Great Title

The Failure of Models that Predict Failure: Distance, Incentives and Defaults.
By Uday Rajan, Amit Seru and Vikrant Vig
Abstract:
Using data on securitized subprime loans issued in the period 1997-2006, we demonstrate that as the degree of securitization increases, interest rates on new loans rely increasingly on hard information about borrowers. As a result, statistical default model fitted in a low securitization period breaks down in the high securitization period in a systematic manner: it underpredicts defaults for borrowers for whom soft information is more valuable (i.e., borrowers with low documentation, low FICO scores and high loan-to-value ratios). We rationalize these findings in a theoretical model that highlights a reduction in lenders' incentives to collect soft information as securitization becomes common, resulting in worse loans being issued to borrowers with similar hard information characteristics. Our results partly explain why statistical default models severely underestimated defaults during the subprime mortgage crisis, and imply that these models are subject to a Lucas critique. Regulations that rely on such models may therefore be undermined by the actions of market participants.