Responsible AI in Historical Mortgage Decision Prediction: Calibration, Group Differences, and Audit Evidence in Delaware

Mortgage lenders need evidence on how prediction models reproduce past decisions and distribute approvals across applicant groups. We study 34,012 Delaware mortgage records from 2024 and compare logistic regression with flexible tree-based models. Three findings answer the research questions. First, the selected flexible model predicts recorded lending decisions more closely and reduces average squared probability error by about 30% relative to penalised logistic regression. Second, racial differences persist when race, sex and ethnicity remain outside the model inputs. At a decision threshold chosen to reproduce the overall historical approval rate, the model approves about 66% of Black applicants and 78% of White applicants. Third, checks of probability reliability, missing debt information and decision thresholds identify issues for audit review. The available evidence does not establish individual reasons for denial. Independent mortgage banks account for 56.0% of sample records, and 57.5% record use of at least one automated underwriting system. The analysis concerns historical decisions and provides no evidence on subsequent loan repayment. Lenders and regulators gain a practical sequence for testing models, reviewing group differences and requesting supporting decision records before operational use.