A1603
Title: Credit risk modeling in the age of machine learning
Authors: Martin Hibbeln - University of Duisburg-Essen (Germany)
Raphael Kopp - University of Duisburg-Essen (Germany)
Noah Urban - University of Duisburg-Essen (Germany) [presenting]
Abstract: Based on the world's largest loss database of non-retail defaults, we perform a comparative analysis of machine learning methods in credit risk modeling across the globe. We identify substantial benefits in using machine learning methods, especially tree-based methods, frequently more than doubling the performance metrics, over both simple and sophisticated benchmarks that particularly consider the specific distributions of credit risk parameters. Superior predictive abilities across many dimensions are primarily attributable to nonlinear relationships between the features and the credit risk parameters traced by methods of the explainable machine learning toolbox. Finally, we highlight important differences regarding the nature of macroeconomic features and implications of the temporal order of defaults. The results are robust to a battery of different specifications.
