Boosting Performance of Traditional Models in Prepayment Modelling

This thesis uses the Freddie Mac Single Family Loan Level Data Set to investigate if a ma­chine learning algorithm called gradient boosting can outperform a multinomial logit model in monthly prepayment predictions for mortgages. If financial institutions can correctly predict prepayments, they can hedge risks and price mortgages better. Additionally, this thesis uses a model interpreter called Shapley Additive exPlanations (SHAP) to interpret the XGBoost model. The XGBoost is better in predicting prepayments than the multinomial logit model. Using SHAP values, this thesis finds that XGBoost is better able to capture the non-linear de­pendencies of prepayment events on explanatory variables. Although prepayment dynamics are better captured with the XGBoost model, both models are not able to discriminate well between a full prepayment and other prepayment classes on a monthly basis.