Machine Learning for Corporate Default Risk: Multi-Period Prediction, Frailty Correlation, Loan Portfolios, and Tail Probabilities

Abstract

We use machine learning methods for modeling multi-period corporate default probabilities and obtain higher prediction accuracy compared to linear models with the differences being larger for longer prediction horizons. Overall, tree-boosting has the highest prediction accuracy. In addition, we introduce a novel hybrid econometric-machine learning model combining tree-boosting with a latent frailty model. This ``LaGaBoost frailty model" results in more accurate predictions of upper tails of portfolio losses compared to both a linear frailty model and machine learning methods ignoring frailty correlation. We also investigate the reasons and find various explanations for the observed differences in prediction accuracy.