WASHINGTON -- The Federal Reserve Bank of Boston's landmark discrimination study provides bank executives with a powerful yet inexpensive tool for detecting bias in their lending procedures, regulators said last week.
For its study, the Boston Fed created a statistical model of mortgage lending behavior. Relying on only 12 variables, researchers were able to predict with surprising accuracy whether a loan application would be accepted or denied.
Moderate Cost per Loan
Fed researchers say the model can be easily applied at any bank. The cost: an estimated $30 or less per loan for banks with at least 200 mortgage applications a year.
"If I were in charge of a large mortgage company, I might want to use this model," said Alicia Munnell, senior vice president of the Boston Fed and primary author of the study, which confirmed that racial discrimination is a significant problem among home lenders.
Regulators themselves will also be using the model to root out bias.
At the Boston Fed, researchers have already put the concept to work analyzing each loan application collected its the study of metropolitan Boston.
In cases where a loan was denied, but the model predicted it should have been approved, the Fed passed on the application to the primary regulator for further examination.
Regulators say they are following up on the Fed's information. They are also developing computer programs for examiners to use with raw loan-approval data compiled under the Home Mortgage Disclosure Act.
Many Sources of Data
The Boston model, which regulators emphasize is not fool-proof, uses data from loan applications, credit reports, and census tracts to predict and evaluate lending decisions.
Crucial variables determining probability of loan acceptance include consumer credit history, history of bankruptcy, private mortgage denials, housing expense-to-income ratio, and race.
After running more than 3,000 applications through the model, the Boston Fed found that, when all the variables were accounted for, black applicants were 60% more likely to be denied than whites.
While the researchers were more interested in lending discrimination on a regional basis, the model could be applied to more limited circumstances, Ms. Munnell said.
"One thing you can do is to target individual loan files or individual institutions," she said.
Using a Model
To use the model, a bank would have to buy inexpensive, off-the-shelf software and plug in the variables determined by the Boston Fed. Most of the cost of operating the system would stem from collecting data. (Details of the model are published in the Boston Fed's study.)
A bank could then run mortgage applications through the model to predict the likelihood of approval. If the model predicts a loan's acceptance, yet the applicant is denied, discrimination could be a factor.
The bank could then go back to the application and look more closely at the reasons for the rejection.
"Every file that's turned down will probably have a smoking gun," Ms. Munnell said.
"I'd be very interested" in making in-house use of the Boston model, said David C. Fynn, vice president of National City Corp., Cleveland. "If we can find a tool that will enable us to give a greater degree of consistency in lending decisions, then that tool has to be of interest to any lending officer."
Mr. Fynn sees ways his company could use the model even before final loan decisions are made.
"If the loans that are headed for rejection make it through the model, maybe they should go through another review," he said.
Tuning Up the Model
Users could replicate the Boston model on a readily available computer program, or tailor it to their own bank's lending policies.
For example, if a bank has a policy of automatically denying an applicant who had filed for bankruptcy, that bank could adjust the model to reflect that policy.
"If you have a hard and fast rule like that, then you wouldn't structure the model the way it was done in the Boston study," said Glenn Canner, a senior economist at the Fed. "But certainly institutions, if they choose, could use a statistical model to look at their behavior."
Variety of Factors
Researchers are careful to explain that no model is perfect or could ever capture every variable used to determine lending decisions.
A variety of factors that could not possibly be quantified go into the lending decision. And a model can be woefully inadequate and misleading, they say.
"We've heard all those criticisms," Ms. Munnell said. "There's a controversy among regulators about the extent to which they want to use equations. But I think if I were an examiner, I would be tempted, with a large institution, to do something like this."