Predicting Short-Term Severe Delinquency Risk for Mortgages Using Fannie Mae Data

Abstract

Mortgage default risk prediction is an important problem in credit risk management. Using the Single-Family Historical Loan Credit Performance Primary Dataset published by Fannie Mae, this study examines whether newly originated mortgage loans become 90 or more days delinquent within the following six months. The sample covers nine quarters from 2023Q3 to 2025Q3. To ensure consistent label definitions, the training set, validation set, and nominal test interval are split in chronological order, and samples are screened under a unified observability rule. Methodologically, Logistic Regression (LR) is used as an interpretable baseline model, while XGBoost serves as a nonlinear benchmark, with classification thresholds selected on the validation set. The results show that LR performs better in temporal stability, whereas XGBoost achieves stronger overall discrimination, better probability estimation quality, and better high-risk sample screening performance. Further analysis based on Shapley Additive Explanations (SHAP) indicates that credit score, debt-to-income ratio (DTI), loan-to-value ratio (LTV), loan amount, and regional variables are closely associated with severe delinquency risk. Overall, under the setting of this study, the nonlinear model shows stronger predictive ability for short-term severe delinquency, while the linear model remains a stable and transparent benchmark.