BRENNEN

Industries / Construction and property / Real estate / PropTech

Valuation models

This work estimates what a property is worth with software, instead of or alongside a person's appraisal, for lenders, tax assessors and the public.

7

Milestones achieved, from 2003 to 2021.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2003

    Assessors set model rules

    Assessors value every property in their area each cycle for property tax, a job called mass appraisal. In 2003 the International Association of Assessing Officers (IAAO) adopted a Standard on Automated Valuation Models, replaced by a full revision in 2018, which guides both public assessors and private model makers and tests a model by ratio studies: comparing estimated values with actual sale prices. Its preferred measure of spread is the coefficient of dispersion (COD), the average gap between each home's ratio and the median ratio, as a share of the median.

    iaao.org

    Credit

    IAAO

  2. 2014

    Instant offers priced by algorithm

    Opendoor, Offerpad and other "iBuyers" buy homes directly from owners through a website, at a price set largely by a valuation model, and resell them. A study by Greg Buchak, Gregor Matvos, Tomasz Piskorski and Amit Seru (NBER working paper 28252) found that iBuyers' share of Phoenix sales grew from about 1% at their entry in 2014 to about 6% by 2018; they paid about 3.1 percentage points less than for similar listed homes, resold about half within three months, and stayed out of areas where models price homes poorly.

    LessonA seller knows things a model cannot see, so a fast buyer that relies on a model risks buying the homes it has overvalued (see Real estate for Zillow's exit from home buying).

    nber.orgnber.org

    Credit

    Opendoor and Offerpad; Buchak, Matvos, Piskorski and Seru

  3. 2015

    Software scores every appraisal

    In 2010 the Federal Housing Finance Agency (FHFA) told Fannie Mae and Freddie Mac, which buy most US home loans, to standardize appraisal data, and from 2012 lenders had to send every appraisal through one portal. In 2015 Fannie Mae launched Collateral Underwriter, which checks each appraisal against its database, now of more than 70 million appraisals by its own count, and gives a risk score for overvaluation and errors; appraisals scoring 2.5 or less free the lender from guaranteeing the value.

    fhfaoig.govsinglefamily.fanniemae.com

    Credit

    FHFA; Fannie Mae

  4. 2017

    Loans approved without appraisals

    Fannie Mae began offering appraisal waivers on refinances in December 2016 and on purchases in August 2017; Freddie Mac's version, Automated Collateral Evaluation (ACE), built on its own valuation model, followed in June and September 2017. FHFA's inspector general judged the risks small as the programs then stood: in March 2018 Fannie Mae offered waivers on about 10% of purchase loans. By Freddie Mac's count, nearly 40% of the loans it funded in the third quarter of 2020 used ACE waivers, and its own 2024 study found ACE loans about 8.9% less likely to fall 90 days or more behind on payments than similar appraised loans.

    fhfaoig.govcapitalmarkets.freddiemac.com

    Credit

    Fannie Mae; Freddie Mac; FHFA

  5. 2019

    Outsiders beat the Zestimate

    Zillow launched the Zillow Prize on 24 May 2017, run on the Kaggle competition site, offering $1 million to a team that could beat its home value model; it said its median error was then 5%, down from 14% in 2006. In the final round, entries were scored against real sale prices. On 30 January 2019 the prize went to team ChaNJestimate, Chahhou Mohamed, Jordan Meyer and Nima Shahbazi, from Morocco, the US and Canada, whose model beat Zillow's benchmark by about 13%; Zillow said it would fold parts of their method into the Zestimate.

    zillow.mediaroom.comventurebeat.com

    Credit

    Zillow; Kaggle; the winning team

  6. 2020

    Valuation errors measured by race

    The Urban Institute compared one large data provider's automated values with sale prices from 2000 to 2018 in counties around Atlanta, Memphis and Washington, DC. Errors in dollars were similar or smaller, but as a share of price they were roughly twice as large in majority-Black neighborhoods as in majority-white ones; in Atlanta in 2009 they were 64% against 24%.

    LessonThe same dollar error weighs more on cheaper homes, which are more common in majority-Black neighborhoods.

    urban.orgurban.org

    Credit

    Michael Neal, Sarah Strochak, Linna Zhu and Caitlin Young, Urban Institute

  7. 2021

    Assessment models made public

    In 2017 the Chicago Tribune reported that the Cook County assessor systematically over-assessed low-value homes on Chicago's south and west sides and under-assessed homes on its affluent north side, and the assessor, Joseph Berrios, was voted out. The office had valued homes with a separate linear regression for each township, in the statistics program SPSS, from 2009 to 2017. The new administration published its model code from 2019, and since 2021 has used one county-wide model built with LightGBM, a method that adds up many small decision trees, each correcting the errors of the ones before; its code, data and fairness measures are public on GitHub. In 2021 Christopher Berry of the University of Chicago, using 26 million US sales from 2007 to 2016, found that within a jurisdiction homes in the cheapest tenth were assessed at twice the share of their price as homes in the dearest tenth.

    LessonWhen a model fits cheap homes worse, their owners pay more of the tax.

    law.yale.edugithub.com

    Credit

    Chicago Tribune; Cook County Assessor's Office; Christopher Berry

Watch for

  1. Savings from skipped appraisals

    Fannie Mae's second-quarter 2026 report cites $3 billion in estimated closing-cost savings for borrowers from appraisal alternatives on 5.34 million loans delivered from January 2018 to June 2026. The Appraisal Institute, a body of appraisers, replies that this is about $550 per loan in fees not paid, not a measure of net savings, and that waivers covered 28% of the two companies' loans in March 2026, against nearly 50% in 2021.

    Achieved whenFHFA, its inspector general or an independent study reports net borrower savings and the default and loss record of loans valued without an appraisal, by year and loan type.

    appraisalinstitute.org

    Claimed by

    Fannie Mae, in its second-quarter 2026 report, as summarized by the Appraisal Institute in August 2026

  2. A volunteer reassessment

    Allegheny County, Pennsylvania, which includes Pittsburgh, has gone decades without regular countywide reassessment. The group Pro-Housing Pittsburgh says a small unpaid team, using the Center for Land Economics' open-source OpenAVMKit toolkit and public data, produced assessments "of similar or higher quality" than the county's own, and estimates a regular reassessment would cost the county about $1 million a year.

    Achieved whenAn independent ratio study, on sales not used to build the model, confirms the quality, or the county adopts values made this way.

    github.com

    Claimed by

    Pro-Housing Pittsburgh, July 2026

Open

  1. 01

    Check online estimates by neighborhood

    An independent audit of free online home value estimates, such as Zillow's Zestimate and the Redfin Estimate, against recorded sale prices, broken down by the income and racial makeup of each home's neighborhood and by whether the home was listed for sale. It should also test whether the price ranges these sites show contain the sale price as often as they should.

    The brief
    Credit · unclaimed
  2. 02

    Reassess a second stale county

    An open reassessment of every home in a US county whose assessed values are years out of date, built from public records with open code, tested with the IAAO's ratio statistics on sales kept out of the model, and compared side by side with the county's current roll. The statistics are the COD, the price-related differential (PRD) and the price-related bias (PRB), which show whether cheaper homes are assessed at a higher share of their price.

    The brief
    Credit · unclaimed

Open 01 · the brief

Check online estimates by neighborhood

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An independent audit of free online home value estimates, such as Zillow's Zestimate and the Redfin Estimate, against recorded sale prices, broken down by the income and racial makeup of each home's neighborhood and by whether the home was listed for sale. It should also test whether the price ranges these sites show contain the sale price as often as they should.

Why it is still open

Zillow and Redfin publish their own error rates by metro area and state (median errors of 1.80% and 1.95% for homes on the market, 7.21% and 7.69% off it), but not by neighborhood. Outside studies are narrow: the Urban Institute's covered one unnamed model in three metro areas with sales to 2018, and a 2025 Marketing Science study of 4,027 Pittsburgh homes found the Zestimate less accurate in lower-income neighborhoods but did not look at race. A GitHub search found no public audit code.

zillow.comredfin.cominforms.org

Built with what exists

Estimates recorded from the public sites by hand, or through the companies' data programs where their terms allow, with the date each was recorded; sale prices from county recorder or assessor open data; American Community Survey tract data for income and race; Python.

  1. First step

    Pick one county that posts sale records free online, record the estimates and ranges for 200 homes just listed for sale, and wait for their sales to be recorded.

  2. Done looks like

    At least 10,000 sales in at least five metro areas, with median absolute percent error and range coverage by neighborhood income fifth and racial majority, for listed and unlisted homes, and the method and code published.

Open 02 · the brief

Reassess a second stale county

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open reassessment of every home in a US county whose assessed values are years out of date, built from public records with open code, tested with the IAAO's ratio statistics on sales kept out of the model, and compared side by side with the county's current roll. The statistics are the COD, the price-related differential (PRD) and the price-related bias (PRB), which show whether cheaper homes are assessed at a higher share of their price.

Why it is still open

Pro-Housing Pittsburgh's Allegheny County study of July 2026 is the only published volunteer reassessment with open code that this search found, and its quality claim has not been checked by an outside party. Berry found regressive assessments common across US jurisdictions, and a 2026 study by Smith, Harvey and Berry, using 26 million sales from 95% of US counties, found in simulations that adding public Census data to assessment models improved both accuracy and fairness.

github.comarxiv.org

Built with what exists

OpenAVMKit, an open-source Python library for mass appraisal; the Cook County Assessor's published LightGBM model code as a reference; the county's parcel and sale records; US Census data; the ratio statistics defined in the IAAO standards.

  1. First step

    Choose a county that publishes parcel and sale data free, run OpenAVMKit's sample-data tutorial, then load one township's records.

  2. Done looks like

    A value for every residential parcel, with COD, PRD and PRB on held-out sales for both the new values and the current roll, code and data published, and a review by an assessor not involved in the work.