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

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

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

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

Saved successfully!

Ooh no, something went wrong!