SUPPORT VECTOR MACHINE VS. LOGISTIC REGRESSION FOR PREDICTING MORTGAGE DEFAULTS

Abstract: Mortgage loan providers estimate the credit risks it caries when approving a mortgage loan to their clients. Further, defaulting a mortgage loan is a risk that has been calculated through decades using statistical models. By using entries at the time of a mortgage application, the goal of the thesis is to com- pare the accuracy between logistic regression and Support Vector Machine in predicting a mortgage loan default. For this purpose, Fannie Mae 30-year- fixed-rate single-family mortgage loans are used for the years; 2000, 2005 and 2010. The models aim is to predict probability of default during five years period from the loan acquiring date. While the result showed that logistic re- gression was both faster and less complex to implement, SVM proved to have a marginally better prediction with the drawback of a longer computational time. The forecast accuracy to compare the two models at hand was ROC and Precision-recall, although precision-recall was favored due to the unbal- anced data.