Delinquency model predictive power among low-documentation loans

Using data from a major mortgage bank, we examine the predictive power of mortgage delinquency models as an aggregate measure of the quality of information recorded at loan origination. We measure model predictive power using an out-of-sample prediction criterion and compare predictive power of delinquency models over time and across loans of different documentation levels and origination channels. We found that model predictive power declined over time throughout the housing boom; further, predictive power is attenuated in delinquency models for low-documentation loans, providing suggestive evidence that loan and borrower characteristics used in the delinquency predictions may be inaccurate or falsified for such loans. We do not find differences in model predictive power when comparing bank- versus third party (broker) originated loans.