One More Model Risk When Using Gaussian Copula for Risk Management

Abstract

Gaussian Copula as a model for default correlation has been recently criticized for a number of fallacies in its application to pricing and risk management of financial liabilities.

Here we point out an element of model risk that appears to be overlooked. When the Gaussian Copula is applied to the computation of the probability of losses concentrated in time, it can give paradoxical and misleading results, where an increase of correlation reduces the model probability of loss concentration. This behaviour is dangerous since, if not taken into account, can lead to completely wrong model stress testing.

Here we show how this behaviour can affect three practical problems: the estimation of future liquidity risk due to clustered losses, the assessment of CDS counterparty risk and the computation of dynamic value-at-risk.