ABSTRACT
Since the implementation of International Financial Reporting Standards 9, several techniques for estimating the risk parameters for calculating expected credit losses have been implemented across financial institutions. The purpose of this study is to present the advantages of using Random Survival Forests and other machine learning techniques to estimate the probability of default (PD) in credit risk given the specificity of the method in estimating time-to-event. Machine learning techniques were selected as modeling to predict the default incorporating time-to-event and the macroeconomic conditions into the model, as only a stable and sustainable economic situation increases the security of credit risk and the financial system. All analyses were performed on a real data sample for the mortgage portfolio of a European bank, with the sample covering the years 2008–2020 and more than 3,000 clients. At the end of our research, we validated the accuracy of the applied methods used over time, including a lifetime perspective. Explainable methods for machine learning models were applied to give insight into and explain the reasons for the high PD, which is an approach required by the regulatory authorities. Results are promising, as the accuracy of such models is higher compared to simple regression techniques.
