Content area
Consumer lending models are critical for making credit decisions, approving loans,and assessing credit risk, which can significantly impact people’s financial health in the US. These opaque "black-box" models can often lead to unfair and discriminatory outcomes for specific groups and to the perpetuation of historical and societal injustices. As a result, many people still cannot achieve financial stability and freedom, and to make data-driven upward social mobility possible, we need to establish fairness guardrails to reduce bias in credit scoring models like FICO and VantageScore. Accordingly, this dissertation offers a comprehensive framework to address the trilemma of ensuring AI fairness, upholding regulatory compliance, and seizing a substantial market opportunity to serve underserved populations. This study begins by examining all past relevant work and its gaps, then establishing the legal and technical foundations of fair lending, including the statutory requirements of the Equal Credit Opportunity Act (ECOA) and the Fair Housing Act (FHA), the mathematical tradeoffs inherent in fairness metrics, and the application of model risk management principles. A quantitative analysis of a decade of enforcement actions reveals the substantial and growing financial risk of non-compliance. These models’ data processing methods must be public and transparent, as biases and errors can arise from both the data source and the processing itself. In contrast, a market-sizing analysis quantifies the multi-billion-dollar opportunity to expand credit access. The central contribution is a proposed governance structure grounded in the principles of Explainable AI (XAI) to develop, validate, and monitor fairness in consumer lending AI models. Through future partnerships among data scientists, regulators, and industry stakeholders, this framework aims to create more equitable financial models that benefit individuals and communities, transforming a compliance burden into a virtuous cycle of strategic growth and proactive inclusion.