Residential‐Mortgage Lending Discrimination and Lender‐Risk‐Compensating Policies
使用波士顿联邦储备银行1990年的贷款申请数据，通过人工智能方法推断贷款人的风险补偿政策，发现这些政策在统计上解释了种族间贷款拒绝率的差异，但无法区分政策是否公平应用。
The Boston Federal Reserve study ( Munnell et al. 1996 ) concluded that illegal discrimination is a statistically significant contributor to the observed gap between white and minority residential‐mortgage rejection rates. The Boston study speculated that discrimination arises because lenders do not equally apply risk compensation or mitigation policies for imperfect loans. Using the same 1990 Boston loan application data, our study specifically examines the relation between compensating policies and discrimination. Since compensating policies are encouraged by secondary‐mortgage‐market sale guidelines, we model both the lender's origination decision and its loan sale decision. Using a rule‐based artificial‐intelligence technique applied to each lender, we infer compensating policies (rules) that equally apply to all races and explain lending decisions. A minority‐race indicator loses its statistical significance when an indicator of compensating‐policy violations appears in the loan accept–reject equation. This result reflects the fact that the risk levels of marginal minority loans tend to be more extreme than those of marginal white loans. However, the result does not necessarily reject the existence of discrimination. Equally applied policies may be empirically indistinguishable from unfairly applied policies. In addition, equally applied policies may fail the adverse‐impact doctrine if they do not serve a business necessity (such as profits). The industry's move away from discretionary, rule‐based decisions to mortgage scoring answers the need for a decision framework that rigorously uses loan performance to evaluate all loan applicants fairly.