Fairness-Aware Credit Risk Assessment Using Alternative Data: An Explainable AI Approach for Bias Detection and Mitigation

Abstract

We present a fairness-aware credit risk framework that fuses tabular and auxiliary signals with adversarial debiasing. On 150,000 applications, the method improves AUROC from 0.742 to 0.823 and achieves a 76.9% reduction in Demographic Parity violations (0.187 → 0.043) and 71.4% in Equalized Odds (0.234 → 0.067). Group-wise calibration (ECE) remains stable, and bootstrap confidence intervals with permutation tests (10,000 iterations) indicate statistical significance (p < 0.001). SHAP-based analyses show consistent feature usage across groups. We model the protected attribute A as binary for the discriminator (chance level ≈ 0.5 under balanced classes). Fairness is enforced via in-processing regularization on Demographic Parity and Equalized Odds; we report group-wise calibration and AUROC to assess trade-offs.