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.

DEFINITION 2 (TAIL DEPENDENCE 1)<br />

L<strong>et</strong> X and Y be random variables with continuous marginals FX and FY . The (upper) tail <strong>de</strong>pen<strong>de</strong>nce<br />

coefficient of X and Y is, if it exists,<br />

lim<br />

u→1<br />

Pr{X > F −1<br />

X<br />

197<br />

−1<br />

(u)|Y > F (u)} = λ ∈ [0, 1] . (3)<br />

In words, given th<strong>at</strong> Y is very large (which occurs with probability 1 − u), the probability th<strong>at</strong> X is very<br />

large <strong>at</strong> the same probability level u <strong>de</strong>fines asymptotically the tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ.<br />

It turns out th<strong>at</strong> this tail <strong>de</strong>pen<strong>de</strong>nce is a pure copula property which is in<strong>de</strong>pen<strong>de</strong>nt of the marginals. L<strong>et</strong><br />

C be the copula of the variables X and Y , then<br />

THEOREM 3<br />

if the bivari<strong>at</strong>e copula C is such th<strong>at</strong><br />

lim<br />

u→1<br />

¯C(u, u)<br />

1 − u<br />

Y<br />

= λ (4)<br />

exists (where ¯ C(u, u) = 1 − 2u − C(u, u)), then C has an upper tail <strong>de</strong>pen<strong>de</strong>nce coefficient λ.<br />

If λ > 0, the copula presents tail <strong>de</strong>pen<strong>de</strong>nce and large events tend to occur simultanously, with the<br />

probabilty λ. On the contrary, when λ = 0, the copula has no tail <strong>de</strong>pen<strong>de</strong>nce in this sense and large<br />

events appear to occur essentially in<strong>de</strong>pen<strong>de</strong>ntly. There is however a subtl<strong>et</strong>y in this <strong>de</strong>finition of tail<br />

<strong>de</strong>pen<strong>de</strong>nce. To make it clear, first consi<strong>de</strong>r the case where for large X and Y the distribution function<br />

F (x, y) factorizes such th<strong>at</strong><br />

lim<br />

x,y→∞<br />

F (x, y)<br />

= 1 . (5)<br />

FX(x)FY (y)<br />

This means th<strong>at</strong>, for X and Y sufficiently large, these two variables can be consi<strong>de</strong>red as in<strong>de</strong>pen<strong>de</strong>nt. It<br />

is then easy to show th<strong>at</strong><br />

lim<br />

u→1<br />

Pr{X > F −1<br />

X<br />

−1<br />

(u)|Y > F (u)} = lim<br />

Y<br />

u→1<br />

so th<strong>at</strong> in<strong>de</strong>pen<strong>de</strong>nt variables really have no tail <strong>de</strong>pen<strong>de</strong>nce, as one can expect.<br />

−1<br />

1 − FX(FX (u)) (6)<br />

= lim<br />

u→1 1 − u = 0, (7)<br />

Unfortun<strong>at</strong>ly, the converse does not holds : a value λ = 0 does not autom<strong>at</strong>ically imply true in<strong>de</strong>pen<strong>de</strong>nce,<br />

namely th<strong>at</strong> F (x, y) s<strong>at</strong>isfies equ<strong>at</strong>ion (5). In<strong>de</strong>ed, the tail in<strong>de</strong>pen<strong>de</strong>nce criterion λ = 0 may<br />

still be associ<strong>at</strong>ed with an absence of factoriz<strong>at</strong>ion of the multivari<strong>at</strong>e distribution for large X and Y .<br />

In a weaker sense, there may still be a <strong>de</strong>pen<strong>de</strong>nce in the tail even when λ = 0. Such behavior is for<br />

instance exhibited by the Gaussian copula, which has zero tail <strong>de</strong>pen<strong>de</strong>nce according to the <strong>de</strong>finition 2<br />

but nevertheless does not have a factorizable multivari<strong>at</strong>e distribution, since the non-diagonal term of the<br />

quadr<strong>at</strong>ic form in the exponential function does not become negligible in general as X and Y go to infinity.<br />

To summarize, the tail in<strong>de</strong>pen<strong>de</strong>nce, according to <strong>de</strong>finition 2, is not equivalent to the in<strong>de</strong>pen<strong>de</strong>nce<br />

in the tail as <strong>de</strong>fined in equ<strong>at</strong>ion (5).<br />

After this brief review of the main concepts un<strong>de</strong>rlying copulas, we now present two special families<br />

of copulas : the Gaussian copula and the Stu<strong>de</strong>nt’s copula.<br />

5

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

Saved successfully!

Ooh no, something went wrong!