dès le départ le proj<strong>et</strong><br />
était conçu <strong>de</strong> manière pluridisciplinaire avec une part <strong>de</strong> physique théorique – ou du moins faisant appel<br />
aux outils <strong>de</strong> la physique théorique – <strong>et</strong> une part <strong>de</strong> <strong>gestion</strong> financière.<br />
Volontairement, le cadre du proj<strong>et</strong> a été défini <strong>de</strong> manière assez générale afin <strong>de</strong> lui laisser le plus <strong>de</strong><br />
souplesse possible pour pouvoir évoluer au fur <strong>et</strong> à mesure <strong>de</strong> l’avancée <strong>de</strong>s recherches. En eff<strong>et</strong>, si<br />
le cadre global <strong>de</strong>s recherches à conduire apparaissait dès le départ assez clairement, il était difficile<br />
d’évaluer très précisément les résult<strong>at</strong>s auxquels nous <strong>de</strong>vions nous <strong>at</strong>tendre. En outre, <strong>de</strong> part la codirection<br />
<strong>de</strong> c<strong>et</strong>te thèse <strong>et</strong> son caractère pluridisciplinaire, il fallait que les termes du proj<strong>et</strong> conviennent<br />
à chacune <strong>de</strong>s parties. Ce point aurait pu être suj<strong>et</strong> à quelques difficultés, mais tout s’est en fait très<br />
bien déroulé <strong>et</strong> il n’y a eu aucun problème. Je dois dire que ceci tient en gran<strong>de</strong> partie au caractère <strong>de</strong><br />
mes encadrants qui ont toujours fait preuve d’une gran<strong>de</strong> ouverture d’esprit <strong>et</strong> m’ont laissé beaucoup <strong>de</strong><br />
liberté quant aux choix <strong>de</strong>s directions <strong>de</strong> recherche à suivre.<br />
A.3.2 Conduite du proj<strong>et</strong><br />
Tout au long <strong>de</strong> ma thèse, j’ai bénéficié d’un encadrement toujours très présent <strong>et</strong> en même temps d’une<br />
gran<strong>de</strong> liberté. C<strong>et</strong> ét<strong>at</strong> <strong>de</strong> fait est lié au choix <strong>de</strong> mes encadrants <strong>et</strong> je savais dès le départ à quoi m’<strong>at</strong>tendre.<br />
En eff<strong>et</strong>, <strong>de</strong> par ses activités, mon directeur <strong>de</strong> thèse n’était en France que six mois par an. Ceci<br />
m’a très rapi<strong>de</strong>ment poussé à <strong>de</strong>voir faire preuve d’autonomie. De plus, j’avais choisi un co-directeur occupant<br />
un poste à l’Université Lyon I, ce qui m’a imposé une certaine mobilité <strong>et</strong> conduit à effectuer <strong>de</strong><br />
nombreux allers-r<strong>et</strong>ours à Lyon ou Paris. Ceci m’a, en contrepartie, permis <strong>de</strong> rencontrer <strong>de</strong>s personnes<br />
extérieures à mon labor<strong>at</strong>oire <strong>de</strong> r<strong>at</strong>tachement, ce qui a conduit à développer <strong>de</strong>s collabor<strong>at</strong>ions très utiles<br />
<strong>et</strong> enrichissantes.<br />
Ceci étant, malgré la distance qui nous séparait parfois, j’ai toujours eu quelqu’un ”à l’autre bout du fil”,<br />
si ce n’est en face <strong>de</strong> moi pour répondre à mes questions <strong>et</strong> me gui<strong>de</strong>r lors <strong>de</strong> mes hésit<strong>at</strong>ions. Ceci a été<br />
extrêmement important tout au long <strong>de</strong> ma thèse <strong>et</strong> plus particulièrement au cours <strong>de</strong>s premiers mois où<br />
je n’ai fait que suivre les directives <strong>de</strong> mes encadrants. Puis, passés les six premiers mois, je me suis mis<br />
à choisir moi-même les directions <strong>de</strong> recherche que je souhaitais suivre, à en évaluer la faisabilité <strong>et</strong> à<br />
les concrétiser. Bien sûr, tout cela se faisait en plein accord avec mes encadrants <strong>et</strong> après discussion <strong>et</strong><br />
concert<strong>at</strong>ion avec eux.<br />
Ponctuellement, j’ai participé à <strong>de</strong>s proj<strong>et</strong>s a priori un peu éloignés du thème central <strong>de</strong> ma thèse. Mais<br />
au final, cela a toujours été stimulant <strong>et</strong> génér<strong>at</strong>eur d’idées nouvelles qui se sont parfaitement intégrées<br />
au corps déjà établi <strong>de</strong> mes recherches, y apportant un complément ou un éclairage différent.<br />
En résumé, s’il s’est toujours maintenu à l’intérieur du cadre originellement établi, le déroulement <strong>de</strong> ma<br />
thèse était loin d’être planifié dès le départ <strong>et</strong> s’est plutôt produit sous l’eff<strong>et</strong> <strong>de</strong>s découvertes <strong>et</strong> avancées<br />
successives.<br />
De plus, l’avancement <strong>de</strong> la thèse s’est effectué <strong>de</strong> manière rapi<strong>de</strong> puisqu’il n’a pas été nécessaire d’aller<br />
au bout <strong>de</strong>s trois ans que dure normalement une thèse. En fait, un an <strong>et</strong> <strong>de</strong>mi après le début du proj<strong>et</strong>, j’ai
A.4. Compétences acquises <strong>et</strong> enseignements personnels 499<br />
émis l’idée <strong>de</strong> soutenir avant terme ce qui est apparu raisonnable à mes encadrants, qui en ont accepté le<br />
principe. J’estimais alors avoir encore besoin <strong>de</strong> six mois pour terminer les proj<strong>et</strong>s en cours <strong>et</strong> quelques<br />
mois pour rédiger le manuscrit <strong>de</strong> thèse. Sur ce point, la rédaction régulière d’articles au fur <strong>et</strong> à mesure<br />
<strong>de</strong> l’avancée <strong>de</strong>s travaux m’a gran<strong>de</strong>ment facilité la tâche. Au final les délais que je m’étais fixés ont<br />
effectivement été tenus <strong>et</strong> j’ai soutenu au bout d’un peu plus <strong>de</strong> <strong>de</strong>ux années <strong>de</strong> thèse.<br />
A.3.3 R<strong>et</strong>ombées scientifiques<br />
Les r<strong>et</strong>ombées scientifiques ont été nombreuses <strong>et</strong> importantes. Les résult<strong>at</strong>s obtenus nous ont notamment<br />
permis d’isoler <strong>et</strong> <strong>de</strong> mieux comprendre certaines <strong>de</strong>s origines <strong>de</strong>s phénomènes extrêmes observés sur les<br />
marchés financiers. Il nous a alors été possible d’établir, sur <strong>de</strong>s bases théoriques soli<strong>de</strong>s, <strong>de</strong> nouvelles<br />
mesures <strong>de</strong>s risques extrêmes, qui ont en r<strong>et</strong>our autorisé le développement concr<strong>et</strong> <strong>de</strong> <strong>théorie</strong>s <strong>de</strong> <strong>portefeuille</strong><br />
prenant en compte ce type <strong>de</strong> risques, ce qui était l’objectif ultime <strong>de</strong> notre proj<strong>et</strong>. Globalement,<br />
tous les objectifs premiers ont été <strong>at</strong>teints <strong>et</strong> dans certains cas largement dépassés.<br />
En outre, les r<strong>et</strong>ombées scientifiques du proj<strong>et</strong> ont été rapi<strong>de</strong>s <strong>et</strong> régulières. En eff<strong>et</strong>, très tôt les premiers<br />
résult<strong>at</strong>s sont apparus. Dès les qu<strong>at</strong>re premiers mois, un article a été soumis. Puis, à partir <strong>de</strong> là, environ<br />
un papier tous les <strong>de</strong>ux ou trois mois était rédigé, si bien qu’après un peu plus <strong>de</strong> <strong>de</strong>ux années <strong>de</strong> thèse,<br />
six articles sont parus à la fois dans <strong>de</strong>s journaux <strong>de</strong> physique <strong>et</strong> <strong>de</strong>s revues <strong>de</strong> finance, <strong>et</strong> environ autant<br />
sont soumis ou à paraître.<br />
La régularité <strong>de</strong> public<strong>at</strong>ion tient avant tout à la diversific<strong>at</strong>ion <strong>de</strong> mes thèmes <strong>de</strong> recherches. C’était le<br />
principal intérêt d’avoir défini un cadre d’étu<strong>de</strong> rel<strong>at</strong>ivement large. Il est intéressant <strong>de</strong> remarquer que<br />
sur les plus <strong>de</strong> dix papiers aujourd’hui produits, seul un quart était <strong>at</strong>tendu, dans le sens où ils portent<br />
<strong>de</strong> manière très directe sur <strong>de</strong>s thèmes <strong>de</strong> recherche parfaitement i<strong>de</strong>ntifiés dès le début <strong>de</strong> la thèse. Les<br />
autres sont les fruits <strong>de</strong> la progression <strong>et</strong> <strong>de</strong>s idées nouvelles qui sont apparues au fur <strong>et</strong> à mesure <strong>de</strong><br />
l’avancement <strong>de</strong>s travaux.<br />
Il est bien évi<strong>de</strong>nt qu’une telle quantité <strong>de</strong> travail n’a pu être réalisée seul, <strong>et</strong> le nombre <strong>de</strong>s public<strong>at</strong>ions<br />
tient aussi aux collabor<strong>at</strong>ions – avec <strong>de</strong>s partenaires publics mais aussi privés – qui se sont tissées tout<br />
au long <strong>de</strong> ma thèse <strong>et</strong> qui se sont toutes révélées très fructueuses.<br />
A.4 Compétences acquises <strong>et</strong> enseignements personnels<br />
La thèse constitue pour moi une expérience riche d’enseignement. Sur le plan scientifique, elle m’a tout<br />
d’abord permis d’appliquer les connaissances que j’avais acquises en physique théorique lors <strong>de</strong> ma<br />
form<strong>at</strong>ion initiale (physique <strong>st<strong>at</strong>istique</strong>, systèmes complexes <strong>et</strong> phénomènes critiques notamment). Ceci<br />
a été pour moi le moyen <strong>de</strong> définitivement m’approprier ce savoir. J’ai aussi dû ponctuellement m<strong>et</strong>tre à<br />
jour <strong>et</strong> compléter ces connaissances lorsque mes recherches me conduisaient à abor<strong>de</strong>r <strong>de</strong>s thèmes que je<br />
n’avais pas eu à considérer lors <strong>de</strong> mes étu<strong>de</strong>s antérieures. Ceci fait qu’aujourd’hui je pense pouvoir dire<br />
que je possè<strong>de</strong> un niveau <strong>de</strong> connaissances élevées sur les métho<strong>de</strong>s employées en physique pour l’étu<strong>de</strong><br />
<strong>de</strong>s phénomènes critiques.<br />
Par ailleurs, <strong>et</strong> c’est là où j’ai dû fournir l’effort le plus important, il m’a fallu acquérir la somme <strong>de</strong><br />
connaissances en <strong>théorie</strong> financière qui me faisait défaut (<strong>théorie</strong> <strong>de</strong> la décision <strong>et</strong> analyse du risque,<br />
<strong>théorie</strong>s <strong>de</strong> <strong>portefeuille</strong> <strong>et</strong> modèles d’équilibre <strong>de</strong> marchés, pour ne citer que quelques exemples), ainsi<br />
qu’en <strong>st<strong>at</strong>istique</strong> <strong>et</strong> <strong>théorie</strong> <strong>de</strong>s probabilités. C<strong>et</strong> investissement a été long (<strong>et</strong> je ne le considère pas comme<br />
totalement achevé), mais j’estime avoir acquis un bon niveau <strong>de</strong> compétences dans ces domaines. Pour<br />
cela, j’ai – entre autre – participé à <strong>de</strong>s écoles organisées par le CNRS. Celles-ci se sont révélées très
500 A. Evalu<strong>at</strong>ion <strong>de</strong> la conduite du proj<strong>et</strong> <strong>de</strong> thèse<br />
form<strong>at</strong>rices <strong>et</strong> m’ont permis tout d’abord <strong>de</strong> combler mes lacunes puis d’acquérir <strong>de</strong>s connaissances <strong>de</strong><br />
haut niveau au contact <strong>de</strong>s personnes les plus expertes dans ces domaines. En outre, les connaissances<br />
que j’ai acquises en m<strong>at</strong>hém<strong>at</strong>ique m’ont permis <strong>de</strong> réaliser la synthèse <strong>et</strong> d’approfondir les savoirs que<br />
j’avais accumulés en physique au cours <strong>de</strong> mes étu<strong>de</strong>s antérieures.<br />
Sur le plan purement technique, j’ai développé notamment <strong>de</strong> bonnes compétences inform<strong>at</strong>iques. J’ai<br />
appris la programm<strong>at</strong>ion en langage C, afin <strong>de</strong> pouvoir réaliser <strong>de</strong>s simul<strong>at</strong>ions numériques évoluées. J’ai<br />
également appris à utiliser <strong>de</strong>s logiciels tels que M<strong>at</strong>Lab pour le traitement <strong>de</strong> données <strong>et</strong> Maple pour<br />
l’ai<strong>de</strong> au calcul formel.<br />
En outre, j’ai dû apprendre à gérer mon temps <strong>de</strong> travail ainsi que le travail en groupe. Ceci me semble<br />
être la clé du succès <strong>de</strong> tout proj<strong>et</strong> (comme la thèse par exemple, mais cela va bien au-<strong>de</strong>là). En eff<strong>et</strong>,<br />
j’ai continuellement travaillé sur plusieurs suj<strong>et</strong>s en même temps, généralement avec la même personne<br />
mais plusieurs fois au sein <strong>de</strong> collabor<strong>at</strong>ion regroupant <strong>de</strong>s personnes géographiquement dispersées sur<br />
l’ensemble du globe. Ceci conduit à s’astreindre à une certaine rigueur dans l’organis<strong>at</strong>ion afin <strong>de</strong> ne pas<br />
se laisser débor<strong>de</strong>r <strong>et</strong> <strong>de</strong> maintenir une progression régulière <strong>de</strong>s différents proj<strong>et</strong>s. En fait, un travail en<br />
collabor<strong>at</strong>ion n’avançant qu’au rythme du plus lent <strong>de</strong> ces membres, le r<strong>et</strong>ard peut parfois s’accumuler <strong>de</strong><br />
manière importante. C’est pourquoi il m’a aussi fallu apprendre à limiter mes ambitions <strong>et</strong> ainsi refuser<br />
<strong>de</strong> participer à certains proj<strong>et</strong>s afin <strong>de</strong> ne pas trop me disperser. Par ailleurs, mes travaux se trouvant à<br />
l’interface <strong>de</strong> <strong>de</strong>ux mon<strong>de</strong>s – celui <strong>de</strong> la physique <strong>et</strong> celui <strong>de</strong> la finance – j’ai eu sans cesse à interagir<br />
avec <strong>de</strong>s interlocuteurs parlant <strong>de</strong>s langages différents <strong>et</strong> ayant <strong>de</strong>s sensibilités bien distinctes. Il a donc<br />
été nécessaire que je m’adapte à cela <strong>et</strong> notamment que j’apprenne à communiquer avec <strong>de</strong>s personnes<br />
ayant à la base <strong>de</strong>s form<strong>at</strong>ions <strong>et</strong> cultures très différentes <strong>de</strong> la mienne, ce qui s’est fait p<strong>et</strong>it à p<strong>et</strong>it mais<br />
sans gran<strong>de</strong>s difficultés.<br />
Enfin, bénéficiant d’un monitor<strong>at</strong> à l’Université <strong>de</strong> Nice, j’ai eu à assurer une charge d’enseignement <strong>de</strong><br />
64 heures par an. C<strong>et</strong>te charge s’est trouvée repartie entre différents niveaux, allant du premier cycle à la<br />
prépar<strong>at</strong>ion à l’agrég<strong>at</strong>ion, <strong>et</strong> s’est déroulée <strong>de</strong>vant <strong>de</strong>s publics divers : étudiants en physique, géologie,<br />
ou encore ponctuellement économie. Ces quelques heures d’enseignement ont été réellement profitables.<br />
Elles ont bien sûr été l’occasion <strong>de</strong> m<strong>et</strong>tre en pr<strong>at</strong>ique les conseils que j’avais reçus lors <strong>de</strong> ma form<strong>at</strong>ion<br />
à l’enseignement au cours <strong>de</strong> mon année <strong>de</strong> prépar<strong>at</strong>ion à l’agrég<strong>at</strong>ion, mais elles ont en fait surtout été<br />
le moyen d’apprendre à transm<strong>et</strong>tre <strong>de</strong>s connaissances. En eff<strong>et</strong>, l’effort <strong>de</strong> communic<strong>at</strong>ion en direction<br />
d’un public étudiant est bien plus important <strong>et</strong> totalement différent <strong>de</strong> celui auquel doit s’habituer le<br />
chercheur qui expose ses travaux lors d’une conférence <strong>de</strong>vant un public <strong>de</strong> spécialistes. En outre, les<br />
questions – parfois naïves – <strong>de</strong>s étudiants imposent une remise en cause, ou du moins une réorganis<strong>at</strong>ion,<br />
permanente <strong>de</strong> ses propres connaissances. Enfin, cela m’a là encore conduit à travailler en coordin<strong>at</strong>ion,<br />
si ce n’est en collabor<strong>at</strong>ion, avec mes collègues enseignants au sein d’équipes pédagogiques, forgeant un<br />
peu plus mon expérience du travail en groupe.<br />
A.5 Conclusion<br />
J’estime que ma thèse s’est extrêmement bien déroulée <strong>et</strong> j’en gar<strong>de</strong>rai le souvenir d’une expérience très<br />
positive. J’<strong>at</strong>tribue cela essentiellement à la très bonne entente qui a toujours régné entre mes encadrants<br />
<strong>et</strong> moi-même. Je ne conçois pas avoir pu <strong>at</strong>teindre les mêmes objectifs si <strong>de</strong>s conditions <strong>de</strong> travail aussi<br />
agréables n’avaient pas été réunies. Cela souligne l’importance <strong>de</strong>s rel<strong>at</strong>ions humaines dans la bonne<br />
marche <strong>de</strong> tout proj<strong>et</strong>. Ceci me parait d’autant plus vrai que les groupes <strong>de</strong> travail sont p<strong>et</strong>its <strong>et</strong> qu’en<br />
conséquence les membres sont plus dépendants les uns <strong>de</strong>s autres.<br />
C<strong>et</strong>te expérience très motivante m’a conforté dans l’intention <strong>de</strong> poursuivre ma carrière dans le secteur
A.5. Conclusion 501<br />
<strong>de</strong> la recherche – académique ou privée. En outre, l’enseignement – ou la form<strong>at</strong>ion au sens large –<br />
me semble désormais indissociable <strong>de</strong> c<strong>et</strong>te activité. Je pense que mes capacités d’enseignant se sont<br />
nourries <strong>de</strong> mon expérience <strong>de</strong> chercheur <strong>et</strong> que réciproquement, les échanges lors <strong>de</strong> conférences ou<br />
séminaires ont fourni généralement m<strong>at</strong>ière à <strong>de</strong> nouvelles recherches. Je dois à ces années <strong>de</strong> thèse<br />
d’avoir découvert la formidable complémentarité entre d’une part une activité quelque peu solitaire qu’est<br />
la recherche <strong>et</strong> d’autre part une activité tournée vers l’extérieur qu’est l’enseignement ou la communic<strong>at</strong>ion.<br />
Le tableau que j’ai dressé <strong>de</strong> mes années <strong>de</strong> thèse peut paraître idyllique, il n’en est pas moins sincère.<br />
J’ai sûrement bénéficié d’une part <strong>de</strong> chance – inhérente à toute entreprise – mais cela ne saurait tout<br />
expliquer. Je pense que l’organis<strong>at</strong>ion méthodique <strong>de</strong> mon proj<strong>et</strong>, ce qui commence bien avant le début<br />
officiel <strong>de</strong> la thèse <strong>et</strong> passe par le choix avisé du suj<strong>et</strong> <strong>et</strong> <strong>de</strong>s encadrants, est la clé <strong>de</strong> la réussite <strong>de</strong> c<strong>et</strong>te<br />
entreprise. Pour ma part, je m’étais astreint à c<strong>et</strong>te démarche plus <strong>de</strong> six mois avant le début <strong>de</strong> ma thèse,<br />
ce qui, me semble-t-il, a porté ses fruits bien au-<strong>de</strong>là <strong>de</strong> mes espérances initiales.
502 A. Evalu<strong>at</strong>ion <strong>de</strong> la conduite du proj<strong>et</strong> <strong>de</strong> thèse
Bibliographie<br />
ACERBI, C. (2002) : “Spectral measures of risk : A coherent represent<strong>at</strong>ion of subjective risk aversion”,<br />
Journal of Banking & Finance 26, 1505–1518.<br />
ACERBI, C. ET D. TASCHE (2002) : “On the coherence of expected shorfall”, Journal of Banking &<br />
Finance 26, 1487–1503.<br />
ACERBI, C. ET P. SIMONETTI (2002) : “Portfolio optimiz<strong>at</strong>ion with spectral measures of risk”, Document<br />
<strong>de</strong> Travail.<br />
ADAMS, M. C. ET A. SZARFARZ (1992) : “Specul<strong>at</strong>ive bubbles and financial mark<strong>et</strong>s”, Oxford Economic<br />
Papers 44, 626–640.<br />
AHN, D., J. BOUDOUKH, M. RICHARDSON ET R. WHITELAW (1999) : “Optimal risk management<br />
using options”, Journal of Finance 54, 359–376.<br />
ALEXANDER, G. J. ET A. M. BAPTISTA (2002) : “Economic Implic<strong>at</strong>ions of Using a Mean-VaR Mo<strong>de</strong>l<br />
for Portfolio Selection : A Comparison with Mean-Variance Analysis”, Journal of Economic<br />
Dynamics & Control 26, 1159–1193.<br />
ALLAIS, M. (1953) : “Le comportement <strong>de</strong> l’homme r<strong>at</strong>ionnel <strong>de</strong>vant le risque : critique <strong>de</strong>s postul<strong>at</strong>s<br />
<strong>de</strong> l’école amricaine”, Econom<strong>et</strong>rica 21, 503–546.<br />
ANDERSEN, J. V. ET D. SORNETTE (2002a) : “Have your cake and e<strong>at</strong> it too : Increasing r<strong>et</strong>urns while<br />
lowering large risks !”, Journal of Risk Finance 2, 70–82.<br />
ANDERSEN, J. V. ET D. SORNETTE (2002b) : “the $-game”, Document <strong>de</strong> Travail 0205423, Cond-m<strong>at</strong>.<br />
ANDERSSON, M., B. EKLUND ET J. LYHAGEN (1999) : “A simple linear time series mo<strong>de</strong>l with misleading<br />
nonlinear properties”, Economics L<strong>et</strong>ters 65, 281–285.<br />
ARDITTI, F. (1967) : “Risk and the required r<strong>et</strong>urn on equity”, Journal of Finance 22, 19–36.<br />
ARNÉODO, A., J. F. MUZY ET D. SORNETTE (1998) : “Causale casca<strong>de</strong> in the stock mark<strong>et</strong> from the<br />
infrared to the ultraviol<strong>et</strong>”, European Physical journal B 2, 277–282.<br />
ARTHUR, W. B. (1987) : “Inductive Reasoning and Boun<strong>de</strong>d R<strong>at</strong>ionality : Self-reinforcing mechanisms<br />
in economics”, Center for Economic Policy Research 111, 1–20.<br />
ARTHUR, W. B. (1994) : “Inductive Reasoning and Boun<strong>de</strong>d R<strong>at</strong>ionality”, American Economic Review<br />
84, 406–411.<br />
ARTHUR, W. B., J. H. HOLLAND, B. LEBARON, R. PALMER ET P. TAYLOR (1997) : “Ass<strong>et</strong> pricing un<strong>de</strong>r<br />
endogeneous expect<strong>at</strong>ions in an artificial stock mark<strong>et</strong>”, in Arthur, W. B., D. Lane <strong>et</strong> S. Durlauf<br />
(eds.), The Economy as an Evolving Complex System II, Addison-Wesley, Redwood City.<br />
ARTZNER, P., F. DELBAEN, J.-M. EBER ET D. HEATH (1999) : “Coherent measures of risk”, M<strong>at</strong>hem<strong>at</strong>ical<br />
Finance 9, 203–288.<br />
AVNIR, D., O. BIHAM, D. LIDAR ET O. MALCAI (1998) : “Is the geom<strong>et</strong>ry of n<strong>at</strong>ure fractal ?”, Science<br />
279, 39–40.<br />
503
504 Bibliographie<br />
BACHELIER, L. (1900) : “Théorie <strong>de</strong> la spécul<strong>at</strong>ion”, Annales Scientifiques <strong>de</strong> l’Ecole Normale<br />
Supérieure 17, 21–86.<br />
BACRY, E., J. DELOUR ET J. F. MUZY (2001) : “Multifractal random walk”, Physical Review E<br />
64(26103).<br />
BACRY, E. ET J. F. MUZY (2002) : “Log-infinitely divisible multifractal processes”, Document <strong>de</strong> Travail<br />
0207094, Cond-m<strong>at</strong>.<br />
BASLE COMMITTEE ON BANKING SUPERVISION (1996) : “Amen<strong>de</strong>ment to the Capital Accord to Incorpor<strong>at</strong>e<br />
Mark<strong>et</strong> Risks”.<br />
BASLE COMMITTEE ON BANKING SUPERVISION (2001) : “The New Basel Capital Accord”.<br />
BAVIERA, R., L. BIFERALE, R. N. MANTEGNA ET A. VULPIANI (1998) : “Transient multiaffine behaviors<br />
in ARCH and GARCH processes”, Intern<strong>at</strong>ional Workshop on Econophysics and St<strong>at</strong>istical<br />
Finance, Italy.<br />
BECK, U. (2001) : La société du risque, Aubier.<br />
BERNOULLI, D. (1738) : “Specimen theoriae novae <strong>de</strong> mensura sortis”, Commentarii Aca<strong>de</strong>miae Scientiarum<br />
Imperialis P<strong>et</strong>ropolitanae .<br />
BINGHAM, N. H., C. M. GOLDIE ET J. L. TEUGEL (1987) : Regular Vari<strong>at</strong>ion, Cambridge University<br />
Press.<br />
BLACK, F. (1976) : “Studies of stock price vol<strong>at</strong>ility changes”, Proceeding of the Business and Economic<br />
St<strong>at</strong>istics Section pp. 177–181.<br />
BLACK, F. ET M. SCHOLES (1973) : “The pricing of options and corpor<strong>at</strong>e liabilities”, Journal of Political<br />
Economy 81, 637–653.<br />
BLANCHARD, O. J. (1979) : “Specul<strong>at</strong>ive bubble, crashes and r<strong>at</strong>ional expect<strong>at</strong>ions”, Economics L<strong>et</strong>ters<br />
3, 387–396.<br />
BLANCHARD, O. J. ET M. W. WATSON (1982) : “Bubbles, r<strong>at</strong>ional expect<strong>at</strong>ions and specul<strong>at</strong>ive mark<strong>et</strong>s”,<br />
in W<strong>at</strong>chel, P. (ed.), Crisis in Economic and Financial Structure : Bubles, Bursts and Shocks,<br />
Lexington Books, Lexington.<br />
BLUM, P., A. DIAS ET P. EMBRECHTS (2002) : “The ART of <strong>de</strong>pen<strong>de</strong>nce mo<strong>de</strong>lling : the l<strong>at</strong>est advances<br />
in correl<strong>at</strong>ion analysis”, in Lane, M. (ed.), Altern<strong>at</strong>ive Risk Str<strong>at</strong>egies, Risk Books, London,<br />
pp. 339–356.<br />
BOLLERSLEV, T. (1986) : “Generalized autoregressive conditional h<strong>et</strong>eroskdasticity”, Journal of Econom<strong>et</strong>rics<br />
31, 307–327.<br />
BOLLERSLEV, T., R. F. ENGLE ET D. B. NELSON (1994) : “ARCH mo<strong>de</strong>ls”, in Engle, R. F. <strong>et</strong> D. Mc-<br />
Fad<strong>de</strong>n (eds.), Handbook of Econom<strong>et</strong>rics, Vol. IV, Elsevier, pp. 2959–3038.<br />
BOLLERSLEV, T., R. Y. CHOU ET K. F. KRONER (1992) : “ARCH mo<strong>de</strong>ling in finance : A review of<br />
the theory and empirical evi<strong>de</strong>nce”, Journal of Econom<strong>et</strong>rics 52, 5–59.<br />
BOUCHAUD, J. P., A. MATACZ ET M. POTTERS (2001) : “The leverage effect in financial mark<strong>et</strong>s :<br />
r<strong>et</strong>ar<strong>de</strong>d vol<strong>at</strong>ility and mark<strong>et</strong> panic”, Physical Review L<strong>et</strong>ters 87(228701).<br />
BOUCHAUD, J. P., D. SORNETTE, C. WALTER ET J. P. AGUILAR (1998) : “Taming large events :<br />
Optimal portfolio theory for strongly fluctu<strong>at</strong>ing ass<strong>et</strong>s”, Intern<strong>at</strong>ional Journal of Theor<strong>et</strong>ical and<br />
Applied Finance 1, 25–41.<br />
BOUCHAUD, J. P., M. POTTERS ET M. MEYER (2000) : “Apparent multifractality in financial time<br />
series”, European Physical Journal B 13, 595–599.<br />
BOUCHAUD, J. P. ET M. POTTERS (2000) : Theory of Financial Risks : From St<strong>at</strong>istical Physics to Risk<br />
Management, Cambridge University Press.
Bibliographie 505<br />
BOUCHAUD, J. P. ET R. CONT (1998) : “A Langevin approach to stock mark<strong>et</strong> fluctu<strong>at</strong>ions and crashes”,<br />
European Physical Journal B 6, 543–550.<br />
BOUYÉ, E., V. DURRLEMAN, A. NIKEGHBALI, G. RIBOULET ET T. RONCALLI (2000) : “Copulas<br />
for finance : A reading gui<strong>de</strong> and some applic<strong>at</strong>ions”, Document <strong>de</strong> Travail, Groupe <strong>de</strong> Recherche<br />
Opér<strong>at</strong>ionelle, Crédit Lyonnais.<br />
BRIEMAN, L. (1960) : “Investment policies for expanding businesses optimal in a long run sense”, Naval<br />
Research Logistics Qu<strong>at</strong>erly 7, 647–651.<br />
BROCK, W. A. (1993) : “P<strong>at</strong>hways to randomness in the economy : Emergent nonlinearity and chaos in<br />
economics and finance”, Estudios Economicos 8.<br />
BROCK, W. A. ET B. LEBARON (1996) : “A dynamic structural mo<strong>de</strong>l for stock r<strong>et</strong>urn vol<strong>at</strong>ility and<br />
trading volume”, Review of Economics and St<strong>at</strong>istics 78, 94–110.<br />
BROCK, W. A. ET C. H. HOMMES (1997) : “R<strong>at</strong>ional route to randomness”, Econom<strong>et</strong>rica 65, 1059–<br />
1095.<br />
BROWN, P., D. WALSH ET A. YUEN (1997) : “The interaction b<strong>et</strong>ween or<strong>de</strong>r imbalance and stock<br />
price”, Pacific-Basin Finance Journal 5, 539–557.<br />
BROZE, L., C. GOURIÉROUX ET A. SZAFARZ (1990) : Reduced Forms of R<strong>at</strong>ional Expect<strong>at</strong>ions Mo<strong>de</strong>ls,<br />
Harwood Aca<strong>de</strong>mic Press.<br />
BURDA, Z., J. JURKIEWICZ, M. NOWAK, G. PAPP ET I. ZAHED (2001a) : “Free Lévy m<strong>at</strong>rices and<br />
financial correl<strong>at</strong>ions”, Document <strong>de</strong> Travail 0103109, Cond-m<strong>at</strong>.<br />
BURDA, Z., J. JURKIEWICZ, M. NOWAK, G. PAPP ET I. ZAHED (2001b) : “Lévy m<strong>at</strong>rices and financial<br />
covariances”, Document <strong>de</strong> Travail 0103108, Cond-m<strong>at</strong>.<br />
BURDA, Z., R. JANIK, J. JURKIEWICZ, M. NOWAK, G. PAPP ET I. ZAHED (2002) : “Free random<br />
Lévy m<strong>at</strong>rices”, Physical Review. E 65(011106).<br />
CAMPBELL, J., A. W. LO ET C. MACKINLAY (1997) : The Econom<strong>et</strong>rics of Financial Mark<strong>et</strong>s, Princ<strong>et</strong>own<br />
University press.<br />
CASADESUS-MASANELL, R., P. KLIBANOFF ET E. OZDENOREN (2000) : “Maxmin expected utility<br />
through st<strong>at</strong>ewise combin<strong>at</strong>ions”, Economics L<strong>et</strong>ters 66, 49–54.<br />
CAVES, C. M., C. A. FLUCHS ET R. SCHACK (2002) : “Quantum probabilities as Bayesian probabilities”,<br />
Physical Review A 65(022305).<br />
CHABAANE, A., E. DUCLOS, J. P. LAURENT, Y. MALEVERGNE ET F. TURPIN (2002) : “Looking for<br />
efficient portofolios : An empirical investig<strong>at</strong>ion”, Document <strong>de</strong> Travail.<br />
CHALLET, D., A. CHESSA, M. MARSILI ET Y. C. ZHANG (2001) : “From minority games to real<br />
financial mak<strong>et</strong>s”, Quantit<strong>at</strong>ive Finance 1, 168–176.<br />
CHALLET, D. ET Y. C. ZHANG (1997) : “Emergence of Cooper<strong>at</strong>ion and Organiz<strong>at</strong>ion in an Evolutionary<br />
Game”, Physica A 246, 407–418.<br />
CHAN, K. ET W.-M. FONG (2000) : “Tra<strong>de</strong> size, or<strong>de</strong>r imbalance, and the vol<strong>at</strong>ility-volume rel<strong>at</strong>ion”,<br />
Journal of Financial Economics 57, 247–273.<br />
CHATEAUNEUF, A. (1991) : “On the use of capacities in mo<strong>de</strong>ling uncertainty aversion and risk aversion”,<br />
Journal of M<strong>at</strong>hem<strong>at</strong>ical Economics 20, 343–369.<br />
CHEKHLOV, A., S. URYASEV ET M. ZABARANKIN (2000) : “Portfolio Optimiz<strong>at</strong>ion with Drawdown<br />
Constraints”, Document <strong>de</strong> Travail, ISE Dept., Univ. of Florida.<br />
CHERUBINI, U. ET E. LUCIANO (2000) : “Multivari<strong>at</strong>e option pricing with copula”, Document <strong>de</strong> Travail,<br />
SSRN.
506 Bibliographie<br />
CHORDIA, T., R. ROLL ET A. SUBRAHMANYAM (2002) : “Or<strong>de</strong>r imbalance, liquidity, and mark<strong>et</strong><br />
r<strong>et</strong>urns”, Journal of Financial Economics 65, 111–130.<br />
CHRISTIE, A. A. (1982) : “The stochastic behavior of common stock variances : Value, leverage and<br />
interest r<strong>at</strong>e effects”, Journal of Financial Economics 10, 407–432.<br />
COHEN, M. ET J. M. TALLON (2000) : “Décision dans le risque <strong>et</strong> l’incertain : l’apport <strong>de</strong>s modèles non<br />
additifs”, Revue d’Economie Politique 110, 631–681.<br />
COLES, S., J. HEFFERNAN ET J. TAWN (1999) : “Depen<strong>de</strong>nce measures for extreme value analysis”,<br />
Extremes 2, 339–365.<br />
COLLETAZ, G. ET J. P. GOURLAOUEN (1989) : “Les bulles r<strong>at</strong>ionnelles : une synthèse <strong>de</strong> la littér<strong>at</strong>ure”,<br />
in Bourguin<strong>at</strong>, H. <strong>et</strong> P. Artus (eds.), Théorie Economique <strong>et</strong> Crises <strong>de</strong>s Marchés Financiers, Economica,<br />
pp. 67–101.<br />
CONSIGLI, G. (2002) : “Tail estim<strong>at</strong>ion and mean-VaR portfolio selection in mark<strong>et</strong>s subject to financial<br />
instability”, Journal of Banking & Finance 26, 1355–1382.<br />
CONSIGLI, G., G. FRASCELLA ET G. SARTORELLI (2001) : “Un<strong>de</strong>rstanding financial mark<strong>et</strong> with<br />
extreme value theory from Value-<strong>at</strong>-Risk to crises correl<strong>at</strong>ion analysis”, Document <strong>de</strong> Travail, Uni-<br />
Credit Banca Mobiliare.<br />
CONT, R. (2001) : “Empirical properties of ass<strong>et</strong> r<strong>et</strong>urns : stylized facts and st<strong>at</strong>istical issues”, Quantit<strong>at</strong>ive<br />
Finance 1, 223–236.<br />
CONT, R., M. POTTERS ET J. P. BOUCHAUD (1997) : “Scaling in stock mark<strong>et</strong> d<strong>at</strong>a : Stable laws and<br />
beyond”, in Dubrulle, Graner <strong>et</strong> Sorn<strong>et</strong>te (eds.), Scale Invariance and Beyond, Springer, Berlin,<br />
pp. 75–85.<br />
CONT, R. ET J. P. BOUCHAUD (2000) : “Herd behavior and aggreg<strong>at</strong>e fluctu<strong>at</strong>ions in financial mark<strong>et</strong>s”,<br />
Macroeconomic Dynamics 4, 170–196.<br />
CORCOS, A., J. P. ECKMANN, A. MALASPINAS, Y. MALEVERGNE ET D. SORNETTE (2002) : “Imit<strong>at</strong>ion<br />
and contrarian behavior : Hyperbolic bubbles, crashes and chaos”, Quantit<strong>at</strong>ive Finance<br />
2, 264–281.<br />
COSSETTE, H., P. GAILLARDETZ, E. MARCEAU ET J. RIOUX (2002) : “On two <strong>de</strong>pen<strong>de</strong>nt individual<br />
risk mo<strong>de</strong>ls”, Insurance : M<strong>at</strong>hem<strong>at</strong>ics and Economics 30, 153–166.<br />
COUTANT, S., V. DURRLEMAN, G. RAPUCH ET T. RONCALLI (2001) : “Copulas, multivari<strong>at</strong>e riskneutral<br />
distributions and implied <strong>de</strong>pen<strong>de</strong>nce functions”, Document <strong>de</strong> Travail, Groupe <strong>de</strong> Recherche<br />
Opér<strong>at</strong>ionelle, Crédit Lyonnais.<br />
CVITANIC, J. ET I. KARATZAS (1995) : “On portfolio optimiz<strong>at</strong>ion un<strong>de</strong>r ‘drawdown’ constraints”, IMA<br />
Lecture Notes in M<strong>at</strong>hem<strong>at</strong>ics & Applic<strong>at</strong>ions 65, 77–88.<br />
DACOROGNA, M. M., R. GENÇAY, U. A. MÜLLER ET O. V. PICTET (2001) : “Effective r<strong>et</strong>urn, risk<br />
aversion and drawdowns”, Physica A 289, 229–248.<br />
DACOROGNA, M. M., U. A. MÜLLER, O. V. PICTET ET C. G. DE VRIES (1992) : “The distribution of<br />
extremal foreign exchange r<strong>at</strong>e r<strong>et</strong>urns in large d<strong>at</strong>a s<strong>et</strong>s”, Document <strong>de</strong> Travail 19921022, Olsen<br />
and Associ<strong>at</strong>es Internal Documents UAM.<br />
DANIELSON, J. ET C. G. DE VRIES (2000) : “Value-<strong>at</strong>-Risk and extreme r<strong>et</strong>urns”, in Embrechts, P.<br />
(ed.), Extremes and Integr<strong>at</strong>ed Risk Management, RISK Books in associ<strong>at</strong>ion with USB Warburg,<br />
pp. 85–106.<br />
DANIELSSON, J., P. EMBRECHTS, C. GOODHART, C. KEATING, F. MUENNICH, O. RENAULT ET H.-<br />
S. SHIN (2001) : “An aca<strong>de</strong>mic response to Basel II”, Document <strong>de</strong> Travail 130, FMG and ESRC,<br />
London.
Bibliographie 507<br />
DE FINETTI, B. (1937) : “La prévision : ses lois logiques, ses sources subjectives”, Annales <strong>de</strong> l’Institut<br />
Henri Poincaré 7, 1–68.<br />
DE VRIES, C. G. (1994) : “Stylized facts of nominal exchange r<strong>at</strong>e r<strong>et</strong>urns”, in van <strong>de</strong>r Ploeg, F. (ed.),<br />
The Handbook of Intern<strong>at</strong>ional Macroeconomics, Blackwell, Oxford, pp. 348–389.<br />
DELBAEN, F. (2000) : “Coeherent risk measures on general probability spaces”, Document <strong>de</strong> Travail,<br />
Risklab.<br />
DENAULT, M. (2001) : “Coherent alloc<strong>at</strong>ion of risk capital”, Journal of Risk 3, 1–34.<br />
DURRLEMAN, V., A. NIKEGHBALI ET T. RONCALLI (2000) : “Copulas approxim<strong>at</strong>ion and new families”,<br />
Document <strong>de</strong> Travail, Groupe <strong>de</strong> Recherche Opér<strong>at</strong>ionelle, Crédit Lyonnais.<br />
EINSTEIN, A. (1905) : “Über die von <strong>de</strong>r molekularkin<strong>et</strong>ishen Theorie <strong>de</strong>r Wärme gefor<strong>de</strong>rte Bewegung<br />
von in ruhen<strong>de</strong>n Flüssigkeiten suspendierten Teilchen”, Annalen <strong>de</strong> Physik 17, 549–560.<br />
ELLSBERG, D. (1961) : “Risk, ambiguity, and the Savage axioms”, Quarterly Journal of Economics<br />
75, 643–669.<br />
ELTON, E. ET M. GRUBER (1995) : Mo<strong>de</strong>rn Portfolio Theory and Investment Analisys, John Willey &<br />
Sons.<br />
EMBRECHTS, P., A. HOEING ET A. JURI (2001) : “Using Copulae to bound the Value-<strong>at</strong>-Risk for<br />
functions of <strong>de</strong>pen<strong>de</strong>nt risk”, Document <strong>de</strong> Travail, Risklab.<br />
EMBRECHTS, P., A. MCNEIL ET D. STRAUMANN (2002) : “Correl<strong>at</strong>ion and <strong>de</strong>pen<strong>de</strong>nce in risk management<br />
: properties and pitfalls”, in Dempster, M. A. H. (ed.), Risk Management : Value <strong>at</strong> Risk<br />
and Beyond, Cambridge University Press, Cambridge, pp. 176–223.<br />
EMBRECHTS, P., C. P. KLÜPPELBERG ET T. MIKOSH (1997) : Mo<strong>de</strong>lling Extremal Events, Springer-<br />
Verlag, Berlin.<br />
ENGLE, R. F. (1982) : “Autoregressive conditional h<strong>et</strong>eroskedasticity with estim<strong>at</strong>e of the variance of<br />
UK infl<strong>at</strong>ion”, Econom<strong>et</strong>rica 50, 987–1008.<br />
ENGLE, R. F. ET A. J. PATTON (2001) : “Wh<strong>at</strong> good is a vol<strong>at</strong>ility mo<strong>de</strong>l ?”, Quantit<strong>at</strong>ive Finance<br />
1, 237–245.<br />
ERDÖS, P. ET A. RENYI (1960) : “On the evolution of random graphs”, Public<strong>at</strong>ions of the M<strong>at</strong>hem<strong>at</strong>ical<br />
Institute of the Hungarian Aca<strong>de</strong>my of Sciences 5, 17–61.<br />
FAMA, E. F. (1963) : “Man<strong>de</strong>lbrot and the stable par<strong>et</strong>ian hypothesis”, Journal of Business 36, 420–449.<br />
FAMA, E. F. (1965a) : “The behavior of stock mark<strong>et</strong> prices”, Journal of Business 38, 29–105.<br />
FAMA, E. F. (1965b) : “Portofolio analysis in a stable par<strong>et</strong>ian mark<strong>et</strong>”, Management Science 11, 404–<br />
419.<br />
FAMA, E. F. (1970) : “Multi-period consumption-investment <strong>de</strong>cision”, American Economic Review<br />
60, 163–174.<br />
FAMA, E. F. (1971) : “Efficient capital mark<strong>et</strong>s : A review of theory and empirical work”, Journal of<br />
Finance 25, 383–417.<br />
FANG, H. B., K. T. FANG ET S. KOTZ (2002) : “The m<strong>et</strong>a-elliptical distributions with given marginals”,<br />
Journal of Multivari<strong>at</strong>e Analysis 82, 1–16.<br />
FARMER, J. D. (1998) : “Mark<strong>et</strong> force, ecology and evolution”, Document <strong>de</strong> Travail, Santa Fe Institute.<br />
FISHER, A., L. CALVET ET B. MANDELBROT (1998) : “Multifractal analysis of USD/DM exchange<br />
r<strong>at</strong>es”, Document <strong>de</strong> Travail, Yale University.<br />
FISHER, T. (2001) : “Coherent risk mesures <strong>de</strong>pending on higher moments”, Document <strong>de</strong> Travail,<br />
University of Hei<strong>de</strong>lberg.
508 Bibliographie<br />
FÖLLMER, H. ET A. SCHIED (2002a) : “Convex measures of risk and trading constraints”, Finance &<br />
Stochastics 6, 429–447.<br />
FÖLLMER, H. ET A. SCHIED (2002b) : “Robust preference and convex measures of risk”, Advances in<br />
Finance & Stochastics .<br />
FREES, W. E. ET E. A. VALDEZ (1998) : “Un<strong>de</strong>rstanding rel<strong>at</strong>ionship using copulas”, North American<br />
Actuarial Journal 2, 1–25.<br />
FREY, R., A. MCNEIL ET M. NYFELER (2001) : “Credit Risk and Copulas”, Document <strong>de</strong> Travail,<br />
Risklab.<br />
FREY, R. ET A. J. MCNEIL (2002) : “VaR and expected shortfall in portfolios of <strong>de</strong>pen<strong>de</strong>nt credit risks :<br />
Conceptual and practical insights”, Journal of Banking & Finance 26, 1317–1334.<br />
FREY, R. ET A. MCNEIL (2001) : “Mo<strong>de</strong>lling <strong>de</strong>pen<strong>de</strong>nt <strong>de</strong>faults”, Document <strong>de</strong> Travail, Risklab.<br />
FRISCH, U. (1995) : Turbulence : The legacy of A.N. Kolmogorov, Cambridge university Press.<br />
FRISCH, U. ET D. SORNETTE (1997) : “Extreme <strong>de</strong>vi<strong>at</strong>ions and applic<strong>at</strong>ions”, Journal <strong>de</strong> Physique I<br />
France 7, 1155–1171.<br />
GIARDINA, I. ET J. P. BOUCHAUD (2002) : “Bubbles, crashes and intermittency in agent based mark<strong>et</strong><br />
mo<strong>de</strong>ls”, Document <strong>de</strong> Travail 0206222, Cond-m<strong>at</strong>.<br />
GILBOA, I. ET D. SCHMEIDLER (1989) : “Maxmin expected utility with non-unique prior”, Journal of<br />
M<strong>at</strong>hem<strong>at</strong>ical Economics 18, 141–153.<br />
GLOSTEN, L. R., R. JAGANNANTHAN ET D. E. RUNKLE (1993) : “On the rel<strong>at</strong>ion b<strong>et</strong>ween the expected<br />
value and the vol<strong>at</strong>ilty of the nominal excess r<strong>et</strong>urns on stocks”, Journal of Finance 48, 779–801.<br />
GOLDBERG, J. ET R. VON NITZSCH (2001) : Behavioral Finance, John Wiley.<br />
GOLDIE, C. M. (1991) : “Implicit renewal theory and tails of solution of random equ<strong>at</strong>ions”, Annals of<br />
Applied Probability 1, 126–172.<br />
GOLDIE, C. M. ET C. P. KLÜPPELBERG (1998) : “Subexponential distributions”, in Adler, R., R. Feldman<br />
<strong>et</strong> M. Taqqu (eds.), A Practical Gui<strong>de</strong> to Heavy Tails : St<strong>at</strong>istical Techniques for Analysing<br />
Heavy Tailed Distributions, Birkhäuser, Boston, pp. 435–459.<br />
GONEDES, N. (1976) : “Capital mark<strong>et</strong> equilibrium for a class of h<strong>et</strong>erogeneous expect<strong>at</strong>ions in a twoparam<strong>et</strong>er<br />
world”, Journal of Finance 31, 1–15.<br />
GOPIKRISHNAN, P., M. MEYER, L. A. N. AMARAL ET H. E. STANLEY (1998) : “Inverse Cubic Law<br />
for the Distribution of Stock Price Vari<strong>at</strong>ions”, European Physical Journal B 3, 139 –140.<br />
GOPIKRISHNAN, P., V. PLEROU, X. GABIAX ET H. E. STANLEY (2000) : “St<strong>at</strong>istical properties of<br />
share volume tra<strong>de</strong>d in financial mark<strong>et</strong>s”, physical Review E 62, 3023–3026.<br />
GOURIÉROUX, C. (1997) : ARCH Mo<strong>de</strong>ls and Financial Applic<strong>at</strong>ions, Springer, Berlin.<br />
GOURIÉROUX, C., J. LAURENT ET O. SCAILLET (2000) : “Sensitivity analysis of values <strong>at</strong> risk”, Journal<br />
of Empirical Finance 7, 225–245.<br />
GOURIÉROUX, C. ET J. JASIAK (1998) : “Trunc<strong>at</strong>ed maximum likelihood, goodness of fit tests and tail<br />
analysis”, Document <strong>de</strong> Travail, CREST.<br />
GOURIÉROUX, C. ET J. JASIAK (2001) : Financial Econom<strong>et</strong>rics, Princ<strong>et</strong>own university Press.<br />
GRANDMONT, J. M. (1998) : “Expect<strong>at</strong>ion form<strong>at</strong>ion and stability of large socioeconomic systems”,<br />
Econom<strong>et</strong>rica 66, 741–782.<br />
GRANGER, C. W. J. ET R. JOYEUX (1980) : “An introduction to long-range time series mo<strong>de</strong>ls and<br />
fractional differencing”, Journal of Time Series Analysis 1, 15–30.
Bibliographie 509<br />
GRANGER, C. W. ET T. TERÄSVIRTA (1999) :“Asimple nonlinear mo<strong>de</strong>l with misleading linear properties”,EconomicsL<strong>et</strong>ters62,161–165.<br />
GROSSMAN, S. J. ET Z. ZHOU (1993) :“Optimal investment str<strong>at</strong>egies for controlling drawdowns”,<br />
M<strong>at</strong>hem<strong>at</strong>icalFinance3,241–276.<br />
GUILLAUME, D. M., M. M. DACOROGNA, R. R. DAVÉ, U. A. MÜLLER, R. B. OLSEN ET O. V.<br />
PICTET (1997) :“From the bird eye to the microscope :asurvey of new stylized facts of the intradayforeignexchange<br />
mark<strong>et</strong>s”,Finance&Stochastics1,95–130.<br />
HAKANSSON,N.(1971):“Capitalgrothandthemean-varianceapproachtoportfolioselection”,Journal<br />
ofFinancialandQuantit<strong>at</strong>iveAnalysis6,517–557.<br />
HAUSMAN, J., A. W. LO ET C. MACKINLAY (1992) :“An or<strong>de</strong>red probit analysis of transaction stock<br />
prices”, JournalofFinancialEconomics31,319–379.<br />
HEATH, D. (2000) :“Back to the future”, Plenary lecture <strong>at</strong> the First World Congress of the Bachelier<br />
Soci<strong>et</strong>y,Paris.<br />
HILL, B. M. (1975):“Asimplegeneralapproachtoinferenceaboutth<strong>et</strong>ailofadistribution”,Annalsof<br />
St<strong>at</strong>istics3,1163–1174.<br />
HOSKING, J. R. M. (1981):“Fractional differencing”,Biom<strong>et</strong>rika65,165–176.<br />
HUISMAN, R., K. G. KOEDIJK ET R. A. J. POWNALL (2001) :“Ass<strong>et</strong> alloc<strong>at</strong>ion in aValue-<strong>at</strong>-Risk<br />
framework”,Document<strong>de</strong>Travail,ErasmusUniversity.<br />
HULT, H. ET F. LINDSKOG (2001) :“Multivari<strong>at</strong>e extremes, aggreg<strong>at</strong>ion and <strong>de</strong>pen<strong>de</strong>nce in elliptical<br />
distributions”, Document<strong>de</strong>Travail,Risklab.<br />
IDE, K. ET D. SORNETTE (2002) :“Oscill<strong>at</strong>ory finite-time singularities in finance, popul<strong>at</strong>ion and rupture”,PhysicaA307,63–106.<br />
JAFFRAY, J. ET F. PHILIPPE (1997) :“On the exitence of the subjective upper lower probabilities”,<br />
M<strong>at</strong>hem<strong>at</strong>icsofOper<strong>at</strong>ionsResearch22,165–185.<br />
JOE, H. (1997):Multivari<strong>at</strong>emo<strong>de</strong>lsand<strong>de</strong>pen<strong>de</strong>nceconcepts,Chapman&Hall,London.<br />
JOHANSEN, A., D. SORNETTE ET O. LEDOIT (1999) :“Predidicting financial crashes using discr<strong>et</strong>e<br />
scaleinvariance”,JournalofRisk1,5–32.<br />
JOHANSEN, A., O. LEDOIT ETD. SORNETTE(2000):“Crashesascriticalpoints”,Intern<strong>at</strong>ionalJournal<br />
oftheor<strong>et</strong>icalandAppliedFinance3,219–255.<br />
JOHANSEN, A. ET D. SORNETTE (1998):“Stockmark<strong>et</strong>crashesareoutliers”,EuropeanPhysicaljournalB1,141–144.<br />
JOHANSEN, A. ET D. SORNETTE (2002):“Largemark<strong>et</strong>pricedrawdownsareoutliers”,JournalofRisk<br />
4,69–110.<br />
JOHNSON, N. F., D. LAMPER, P. JEFFERIES, M. L. HART ET S. HOWISON (2001) :“Applic<strong>at</strong>ion<br />
of multi-agent games to the prediction of financial time series”, Document <strong>de</strong> Travail 0105303,<br />
Cond-m<strong>at</strong>.<br />
JONDEAU, E. ET M. ROCKINGER (2000) :“Conditional vol<strong>at</strong>ility,skewness and kurtosis :exitence and<br />
persistence”,Document<strong>de</strong>Travail,HEC.<br />
JONDEAU, E. ET M. ROCKINGER (2001):“Testingfordifferencesinth<strong>et</strong>ailsofstock-mark<strong>et</strong>r<strong>et</strong>urns”,<br />
Document<strong>de</strong>Travail,HEC.<br />
JURCZENKO, E. ET B. MAILLET (2002) :“Multi-moment kernel ass<strong>et</strong> pricing mo<strong>de</strong>l (KAPM) :Some<br />
basicresults”,in Jurczenko,E. <strong>et</strong>B.Maill<strong>et</strong>(eds.), Multi-Moment Capital Pricing Mo<strong>de</strong>ls, Springer.<br />
JURI, A. ET M. V. WÜTHRICH (2002) : “Copula convergence theorem for tail events”, Insurance :<br />
M<strong>at</strong>hem<strong>at</strong>ics and Economics 30, 405–420.
510 Bibliographie<br />
KAHNEMAN, D. ET A. TVERSKY (1979) : “Prospect theory : An analysis of <strong>de</strong>cision un<strong>de</strong>r risk”, Econom<strong>et</strong>rica<br />
47, 263–291.<br />
KAPLANSKI, G. ET Y. KROLL (2001a) : “Efficient VaR Portfolios”, Document <strong>de</strong> Travail, Hebrew<br />
University of Jerusalem.<br />
KAPLANSKI, G. ET Y. KROLL (2001b) : “Value-<strong>at</strong>-Risk equilirium Pricing Mo<strong>de</strong>l”, Document <strong>de</strong> Travail,<br />
Hebrew University of Jerusalem.<br />
KARLEN, D. (1998) : “Using projection and correl<strong>at</strong>ion to approxim<strong>at</strong>e probability distributions”, Computer<br />
in Physics 12, 380–384.<br />
KEARNS, P. ET A. R. PAGAN (1997) : “Estim<strong>at</strong>ing the <strong>de</strong>nsity tail in<strong>de</strong>x for financial time series”,<br />
Review of Economics and St<strong>at</strong>istics 79, 171–175.<br />
KEMPF, A. ET O. KORN (1999) : “Mark<strong>et</strong> <strong>de</strong>pth and or<strong>de</strong>r size”, Journal of Financial Mark<strong>et</strong>s 2, 29–48.<br />
KESTEN, H. (1973) : “Random difference equ<strong>at</strong>ion and renewal theory for products of random m<strong>at</strong>rices”,<br />
Acta M<strong>at</strong>hem<strong>at</strong>ica 131(207-253).<br />
KIMBALL, M. (1993) : “Standard risk aversion”, Econom<strong>et</strong>rica 61, 573–589.<br />
KIRMAN, A. (1983) : “Communic<strong>at</strong>ion in mark<strong>et</strong>s : a suggested approach”, Economics L<strong>et</strong>ters 12, 101–<br />
108.<br />
KIRMAN, A. (1991) : “Epi<strong>de</strong>mincs of opinion and specul<strong>at</strong>ive bubbles in financial mark<strong>et</strong>s”, in Taylor,<br />
M. P. (ed.), Money and Financial Mark<strong>et</strong>s, Blackwell, Cambridge, chapter 17.<br />
KLUGMAN, S. ET R. PARSA (1999) : “Fitting bivari<strong>at</strong>e loss distributions with copulas”, Insurance :<br />
M<strong>at</strong>hem<strong>at</strong>ics and Economics 24, 139–148.<br />
KRAUSS, A. ET R. LINTZENBURGER (1976) : “Skewness preference and the valu<strong>at</strong>ion of risky ass<strong>et</strong>s”,<br />
Journal of Finance 21, 1085–1094.<br />
KRZYSZTOFOWICZ, R. ET K. S. KELLY (1996) : “A m<strong>et</strong>a-Gaussian distribution with specified marginals”,<br />
Document <strong>de</strong> Travail, University of Virginia.<br />
KUSUOKA, S. (2001) : “On law invariant coherent risk measures”, Advances in M<strong>at</strong>hem<strong>at</strong>ical Economics,<br />
Springer, Tokyo, pp. 83–95.<br />
LAHERRÈRE, J. ET D. SORNETTE (1999) : “Streched exponential distributions in n<strong>at</strong>ure and economy :<br />
F<strong>at</strong> tails with characteristic scales”, European Physical Journal B 2, 525–539.<br />
LALOUX, L., P. CIZEAU, J. BOUCHAUD ET M. POTTERS (1999) : “Noise dressing of financial correl<strong>at</strong>ion<br />
m<strong>at</strong>rices”, Physical Review L<strong>et</strong>ters 83, 1467–1470.<br />
LALOUX, L., P. CIZEAU, J. BOUCHAUD ET M. POTTERS (2000) : “Random m<strong>at</strong>rix theory and financial<br />
correl<strong>at</strong>ions”, Intern<strong>at</strong>ional Journal of Theor<strong>et</strong>ical and Applied Finance 3, 391–397.<br />
LEBARON, B. (2001) : “Stochastic vol<strong>at</strong>ility as a simple gener<strong>at</strong>or of apparent financial power laws and<br />
long memory”, Quantit<strong>at</strong>ive Finance 1, 621–631.<br />
LEBARON, B., W. ARTHUR ET R. PALMER (1999) : “Time series properties of an artificial stock mark<strong>et</strong>”,<br />
Journal of Economic Dynamics and Control 23, 1487–1516.<br />
LEVY, H. (1998) : Stoch<strong>at</strong>ic Dominance : Investment <strong>de</strong>cision Making un<strong>de</strong>r Uncertainty, Kluwer Aca<strong>de</strong>mic<br />
Press.<br />
LÉVY, M., H. LÉVY ET S. SOLOMON (1995) : “Microscopic simul<strong>at</strong>ion of the stick mark<strong>et</strong> : the effect<br />
of miscroscopic diversity”, Journal <strong>de</strong> Physique I France 5, 1087–1107.<br />
LI, D., P. MIKUSINSKI ET M. TAYLOR (1998) : “Strong approxim<strong>at</strong>ion of copulas”, Journal of M<strong>at</strong>hem<strong>at</strong>ical<br />
Analysis and Applic<strong>at</strong>ions 225, 608–623.<br />
LILLO, F., J. D. FARMER ET R. N. MANTEGNA (2002) : “Single curve collapse of the price impact for<br />
the New York Stock Exchange”, Document <strong>de</strong> Travail 0207428, Cond-m<strong>at</strong>.
Bibliographie 511<br />
LILLO, F. ET R. N. MANTEGNA (2002) : “Omori law after a financial mark<strong>et</strong> crash”, Physical Review<br />
L<strong>et</strong>ters .<br />
LINDSKOG, F. (2000) : “Mo<strong>de</strong>lling Depen<strong>de</strong>nce with Copulas”, Document <strong>de</strong> Travail, RiskLab.<br />
LINDSKOG, F., A. MCNEIL ET U. SCHMOCK (2001) : “Kendall’s tau for elliptical distributions”, Document<br />
<strong>de</strong> Travail, Risklab.<br />
LINTNER, J. (1965) : “The valu<strong>at</strong>ion of risk ass<strong>et</strong>s and the selection of risky investments in stock portfolios<br />
and capital budg<strong>et</strong>s”, Review of Economics and St<strong>at</strong>istics 13, 13–37.<br />
LINTNER, J. (1969) : “The aggreg<strong>at</strong>ion of investors divers judgements and preferences in purely comp<strong>et</strong>itive<br />
security mark<strong>et</strong>s”, Journal of Financial and Quantit<strong>at</strong>ive Analysis 4, 347–400.<br />
LIU, Y., P. CIZEAU, M. MEYER, C. K. PENG ET H. E. STANLEY (1997) : “Correl<strong>at</strong>ions in economic<br />
time series”, Physica A 245, 437–440.<br />
LONGIN, F. M. (1996) : “The asymptotic distribution of extreme stock mark<strong>et</strong> r<strong>et</strong>urns”, Journal of<br />
Business 96, 383–408.<br />
LONGIN, F. M. (2000) : “From VaR to stress testing : the extreme value approach”, Journal of Banking<br />
& Finance 24, 1097–1130.<br />
LUX, T. (1995) : “Herd behaviour, bubbles and crashes”, Economic Journal 105, 881–896.<br />
LUX, T. (1996) : “The stable Par<strong>et</strong>ian hypothesis and the frequency of large r<strong>et</strong>urns : an examin<strong>at</strong>ion of<br />
major German stocks”, Applied Financial Economics 6, 463–475.<br />
LUX, T. (1997) : “The limiting extremal behavior of specul<strong>at</strong>ive r<strong>et</strong>urns : an annalysis of intradaily d<strong>at</strong>a<br />
from Frankfurt Stock exchange”, Document <strong>de</strong> Travail, University of Bonn.<br />
LUX, T. (1998) : “The socio-economic dynamics of specul<strong>at</strong>ive mark<strong>et</strong>s : interacting agents, chaos, and<br />
the f<strong>at</strong> tails of r<strong>et</strong>urn distributions”, Journal of Economic Behavior and Organiz<strong>at</strong>ion 33, 143–165.<br />
LUX, T. ET D. SORNETTE (2002) : “On r<strong>at</strong>ional bubbles and f<strong>at</strong> tails”, Journal of Money, Credit and<br />
Banking 34, 589–610.<br />
LUX, T. ET M. MARCHESI (1999) : “Scaling and criticality in a stochastic multi-agent mo<strong>de</strong>l of a<br />
financial mark<strong>et</strong>”, N<strong>at</strong>ure 397, 498–500.<br />
MALEVERGNE, Y. ET D. SORNETTE (2001a) : “Multi-dimensional r<strong>at</strong>ional bubbles and f<strong>at</strong> tails”, Quantit<strong>at</strong>ive<br />
Finance 1, 533–541.<br />
MALEVERGNE, Y. ET D. SORNETTE (2001b) : “Testing the Gaussian copula hypothesis for mo<strong>de</strong>lling<br />
financial ass<strong>et</strong>s <strong>de</strong>pen<strong>de</strong>nce”, Submitted to Quantit<strong>at</strong>ive Finance .<br />
MALEVERGNE, Y. ET D. SORNETTE (2002a) : “Collective origin of the coexistence of apparent RMT<br />
noise and factors in large sample correl<strong>at</strong>ion m<strong>at</strong>rices”, Submitted to Physical Review L<strong>et</strong>ters .<br />
MALEVERGNE, Y. ET D. SORNETTE (2002b) : “Minimizing Extremes”, Risk 15, 129–132.<br />
MALEVERGNE, Y. ET D. SORNETTE (2002c) : “Multi-moment m<strong>et</strong>hods for portfolio management :<br />
Generalized capital ass<strong>et</strong> pricing mo<strong>de</strong>l in homogeneous and h<strong>et</strong>erogeneous mark<strong>et</strong>s”, in Maill<strong>et</strong>,<br />
B. <strong>et</strong> Jurzenko (eds.), Proceeding of the workshop on Multi-Moment Capital Pricing Mo<strong>de</strong>ls and<br />
Rel<strong>at</strong>ed Topics, Springer.<br />
MALEVERGNE, Y. ET D. SORNETTE (2002d) : “VaR efficient portofolios for super and sub exponentially<br />
<strong>de</strong>caying ass<strong>et</strong>s’ r<strong>et</strong>urns distributions”, Document <strong>de</strong> Travail, University of Nice-Sophia Antipolis.<br />
MALEVERGNE, Y., V. PISARENKO ET D. SORNETTE (2002) : “Distribution of r<strong>et</strong>urns : Exponential<br />
versus power-like ?”, Document <strong>de</strong> Travail, University of Nice-Sophia Antipolis.<br />
MANDELBROT, B. (1963) : “The vari<strong>at</strong>ion of certain specul<strong>at</strong>ive prices”, Journal of Business 36, 392–<br />
417.
512 Bibliographie<br />
MANDELBROT, B. (1971) : “When can prices be arbitraged efficiently ? A limit to the validity of random<br />
walk and martigale mo<strong>de</strong>ls”, Review of Economics and St<strong>at</strong>istics 53, 225–261.<br />
MANTEGNA, R. N. ET H. E. STANLEY (1994) : “Stochastic process with ultraslow convergence to a<br />
Gaussian : The trunc<strong>at</strong>ed Lvy flight”, Physical Review L<strong>et</strong>ters 73, 2946–2949.<br />
MANTEGNA, R. N. ET H. E. STANLEY (1995) : “Scaling behavior of an economic in<strong>de</strong>x”, N<strong>at</strong>ure<br />
376, 46–55.<br />
MANTEGNA, R. N. ET H. E. STANLEY (1999) : An Introduction to Econophysics : Correl<strong>at</strong>ions and<br />
Complexity in Finance, Cambridge University Press.<br />
MARKOVITZ, H. (1959) : Portfolio Selection : Efficient Diversific<strong>at</strong>ion of Investments, John Wiley and<br />
Sons, New York.<br />
MAURER, S. M. (2001) : “Portfolios of quantum algorithms”, Physical Review L<strong>et</strong>ters 97(257901).<br />
MEERSCHAERT, M. ET H. SCHEFFLER (2001) : “Sample cross-correl<strong>at</strong>ions for moving averages with<br />
regularly varying tails”, Journal of Time Series Analysis 22, 481–492.<br />
MERTON, R. C. (1973) : “An intertemporal capital ass<strong>et</strong> pricing mo<strong>de</strong>l”, Econom<strong>et</strong>rica 41, 867–888.<br />
MERTON, R. C. (1992) : Continuous Time Finance, Blackwell, Cambridge, Massaschus<strong>et</strong>ts.<br />
MONTESSANO, A. ET F. GIOVANNONI (1996) : “Uncertainty aversion and aversion to increasing uncertainty”,<br />
Theory and Decision 41, 133–148.<br />
MOSSIN, J. (1966) : “Equilibrium in a capital mark<strong>et</strong>”, Econom<strong>et</strong>rica 34, 768–783.<br />
MUZY, J. F., D. SORNETTE, J. DELOUR ET A. ARNÉODO (2001) : “Multifractal r<strong>et</strong>urns and hierarchical<br />
portfolio theory”, Quantit<strong>at</strong>ive Finance 1, 131–148.<br />
MUZY, J. F., J. DELOUR ET E. BACRY (2000) : “Mo<strong>de</strong>lling fluctu<strong>at</strong>ions of financial times series : From<br />
casca<strong>de</strong> process to stochastic vol<strong>at</strong>ility mo<strong>de</strong>l”, European Physical Journal B 17, 537–548.<br />
MUZY, J. F. ET E. BACRY (2002) : “Multifractal st<strong>at</strong>ionary random measure and multifractal random<br />
walks with log-infinitely divisible scaling laws”, Document <strong>de</strong> Travail 0206202, Cond-m<strong>at</strong>.<br />
NAKAMURA, Y. (1990) : “Subjective expected utility with non-additive probabilities on finite st<strong>at</strong>e spaces”,<br />
Journal of Economic Theory 51, 346–366.<br />
NELSEN, R. (1998) : An Introduction to Copulas, Lectures Notes in st<strong>at</strong>istic 139, Springer Verlag, New<br />
York.<br />
NELSON, D. B. (1991) : “Conditional h<strong>et</strong>eroskedasticity in ass<strong>et</strong> pricing : a new approach”, Econom<strong>et</strong>rica<br />
59, 347–370.<br />
ORLÉAN, A. (1989) : “Comportements minétiques <strong>et</strong> diversité <strong>de</strong>s opinions sur les marchés financiers”,<br />
in Bourguin<strong>at</strong>, H. <strong>et</strong> P. Artus (eds.), Théorie Economique <strong>et</strong> Crises <strong>de</strong>s Marchés Financiers, Economica,<br />
pp. 45–65.<br />
ORLÉAN, A. (1992) : “Contagion <strong>de</strong>s opinions <strong>et</strong> fonctionnements <strong>de</strong>s marchés financiers”, Revue Economique<br />
43, 685–697.<br />
PAFKA, S. ET I. KONDOR (2001) : “Noisy covariance m<strong>at</strong>rices and portfolio optimiz<strong>at</strong>ion”, Document<br />
<strong>de</strong> Travail 0111503, Cond-m<strong>at</strong>.<br />
PAFKA, S. ET I. KONDOR (2002) : “Noisy covariance m<strong>at</strong>rices and portfolio optimiz<strong>at</strong>ion II”, Document<br />
<strong>de</strong> Travail 0205119, Cond-m<strong>at</strong>.<br />
PAGAN, A. R. (1996) : “The econom<strong>et</strong>rics of financial mark<strong>et</strong>s”, Journal of Empirical Finance 3, 15 –<br />
102.<br />
PALMER, R., W. B. ARTHUR, J. H. HOLLAND, B. LEBARON ET P. TAYLOR (1994) : “Artificial economic<br />
life - A simple mo<strong>de</strong>l of a stock mark<strong>et</strong>”, Physica D 75, 264–274.
Bibliographie 513<br />
PATTON, A. (2001) : “Estim<strong>at</strong>ion of Copula Mo<strong>de</strong>ls for Time Series of Possibly Different Lengths”,<br />
Document <strong>de</strong> Travail 01-17, University of California, San Diego.<br />
PFLUG, G. (2000) : “Some remarks on the value-<strong>at</strong>-risk and the conditional value-<strong>at</strong>-risk”, in Uryasev, S.<br />
(ed.), Probabilistic Constrained Optimiz<strong>at</strong>ion : M<strong>et</strong>hodology and Applic<strong>at</strong>ions, Kluwer Aca<strong>de</strong>mic<br />
Publisher.<br />
PICKANDS, J. (1975) : “St<strong>at</strong>istical inference using extreme or<strong>de</strong>r st<strong>at</strong>istics”, Annals of St<strong>at</strong>istics 3, 119–<br />
131.<br />
PLEROU, V., P. GOPIKRISHNAN, B. ROSENOW, L. N. AMARAL ET H. STANLEY (1999) : “Universal<br />
and nonuniversal properties of cross correl<strong>at</strong>ions in financial time series”, Physical Review L<strong>et</strong>ters<br />
83(1471).<br />
PLEROU, V., P. GOPIKRISHNAN, X. GABAIX ET H. E. STANLEY (2001) : “Quantifying stock price<br />
response to <strong>de</strong>man<strong>de</strong> fluctu<strong>at</strong>ions”, Document <strong>de</strong> Travail 0106657, cond-m<strong>at</strong>.<br />
POCHART, B. ET J. P. BOUCHAUD (2002) : “The skewed multifractal random walk with applic<strong>at</strong>ions to<br />
option smiles”, Quantit<strong>at</strong>ive Finance 2, 303–314.<br />
POON, S., M. ROCKINGER ET J. TAWN (2001) : “New extreme-value <strong>de</strong>pen<strong>de</strong>nce measures and finance<br />
applic<strong>at</strong>ions”, Document <strong>de</strong> Travail, HEC.<br />
QUIGGIN, J. (1982) : “A theory of anticiped utility”, Journal of Economic Behavior and Organiz<strong>at</strong>ion<br />
3, 323–343.<br />
RICHARDSON, M. ET T. SMITH (1993) : “A test for multivari<strong>at</strong>e normality in stocks”, Journal of Business<br />
66, 295–321.<br />
ROCKAFELLAR, R. ET S. URYASEV (2000) : “Optimiz<strong>at</strong>ion of conditional value-<strong>at</strong>-risk”, Journal of<br />
Risk 2, 21–41.<br />
ROCKAFELLAR, R. T. ET S. URYASEV (2002) : “Conditional value-<strong>at</strong>-risk for general loss distributions”,<br />
Journal of Banking & Finance 26, 1443–1471.<br />
ROCKINGER, M. ET E. JONDEAU (2001) : “Conditional <strong>de</strong>pen<strong>de</strong>ncy of financial series : An applic<strong>at</strong>ion<br />
of copulas”, Document <strong>de</strong> Travail, HEC.<br />
ROCKINGER, M. ET E. JONDEAU (2002) : “Entropy Densities with an applic<strong>at</strong>ion to autoregressive<br />
conditional skewness and kurtosis”, Journal of Econom<strong>et</strong>rics 106, 119–142.<br />
ROEHNER, B. M. (2001) : Hid<strong>de</strong>n Collective Factors in Specul<strong>at</strong>ive Trading, Springr Verlag, New York.<br />
ROEHNER, B. M. (2002) : P<strong>at</strong>terns of Specul<strong>at</strong>ion : A Study in Observ<strong>at</strong>ional Econophysics, Cambridge<br />
University Press.<br />
ROEHNER, B. M. ET D. SORNETTE (1998) : “The sharp peak-fl<strong>at</strong> trough p<strong>at</strong>tern and critical specul<strong>at</strong>ion”,<br />
European Physical Journal B 4, 387–399.<br />
ROLL, R. (1973) : “Evi<strong>de</strong>nce on the ‘Groth-Optimum’ mo<strong>de</strong>l”, Journal of Finance 28, 551–556.<br />
ROOTZÈN, H., M. R. LEADBETTER ET L. DE HAAN (1998) : “On the distribution of tail array sums for<br />
strongly mixing st<strong>at</strong>ionnary sequences”, Annals of Applied Probability 8, 868–885.<br />
ROSENOW, B., V. PLEROU, P. GOPIKRISHNAN ET H. STANLEY (2001) : “Portfolio optimiz<strong>at</strong>ion and<br />
the random magn<strong>et</strong> problem”, Document <strong>de</strong> Travail 0111537, Cond-m<strong>at</strong>.<br />
ROTSCHILD, M. ET J. E. STIGLITZ (1970) : “Increasing Risk I : A <strong>de</strong>finition”, Journal of Economic<br />
Theory 2, 225–243.<br />
ROY, A. D. (1952) : “Saf<strong>et</strong>y-First and the holding of ass<strong>et</strong>s”, Econom<strong>et</strong>rics 20, 431–449.<br />
RUBINSTEIN, M. (1973) : “The fundamental theorem of param<strong>et</strong>er-preference security valu<strong>at</strong>ion”, Journal<br />
of Financial and Quantit<strong>at</strong>ive Analysis 8, 61–69.
514 Bibliographie<br />
SAMUELSON, P. A. (1941) : “The stability of equilibirum”, Econom<strong>et</strong>rica 9, 97–120.<br />
SAMUELSON, P. A. (1958) : “The fundamental approxim<strong>at</strong>ion theorem ofportfolio analysis in terms of<br />
means variances and higher moments”, Review of Economic Studies 25, 65–86.<br />
SAMUELSON, P. A. (1965) : “Proof th<strong>at</strong> properly anticip<strong>at</strong>ed prices fluctu<strong>at</strong>es randomly”, Industrial<br />
management Review 6, 41–50.<br />
SAMUELSON, P. A. (1973) : “M<strong>at</strong>hem<strong>at</strong>ics of specul<strong>at</strong>ive price”, SIAM Review 15, 1–42.<br />
SAVAGE, L. (1954) : The fund<strong>at</strong>ions of st<strong>at</strong>istics, John Wiley, New York.<br />
SCAILLET, O. (2000a) : “Nonparam<strong>et</strong>ric estim<strong>at</strong>ion and sensitivity analysis of expected shortfall”, Document<br />
<strong>de</strong> Travail, HEC Genève.<br />
SCAILLET, O. (2000b) : “Nonparam<strong>et</strong>ric estim<strong>at</strong>ion of copulas for time series”, Document <strong>de</strong> Travail,<br />
Université C<strong>at</strong>holique <strong>de</strong> Louvain.<br />
SCHMEIDLER, D. (1986) : “Integral represent<strong>at</strong>ion without additivity”, Proceedings of the American<br />
M<strong>at</strong>hem<strong>at</strong>ical Soci<strong>et</strong>y 97, 255–261.<br />
SCHMEIDLER, D. (1989) : “Subjective probability and expected utility without additivity”, Econom<strong>et</strong>rica<br />
57, 571–587.<br />
SHAPIRO, A. ET S. BASAK (2000) : “Value-<strong>at</strong>-Risk based risk management : Optimal policies and ass<strong>et</strong><br />
prices”, Review of Financial Studies 14, 371–405.<br />
SHARPE, W. (1964) : “Capital ass<strong>et</strong> prices : A theory of mark<strong>et</strong> equilibrium un<strong>de</strong>r conditions of risk”,<br />
Journal of Finance 19, 425–442.<br />
SHEFFRIN, H. (2000) : Beyond Greed and Fear : Un<strong>de</strong>rstanding Behavioral Finance and the Psychology<br />
of Investing, Havard Business School Press.<br />
SHILLER, R. J. (2000) : Irr<strong>at</strong>ional Exuberance, Princ<strong>et</strong>own university Press.<br />
SHLEIFFER, A. (2000) : Inefficient Mark<strong>et</strong> : an Introduction to Behavioral Finance, Oxford University<br />
Press.<br />
SKLAR, A. (1959) : “Fonction <strong>de</strong> répartition à n dimensions <strong>et</strong> leurs marges”, Publ. Inst. St<strong>at</strong>ist. Univ.<br />
Paris 8, 229–231.<br />
SMITH, A. (1776) : An Inquiry into the N<strong>at</strong>ure and Causes of the Wealth of N<strong>at</strong>ions, R. H. Campbell and<br />
A. S Skinner (eds.), Oxford, UK : Clarenton Press, 1976.<br />
SMITH, V. L. (1994) : “Economics in the labor<strong>at</strong>ory”, Journal of Economic Persectives 8, 113–131.<br />
SMITH, V. L. (1998) : “The two faces of Adam Smith”, Southern Economic Journal 95, 1–19.<br />
SORNETTE, D. (2000) : Critical Phenomena in N<strong>at</strong>ural Sciences, Springer.<br />
SORNETTE, D., J. V. ANDERSEN ET P. SIMONETTI (2000) : “Portfolio theory for “F<strong>at</strong> Tails””, Intern<strong>at</strong>ional<br />
journal of Theor<strong>et</strong>ical and Applied Finance 3, 523–535.<br />
SORNETTE, D., P. SIMONETTI ET J. ANDERSEN (2000) : “φ q -field theory for portfolio optimiz<strong>at</strong>ion :<br />
”f<strong>at</strong>-tails” and non-linear correl<strong>at</strong>ions”, Physics Reports 335, 19–92.<br />
SORNETTE, D. ET J. V. ANDERSEN (2002) : “A nonlinear super-exponential r<strong>at</strong>ional mo<strong>de</strong>l of specul<strong>at</strong>ive<br />
financial bubbles”, Intern<strong>at</strong>ional Journal of Mo<strong>de</strong>rn Physics 13, 171–188.<br />
SORNETTE, D. ET Y. MALEVERGNE (2001) : “From r<strong>at</strong>ional bubbles to crashes”, Physica A 299, 10–59.<br />
SORNETTE, D., Y. MALEVERGNE ET J. F. MUZY (2002) : “Vol<strong>at</strong>ility fingerprints of large shocks :<br />
Endogeneous versus exogeneous”, forthcoming Risk .<br />
STAUFFER, D. ET D. SORNETTE (1999) : “Self-organized percol<strong>at</strong>ion mo<strong>de</strong>l for stock mark<strong>et</strong> fluctu<strong>at</strong>ions”,<br />
Physica A 271, 496–506.
Bibliographie 515<br />
SUSMEL, R. (1996) : “Switching vol<strong>at</strong>ility in intern<strong>at</strong>ional equity mark<strong>et</strong>”. Dept. of Finance, University<br />
of Houston.<br />
SZERGÖ, G. (1999) : “A critique to Basel regul<strong>at</strong>ion, or how to enhance (im)moral hazards”, Proceedings<br />
of the Intern<strong>at</strong>ional Conference on Risk Management and Regul<strong>at</strong>ion in Banking, Bank of Israel,<br />
Kluwer.<br />
TASCHE, D. (2000) : “Risk contribrutions and performance measurement”, Document <strong>de</strong> Travail, TU<br />
München.<br />
TASCHE, D. (2002) : “Expected shortfall and beyond”, Journal of Banking & Finance 26, 1519–1533.<br />
TASCHE, D. ET L. TIBILETTI (2001) : “Approxim<strong>at</strong>ions for the value-<strong>at</strong>-risk approach to risk-r<strong>et</strong>urn<br />
analysis”, Document <strong>de</strong> Travail 269733, SSRN.<br />
THALLER, R. H. (ed.) (1993) : Advances in Behavioral Finance, Russel Sage Found<strong>at</strong>ion, New York.<br />
VALDEZ, E. A. (2001) : “Bivari<strong>at</strong>e analysis of survivorship and persistency”, Insurance : M<strong>at</strong>hem<strong>at</strong>ics<br />
and Economics 29, 357–373.<br />
VENEZIANO, D., G. E. MOGLEN ET R. L. BRAS (1995) : “Multifractal analysis : pitfalls of standard<br />
procedures and altern<strong>at</strong>ives”, Physical Review E 52, 1387–1398.<br />
VON NEUMANN, J. ET O. MORGENSTERN (1947) : Theory of games and economic behavior, Princ<strong>et</strong>own<br />
University Press.<br />
WANG, T. (1999) : “A class of dynamic risk measures”, Document <strong>de</strong> Travail, University of British<br />
Columbia.<br />
WESTERHOFF, F. H. (2001) : “Specul<strong>at</strong>ive mark<strong>et</strong>s and the effectiveness of price limits”, Document <strong>de</strong><br />
Travail, Universität Osnabrück.<br />
WESTERHOFF, F. H. (2002) : “H<strong>et</strong>erogeneous tra<strong>de</strong>rs and the Tobin tax”, Document <strong>de</strong> Travail, Universität<br />
Osnabrück.<br />
WISHART, J. (1928) : “The generalized product moment distribution in samples from a multinormal<br />
popul<strong>at</strong>ion”, Biom<strong>et</strong>rika 20, 32.<br />
ZAKOIAN, J. M. (1994) : “Threshold h<strong>et</strong>eroskedastic mo<strong>de</strong>ls”, Journal of Economic Dynamics and<br />
Control 18, 931–986.<br />
ZHANG, Y. C. (1999) : “Toward a theory of marginally efficient mark<strong>et</strong>s”, Physica A 269, 30–44.
Risques extrêmes en finance : St<strong>at</strong>istique, <strong>théorie</strong> <strong>et</strong> <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong><br />
Résumé : C<strong>et</strong>te thèse propose une étu<strong>de</strong> <strong>de</strong>s risques extrêmes sur les marchés financiers, considérés<br />
comme un exemple typique <strong>de</strong> système complexe auto-organisé. Nous commençons par décrire <strong>et</strong> modéliser<br />
les propriétés <strong>st<strong>at</strong>istique</strong>s individuelles <strong>de</strong>s actifs financiers afin d’en déduire une estim<strong>at</strong>ion précise<br />
<strong>de</strong>s grands risques <strong>et</strong> d’en comprendre les mécanismes sous-jacents en rel<strong>at</strong>ion avec la micro-structure<br />
<strong>de</strong>s marchés <strong>et</strong> le comportement <strong>de</strong>s agents économiques. A l’ai<strong>de</strong> <strong>de</strong>s copules <strong>et</strong> modèles à facteurs,<br />
nous analysons ensuite les propriétés <strong>de</strong> dépendance extrêmes <strong>de</strong>s actifs financiers afin <strong>de</strong> mieux cerner<br />
les possibilités <strong>et</strong> les limites <strong>de</strong> la diversific<strong>at</strong>ion <strong>de</strong>s grands risques. Enfin, nous étudions les mesures<br />
<strong>de</strong> risques les plus à mêmes <strong>de</strong> quantifier <strong>de</strong>s risques extrêmes <strong>et</strong> appliquons l’ensemble <strong>de</strong> ces résult<strong>at</strong>s<br />
à l’obtention <strong>de</strong> <strong>portefeuille</strong>s les moins sensibles à ce type <strong>de</strong> risques. Parallèlement, nous exposons<br />
certaines conséquences <strong>de</strong> ce type d’alloc<strong>at</strong>ion sur les équilibres <strong>de</strong> marché.<br />
Mots-clés : système complexe auto-organisé, risques extrêmes, marchés financiers, <strong>gestion</strong> <strong>de</strong> <strong>portefeuille</strong>,<br />
éconophysique, bulles spécul<strong>at</strong>ives, distributions à queues épaisses, copules.<br />
Extreme risks in finance : St<strong>at</strong>istics, theory and portfolio management<br />
Summary : This thesis proposes a study of extreme risks observed on financial mark<strong>et</strong>s, consi<strong>de</strong>red as<br />
a typical example of self-organized complex systems. We start by <strong>de</strong>scribing and mo<strong>de</strong>ling individual<br />
st<strong>at</strong>istical properties of financial ass<strong>et</strong>s in or<strong>de</strong>r to obtain the most accur<strong>at</strong>e estim<strong>at</strong>ion of large risks<br />
and to provi<strong>de</strong> a b<strong>et</strong>ter un<strong>de</strong>rstanding of the un<strong>de</strong>rlying mechanisms in rel<strong>at</strong>ion with the mark<strong>et</strong>s microstructure<br />
and the economic agents’ behaviors. Using copulas as well as factor mo<strong>de</strong>ls, we then analyze<br />
the <strong>de</strong>pen<strong>de</strong>nce properties b<strong>et</strong>ween ass<strong>et</strong>s, and more specifically their extreme counterparts, in or<strong>de</strong>r to<br />
un<strong>de</strong>rstand the opportunities for diversific<strong>at</strong>ion of large risks, but also their limits. Finally, we study the<br />
risk measures th<strong>at</strong> are the most appropri<strong>at</strong>e for the assessment of extreme risks and apply all these results<br />
to <strong>de</strong>fine the portofolios th<strong>at</strong> are the least sensitive to extreme risks. In parallel, we <strong>de</strong>velop consequences<br />
of such ass<strong>et</strong>s alloc<strong>at</strong>ions with respect to mark<strong>et</strong> equilibria.<br />
Key-words : self-organized complex systems, extreme risks, financial mark<strong>et</strong>s, portfolio management,<br />
econophysics, specul<strong>at</strong>ive bubbles, heavy tail distributions, copulas.