Our near and dear old-economy business of real estate - which makes up more than 10% of the nation’s GDP and $32 trillion (!) of value globally - has plenty of entrenched norms and processes that technology could help make more efficient. Appraisals - especially the analysis behind them - are one such corner of the industry. Their weight is hard to overstate: whenever an asset is financed or an owner needs to value an asset or portfolio without selling, owners turn to appraisers. Regulators and insurers lean on them too. Everyone knows the appraisal process is flawed but recent research confirms that a dose of machine learning could make appraisals materially more accurate.¹

Specifically, using machine learning algorithms improves appraisal accuracy by 20% for apartments, 18% for industrial, 16% for office and 14% for retail assets. Also, the dispersion of ML-derived values around the actual sale value is tighter than that of traditional appraisals, and while traditional appraisals systematically overvalue assets, the ML valuations were not biased in either direction. This is better in so many ways.

Research has long shown that appraisals have a “consistent tendency of structural bias and inaccuracy.” Despite that truth, appraisers’ methods “have largely remained unchanged for the past decades.” Some of that is because appraising commercial assets is hard, or at least harder than appraising houses, where you find abundant data on assets with similar property characteristics. But this new study, by researchers at the University of Regensberg in Germany and Florida International University, shows that modern ML can handle that complexity.

The study analyzed 24 years of property-level transaction data for 7,100 assets, with information provided by NCREIF. Each asset was appraised at least every quarter until sale. On average, appraisals missed the sales price by 11% and tended to overvalue assets, but interestingly the errors swung with the cycle: appraisals structurally undervalued assets by 5% during rising markets and overvalued assets 13% during 2008-2009, the only real falling market in their study. Critically, this underscores how “market cycles have an impact on the reliability of real estate appraisals” since appraised values lag sales prices when markets are changing.

Appraisals are a natural candidate for technological improvement because machine learning excels at finding hidden and non-linear relationships between value-driving factors. In an illiquid market where sales are rare, those subtle relationships between comparable assets - exactly what ML can find that mere mortals can not - are very valuable. And real estate is not getting more liquid anytime soon.

The researchers first mapped the differences between appraised values and subsequent sale prices, identified the drivers of those deviations and then built and trained an ML algorithm on 50 explanatory variables to value assets in a more comprehensive way. The study found the largest differences between appraisers’ valuations and those from actual investors stemmed from how each viewed building occupancy and recent capex - two factors subject to highly idiosyncratic underwriting.

As an aside, appraisals historically value real estate in three classic ways: on income, on replacement cost and on comparable sales. The income approach, which usually produces a valuation based on a discounted cash flow model, is the most common. The income model relies on NOI, which is easy to determine, and a capitalization rate, which is “not straightforward and depends on many assumptions,” the paper says, dramatically understating the truth. Each method has flaws, among which are backward-looking data (NOI), subjective assumptions (cap rate, comparable sales) and relying on data that’s mostly unavailable to appraisers (true land and building costs). These limitations set the stage for why machine learning can do better.

It’s no surprise that machine learning, smartly applied, would limit errors in appraisal valuations, but the scale of improvement is striking. It fixes problems that have dogged the industry for decades and that create real inefficiencies that no one has tried hard to fix; an appraised value is still largely influenced by individual biases.

This research—the first to rigorously test ML’s impact on appraisals—shows how existing tools could dramatically improve a critical, billion-dollar real estate service. The inefficiencies are well known, the technology is proven, and the opportunity is waiting. The only question is who will move first.

Who: Cherre

What: A frontrunner in the technical race to facilitate better investment- and asset-management decision making.

The Sparkle: Lots of investment-firm-facing proptech offerings get you dashboards, and as AI can digest unstructured data and data from disparate sources (including external feeds like CoStar), those dashboards are becoming more valuable. What’s new and what Cherre is pioneering is benchmarking measures - seeing how your assets perform against market averages, etc. Soon, with the recent release of agentic AI tools, we suspect you’ll get AI-generated, idiosyncratic guidance that helps drive NOI at each individual asset.

What if you could search every visible word on New York City’s streets?

The word “pizza” can be found more than 111,000 times walking the streets of New York City. That’s a lot of slices. A new mapping project shows how certain words light up the city: the location and occurrence data on text in “The Big Apple” (3,747 matches) leads to beautiful and insightful maps.

Some, like “Stop” and “One Way” are everywhere equally. “Broadway” lights up thoroughfares, especially in Manhattan, Brooklyn and Queens. “Jerk” (the chicken) lights up Jamaican enclaves. Have a look:

How many times does the word “pizza” appear on NYC streets? Hint: It’s more than 1, and less than 111,291.

Editor’s Note: The Real Estate Haystack believes in sharing valuable information. If you enjoyed this week's newsletter, subscribe for regular delivery and forward it to a friend or colleague who might find it useful. It's a quick and easy way to spread the word.