Segment-Aware Credit Risk Modelling: A Multi-Algorithm Benchmark

Abstract

Today, most lenders use the same predictive model across their whole loan portfolio, whether the applicant is a salaried employee with a regular income or a gig worker whose earnings are unpredictable. This homogeneous approach does reasonably well at the aggregate level but systematically underpredicts default risk for subgroups that differ from the typical borrower profile. In this paper we pose a simple question: does training a separate machine learning model for each distinct group of borrowers improve the ability to detect defaults within those groups? We classify the 733,451 leakage-free LendingClub loans (2013–2015) into four groups, based on two binary dimensions: financial capacity (annual income and debt-to-income ratio) and behavioral stability (employment length above or below three years). The resulting groups are Prime, Transitional, Constrained, and High-Risk. For each segment, we trained five algorithms (Logistic Regression, Random Forest, XGBoost, CatBoost, LightGBM) independently, and benchmarked against a non-segmented baseline. Recall improves in 13 out of 20 algorithm-segment combinations with the largest gains of +0.0851 (Prime, Random Forest), +0.0564 (Constrained, Logistic Regression) and +0.0555 (High-Risk, LightGBM). All four best-per-segment gains are statistically significant in both DeLong and McNemar’s tests (p < 0.001). Besides the headline numbers, we document a reproducible pattern, which we call algorithm-segment affinity, in which the structure of a borrower segment reliably predicts which algorithm will be best within that segment. We also see that gradient boosting models require about 100,000-150,000 training records per segment to consistently outperform simpler alternatives. Implications for Indian FinTech and the NBFCs functioning under the RBI Model Risk Management Guideines also discussed.