Machine Learning Techniquesin Joint Default Assessment

Abstract

This paper empirically compares logistic regression with machine learning techniques in order to estimate the default risk measures and their bounds in large portfolios of identically distributed obligors. The methods compute different predictions of the probability of individual default and default correlation, so they are compared in various settings using increasing amounts of information: first the marginal probability, then the marginal probability and correlation, and lastly a specific model, the beta-binomial distribution.We make this evaluation using Value at Risk as well as Expected Shortfall in two settings: one synthetic and one real. In the synthetic setting, we construct portfolios of up to 10,000 obligors and test the performance of each method on 200 datasets. In the real setting, we use a publicly available credit card dataset of 30,000 obligors.

Keywords: Model Risk, Credit risk, Bernoulli mixture model, ML methods, credit cards.