Machine Learning-Driven Risk Scoring Models for Retail Credit Assessment: Comparative Evaluation of Neural Networks, XGBoost, and Logistic Regression in the Indian Lending Sector

Conventional credit risk scorecards, such as logistic regressions-based models, do not have the predictive power to adapt to the changing financial habits of no file and thin file market segment. In counteraction to these limitations, ML algorithms have evolved as reasonable options providing adaptive, data-driven ones for retail credit scoring. In this paper, we evaluated three state-of-the-art modeling techniques—Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), and Neural Networks (NN)—on a real-world anonymized dataset that we collected from Indian NBFCs and fintech lenders. Our data set includes borrower’s demographic, financial and transactional attributes. Standardized pre-processing, class balancing and stratified 5-fold cross-validation were used when training and validating models. Performance was evaluated through several measures including the Accuracy, Precision, Recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC). The empirical results, from which we observe that, no matter which evaluation criteria are adopted, XGBoost significantly outperforms LR and NN, with AUC reaching up to 0.901 and the best F1-score. Neural Networks produced good results, though not as understandable, to be used in practice. Logistic Regression, while less predictive, is still the most interpretable and auditable model. These results indicate that ML models lead to material improvements in credit risk scoring in India, but that the interpretability and regulatory considerations are essential to their implementation.