Temporal Stability of Probability of Default Models in Fintech Lending: An Empirical Comparison with Traditional Credit Scoring Benchmarks

Abstract

This paper presents an empirical validation study of a production-grade Probability of Default (PD) model deployed in a digital consumer lending platform, and benchmarks its temporal stability metrics against values reported in the traditional credit scoring literature. Fintech lenders operate under conditions that differ structurally from traditional retail banking: faster borrower acquisition cycles, thinner credit files, limited credit bureau penetration, and shorter loan tenors. Despite these constraints, the model under review achieves a Gini coefficient of 0.487 and a Kolmogorov-Smirnov statistic of 0.421, placing it in the upper range of traditional bank scorecard benchmarks. Monthly Gini volatility is 0.031, with no statistically significant downward trend over the multi-year observation window. However, a systematic upper-tail calibration bias of +0.048 percentage points is identified as a fintech-specific failure mode, attributable to the absence of credit bureau cycle features. Additionally, this study demonstrates that the Population Stability Index (PSI), the industry standard for distribution shift detection, is insufficient as a standalone early-warning signal: it remained below alert thresholds in months when class-conditional drift for defaulters was already in the investigation zone. To address this limitation, a Composite Model Health Score (CMHS) is proposed, which aggregates normalised deviations across six validation metrics into a single monitoring index. Empirical and simulation results show that CMHS detects model degradation two to eight months earlier than standalone PSI. The complete validation pipeline is implemented in open-source Python and presented as a reproducible framework for fintech model risk management.