A generalised latent Poisson factor modelling approach for default correlations in credit portfolios

Abstract

Default risk is one of the major concerns for lending institutions and banking regulators. This paper focuses on the analysis of default data, using a new approach based on generalised latent Poisson factor models. In this case, the correlation structure of the default events is driven by a small number of common latent factors. Conditional to these factors, the defaults become independent and each default sequence is fitted to a generalised linear model with Poisson response and log-link function. This model provides a flexible framework for the computation of the value-at-risk and the expected shortfall of a credit portfolio. The practical implementation of the proposed local Fisher scoring estimation algorithm is illustrated by a Monte Carlo simulation study. Then, a real scenario, with default data taken from a large database provided by Standard & Poor's, is used to analyse the empirical behaviours of the different risk measures. The achieved results show promising performance.