25.06.2013 Views

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

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

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

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!