Credit risk modeling: combining classification and regression algorithms to predict expected loss

Credit risk assessment is of paramount importance in the financial industry. Machine learning techniques have been used successfully over the last two decades to predict the probability of loan default (PD). This way, credit decisions can be automated and risk can be reduced significantly. In the more recent parts, intensified regulatory requirements led to the need to include another parameter loss given default (LGD), the share of the loan which cannot be recovered in case of loan default in risk models. We aim to build a unified credit risk model by estimating both parameters jointly to estimate expected loss. A large, highdimensional, real world dataset is used to benchmark several combinations of classification, regression and feature selection algorithms. The results indicate that non-linear techniques work especially well to model expected loss.