Balancing Fairness and Accuracy in Machine Learning-Based Probability of Default Modeling via Threshold Optimization

This study presents a fairness-aware framework for modeling the Probability of Default (PD) in individual credit scoring, explicitly addressing the trade-off between predictive accuracy and fairness. As machine learning (ML) models become increasingly prevalent in financial decision-making, concerns around bias and transparency have grown, particularly when improvements in fairness are achieved at the expense of predictive performance. To mitigate these issues, we propose a model-agnostic, post-processing threshold optimization framework that adjusts classification cut-offs using a tunable parameter, enabling institutions to balance fairness and performance objectives. This approach does not require model retraining and supports a scalarized optimization of fairness–performance trade-offs. We conduct extensive experiments with logistic regression, random forests, and XGBoost, evaluating predictive accuracy using Balanced Accuracy alongside fairness metrics such as Statistical Parity Difference and Equal Opportunity Difference. Results demonstrate that the proposed framework can substantially improve fairness outcomes with minimal impact on predictive reliability. In addition, we analyze model-specific trade-off behaviors and introduce diagnostic tools, including quadrant-based and ratio-based analyses, to guide threshold selection under varying institutional priorities. Overall, the framework offers a scalable, interpretable, and regulation-aligned solution for deploying responsible credit risk models, contributing to the broader goal of ethical and equitable financial decision-making.