30 Year Fixed-Rate Agency Mortgage Default Risk Prediction

Prior to 1980’s, people evaluated mortgage default risk established on rule of thumb and their experience towards risk ratings. The collapse of mortgage market in 2008 stimulated people to quantitatively assess mortgage default risk hence different statistical models are applied to consider different and specific business requirements. A good measure of Probability of Default (PD) benefits financial institutions in assessing loan loss and some insights from modelling default risk can guide a competitive mortgage pricing and better underwriting practice. This project provides some quantitative methods to help institutions assess the default risk on a pool of mortgage loans using statistical tools. Several models such as logistic regression, gradient boosting and decision trees were used to evaluate risks. HPI files were matched with original data to provide better predict ability. Through GBM feature selection and covariance matrix analysis, eight variables were totally selected. Cross validation was conducted to choose best hyper parameters and rare event model was applied in logistic regression as well to reduce suffering from small-sample bias. The ROC tests for these models reached 0.81 and KS scores were greater than 0.45. 3.5 million Mortgage observations were used from Freddie Mac (issued between Q42009 to Q32012 with five years’ performance). HPI (Housing Price Index) file from FRED is also used to combine with mortgage data. This paper adds machine learning techniques to traditional statistical models to predict single-family mortgages’ default risk within five years of the issuance. Strategies such as mortgages swap in/out can be applied on application side together with models, which can improve profitability and system efficiency. We can also have a better intuition towards how credit markets evolve after the financial crisis from feature importance of different models.No embargoAcademic Major: Financ