Explainable Artificial Intelligence in Credit Risk: Evaluating Machine Learning and Traditional Models in Buy Now Pay Later under EU Regulation

Abstract

This paper examines the trade-off between predictive performance and interpretability in credit risk modeling within a buy-now-pay-later (BNPL) setting. Using a proprietary dataset from Walley, the study compares logistic regression with random forest and XG Boost. To assess interpretability, SHAP and LIME are applied as post-hoc explainability methods. The study is motivated by the growing use of machine learning in credit scoring, where improved predictive performance must be balanced against transparency, validation, and supervisory requirements. The results show that XGBoost provides the strongest predic tive performance across all evaluations and the XAI methods used provided insights into the black-box models. LIME delivers case-specific, local explanations for individual pre dictions while SHAP provided an overview of the global characteristics and how different features interacted in the framework. Overall, the findings suggest that machine learning models can improve predictive per formance in credit risk modeling, but that their practical usefulness depends on whether their outputs can be explained in a meaningful and transparent way.