CMStatistics 2018: Start Registration
View Submission - CFE
Title: Identifying hidden patterns in credit risk survival data using generalised additive models Authors:  Jonathan Crook - University of Edinburgh (United Kingdom) [presenting]
Viani Djeundje - University of Edinburgh (United Kingdom)
Abstract: Survival models for credit card default have several advantages over cross sectional credit scoring models. They provide more information to analysts, they can be used to model provisions for IFRS9 and the inclusion of macroeconomic variables means they can be used for stress testing. However, in such models the hazard of default is typically expressed as a simple linear combination or covariates. We investigate the predictive accuracy of survival models if the hazards are expressed as GAMs. Specifically, we parameterise hazard models in terms of penalised splines. We estimate the parameters using frequentist and Bayesian methods applied to a large portfolio of credit card accounts. We compare the predictive accuracy of models without splines with those with splines, applied in turn to application, behavioural, to macroeconomic variables, and to models with all of these types of variables. We find that GAM specifications have higher predictive accuracy than linear models. The results suggest that some applications for accounts may be less attractive to a lender when GAMs are used rather than linear models. Similarly, expected profit may be more accurately estimated with a GAM model than a simple linear model.