Explainable machine learning models for the probability of credit default

Abstract

This dissertation addresses the challenge of balancing predictive accuracy and interpretability in credit risk assessment models within regulated financial environments. Using annual panel data from a company’s credit application in 2024, the study conducts a comparative analysis of traditional logistic regression and two advanced machine learning models: Catboost and Explainable Boosting Machine, EBM. Through empirical evaluation using standardized metrics for both accuracy and interpretability, the research demonstrates that modern machine learning models can effectively overcome the conventional trade-off between accuracy and interpretability. This study’s results revealed that both Catboost and Explainable Boosting Machine outperform logistic regression in predictive accuracy while meeting the interpretability standards required by regulatory frameworks. Catboost achieved superior predictive performance and utilized SHAP values for post-hoc interpretation. Meanwhile, EBM demonstrated competitive accuracy with built-in interpretability features. These findings indicate that financial institutions can modernize their credit risk assessment frameworks without compromising regulatory compliance or stakeholder trust. The present work contributes to the broader discourse on responsible machine learning in financial services, providing empirical evidence that advanced models can simultaneously achieve high accuracy and interpretability in credit risk assessment applications.