Case Study 1 — The Mid-1990s Arrival of Automated Underwriting

What changed when the first pass stopped being a person

Background

For most of the twentieth century, a residential mortgage was underwritten the way a bank loan had always been underwritten: a human being read a file.

The underwriter had a guide, a set of benchmark ratios, and their own accumulated judgment. They read the application, the verifications, the credit report, and the appraisal, and they decided. Then the file went back to the branch, or to the correspondent, or to the broker's account executive, and everybody waited. Turn times measured in weeks were unremarkable. A file that needed a second look could sit for days. If a borrower failed a benchmark ratio, the file's survival depended on whether a particular underwriter, on a particular Tuesday, was persuaded by the compensating factors in front of them.

Two things about that arrangement are worth holding onto before we describe what replaced it, because both of them made the change inevitable.

It was slow and expensive. Every file consumed skilled human hours regardless of how obviously good or obviously hopeless it was. A borrower with a 780 score putting thirty percent down got the same multi-day treatment as a marginal file, because nothing in the process could tell the difference cheaply.

It was inconsistent, and the inconsistency had a documented pattern. In 1992 the Federal Reserve Bank of Boston published a study of Boston-area mortgage lending — Munnell, Browne, McEneaney, and Tootell, Mortgage Lending in Boston: Interpreting HMDA Data, later published in revised form in the American Economic Review in 1996 — which examined denial rates using a far richer set of financial variables than the publicly reported HMDA data contained. It found racial disparities in denial rates that persisted after controlling for those variables. The study's methodology was vigorously contested at the time and afterward, and the debate over it is genuinely technical. What is not contested is that it made a question unavoidable: what exactly is the discretionary part of manual underwriting doing?

Into that context, in the mid-1990s, both government-sponsored enterprises deployed automated underwriting systems: Fannie Mae's Desktop Underwriter and Freddie Mac's Loan Prospector (now Loan Product Advisor). Each was built on something neither the enterprises' customers nor anyone else possessed — a database of the actual performance of millions of mortgage loans. Adoption was rapid. Within a few years, submitting a conventional conforming file to an AUS had gone from novelty to default, and by the end of the decade an originator who did not run one was working against the grain of the entire industry.

The issue

The systems did four things at once, and it is worth separating them because the industry has consistently confused them.

1. They collapsed the calendar

The first-pass evaluation went from days or weeks to minutes. That is the change everybody noticed and the one that sold the systems.

But the more important consequence is where in the process the evaluation moved. In the manual world, the underwriting decision came near the end: application, processing, verification, appraisal, and then — after real money and real weeks had been spent — a decision. In the automated world, the first-pass evaluation comes at the beginning, on day 5 or day 6, before the appraisal is ordered and before the borrower has been asked for a single verification.

That is the entire architecture of modern origination, and it is the reason Chapter 15 exists. A loan officer today finds out on day 6 what a loan officer in 1990 found out on day 30.

2. They tiered the documentation

Because the system evaluated the file's risk, it could scale the documentation to it. A file the model regarded as strong could be documented more lightly than a file it regarded as marginal — fewer paystubs, fewer months of statements, streamlined verification paths.

This was a genuine and legitimate innovation, and it is also the one that got mangled later. The AUS documentation tiers reduced the amount of paper required to verify facts the file still had to state and the system still had to evaluate. They did not dispense with the facts. The distinction matters enormously, and Part 2 of this case study is about what happened when the industry lost it.

3. They replaced sequential thresholds with a combined evaluation

This is the substantive change and the one least understood outside underwriting.

A benchmark-driven manual process asks a series of yes-or-no questions. Is the housing ratio under the benchmark? Is the total ratio under the benchmark? Is the score above the floor? A borrower who failed one of them needed an underwriter to be affirmatively persuaded by compensating factors — and whether that happened depended on the underwriter.

A statistical model does not work that way. It evaluates the characteristics together, with strength in one dimension genuinely offsetting weakness in another as a matter of the model's structure rather than an individual's disposition. The result, systematically, is that some borrowers who would have failed a threshold test receive an approval.

The Harlow Street file in this book is exactly that borrower: a 641 representative score, \$4,150.00 of income, ratios of 41.48% front and 51.00% back against an FHA manual benchmark of 31% / 43%. Under a benchmark, that borrower's payment would have to be roughly a quarter smaller. With an Accept from FHA's TOTAL Scorecard, the loan is approvable as submitted.

4. They made underwriting a data problem

And this is the consequence nobody advertised. Once a machine makes the first pass, the quality of the input becomes the binding constraint on the quality of the output. A manual underwriter reading a file will notice that the stated income and the paystub disagree. A model reading a data record will not, because it never sees the paystub.

The industry took roughly fifteen years to fully absorb that fact, and it absorbed it the hard way.

What it shows

The access argument, honestly stated

The GSEs argued that automated underwriting would expand credit access, particularly for borrowers whom benchmark-driven manual underwriting had been declining. Freddie Mac researchers Susan Gates, Vanessa Perry, and Peter Zorn published a version of this argument in Housing Policy Debate in 2002 — Automated Underwriting in Mortgage Lending: Good News for the Underserved? — reporting that automated evaluation increased approval rates, including for lower-income and minority applicants.

Read that with two things in mind. First, the authors were affiliated with one of the enterprises whose system was being evaluated, which does not make the work wrong but is a fact a professional reader should know. Second, the underlying mechanism is real and is visible in every findings report you will ever pull: a model that weighs reserves, residual credit history, and employment stability against a high ratio will approve files a two-number benchmark declines.

The honest summary is that automated underwriting genuinely did approve borrowers a benchmark would have refused, and that this is a real expansion of access, and that it is not the same thing as having eliminated bias from mortgage lending. Case Study 2 takes up the second half of that sentence.

The 2004–2007 counterpoint, and what it actually proves

The obvious objection is: automated underwriting was in universal use by 2005, and the American mortgage market then produced the worst underwriting outcomes in its history. Does that not settle the question?

No — and getting this right matters, because the wrong lesson has been drawn from it constantly.

The loans that failed catastrophically were, in very large part, originated outside the agency AUS channel. Subprime and Alt-A products, stated-income and low-documentation underwriting, payment option ARMs, and 100% combined-LTV piggyback structures were originated to private guidelines and sold into private-label securitizations, not evaluated against the Fannie Mae Selling Guide by Desktop Underwriter. The GSEs did loosen during the period, and they did buy loans that performed badly, and both entered conservatorship in 2008. But the specific catastrophe was not "a model approved bad loans." It was the abandonment of underwriting as such.

The precise lesson is the one this chapter teaches in §15.9:

An automated underwriting system is only as good as (a) the guideline set behind it and (b) the data entered into it. Neither of those is the model's responsibility. Both are the industry's.

"Stated income" is the cleanest illustration available. An AUS documentation tier reduces the paper needed to verify an income figure that the file still asserts and the system still evaluates. A stated-income product removes the verification of a figure nobody intends to check. The two look similar on a checklist and are opposite in kind, and the industry spent several years pretending otherwise.

What the enterprises did afterward

The post-crisis response is a useful tell about where the real problem was. Both enterprises tightened and repeatedly recalibrated their engines, and both eventually built out programs tying representation-and-warranty relief to data validated by approved third-party sources rather than to data typed by a loan officer. Fannie Mae's and Freddie Mac's validation services do exactly that for income, employment, and assets.

Read that as an institutional admission: after a decade of arguing about models, the enterprises put their money on data quality. Which is where a loan officer's leverage has been the whole time.

Outcome

Automated underwriting is now the default first pass for essentially all agency and government-insured residential lending in the United States. Manual underwriting still exists, still closes loans every day, and is the required path whenever the automated one is closed — but it is the exception rather than the process.

For the person taking applications, four things follow:

1. You find out early. The evaluation is available on day 5 or 6, before real money is spent. That is a gift, and it is wasted by anyone who does not read past page one.
2. The document list is generated for you. The verification messages are the single most under-used artifact in origination.
3. The judgment did not disappear; it moved. It moved upstream, into which income figure is countable, which debt is excludable, whether a deposit is sourced, how a property is classified. The machine evaluates those judgments. It does not make them.
4. The discretion that was removed was the underwriter's, not yours. A loan officer has more influence over the outcome of a modern file than an underwriter had over a 1990 file — because the loan officer supplies the data.

The lesson

Automation moved the decision earlier and made it cheaper. It did not make it self-executing.

The mid-1990s change is usually told as a story about speed. It is better understood as a story about relocation: the analytical work did not vanish, it moved from the underwriter's desk on day 30 to the loan officer's keyboard on day 5. That is why a book about being a loan officer spends a chapter on reading findings reports. The job absorbed a piece of underwriting, permanently, and mostly without telling anyone.

And the second lesson, which the 2000s taught at enormous cost: a good model on bad data is worse than no model at all, because it produces a confident answer. A manual underwriter who does not know something usually knows they do not know it. A model does not have that capacity. It answers the question you asked, in the format you expected, to two decimal places.

Discussion questions

1. The chapter argues that the most important consequence of automated underwriting was not speed but when in the process the first-pass evaluation happens. Restate that argument in your own words and say what it implies about how a loan officer should structure their first week on a file.

2. The Boston Fed study raised a question about discretion in manual underwriting. Automated underwriting removes discretion from the first pass. Does removing discretion remove the problem the study identified? Argue both sides before you answer.

3. Distinguish, in two sentences a new loan officer would understand, between an AUS documentation tier and a stated-income product. Why did the industry conflate them?

4. The GSEs' post-crisis move toward third-party-validated data is described here as "an institutional admission." What was being admitted? Do you agree with that reading?

5. If automated underwriting genuinely approves borrowers a benchmark would decline, and the model's reasoning is not published, what should a loan officer say to a borrower who asks why they were approved when their ratios look bad? Write the actual sentences.

6. Suppose the systems had never been built and every file were still manually underwritten today. Name three specific things about your job that would be different — and one thing that would be better.

Sources for this case study are documented public history: the Federal Reserve Bank of Boston's 1992 Boston mortgage lending study and its 1996 revision; Gates, Perry, and Zorn in Housing Policy Debate (2002); the Fannie Mae Selling Guide and Freddie Mac Seller/Servicer Guide; and the well-documented 2008 financial crisis and conservatorship of Fannie Mae and Freddie Mac. Exact launch dates, adoption percentages, and market-share figures are deliberately not asserted here. The Linden Street and Harlow Street files are constructed teaching examples.