23.01.2014 Views

Modelling Dependence with Copulas - IFOR

Modelling Dependence with Copulas - IFOR

Modelling Dependence with Copulas - IFOR

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.

7 Extreme Value <strong>Copulas</strong><br />

7.1 Univariate Extreme Value Theory<br />

Let X 1 ,X 2 ,... be iid random variables <strong>with</strong> continuous distribution function F .<br />

Let S n = X 1 + ... + X n ,M n =max(X 1 ,... ,X n )andL n =min(X 1 ,... ,X n ).<br />

Furthermore let x F ,givenby<br />

x F =sup{x ∈ R : F (x) < 1} ≤∞<br />

x<br />

denote the right endpoint of F . It can be shown that M n → x F as n →∞. If X i<br />

has finite mean and variance µ and σ 2 respectively, we know from the central limit<br />

theorem that<br />

P[(S n − nµ)/ √ nσ 2 ≤ x] → Φ(x) asn →∞,<br />

where Φ is the distribution function of the standard normal distribution. What<br />

about normalized maxima and minima? Consider possible limiting distributions<br />

for (M n − a n )/b n and (L n − c n )/d n as n → ∞ for suitably chosen sequences<br />

{a n }, {b n }, {c n }, {d n }. Since<br />

min(X 1 ,... ,X n )=− max(−X 1 ,... ,−X n )<br />

it is sufficient to study the case of maxima. The sequences {a n } and {b n } we are<br />

interested in are such that<br />

P[(M n − a n )/b n ≤ x] =P[M n ≤ a n + b n x]=F n (a n + b n x) → H(x),<br />

as n →∞for some non-degenerate distribution function H, i.e. {H(x)|x ∈ R} ̸=<br />

{0, 1}. The three possible types for H are called univariate (max) extreme value<br />

distributions and are given by:<br />

• Fréchet<br />

{<br />

0, x ≤ 0<br />

φ α (x) =<br />

exp(−x −α ), x > 0 α>0.<br />

• Gumbel<br />

Λ(x) =exp(−e −x ),<br />

x ∈ R<br />

• Weibull<br />

Ψ α (x) =<br />

{<br />

exp(−(−x) α ), x ≤ 0<br />

1, x > 0 α>0.<br />

We say that two random variables U and V are of the same type if and only if<br />

This means that<br />

U = d aV + b for b>0,a∈ R.<br />

F U (x) =F V ((x − b)/a)<br />

and hence random variables of the same type have the same distribution function<br />

up to changes of scale and location. Thus in the non-degenerate case the only<br />

possible limits of (M n − a n )/b n are the location-scale families based on the above<br />

three distribution functions. In this case we say that F is in the maximum domain<br />

of attraction of H,<br />

F ∈ MDA(H).<br />

53

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

Saved successfully!

Ooh no, something went wrong!