How Normalization Shapes Fairness–Performance Trade-offs in Threshold-Based Credit Decision Systems

Machine Learning (ML) models are increasingly used in credit risk assessment, creating a growing need for decision mechanisms that balance predictive performance and fairness. Fairness-aware threshold optimization has emerged as an attractive post-processing approach because it enables explicit control of fairness—performance trade-offs without modifying the underlying predictive model. These approaches typically aggregate heterogeneous performance and fairness metrics through scalarized objective functions, implicitly relying on normalization to ensure comparability across metrics. Despite its widespread use, normalization is generally treated as a neutral preprocessing step whose influence on optimization outcomes is assumed negligible. This paper investigates the impact of normalization on fairness-aware threshold optimization in credit risk modeling. Using a common optimization framework, three representative normalization strategies, including empirical min-max, theoretical-bound, and distance-to-ideal normalization, are evaluated across Logistic Regression, Random Forest, and XGBoost classifiers on both synthetic and real-world credit datasets. The analysis reveals that normalization substantially influences optimization outcomes. Different normalization schemes reshape the geometry of the objective landscape, alter optimal decision thresholds, and modify fairness—performance trade-offs. As a consequence, identical policy preferences may lead to materially different operational decisions depending solely on the normalization strategy employed. The observed effects are consistent across models and datasets, indicating that normalization acts as a structural component of the optimization process rather than a purely technical scaling operation. These findings highlight the importance of explicitly assessing normalization choices in machine learning-based risk modeling. More broadly, they demonstrate that normalization should be regarded as a modeling decision that directly affects fairness-aware credit decision systems and their governance.