Global Sensitivity Analysis for Latent Factor Credit Risk Models

This paper proposes the use of a global sensitivity analysis to evaluate the risk associated with a credit portfolio model. The main features of this approach are its ability to assess sensitivities in the presence of non-linearities and to rank the input factors with respect to their relevance for the output variable. The commonly used local sensitivity analysis which is nested in the global model cannot provide this information. We analyze the static and time-varying uncertainties of three key input factors in a latent factor credit risk model, i.e. the multivariate distribution (copula), the default correlation and the default probability. Results show that the importance of the factors strongly depends on the average default probability of the portfolio and the analyzed quantiles of the default distribution. The proposed technique also provides information about trade-offs between the factors, e.g. between the default correlation and the copula