5.

5.5.1

5.5.2

5.5.3

5.5 Fitting Copulas to Data

Presenter: Yen ju Chao April 10, 2012

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Fitting Copulas to Data

We have data vectors X1 , . . . , Xn with identical distribution function F , describing nancial loss or nancial risk factor returns. We write Xt = (Xt ,1 , . . . , Xt ,d ) for an individual data vector, and X = (X1 , . . . , Xd ) for a generic random vector with df F . F has continuous margins F1 , . . . , Fd and by Sklars Theorem, a unique representation F (x ) = C (F1 (x1 ), . . . , Fd (xd )).

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

It may be easier to use empirical estimates of either Spearmans or Kendalls rank correlation to infer an estimate for the copula parameter.(For example:Table 5.5) Recall Denition 5.28: For rvs X1 and X2 with marginal dfs F1 and F2 Spermans rho is given by s (X1 , X2 ) = (F1 (X1 ), F2 (X2 )). We could estimate S (Xi , Xj ) by calculating the usual correlation coecient for the pseudo-observations: {(Fi ,n (Xt ,i ), Fj ,n (Xt ,j )) : t = 1, . . . , n}, where Fi ,n denotes the standard empirical df for the i th margin.

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

We use rank (Xt ,i ) to denote the rank of Xt ,i in X1,i , . . . , Xn,i , we can calculate the correlation coecient for the rank data {(rank(Xt ,i ), rankXt ,j )}, and this gives us the Spermans rank correlation coecient: 12 2 n(n 1) 1 1 (rank (Xt ,i ) (n + 1))(rank (Xt ,j ) (n + 1)) 2 2 t =1 (5.49)
n

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

Recall Denition 5.27: For rvs X1 and X2 Kendalls tau is given by 1 )(X2 X 2 ))), where (X 1 , X 2 ) is (X1 , X2 ) = E (sign((X1 X independent copy of (X1 , X 2). The standard estimator of Kendalls tau (Xi , Xj ) is Kendalls rank correlation coecient : n 2
1

sign((Xt ,i Xs ,i )(Xt ,j Xs ,j ))
1t <s n

(5.50)

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

Example 5.52(bivariate Archimedean copulas with a single parameter). We assumed model is of the form F (x1 , x2 ) = C (F1 (x1 ), F2 (x2 )), where is a single parameter to be estimated. We have simple relationships of the form (X1 , X2 ) = f ().(as shown in Table 5.5) We can calculate a sample value r for Kendalls tau rst, ) for . Then solving the equation r = f ( For example, Gumbels copula is calibrated by taking = (1 r )1 , provided that r 0.

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

Example 5.53(calibrating Gauss copulas using Spearmans rho).

Ga We assumed a meta-Gaussian model for X with Cp and we wish to estimate the correlation matrix P . It follows from Theorem 5.36 that

1 S (Xi , Xj ) = (6/ )arcsin ij ij , 2 where the nal approximation is very accurate. This suggests we estimate P by the matrix of pairwise Spearmans rank correlation coecient R S .

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

Example 5.54(calibrating t copulas using Kendalls tau).

t We assumed a meta-t model for X with copula C, P and we wish to estimate the correlation matrix P . It follows

(Xi , Xj ) = (2/ )arcsinij so that a possible estimator of P is the matrix R with component 1 given by rij = sin( 2 rij ).

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

Algorithm 5.55(eigenvalue method). Let R be a so-called pseudo -correlation matrix, i.e. a symmetric matrix if pairwise correlation estimate correlation estimates with unit diagonal entries and o diagonal entries in [-1,1] that is not positive semidenite. Calcukate the spectral decomposition R = GLG where L is the matrix of eigenvalues and G is an orthogonal matrix whose columns are eigenvectors of R . Replace all negative eigenvalues in L by small value > 0 to . obtain L

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Method-of-Moments using Rank Correlation

Algorithm 5.55(eigenvalue method). , which will be symmetric and positive Calculate Q = G LG denite but not a correlation matrix, since its diagonal elements will not necessarily equal one. Return the correlation matrix R = (Q ), where denotes the correlation matrix operator.,where dened ( ) = ( ( ))1 ( ( ))1 , ( ( )) := diag( 11 , . . . , dd )

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Forming a Pseudo-Sample from the Copula

We now turn to the estimation of parameteric copulas by maximum likelihood (ML). In this section we describe brifely some general approaches to the rst step of estimating margins and constructing a pseudo sample of observations from the copula. In the following section we describe how the copula parameters are estimated by ML from pseudo-sample.

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Forming a Pseudo-Sample from the Copula

1 , . . . , F d denote estimates of the marginal dfs. The Let F pseudo-sample from the copula consists of the vectors 1 , . . . , U n , where U t = (Ut ,1 , . . . , Ut ,d ) = (F 1 (Xt ,1 ), . . . , F d (Xt ,d )) U

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Forming a Pseudo-Sample from the Copula

i include Possible methods for obtaining the marginal estimate F the following.
Parameteric estimation. We choose an appropriate parametric model for the data in question and t it. Non-parametric estimation with of empirical df. We could estimate Fj using Fi ,n (x ) 1 = n+1
n

I { X t ,i x }
t =1

Extreme value theorey for the tails. Empirical distribution functions are known to be poor estimators of the underlying distribution in the tails, and the tail are model using a generalized Pareto distribution, the body of distribution may be modelled empirically.
Y-J, Chao 5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Forming a Pseudo-Sample from the Copula

Example 5.57. Five years of daily log-return data (1996-2000). Intel, Microsoft and General Electric stocks. The marginal distributions are estimated empirically (method(2)). The pseudo-sample from copula is shown in Figure 5.14.

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Multivariate Archimedean Copulas

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Maximum Likelihood Estimation

Let C denote a parameter copula, where is the vector of parameters to be estimated. The MLE is obtained by maximizing
n

1 , . . . , U n) = ln L(; U
t =1

t) ln c (U

One could envisage Using the two-stage method to decide on the most appropriate cpula family and then estimating all parameters (marginal and copula) in a nal fully parametric round of estimation.

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Maximum Likelihood Estimation

Example 5.58 (tting the Gaussian copula). In the case of a Gaussina copula implies that the log-likelihood is 1 , . . . , U n) ln L(P ; U
n n d

=
t =1

t ,1 ), . . . , 1 (U t ,d )) ln fP ( (U
1 t =1 j =1

t ,j )) ln (1 (U

where f will be used to denote the joint density of a random vector with Nd (0, ) distribution.

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Maximum Likelihood Estimation

Example 5.58 (tting the Gaussian copula). = argmaxP n ln f (Yt ), where Yt ,j = 1 (U i ,j ) for P t =1 j=1,. . . , d and P denotes the set of all possible linear correlation matricses. The set P can be constructed as P = {P = (Q ) : Q = AA , A lower triangular with ones on the diagonal} An approximate solution to the maximization may be obtained as follows:
= (1/n) = () P
n t =1

Yt Yt

Y-J, Chao

5.5 Fitting Copulas to Data

5.5

5.5.1

5.5.2

5.5.3

Maximum Likelihood Estimation

Example 5.58 (tting the Gaussian copula). For example 5.57 by full ML, the estimated correlation matrix has 0.58(INTC-MSFT), 0.34(INTC-GE) and 0.4(MSFT-GE); the log- likelihood at the maximum is 376.65, using the alternative method gives a log likelihood value of 376.62

Y-J, Chao

5.5 Fitting Copulas to Data