Comparative Analysis of Machine Learning Algorithms for Consumer Credit Risk Assessment

Abstract

In the rapidly evolving landscape of financial technology, machine learning algorithms are increasingly supplanting traditional methodologies for evaluating consumer credit risk. This study leverages a comprehensive dataset comprising 10,000 credit accounts to conduct a comparative analysis of four prevalent machine learning algorithms: Logistic Regression, Decision Tree, Random Forest, and Gradient Boosting Machine (GBM). The results distinctly favor GBM, which achieves an AUC of 0.87, closely followed by Random Forest with an AUC of 0.85. In stark contrast, Logistic Regression and Decision Tree recorded lower AUCs of 0.78 and 0.72, respectively. GBM and Random Forest significantly outperform in classification accuracy, attaining 92% and 90%, respectively, far exceeding the 86% by Logistic Regression and 80% by Decision Tree. Notably, GBM exhibited 95% specificity and 90% sensitivity, efficiently identifying high-risk accounts while minimizing false positives among low-risk categories. Furthermore, the study delves into the handling of imbalanced datasets, interpretability, and computational demands of each algorithm, offering quantifiable insights that inform future directions for optimizing credit risk models, particularly in enhancing transparency and scalability.