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.

198 8. Tests <strong>de</strong> copule gaussienne<br />

2.3 The Gaussian copula<br />

The Gaussian copula is the copula <strong>de</strong>rived from the multivari<strong>at</strong>e Gaussian distribution. L<strong>et</strong> Φ <strong>de</strong>note<br />

the standard Normal (cumul<strong>at</strong>ive) distribution and Φρ,n the n-dimensional Gaussian distribution with<br />

correl<strong>at</strong>ion m<strong>at</strong>rix ρ. Then, the Gaussian n-copula with correl<strong>at</strong>ion m<strong>at</strong>rix ρ is<br />

whose <strong>de</strong>nsity<br />

reads<br />

−1<br />

Cρ(u1, · · · , un) = Φρ,n Φ (u1), · · · , Φ −1 (un) , (8)<br />

cρ(u1, · · · , un) =<br />

cρ(u1, · · · , un) = ∂Cρ(u1, · · · , un)<br />

∂u1 · · · ∂un<br />

<br />

1<br />

√ exp −<br />

d<strong>et</strong> ρ 1<br />

2 yt (u) (ρ−1 <br />

− Id)y (u)<br />

with yk(u) = Φ −1 (uk). Note th<strong>at</strong> theorem 1 and equ<strong>at</strong>ion (2) ensure th<strong>at</strong> Cρ(u1, · · · , un) in equ<strong>at</strong>ion (8)<br />

is a copula.<br />

As we said before, the Gaussian copula does not have a tail <strong>de</strong>pen<strong>de</strong>nce :<br />

lim<br />

u→1<br />

¯Cρ(u, u)<br />

1 − u<br />

(9)<br />

(10)<br />

= 0, ∀ρ ∈ (−1, 1). (11)<br />

This results is <strong>de</strong>rived for example in (Embrechts <strong>et</strong> al. 2001). But this does not mean th<strong>at</strong> the Gaussian<br />

copula goes to the in<strong>de</strong>pen<strong>de</strong>nt (or product) copula Π(u1, u2) = u1 · u2 when (u1, u2) goes to one.<br />

In<strong>de</strong>ed, consi<strong>de</strong>r a distribution F (x, y) with Gaussian copula :<br />

Its <strong>de</strong>nsity is<br />

where fX and fY are the <strong>de</strong>nsities of X and Y . Thus,<br />

F (x, y) = Cρ(FX(x), FY (y)). (12)<br />

f(x, y) = cρ(FX(x), FY (y)) · fX(x) · fY (y), (13)<br />

f(x, y)<br />

lim<br />

= lim<br />

(x,y)→∞ fX(x) · fY (y) (x,y)→∞ cρ(FX(x), FY (y)), (14)<br />

which should equal 1 if the variables X and Y were in<strong>de</strong>pen<strong>de</strong>nt in the tail. Reasoning in the quantile<br />

space, we s<strong>et</strong> x = F −1<br />

−1<br />

X (u) and y = FY (u), which yield<br />

f(x, y)<br />

lim<br />

= lim<br />

(x,y)→∞ fX(x) · fY (y) u→1 cρ(u, u). (15)<br />

Using equ<strong>at</strong>ion (10), it is now obvious to show th<strong>at</strong> cρ(u, u) goes to one when u goes to one, if and<br />

only if ρ = 0 which is equivalent to Cρ=0(u1, u2) = Π(u1, u2) for every (u1, u2). When ρ > 0, cρ(u, u)<br />

goes to infinity, while for ρ neg<strong>at</strong>ive, cρ(u, u) goes to zero as u → 1. Thus, the <strong>de</strong>pen<strong>de</strong>nce structure<br />

<strong>de</strong>scribed by the Gaussian copula is very different from the <strong>de</strong>pen<strong>de</strong>nce structure of the in<strong>de</strong>pen<strong>de</strong>nt<br />

copula, except for ρ = 0.<br />

The Gaussian copula is compl<strong>et</strong>ly d<strong>et</strong>ermined by the knowledge of the correl<strong>at</strong>ion m<strong>at</strong>rix ρ. The<br />

param<strong>et</strong>ers involved in the <strong>de</strong>scription of the Gaussian copula are very simple to estim<strong>at</strong>e, as we shall<br />

see in the following.<br />

6

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

Saved successfully!

Ooh no, something went wrong!