Bias in Black Boxes: A Framework for Auditing Algorithmic Fairness in Financial Lending Models

This study presents a comprehensive and practical framework for auditing algorithmic fairness in financial lending models, addressing the urgent concern of bias in machine-learning systems that increasingly influence credit decisions. As financial institutions shift toward automated underwriting and risk scoring, the opacity of complex models creates challenges for regulators, borrowers, and internal risk teams who must ensure that lending processes remain fair, transparent, and compliant with legal standards. The proposed framework begins by recognizing the structural and historical nature of discrimination embedded in lending data, where past approval patterns, socio-economic disparities, and proxy variables can produce unintended disadvantages for protected groups such as racial minorities, women, younger or older borrowers, and marginalized socio-economic classes. To address these issues early in the development pipeline, the framework incorporates a data-level audit that includes detailed profiling of input variables, assessment of representation imbalances, analysis of missingness, correlation structures, and detection of historical skew. This stage also requires testing for proxy variables, because attributes such as geography, employment type, or credit history length may indirectly encode sensitive information. Beyond data diagnostics, the framework moves into a model-centric audit that evaluates predictive behavior using multiple fairness metrics, including statistical parity, equal opportunity, equalised odds, disparate impact ratios, and calibration across demographic groups. These metrics help auditors detect whether the model treats similar applicants differently based on protected characteristics, whether error rates are uneven, or whether disparities appear in approval thresholds. The inclusion of explainability tools such as SHAP, LIME, and partial dependence plots makes it.