An application of survival analysis in credit risk management

Abstract

Banks, credit unions, and finance companies use credit risk models to estimate the degree of credit risk associated with lending money to borrowers. Credit scoring models form the preponderance of these models, and their function is to calculate the probability of default (PD) for borrowers’ credit facilities. The global financial crisis that lasted from mid-2007 to early 2009 gave birth to a new capital provisioning approach that requires the use of forward-looking PD estimates. This research focuses on applying survival analysis to provide PD estimates throughout the lifetime of a loan facility. The study covers a variety of survival analysis models that encompass semi-parametric, parametric, and mixture cure models. A large sample of personal loans from Prosper, a peer-to-peer lending marketplace was analysed. Loans are monitored monthly from entry to exit. The study reveals that survival models are suitable for providing point-in-time (PIT) PD estimates. The findings from the top-performing models, namely, Log-normal, Logistic-Log-normal mixture cure, and stratified Cox regression, were satisfactory based on their probabilistic statistical measures, ability to discriminate prognosis in the future, and calibration performance.