Quantifying Algorithmic Bias in AI-Based Credit Scoring Models: An Explainable Fairness Assessment

This study evaluates algorithmic bias in AI-based credit scoring by jointly analyzing predictive utility, probability calibration, group fairness disparities, and explainability-based disparity attribution under deployment-like underwriting constraints. Using a retail credit dataset with 48,750 eligible applications, the gradient boosting model achieved the highest test AUC of 0.792 and PR-AUC of 0.271, compared with 0.748 and 0.214 for logistic regression, and 0.781 and 0.259 for random forest. Overall calibration quality was acceptable (ECE 0.014–0.021), yet group-conditional miscalibration remained nontrivial, with ECE gaps of 0.006–0.010 across protected groups. At the primary operating point (approval rate 0.55), statistical parity differences of 0.042–0.050 and equal opportunity differences of 0.028–0.036 were observed across sex, age band, and intersectional groupings, indicating persistent access and qualified-approval disparities under a single global threshold. Policy sensitivity analysis showed that disparity increased in more permissive regimes, with SPD rising from approximately 0.030 at approval 0.40 to approximately 0.056 at approval 0.70, while AUC remained within 0.784–0.792. Explainability-based audits decomposed score gaps into feature contributions and identified utilization (0.016) and prior delinquency history (0.013) as dominant stable drivers, while geographic region exhibited a smaller but high-governance-risk contribution (0.010) that intensified under intersectional grouping. Robustness tests revealed fairness regression under temporal shift, where test AUC declined from 0.792 to 0.776 while SPD increased from 0.042 to 0.061. Mitigation experiments indicated that recalibration reduced SPD by about 0.004 with negligible AUC change, region coarse-graining reduced SPD by about 0.009 with an AUC decrease of about 0.002, and a combined intervention achieved the largest reduction (about 0.013) with limited utility loss (about 0.004). These results demonstrate that fairness risk is policy-contingent, drift-sensitive, and partially mediated by proxy pathways, and that explainable fairness attribution enables targeted governance interventions with measurable trade-offs.