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
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
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