1625. Bias and Fairness in Credit Scoring: Challenges, Trade-offs, and Business Impact
Invited abstract in session MD-4: EJOR: policy, facts and highlights in stream OR Journals, stream OR Journals.
Monday, 14:30-16:00Room: Rupert Beckett LT
Authors (first author is the speaker)
| 1. | Stefan Lessmann
|
| School of Business and Economics, Humboldt-University of Berlin |
Abstract
Algorithmic decision-making is reshaping credit risk assessment. Machine learning (ML)-based scorecards promise unprecedented predictive power but raise critical concerns about fairness, reliability, and bias. This talk synthesizes insights from recent credit scoring research. We examine how biases—ranging from algorithmic discrimination to selection bias—can undermine the effectiveness of ML-driven credit scoring systems and explore strategies to mitigate these risks.
First, we examine fair lending, assessing the adequacy of statistical fairness criteria and evaluating algorithmic options for incorporating fairness goals in scorecard development. We demonstrate that while fairness and profitability often seem at odds, multiple fairness criteria can be satisfied simultaneously without significant decreases in predictive accuracy and cost. However, we also show how established practices in algorithmic fairness research may severely underestimate the true degree of discrimination.
Second, we address the sampling bias problem arising from using labeled data from previously accepted clients. By introducing bias-aware training and evaluation techniques, we show how ML models can mitigate the adverse effects of selection bias, improving both predictive accuracy and profitability. Our empirical findings highlight that addressing bias at the scorecard evaluation stage holds greater potential than training-time adjustments, which prevail in the literature.
Through these discussions, we emphasize that the effectiveness of ML-based risk prediction systems hinges on a comprehensive understanding of the decision problem and its operational context. Biases from algorithmic discrimination, sampling issues, or unintended model behavior can undermine system effectiveness and must be addressed with targeted interventions. Mitigating bias requires balancing fairness, predictive performance, and business impact. Researchers and system developers must go beyond predictive accuracy and holistically assess ML-based credit scoring models, ensuring alignment with institutional goals, regulatory expectations, and ethical standards.
Keywords
- Machine Learning
- Decision Support Systems
- Finance and Banking
Status: accepted
Back to the list of papers