Automated Underwriting
of Residential Mortgage Applications

Which is the better way to appraise applications for residential mortgages: human judgment or computer programs? In a recent issue of Housing Policy Debate [volume 13, issue 2] three authors examine the question: "Automated Underwriting in Mortgage Lending: Good News for the Underserved?" (The authors: Susan Wharton Gates, Freddie Mac; Vanessa Gail Perry, George Washington University; Peter M. Zorn, Freddie Mac). The issue is especially important since some advocates for "underserved" credit applicants assert, "electronic underwriting is inherently discriminatory against African Americans and other minorities." (See "Class Action Suit Filed Against Fannie Mae" in the section on Legislation and Litigation".)

Credit scoring had been used for many years by grantors of consumer credit, such as issuers of credit cards, automobile loans and small loans. Back in 1996, Robert Avery and other economists at the Federal Reserve published an article in the Federal Reserve Bulletin pointing out that the use of credit scoring could result in more mortgage loan approvals. Their predictions have been realized. Gates, Perry and Zorn note, "Over the past six years, automated underwriting (AU) systems have become the tool of choice in mortgage lending decisions." The basic issue is whether or not automated evaluation systems are good or bad for underserved consumers.

This article addresses a number of important concerns related to AU. However, our focus is on two issues: (1) Does AU using Freddie Mac's automated underwriting system (Loan Prospector) efficiently sort good from bad loans? (2) Does the Loan Prospector do better at sorting good from bad than human judgment?

Sorting "good" from "bad"

Using date from 1994 and 1995, the authors found that loans ranked as "caution" by the Loan Prospector experienced default rates that were four times the average rate for all loans. By contrast, loans rated as "accept-plus" had default rates that were only one-fifth of the overall average. "Moreover, Loan Prospector predicted default for both low income and minority borrowers. Low-income borrowers rated caution default at four times the average, while minority borrowers rated caution default at roughly five times the average. Clearly, there is substantial evidence that Loan Prospector accurately predicts default."

Judging human decisions vs. computer decisions

It is evident that decisions made by the computer were very accurate in predicting the performance of mortgage loans. However, were its predictions any better than those made by human judgment? Surely, experienced humans would outperform a heartless computer. Sadly, this is not the case. This question was examined by using a random sample of about 1,000 residential mortgage loans that had been purchased by Freddie Mac in 1994 and 1995 under affordable housing initiative. As part of the program, Freddie Mac then had its loan analysts rate the loans. We focus on differences in the outcomes of the human decisions and those that would have been made by the computer. There are several possible outcomes. On the one hand, we have the "swap in," loans that had been rated "caution" by humans but rated "accept" by Loan Prospector. In reality, the 90-day delinquency rates on the loans that would have been approved by Loan Prospector, in spite of a negative human appraisal, turned out to be only one-fifth the overall delinquency rate on all loans. They performed as well as mortgage loans that had been rated "accept" by both the humans and Loan Prospector.

Case 2. On the other hand, we have the "swap outs," those loan applications that had been accepted by humans, but would have been rejected by Loan Prospector. Again, the Loan Prospector out-performed the humans. The accepted loans that would have been rejected by the Loan Prospector had an average delinquency rate that was 1.75 times the overall average. Moreover, these accepted loans performed as poorly as those rated "caution" by both manual underwriters and Loan Prospector.

Finally, using the same random sample of 1,000 loans, the authors address the important issue of whether or not computer credit scoring is unfair to minorities and households with low incomes. As shown in the table below, the 2000 version of Loan Prospector appears to have substantially increased the likelihood that minority borrowers would be able to obtain home loans. Over the period from 1995 to 2000, the proportion of applications from low- and moderate-income that was accepted by Loan Prospector grew from 31.9 percent to 41.0 percent. Over the same period, the proportion of applications from minority borrowers that were approved rose significantly, in part as a result of the improvements in the Loan Prospector.

Source: Cited above.

Conclusions

The authors' conclusions are worth quoting: "We draw two conclusions from this analysis.

1. Compared with traditional manual underwriting, AU more accurately predicts default, and

2. AU's greater accuracy results in higher borrower approval rates, especially for underserved applicants."