FHFA to make Fannie, Freddie appraisal data public

Looking to provide more transparency into home valuations, the Federal Housing Finance Agency is making available to the public the Uniform Appraisal Dataset records compiled by the government-sponsored enterprises.

The data is drawn from 47.3 million appraisal records collected from 2013 through the second quarter of 2022 on single-family properties in a manner that protects borrower privacy. In addition, FHFA is offering UAD Aggregate Statistics Dashboards on its website to provide user-friendly visualizations of the newly available data.

At the Mortgage Bankers Annual convention in Nashville on Monday, FHFA Director Sandra Thompson announced these new features as part of the regulator’s efforts to reduce appraisal bias.

Less than a week ago, the MBA and other groups called on the FHFA to make public more data from Fannie Mae and Freddie Mac, including that for appraisals.

“We view this as a significant first step in sharing the vast amount of valuation data that’s retained by the enterprises,” Thompson said. “With the more than 23 million statistics about single family home appraisals, the public will be able to better monitor industry trends, compare appraisal gaps in minority neighborhoods across states and metropolitan areas, evaluate national state, regional and local trends in appraised values and gain a better understanding for how appraised values differ among neighborhoods and housing features.”

While the UAD Aggregate Statistics Data File is intended for those capable of using statistical software to extract and analyze the records, the UAD dashboards are for all interested in examining the information through customized maps and charts.

The government-sponsored enterprises have made a couple of innovations in the past couple of years to address appraisal bias, a later panel discussion said.

The first was to create objective data fields that use dropdowns and enumerations to have more consistency in the way property data is gathered, said Jake Williamson, senior vice president, single-family collateral at Fannie Mae. This eradicates any room for subjectivity in the report.

“That’s the first innovation that I would say we need to keep pushing as an industry,” Williamson said. The second involves undervaluation risk.

The housing industry has gotten good at looking for overvaluation risk, learning its lesson from the housing crisis. It built tools to guard against that. But now there  is risk of undervaluation, Williamson continued.

In June, Fannie Mae created in its Collateral Underwriter technology an undervaluation message. “Behind it, it has 16 statistical based reason codes that point to the root cause of the undervaluation risk,” said Williamson. “So we can actually start pinpointing what are the most common causes of undervaluation and how do we start to tackle it.”

A third innovation, which is in early stages, is image data collection at the property. 

“I think the next horizon is what do we do with all that image data?” Williamson continued. “I really like the whole concept of image recognition as a way to capture the rest of the appraisal in a very objective way. How can you start using machine learning algorithms to recognize those images to classify condition and quality?”

Danny Wiley, senior director, property valuation for Freddie Mac, brought up the use of inspection technology to help eliminate bias.

“The data tells us we actually get more accurate condition ratings with a third party inspection,” said Wiley.  “I don’t think people see that as potentially something to help with the bias but the data clearly shows that that is a contributor to improving our results and getting less bias in those condition ratings.”

Another innovation allows users to measure the bias in appraisals in a scalable and repeatable way, said Jeremy Sicklick the CEO of HouseCanary. “Being able to now have that information to say, every month, every day, every quarter, are we seeing those biases, where are we seeing, how are we seeing and how do you fix them, that was a big innovation.”

Many are concerned about algorithmic bias in the use of automated valuation models and artificial intelligence. To check for this possibility, Veros looked at its own AVM to detect any forms of bias in its data in certain markets, starting with Chicago, said Jeffrey Hogan, its vice president of valuations.

The area has a very diverse population with many races in close proximity. It was easy for Veros to take the census information and the zip codes and compare those in minority predominantly black communities with predominantly white communities, using sales information as a benchmark.

It started off in minority communities, noting variations of 15% or more in price compared with its AVM.

“We came to the conclusion that the percentage of undervaluation in each one of those markets didn’t vary according to the race or the demographics of that area,” said Hogan. “Now, obviously, that would make sense because AVM doesn’t know the race, it’s pretty much blind to those things.”

It repeated the exercise for a number of other metro areas and after that “we came to the conclusion that basically those differences are insignificant,” Hogan said.

Veros did look at the possibility of overvaluation in white communities and made a similar finding.

“I will tell you that AVMs generally…I believe can be part of the solution,” Hogan said. “But as we look at things that can be nonbiased, so to speak, the AVM does meet that criteria because it doesn’t know anything about the dynamics of who the buyer is [or] who the appraiser is.”