Machine learning for consumer loan default prediction in retail banking

Consumer loan defaults represent consistent source of cost pressures within retail banks, eroding capital reserves and necessitating provisioning that prevents further loaning. While machine learning is now widely adopted in credit risk modelling, empirical studies report mixed evidence regarding the extent of its performance improvements over Logistic Regression. These differences may partly reflect variation in datasets, pre-processing methods, and evaluation metrics. In this paper, a comparison of Logistic Regression, Random Forest, and Gradient Boosting using a publicly available consumer lending dataset containing 32,575 rows and eleven variables is conducted. Evaluation was carried out using an 80/20 holdout validation and additionally five-fold cross-validation. Gradient Boosting outperformed both alternatives across all evaluation metrics, with ensemble methods demonstrating higher recall for borrowers than the Logistic Regression baseline. Loan grade, interest rate, and the ratio of the loan to income accounted for prediction in both ensemble classifiers. The recall gap between Gradient Boosting and Logistic Regression carries direct implications for the design of automated credit approval systems.