Modeling subprime mortgage delinquency, default, prepayment and loss

The current financial environment presents significant challenges for the mortgage industry. Declining house prices have surfaced the importance of delinquency, loan default and loss predictions. Simple models of prepayment behavior are no longer applicable. Investors, originators, servicers and regulators are in need of more accurate predictions for their portfolios of interest.;This dissertation focuses on two topics relevant to modeling residential mortgages. The first topic provides a framework for modeling delinquencies, prepayments, defaults and losses that represents an enhancement over previous studies. A total of nine loan payment statuses are used (current, thirty-days delinquent, sixty-days delinquent, ninety-days delinquent, early foreclosure, late foreclosure, real estate owned, paid in full, and terminated with loss). This framework is compared to the previous framework discussed in the literature that used seven statuses.;The second topic applies reconstructability analysis (RA) to residential mortgage data in order to find new and interesting models. Many statistical methods are unable to reflect non-linearities and significant high-level interactions. RA is capable of doing both. The study explores the hypothesis that the inclusion of RA-suggested interaction terms would improve the accuracy of the logistic regression (LR) models used to forecast loan status changes within mortgage portfolios.;The first topic's result made two unique and important contributions to the mortgage management literature. First, it finds that the nine-state framework yields more accurate results than the seven-state framework. It also introduces a new state 'terminated with loss' that enables the framework to predict losses.;The second topic's results confirm the hypothesis that RA suggested interaction terms improve the performance of LR model. This is a useful contribution to the data mining literature since it enhances the performance of LR which is a widely used data mining methodology.