Interpretable Machine Learning for Credit Risk Decision Support: Balancing Interpretability and Predictive Performance

Abstract

With the growing prevalence of machine learning in making institutional decisions, there is a need for careful balancing of predictive accuracy and interpretability of such models. In this research work, the performance and interpretability of three machine learning models in helping institutions make decisions were investigated, particularly the issue of credit risk assessment. In this process, the German Credit Data dataset, which is open source (N = 1,000), was used and subjected to 5-

Fold Cross-Validation. The results show that while Random Forest attained the highest accuracy (76.3%), Logistic Regression had the highest discriminating power (AUC: 0.79), just marginally better than the complex ensemble approach (AUC: 0.78). It also has the added advantage of better interpretability due to clear coefficient and decision rule visualization. This clearly shows that inherently interpretable models serve as a very good alternative to the complex ensemble approach in decision-making within organizations, without falling prey to the “black-box” problem, and in some cases even performing better in discriminatory power.

However, the study notes relatively low minority-class recall (36.3–46.3%), highlighting a practical limitation. Conclusions are based on a benchmark dataset and the specified model configurations.