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.

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

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

Saved successfully!

Ooh no, something went wrong!