Dependence estimation and visualization in multivariate extremes with applications to financial data (Q2488446): Difference between revisions

Nonparametric estimation of extreme copula for multivariate i.i.d. data \((x_{j,1},\dots,x_{j,m})\), \(j=1,\dots,m\), is considered. It is based on estimation of the measure \[ \Lambda(A)=\lim_{t\to\infty} t{\mathbf P} \{t(\bar G_1(x_{1,1}),\dots,\bar G_m(x_{1,m}))\in A\} \] \[ \text{by}\quad \hat\Lambda_{\varepsilon,n}(A)=\varepsilon\sum_{i=1}^n I\{\varepsilon(R_{i,1},\dots,R_{i,m})\in A\}, \] where \(\bar G_k(x)={\mathbf P}\{x_{1,k}>x\}\), and \(R_{i,k}\) is the rank of the \(i\)-th observation by the \(k\)-th variable. The spectral measure in Pickands representation of the copula is estimated with the aid of \(\hat\Lambda_{\varepsilon,n}\). The authors introduce a tail dependence function, which in the bivariate case can be defined as \[ \rho(\vartheta)=[1+\cot\vartheta-\Lambda(([1,\infty]\times[\cot\vartheta,\infty])^c)](1 \wedge\cot\vartheta)^{-1} \] and discuss it's properties. A plot of \(\rho(\vartheta)\) is proposed as a visualization tool for the extremal dependence of \(x_{1,1}\) and \(x_{1,2}\). Application to data of returns of the Annually Compounded Zero Coupon Swap Rates is considered.