Automated Fairness Audit and Bias Mitigation Pipeline for Loan Approval Models with Regulatory Compliance Reporting

Machine learning is widely used in loan approval systems today. It helps financial institutions process applications faster and more efficiently. However, this shift has also raised concerns about fairness. Models trained on historical financial data can unintentionally reflect existing social biases, which can result in unequal decisions for applicants based on factors such as gender, age, or socioeconomic background. These issues can conflict with regulations like the Equal Credit Opportunity Act (ECOA), the General Data Protection Regulation (GDPR), and the European Union Artificial Intelligence Act (EU AI Act). In this work, we present an automated fairness audit and bias mitigation pipeline designed specifically for loan approval models. The system focuses on identifying, measuring, and reducing bias in a practical way. At its core, it uses a Dual-Layer Hybrid Mitigation Engine (DLHME), which combines datalevel correction through IBM AI Fairness 360 Reweighing with model-level fairness optimization using Fairlearn’s Exponentiated Gradient approach. To make the system more transparent, SHAP-based explanations are included to show how each feature contributes to individual decisions. In addition, an interactive Streamlit dashboard allows users to perform real-time “What-If” analysis, making it easier for compliance teams to understand and evaluate model behavior without deep technical knowledge. We tested the system on a dataset of 45,000 loan applications, where it significantly improved fairness metrics. The Disparate Impact increased from 0.61 to 0.95, while the Demographic Parity Difference was reduced to below 0.04. The pipeline also generates automated compliance reports aligned with regulatory standards. These results show that it is possible to improve fairness in loan approval systems without sacrificing model performance.