AI- Driven Credit Scoring in Indian Financial Institutions: Opportunities, Risks, and Regulatory Implications

AI- Driven Credit Scoring in Indian Financial Institutions: Opportunities, Risks, and Regulatory Implications

Abstract

This paper examines whether machine learning models can meaningfully improve upon logistic regression for credit default prediction in the Indian banking context—and if so, whether those improvements come at the cost of interpretability that regulators require. We work with a panel of approximately 1.24 million loan accounts drawn from three scheduled commercial banks, covering the period Q1 2015 through Q4 2023. The sample is intentionally constructed to span the COVID-19 disruption, which offers an unusually demanding test of model robustness under distribution shift. Our central finding is that Gradient Boosting Machines—specifically Boost—outperform logistic regression by a substantial margin on both discrimination (AUROC: 0.926 vs. 0.739) and calibration (Brier Score: 0.053 vs. 0.081). More interestingly, when we augment the Boost specification with alternative digital data—GST filing regularity, UPI transaction frequency, and utility payment histories sourced through the Account Aggregator framework—the AUROC rises further to 0.934. For borrowers with thin credit files, the alternative data variables account for over 40 percent of the model’s explanatory weight, which has direct implications for financial inclusion policy. A recurring objection to deploying ML models in credit decisioning is the so-called black-box problem. We address this directly by applying SHAP (Shapley Additive explanations) decompositions at both global and instance levels. The SHAP-LIME concordance across a validation subsample (rank correlation = 0.87) suggests the explanations are method-robust, not artefacts of a particular XAI technique. Fairness analysis reveals no statistically detectable disparate impact along gender or rural-urban lines, after conditioning on financial fundamentals—a finding relevant to the RBI’s emerging AI governance framework.

Authors

Praveen Soneja, Abhijeet Raj, Abhilasha Kumari, Kusum Lata, Mayank Raj

Institution

Noida Institute of Engineering & Technology (MCA Institute), Greater Noida

PDF Link

View Full Paper