Predicting loss given default of unsecured consumer loans with time-varying survival scores

Loss Given Default (LGD) is an essential element in effective banking supervision, as set out in the Basel Accords. In this paper, we focus on improving LGD predictions with the help of time-varying covariates. Based on online unsecured consumer loan data, we first build application scores with a Cox proportional hazard model, and behavioral scores with a multiplicative hazard model. We add these time-varying survival scores to fit the specifications of four separate LGD models - Tobit regression, decision trees, Logit-transformed linear regression and Beta regression. It is shown that better LGD predictions can be achieved when both application and behavioral scores are incorporated. Our framework further facilitates the prediction of expected loss, which can produce loss estimates at any time during the repayment period. Our experiment shows that the loss estimates are accurate, though some inherent errors cannot be avoided.