Industries / Construction and property / Real estate / Appraisal
Automated valuation
Automated valuation is computer models that estimate what a property is worth from records of its features and of nearby sales, with nobody visiting, for lenders, tax offices and property websites.
5
Milestones achieved, from 1974 to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1974
A theory of hedonic prices
Sherwin Rosen, then at the University of Rochester, worked out the economics of treating a product's price as a bundle of "implicit" prices for its features, so that a house's price splits into what buyers pay for its floor space, its bathrooms or the clean air around it. Hedonic models, which value a property from its characteristics, rest on this idea.
Credit
Sherwin Rosen, Journal of Political Economy 82(1), 34–55
1975
Assessors take up computers
Tax assessors value every property in their area for property tax, a job called mass appraisal. In 1975 the Lincoln Institute of Land Policy held its first Colloquium on Computer Assisted Mass Appraisal (CAMA), when, by Richard Almy's estimate at the time, "probably no more than 400 and possibly fewer than 200 jurisdictions" used computers for it. In the early 1980s the Institute gave its SOLIR software free to assessors who took a week-long course, and by the late 1980s private firms had built it into their own products.
Credit
Lincoln Institute of Land Policy
1999
Freddie Mac blends two models
In 1999 Freddie Mac began offering "appraisal relief", using an early version of its Home Value Explorer (HVE) model to decide when a loan could skip a full appraisal. HVE blends a repeat-sales model, which works from past prices of the same home, with a hedonic model. Freddie Mac withdrew HVE from outside commercial use on 31 July 2025.
Credit
Freddie Mac; described by FHFA's Office of Inspector General (2018)
2013
Denmark suspends its valuations
Denmark's national audit office, Rigsrevisionen, compared the tax values of single-family houses sold in late 2011 with their sale prices: 41% were valued above the price and 34% more than 15% below it, although the law says a valuation should sit somewhat below the market price. The tax ministry disputed the audit's method. In November 2013 the government suspended the 2013 valuation and kept the 2011 values, and planned a new system for 2017 at DKK 96 million; in 2021 the audit office put its total budgeted cost at about DKK 3.6 billion. By June 2026 only "a small handful" of nearly 500,000 commercial properties had received an individual valuation under the new system, by the account of the Danish law firm Buus Jensen.
LessonTest mass valuations against sale prices from the start; while a replacement runs late, the tax base stays frozen at old values.
cdn.rigsrevisionen.dkcphpost.dkcdn.rigsrevisionen.dkbuusjensen.dk
Credit
Rigsrevisionen; the Danish Ministry of Taxation
2024
Testers catch springing estimates
An automated valuation model (AVM) is a program that estimates a home's value. The American Enterprise Institute's Housing Center scored five unnamed AVMs on the same 10,000 random properties in 34 US counties, using data from 2018 to 2023. One test, "springiness", measured how far an estimate jumps toward the asking or sale price once that price is public; it produced the widest spread of scores, and some vendors failed it.
LessonAn AVM can look accurate on homes for sale simply by copying the asking price, so it should be scored on values recorded before the home was listed.
Credit
AEI Housing Center, as reported by First American
Watch for
IAAO's new fairness test
The International Association of Assessing Officers (IAAO) is rewriting its 2013 Standard on Ratio Studies, which compare assessed values with sale prices. Its drafts of August 2025 and May 2026 make a new Vertical Equity Indicator (VEI) the main measure of compliance; it tests whether cheap and dear homes are assessed at the same share of their value. In July 2026 the IAAO board agreed that the VEI is "the preferred method", ahead of a third draft.
Achieved whenThe board adopts the final standard and a state agency or assessor publishes VEI results.
Claimed by
IAAO's Research and Standards Committee, August 2025 to July 2026
ATTOM's rebuilt AI model
ATTOM, a property data firm, says its rebuilt "AI-first" AVM has a median absolute percentage error of 2.9%, with more than 80% of estimates within 10% of the sale price, in its own tests on sales from the past decade that the model had not been trained on.
Achieved whenAn independent tester reports similar figures on values recorded before each home was listed.
Claimed by
ATTOM, May 2026
Testing without the list price
AVMetrics, described as "the nation's only independent AVM testing firm", has reportedly stopped letting vendors use list prices in its tests, and in 2024 was building blind testing, in which vendors would value every US property each month before any sale.
Achieved whenAVMetrics or a lender publishes accuracy by vendor on values recorded before listing.
Claimed by
Mark Sennott, in a HousingWire column, September 2025
Open
01
Test AVMs against edited records
A public test bench that measures how far a model's estimate moves when someone edits the facts a seller or broker could change, such as adding a bathroom, raising the floor area by 10% or marking the home renovated, together with a detector that flags edited records by checking them against independent public records.
The briefCredit · unclaimed02
Honest price ranges from AI
A test, on Taiwan's register of actual sale prices, of whether AI language models, and a small open model fine-tuned on one graphics chip, give price ranges that contain the sale price as often as they claim, compared with a LightGBM baseline whose ranges are set by conformal prediction, a method that sizes a range from past errors. Every model and prompt is frozen, and its hash published, before the test sales appear in the register.
The briefCredit · unclaimed
Open 01 · the brief
Test AVMs against edited records
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test bench that measures how far a model's estimate moves when someone edits the facts a seller or broker could change, such as adding a bathroom, raising the floor area by 10% or marking the home renovated, together with a detector that flags edited records by checking them against independent public records.
Why it is still open
The 2024 US rule on AVMs tells lenders to "protect against data manipulation", but it "does not provide prescriptive requirements", and the agencies "have declined to issue additional guidance". The only public step against manipulation in AVM testing found was AVMetrics dropping list prices, and this research found no published test of how AVMs respond to edited property records.
Built with what exists
The Cook County Assessor's open residential model (model-res-avm), a LightGBM model in R under the AGPL-3.0 license, and its public sales and property data; Meta's Muse Ember to pull stated facts out of listing-style text for the detector (see the Note "The jobs agents repeat").
First step
Download the latest assessment year's public data, reproduce the published model run, and save baseline estimates for a fixed random sample of 10,000 properties.
Done looks like
On at least 10,000 held-out properties and at least eight kinds of edit, the median and 90th-percentile change in estimated value and the share of edits that move it by more than 5%; for the detector, its catch rate and precision while wrongly flagging no more than 1% of unedited records; code that anyone can rerun to get the same numbers.
Open 02 · the brief
Honest price ranges from AI
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test, on Taiwan's register of actual sale prices, of whether AI language models, and a small open model fine-tuned on one graphics chip, give price ranges that contain the sale price as often as they claim, compared with a LightGBM baseline whose ranges are set by conformal prediction, a method that sizes a range from past errors. Every model and prompt is frozen, and its hash published, before the test sales appear in the register.
Why it is still open
A June 2025 study (ECML-PKDD 2025) found that language models' 90% ranges contained the true price only 1.2% to 64.0% of the time across four datasets, against 85.1% to 90.5% for LightGBM's conformal ranges. It tested only Llama 3.2 3B, Llama 3.1 70B and GPT-4o-mini, on text without images and with no Taiwan data, on sales the models might have seen.
Built with what exists
Taiwan Ministry of the Interior sale-price data, free, non-exclusive and sublicensable with attribution once the ministry's terms are accepted, with personal-data law applying; LightGBM with MAPIE, an open-source library for conformal prediction; Claude Fable 5.2 (see the Note "Three working devices"); Liquid AI's LFM3 fine-tuned on one GPU (see the Note "One GPU to fine-tune") and run locally with ggml (see the Note "A first stable number for local AI"); Gemini Decision 1 to choose among price bands with a probability for each (see the Note "A thousand questions, one pass"); each prompt run 20 times (see the Note "All twenty times").
First step
Download a year of Taipei sales, train the baseline, then freeze and publish the hashes of all models and prompts.
Done looks like
On at least 5,000 residential sales first published after the freeze date, each method's median error, share within 10% of the price, and how often its 80% and 90% ranges contain the price, with a range counted as honest if its coverage is within 3 percentage points of the stated level.