Machine learning techniques in joint default assessment

This paper studies the consequences of capturing non-linear dependence among the covariates that drive the default of different obligors and the overall riskiness of their credit portfolio. Joint default modeling is, without loss of generality, the classical Bernoulli mixture model. Using an application to a credit card dataset we show that, even when Machine Learning techniques perform only slightly better than Logistic Regression in classifying individual defaults as a function of the covariates, they do outperform it at the portfolio level. This happens because they capture linear and non-linear dependence among the covariates, whereas Logistic Regression only captures linear dependence. The ability of Machine Learning methods to capture non-linear dependence among the covariates produces higher default correlation compared with Logistic Regression. As a consequence, on our data, Logistic Regression underestimates the riskiness of the credit portfolio