Comparing ensemble tree-based machine lerarning mod-els to the regression for credit scorning – An emprircial analysis on home credit data

Abstract

In the view of recent developments in Machine Learning techniques for Credit Scoring, a range of promising approaches have emerged. These advancements have challenged tradi-tional methods, such as the Logistic Regression model, and highlighted the effectiveness of ensemble tree-based models in achieving strong predictive outputs. However, it is crucial to consider factors beyond performance metrics, such as interpretability and business case criteria, when comparing these models. This thesis takes a comprehensive approach to evaluate selected tree-based models against the Logistic Regression across multiple evalua-tion metrics, including performance, interpretability and business case examination. Re-sults indicate that ensemble tree-based models outperform the Logistic Regression in all aspects, particularly in achieving higher recall values. Hence, it is suggested to explore newer Machine Learning methods as an alternative to traditional Credit Scoring tech-niques, depending on specific business requirements and data availability, with the aim of improving predictive performance and enabling more effective risk management in credit evaluation.