Supporting the Age-Period-Cohort model of default rate prediction with interpretable machine learning

Abstract

Regular short-term forecasting of defaults is a basic activity of a retail portfolio risk manager. From a business perspective, not only the quality of the forecast is significant, but also the understanding of the trends and their driving factors. The vintage analysis and a more advanced Age-Period-Cohort approach are popular tools used for the purpose. The aim of this article is to demonstrate that interpretable machine learning can support the Age-Period- Cohort approach, facilitating forecasting beyond the time range of training data, eliminating the model identification problem and attributing cohort quality to the specific characteristics of loans approved in a given month. The study is based on real consumer finance portfolios from the Polish market.