Applications of Machine Learning Techniques to Mortgage Prepayment Modeling

Abstract

This thesis will test the performance of non-linear modeling in predicting prepayment risk for mortgage-backed securities relative to traditional linear models. Recent studies indicate promising results relative to linear models for machine learning algorithms and deep learning programs when predicting future prepayments. This thesis will compare linear regressions (i.e. logistic, Ridge, and LASSO) to a nonlinear machine learning system (random forest) using Fannie Mae’s Single-Family Loan Performance Dataset. The results of this experiment favor the use of random forest models when compared to the logistic and penalized regression models, but limitations in computing power as well as potential data under-fitting issues may understate the true accuracy of random forests. It is the hope that these results can provide valuable insight into the use of machine learning systems to optimally forecast prepayment risk when pricing Agency MBS assets. A combination of both linear and non-linear models may be the most optimal solution for modeling in this sector.