Office of the Vice President for Global Communications

Tuesday, September 8, 2009

Subprime mortgage crisis: Failure to predict failure

There's plenty of blame to go around for America's subprime mortgage crisis among borrowers and lenders, but there may be another culprit beyond those already noted.

A researcher at the Stephen M. Ross School of Business says statistical models that predict loan defaults failed to warn lenders about risky borrowers, because these models relied too much on hard information, such as credit scores and loan-to-value ratios, and not enough on soft information from personal contact with borrowers, such as a person’s job security, upcoming expenses or other observable behaviors that may help predict the likelihood of default.

“A fundamental cause for this failure was that the models ignored changes in the incentives of lenders to collect soft information about borrowers and residential properties,” says Uday Rajan, associate professor of finance at the Ross School. “When incentives change, the link between the data and predicted outcomes changes in a fundamental manner.”

Rajan and colleagues Amit Seru of the University of Chicago and Vikrant Vig of the London Business School examined data on securitized subprime loans issued from 1997-2006. They found that in a high-securitization period — a lending environment where greater numbers of loans are sold to third parties — interest rates on new loans relied increasingly on hard information about borrowers (e.g., FICO scores and loan-to-value ratios) rather than more personalized soft information.

However, statistical models designed for low-securitization periods — where the original lender holds the loan — rely on more personal information. Those models break down when applied in a high-securitization period, the research shows. The result is that defaults are underpredicted for borrowers for whom soft information is more valuable, such as those with little documentation, low FICO scores and high loan-to-value ratios.

The researchers say lenders’ incentives to collect soft information changed because of the tremendous growth in securitization in the subprime sector after 2000. When a lender sells the loan to a third party, the original lender no longer bears the risk of default on the loan.

Conversely, in a world without securitization, default by the borrower directly hurts the lender, they say. That increases the incentive to analyze soft information in such instances.

“In addition to collecting hard data about a borrower, such as a credit score, a lender also has an incentive to verify undocumented information, or soft information, about the borrower,” Rajan says. “In particular, the lender screens out borrowers who are poor credit risks based on their soft information.

“But the incentive to acquire soft information about borrowers is lost under securitization, since only hard data can be transmitted credibly to the (third-party) investor. As a consequence, borrowers who are poor credit risks on the dimension of soft information, but apparently creditworthy based on their hard information, also receive loans. Thus, when one examines loans that have been approved, the same hard data have very different implications for borrower creditworthiness with and without securitization. That is, the hard information can mean something very different across these two worlds.”

Rajan and colleagues say their results partly explain why statistical default models severely underestimated defaults during the subprime mortgage crisis. In other words, the models failed to account for the change in the relationship between observable borrower characteristics and default likelihood caused by a fundamental change in lender behavior.

One broad implication of their findings is that regulations that rely on such models to assess default risk may be undermined by the actions of market participants. For example, current guidelines identify default risk as a key factor in setting capital requirements for banks and allow for the use of models by external institutions such as rating agencies in determining default risk.

“Even sophisticated agents such as regulators setting capital requirements or rating agencies will take some time to learn the exact magnitudes of relevant variables following a regime change,” Rajan says. “The assessment of default risk must be extra conservative during this period, and the true challenge for market participants is to recognize such shifts in real time.”

To read the study go to