Distortion copula models

From FinMath

Jump to: navigation, search

This page relates to the paper "Using Distortions of Copulas to price CDOs". Glenis Crane and John van der Hoek. (pdf) University of Adelaide. Matlab code is available at Code:distortion_copula_toolbox.

[edit] Definition: Distortion of a Copula

Let C be a bivariate copula and \psi : [0,1] \rightarrow [0,1] be a bijective map with inverse ψ − 1 then C^{\psi}(u,v) = \psi^{-1}\left(C(\psi(u),\psi(v))\right),\ \ u,v \in I= [0,1] is a strict distortion of C. Technical conditions on ψ described in the paper (pdf) ensure Cψ(u,v) is also a copula, and we may generalize to bijective maps having only pseudo-inverses.

[edit] Density

The density of a distorted copula is given by

\frac{\partial^2}{\partial u \partial v} C^{\psi}(u,v) =     \frac{\psi'(u) \psi'(v) }{[\psi'(C^{\psi})]^3 }             \left\{                 [\psi'(C^\psi)]^2 \frac{\partial^2}{\partial u \partial v} C(\psi(u),\psi(v)) - \psi''(C^{\psi}) \frac{\partial}{\partial u}                   C(\psi(u),\psi(v)) \frac{\partial}{\partial v}  C(\psi(u),\psi(v))                   \right\}

[edit] Application to Gaussian One Factor Model

Denoting the one and two dimensional cumulative normal distributions by Φ and Φ2 respectively we may relate two variables with cumulative marginals F1(z) and F2(y) by:

C^{\psi}(F_1(z),F_2(y)) = \psi^{-1} \left[        \Phi_2(\Phi^{-1}(\psi[F_1(z)]),\Phi^{-1}(\psi[F_2(y)];\rho)     \right]

where as customary ρ is the common pairwise gaussian correlation. As an application, we may model the conditional probability that a firm defaults as:

p^{(V)} = \frac{ \Phi \left( \frac{\Phi^{-1} (\psi[F_1(z)]) - \rho \Phi^{-1}(\psi[F_2(y)])}  { \sqrt{1-\rho^2} }  \right) \psi'(F_2(y))                   }                  {     \psi'\left(\psi^{-1}\left[\Phi_2(\Phi^{-1}(\psi[F_1(z)), \Phi^{-1}(\psi[F_2(y)]);\rho)\right]\right)                                    }

This reduces to the familiar expression

p^{(V)}(y) = \Phi \left( \frac{\Phi^{-1} (\bar{p}) - \rho y}  { \sqrt{1-\rho^2} }  \right)

in the absence of any distortion (ψ = identity) and normally distributed common factor y (i.e. F2 = Φ). Here \bar{p} = F_1(z) is the unconditional probability of default.

Personal tools