A Profit-Based Measure of Lending Discrimination
Abstract
Algorithmic lending has transformed the consumer credit landscape, with complex machine learning models now commonly used to make or assist underwriting decisions. To comply with fair lending laws, these algorithms typically exclude legally protected characteristics, such as race and gender. Yet algorithmic underwriting can still inadvertently favor certain groups, prompting new questions about how to audit lending algorithms for potentially discriminatory behavior. Building on prior theoretical work, we introduce a profit-based measure of lending discrimination in loan pricing. Applying our approach to approximately 80,000 personal loans from a major U.S. fintech platform, we find that loans made to men and Black borrowers yielded lower profits than loans to other groups, indicating that men and Black applicants benefited from relatively favorable lending decisions. We trace these disparities to miscalibration in the platform's underwriting model, which underestimates credit risk for Black borrowers and overestimates risk for women. We show that one could correct this miscalibration -- and the corresponding lending disparities -- by explicitly including race and gender in underwriting models, illustrating a tension between competing notions of fairness.
Summary
This paper addresses the critical issue of algorithmic discrimination in lending, focusing on "disparate impact," where facially neutral algorithms inadvertently disadvantage protected groups. The authors introduce a novel profit-based measure to detect such discrimination in loan pricing, building on the idea that in a fair lending environment, loans should yield similar profits across all groups after accounting for risk. They analyze approximately 80,000 personal loans from a major U.S. fintech platform, inferring race and gender using the BISG method. Their analysis reveals that loans to Black borrowers and men generate lower profits compared to other groups, suggesting these groups benefit from relatively favorable loan terms. This disparity is traced back to a miscalibration in the platform's underwriting model, which underestimates the credit risk for Black borrowers and overestimates it for women. The authors demonstrate that correcting this miscalibration by explicitly including race and gender in the model could eliminate the profit gaps, but this would violate fair lending laws, highlighting a tension between predictive accuracy and legal fairness. The paper contributes a practical and readily implementable audit tool for lenders, regulators, and researchers to identify potentially discriminatory lending practices without needing access to the lender's proprietary risk models.
Key Insights
- •Profit-Based Discrimination Measure: The core contribution is a straightforward measure of lending discrimination based on comparing the annualized internal rate of return (IRR) across demographic groups. Systematically lower profits for loans to a group indicate potentially favorable (discriminatory) loan terms.
- •Miscalibration as a Driver of Disparities: The study identifies model miscalibration, specifically underestimation of risk for Black borrowers and overestimation for women, as a key factor driving the observed profit disparities.
- •Tension Between Fairness and Accuracy: The paper highlights a stark trade-off: improving predictive accuracy by including protected characteristics in the model can eliminate disparities but violates fair lending laws prohibiting disparate treatment.
- •Strategic Shopping Ruled Out: The authors tested and rejected the hypothesis that strategic shopping behavior by certain groups (men and Black borrowers) explains the observed profit gaps, using a counterfactual analysis.
- •Robustness to Alternative Race/Gender Inference: The results are robust to using both weighted averages based on BISG probabilities and assigning individuals to a single race/gender category based on the argmax of BISG probabilities.
- •Quantified Disparities: Loans made to men had an annualized IRR of 8.3% compared to 9.1% for women. Loans made to Black borrowers had an annualized IRR of 7.7% compared to 8.3-8.8% for White, Hispanic, and Asian borrowers.
- •Impact of Aware Model: Switching to a race/gender-aware risk model is predicted to decrease loan approval rates for Black borrowers by 3 percentage points and increase approval rates for White, Hispanic, and Asian borrowers by approximately 1 percentage point each.
Practical Implications
- •Regulatory Audit Tool: The profit-based IRR comparison provides a readily implementable tool for regulators to audit lending algorithms for potential disparate impact, even without access to proprietary model details.
- •Internal Audit for Lenders: Lenders can use this method to proactively identify and address potential sources of algorithmic bias in their underwriting models, improving both fairness and potentially profitability by better calibrating risk models.
- •Informed Policy Discussions: The paper informs policy discussions around balancing predictive accuracy, fairness, and legal compliance in algorithmic lending. It highlights the need for regulatory frameworks that address the complexities of algorithmic pricing systems.
- •Future Research: The authors suggest replicating the analysis in different credit markets and with intersectional analysis (e.g., Black women) to assess the generality of the findings. Further research should also explore alternative fairness metrics and mitigation strategies that do not involve direct use of protected characteristics.
- •Risk Model Calibration: The study emphasizes the importance of careful risk model calibration, especially for different demographic groups, to avoid inadvertently creating or perpetuating discriminatory lending practices.