Macroprudential Implications of AI-Driven Credit Scoring: Stress-Testing Machine Learning Default Models Across Business Cycles to Safeguard Financial Stability in U.S. Consumer Credit Markets

• Abbas Ahsun

  Texas University

  Author

The proliferation of artificial intelligence and machine learning in consumer credit scoring presents both opportunities for enhanced predictive accuracy and novel risks to financial stability, particularly when these models operate across varying business cycle conditions without adequate stress-testing frameworks. Despite widespread adoption of ML-based credit models by fintech lenders and traditional financial institutions, limited empirical evidence exists regarding their performance during economic downturns and their potential to amplify systemic risk through correlated lending behaviors. This study addresses this gap by developing and stress-testing a hybrid machine learning default prediction framework using a retrospective analysis of 450,000 consumer credit records from the Lending Club platform (2018–2025), combined with prospective simulation under macroeconomic stress scenarios aligned with Federal Reserve supervisory frameworks. The proposed framework integrates XGBoost and LSTM neural network architectures with a reduced-form default rate forecasting component, achieving a predictive accuracy of 89.4% (AUC-ROC) during stable economic conditions, which declined to 82.1% during recessionary stress periods when trained exclusively on expansion-phase data. However, incorporating macroeconomic indicators and cycle-informed training techniques improved recession-period performance to 87.3%, representing a 5.2 percentage point improvement over baseline models. The findings demonstrate that ML default models exhibit significant cyclical sensitivity, with an estimated procyclical amplification effect of 12-18% during severe downturns, underscoring the necessity of integrating cycle-robust validation protocols and countercyclical capital buffers into AI-based credit underwriting frameworks. These results carry substantial implications for macroprudential regulators, financial institutions, and credit risk model validators seeking to harness AI's predictive power while safeguarding financial stability.

• pdf

Copyright (c) 2026 Abbas Ahsun (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.