An Ensemble Based Framework for Loan Default Prediction with Bias Mitigation and Explainability

Abstract

With some 190 million adults in India classified as credit - invisible, the country's formal lending institutions systematically ignore this group from formal retail loans. Existing single-model deployments rely exclusively on historical credit bureau data, failing to accommodate alternative behavioral metrics. More critically, these models operate on the assumption that historical training labels accurately reflect default risk. In this paper, we identify and mathematically quantify a severe "policy-label inversion" in a dataset of 51,336 Indian retail banking records. We demonstrate that historically rejected applicants exhibit a median credit score 30 points higher than approved applicants (p- value = < 2.2e-16) proving that risk-naive algorithms memorize commercial targeting policies rather than genuine credit risk. To address this embedded bias and the thin-file exclusion problem simultaneously, we propose a dual-model architecture. Bureau - history applicants are assessed by an ensemble with three layers. It takes the outputs of a Random Forest and ExtraTrees and inputs it into a GradientBoosting classifier that functions as a meta - learner which itself inputs its output into a WoE Scorecard. Using such a stacked ensemble yields a Macro F1 score of 0.9858, reducing incorrect high - risk customers classified as low - risk by 95.7\% relative to its constituent base classifiers. First - time borrowers who don't have Bureau history are instead passed to an alternative data model that uses Reserve Bank of India Account Aggregator behavior features to achieve an ROC - AUC of 0.7581. Analyses of the fairness across gender show gender fairness. So, it turns out we can reach the levels of traditional discrimination capabilities with alternative data without the demographic unfairness that comes along with that.