Predicting Mortgage Default Risk Using Cost-Sensitive Learning and Resampling Techniques: An Empirical Classification Analysis with Single Family Loan-Level Data

Abstract

This thesis performs an extensive analysis on predicting residential mortgage default risk using a binary classification framework. The dataset used in this thesis comes from CoreLogic which contains detailed borrower and loan property characteristics for US residential mortgages which has yet to be explored thoroughly within the current credit scoring literature. In accordance with the existing literature, we observe a class imbalance within our dataset. In order to tackle our imbalanced classification problem, we explore both algorithmic and data-level approaches. For our algorithmic-level approach we adopt cost-sensitive learning techniques. Additionally, we employ strategic resampling techniques using Synthetic Minority Oversampling Technique (SMOTE) and Edited Nearest Neighbors (ENN) as our data-level approach. The five classifiers compared in our analysis are logistic regression, decision tree, random forest, extreme gradient boosting and stacking. Furthermore, in efforts to provide a comprehensive framework when assessing the predictive performance of each classifier within our imbalanced learning task, we use the following four evaluation metrics: F\(_{2}\)-measure, MCC, G-mean and ROC AUC. Lastly, we provide an analysis on the relative feature importance of each variable in order to provide a deeper understanding into feature selection.