Machine learning in default Prediction : the incremental power of machine learning techniques in mortgage default prediction

In this thesis, alternative machine learning techniques have been used to test if these perform better than a Logistic Regression in predicting default on retail mortgages. It is found that the ROC AUC statistic is slightly better for the advanced machine learning techniques, i.e. the Neural Networks, Support Vector Machines and Random Forests. Importantly, all classifiers are trained on the same variables, which are all Weight of Evidence transformed. This enables us to compare the results and view the incremental predictive power as solely a result of the classifiers. Also, it enables us to use the same methodology for probability of default modelling as practitioners currently use, i.e. with Weight of Evidence transformed variables. The analysis is based on a dataset with observations on each loan issued from a financial services firm in the market for retail mortgages in the years 2009-2017. After univariate and multivariate analysis, the number of candidate variables are reduced from 549 to 19. The best model is the deep Neural Network, with an impressive ROC AUC of 0,902. This is very high for prediction of default. Still, the Logistic Regression model also has a very high statistic of 0,882. A more primitive machine learning technique is also included in the analysis, the Decision Tree. As expected, this classifier has the lowest ROC AUC of 0,732. Through the exploratory analysis with WoE variables interesting relationships are found, which may enjoy some readers. Keywords – Probability of Default, PD, Mortgage default, Bankruptcy prediction, Weight of Evidence, Basel, IRB, Neural Network, Support Vector Machine, Random Forest, K-Nearest Neighbor, Decision Tree, Logistic Regression, ROC, Confusion Matri