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.
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