From Cure Rate Models to Explainable Survival Analysis: SurvSHAP Based Insights for Credit Risk

Abstract

This study proposes a unified credit risk framework that combines mixture cure rate models with time-varying SurvSHAP(t) explanations. The model separates the risk of default into incidence submodel (whether default occurs) and latency submodel (when default occurs among susceptibles). A simulation study shows accurate parameter recovery and reliable coverage in low- and high-cure-rate settings, with SurvSHAP(t) correctly identifying incidence, latency, and noise variables. Empirical results for Prosper P2P lending and mortgage data show that the lognormal mixture cure rate model outperforms standard survival models, reflecting non-monotonic default hazard and substantial cured subpopulations. Credit score drives the incidence submodel, while behavioural and macroeconomic variables govern the latency submodel. SurvSHAP(t) reveals consistent temporal structure across simulation and real data: incidence submodel effects dominate long-term survival, while latency submodel effects are strongest early in time. The framework provides an interpretable and statistically sound tool for credit risk modelling, supporting risk stratification and regulatory transparency.