Automated Property Valuation: Widely Used, Costly When Wrong, Less Accurate in Majority-Black Neighbourhoods

The use case

On 2 November 2021 Zillow Group reported a write-down of about $304 million on homes it had bought in the third quarter at prices above its own estimates of what they would later sell for (operator-reported, company press release)¹. The episode shows what is at stake when a valuation model moves from informing a decision to making one.

An automated valuation model (AVM) is software that estimates a property's market value from data such as past sales, property characteristics and location, without a valuer visiting. Lenders use AVMs to check collateral and monitor loan books and, in some countries, to set the value for a new loan. Consumers meet them as online home-value estimates. Home-buying firms use them to set purchase offers.

Scope. Most AVMs predate machine learning: comparable-sales models, regression on property characteristics and house-price indices (HPIs, which move an earlier valuation forward using a regional price trend) have long been used. Here "AI" means machine-learning AVMs. The only deployment in the evidence confirmed as machine learning is Zillow's Zestimate, which Zillow describes as a neural network–based model (vendor-reported, page last modified 24 September 2026)². The European survey and the US rule cover AVMs of every kind, and the survey's glossary states that comparables-based AVMs can count as advanced statistical models³. Figures for "AVMs" below are therefore not figures for AI alone, and are marked as such. Price forecasting, tax assessment and rental estimates are out of scope.

Where it is running

Opendoor. As of 31 December 2023 all of Opendoor's preliminary offers were algorithmically generated, and each offer was reviewed and finalised by its pricing team (operator-reported)⁴. The text read confirms algorithmic pricing but does not confirm that the algorithms are machine learning.

Europe. The May 2025 industry survey, which covers AVMs of every kind, lists AVM use in banking valuation in Denmark, Germany, Greece, Hungary, Ireland, Italy, the Netherlands, Norway, Romania, Serbia, Spain, Slovenia, Sweden and the United Kingdom, and not in Montenegro or Ukraine (self-reported)³. Respondents reported that AVMs cover 100 percent of the book in Ireland, 80 percent of portfolio valuation in Sweden (the other 20 percent by HPIs), 80–90 percent of residential portfolio coverage in Greece and 50–60 percent in Italy (May 2025, self-reported)³. Use is widest in monitoring and portfolio revaluation. For the original loan value it is narrower: seven countries report AVMs as the sole or primary valuation for purchase underwriting, in some cases only for some banks or in certain circumstances³. Germany has permitted AVMs since October 2022³. All European evidence is from one region-wide survey, so it describes policy and practice as reported, not measured results.

What the evidence shows

Percentage errors are larger in majority-Black neighbourhoods. An Urban Institute analysis of property-records sales and AVM data for selected counties in Atlanta, Memphis and Washington, DC (sales data 2000–2018, observational, no control group) found that AVMs in majority-Black neighbourhoods produced larger errors, relative to sale price, than in majority-white neighbourhoods⁵. In Atlanta in 2009 the percentage magnitude of inaccuracy was 64 percent in majority-Black areas against 24 percent in majority-white areas, and as of 2019 it was still more than twice the size⁵. The sample helps explain why: average home value was $169,855 in majority-Black neighbourhoods against $424,810 in majority-white ones, so a given dollar miss is a larger share of a cheaper home⁵. When the authors controlled for home value, neighbourhood conditions and turnover, the neighbourhood coefficient fell from 20.091 to 3.066 in their regression table, which is a large reduction but not zero⁵. These data are older than three years, cover three metropolitan areas, and no newer figure on the same measure was found.

Better data and machine learning narrowed the gap only partly. A 2022 follow-up using property-condition data and machine learning reported that the disparity persisted (abstract)⁷. A companion post (15 June 2022, modelled analysis) found poor exterior condition in 46 percent of majority-Black properties against 34 percent of majority-white ones in Atlanta, and 44 against 34 percent in Memphis⁶. A LightGBM model (a machine-learning method) of AVM error reduced its prediction error from 46.2 to 40.4 percent. In a modelled scenario, moving 60 percent of majority-Black properties into majority-white neighbourhoods lowered predicted error from 36.2 to 31.8 percent, a 4.4-point difference⁶. The mechanism is that record-based AVMs do not see condition, and condition differs systematically; the scenario suggests that most of the error remains once this is accounted for.

AVMs are also used as a yardstick for human appraisals, which inherits their bias. A September 2022 working paper on over 220,000 mortgages refinanced in 2000–2007 (observational, peer review not stated) found appraisals of minority-owned homes lower relative to AVM estimates: 0.9 percentage points lower for Black owners and 0.7 for Hispanic owners than for comparable White owners⁸. After adjusting for potential bias in the AVM itself, the discounts were larger: 3.6 points for Black owners and 2 points for Hispanic owners⁸. The adjustment is model-based and the data are old.

Accuracy: only vendor-reported figures, and old ones. A Zillow spokesperson told CNN in November 2021 that the Zestimate had a median error of 1.9 percent for homes on the market and 6.9 percent for homes off the market (operator-reported)⁹. A peer-reviewed teaching case cites 2.4 percent and 7.49 percent from Zillow's 2022 page (operator-reported)¹⁰. The Zillow page read for this paper says regional accuracy statistics are available, but they did not load, so no current figure was read². Every accuracy figure found for a machine-learning AVM comes from the vendor and is at least four years old. A median error means half of estimates miss by more: on a $400,000 off-market home, a 6.9 percent error is $27,600 (arithmetic illustration). No independent measurement of current machine-learning AVM accuracy was found.

What it takes: cost, integration, governance

US rules. Six agencies (the OCC, Federal Reserve Board, FDIC, NCUA, CFPB and FHFA) published a final rule on quality control standards for AVMs on 7 August 2024, effective 1 October 2025¹¹. Institutions must adopt policies, practices, procedures and control systems so that AVMs ensure a high level of confidence in the estimates, protect against manipulation of data, seek to avoid conflicts of interest, require random sample testing and reviews, and comply with nondiscrimination laws¹¹. The rule defines an AVM as any computerized model used by mortgage originators and secondary market issuers to determine the collateral worth of a mortgage on a consumer's principal dwelling, so the text read is neutral about technique¹¹. The compliance date was not verified.

European rules. The European Banking Authority's guidelines on loan origination and monitoring (EBA/GL/2020/06) introduced criteria for advanced statistical models; the survey glossary lists them as uncertainty, granularity, accuracy, validity, property-specific characteristics, representativeness, data quality and regular quality assurance, and says the capital requirements regulation (CRR3) adapted the same criteria³. National rules differ (May 2025, as reported by survey respondents). Spain forbids using an AVM as the primary property value for mortgage origination. Ireland does not permit standalone AVMs, requiring an accredited valuer to sign off. The Netherlands allows an AVM combined with a valuer's judgement and caps loans at 90 percent of the established value. Denmark requires valuer approval of model valuations. Inspections are required in eight of the 16 countries³.

Data and people. Data quality sets the limit. Greece uses surveyor valuations as benchmarks because of a lack of quality sale-price data, the United Kingdom notes that sale prices lag the valuation date, and Germany has no public national database³. Spain requires back-testing of each AVM valuation exercise against individual valuations³. At Opendoor, human pricing staff review every algorithmic offer⁴. No public price per valuation was found on an official pricing page, so cost evidence here is limited to the losses described below.

Where it falls short

Zillow Offers. Zillow's 2 November 2021 release reported a write-down of about $304 million in the third quarter, "as a result of purchasing homes in Q3 at higher prices than the company's current estimates of future selling prices", expected a further $240–265 million of losses in the fourth quarter (projected), reported a consolidated GAAP net loss of $328 million for the quarter, and announced a wind-down with a workforce cut of about 25 percent (all operator-reported)¹. The chief executive said that unpredictability in forecasting home prices "far exceeds what we anticipated"¹. The mechanism is that an estimate of what a home is worth today carries an error of several percent, and a firm that buys at its own estimate and sells months later also carries the market's movement. A peer-reviewed teaching case argues the Zestimate was not the sole cause: it also cites growth prioritised over profitability, pricing changes known as Project Ketchup, and overpayment¹⁰. The closure shows the cost of misplaced trust in a model, not that the model was the only fault.

Uneven errors. The Urban Institute findings above show larger percentage errors in majority-Black neighbourhoods in the three metropolitan areas studied, persisting after condition data and machine learning were added⁵⁷. Findings from other cities were not found.

Challenge question: what makes the use case look better than it is? Four things. The accuracy figures are vendor-reported, medians that hide the tail, and four or more years old⁹¹⁰. The independent bias studies rest on data from 2000–2018 and three metropolitan areas or one lender dataset⁵⁸. No controlled comparison of AVM against valuer outcomes, over time or by price band, was found. And much of the European use depends on valuer sign-off, inspections, caps on loan size or monitoring-only roles, so gains rest on process as much as on the model³. Searches for withdrawals and legal action found no other withdrawn programme and no enforcement action specific to AVM accuracy in the sources opened.

Phase: In early use — why

The row's hypothesis was In wide use. AVMs of all kinds are widely used: 14 of 16 surveyed European countries report them in banking valuation, and US lenders are now covered by a dedicated rule³¹¹. But only confirmed machine learning counts towards the phase. The one confirmed deployment is the Zestimate, a consumer-facing estimate (vendor-reported)²; Opendoor's text read confirms algorithms but not machine learning⁴; and the survey and the rule do not separate machine-learning models from older ones³¹¹. The share of machine learning in lender use cannot be established, so the lower plausible phase is chosen. This is a judgement about machine-learning AVMs only.

What to watch next

• Supervisory use of the US rule, which took effect on 1 October 2025: any agency guidance or enforcement that cites it¹¹.
• Publication of current regional Zestimate accuracy statistics, which the Zillow page says exist but which did not load, and any independent test of them².
• Changes to national limits on AVMs as the primary value at origination, currently forbidden in Spain and not permitted on a standalone basis in Ireland³.
• Opendoor's annual report for FY2026, filed after 31 December 2026, for whether offers remain algorithmic with human review⁴.

Questions to ask before adopting

• Which part of the estimate is machine learning and which is an older comparables or index method, and is that documented?
• What are the median and worst-decile errors on our own property types, price bands and neighbourhoods, measured against recent sales?
• Is percentage error larger for lower-priced homes or for any protected group in our portfolio, and who tests that?
• Where does a human valuer sign off, and what triggers an inspection instead of an estimate?
• If the model sets a price we pay rather than checks one, who bears the loss when its estimate is wrong?