Analyzing the Risk of Mortgage Default

This paper analyzes the risk of mortgage default and prepay for single-family, 30year fixed rate mortgages using a variety of machine learning and survival analysis methods. Predictions are made for homeowner choices to continue payment, default, or prepay using both parametric and non-parametric models. These models include Binary Logit, Multinomial Logit, K-Nearest Neighbors, K-fold Cross Validation, and Random Forest. Replications of each model, with various combination of parameters, were performed in order to identify the best model; the Random Forest model with 150 trees and 4 entries yielded the highest accuracy at 93%. Difference in survival time (months of mortgage payment until termination) are then compared for owneroccupied homes versus investment homes. Investors are found to pay mortgages longer than primary homeowners. Meanwhile, primary homeowners are less likely to default and more likely to prepay than their investment counterparts, suggesting the presence of an endowment effect. Survival curves are also plotted for six Cox Proportional Hazard models after checking relevant assumptions.