How Do Automated Valuation Models Work? Inside an AI Home Valuation

How do automated valuation models work? Every portal estimate and AI home value estimator is one, and every instant cash offer and every appraisal waiver a bank grants rests on one: an AVM. The number appears in a second, which is exactly why it deserves a look inside. This article takes one apart step by step, explains the federal rule that has governed lender AVMs since October 2025, and follows the Los Angeles sample report from inputs to range. For what "market value" means before the machinery, see the complete guide to home value.

How do automated valuation models work? The four families behind every estimate

Federal law defines the term narrowly: a computerized model relied on by mortgage originators and secondary-market issuers when they determine the collateral worth of a mortgage on a consumer's principal dwelling². Everyday speech is wider and covers any software that turns an address into a value without a person walking through the door: portal estimators, lender models, AI valuation reports such as CheckValue. Behind all of them, four families do the work, and most systems blend at least two.

A bank's collateral model is typically hedonic with an index for the time step; a report like CheckValue works comparables-first and uses the index to bring older sales forward. None of them measures your house; they infer it from what the records say about houses like it.

Step 1: what data goes into an AI valuation

Four streams feed the model, and the quality of the first decides the quality of everything after it.

Public records. The assessor holds living area, lot size, year built and room counts; the recorder holds sale prices, dates and deed types. Their mistakes become the model's mistakes: a 1938 house whose permitted area differs from its usable area is valued on the permit.

Listings. MLS and portal feeds add asking prices, price cuts, days on market and photographs. Redfin, a brokerage, builds its estimate from MLS records of recently sold homes and publishes an on-market error about a quarter the size of its off-market error¹. A listing tells the model the current condition and the seller's expectation; an off-market record adds nothing beyond the county file.

Indexes. The FHFA House Price Index comes out for the nation, for states and census divisions, for metro areas and counties, and down to ZIP codes and census tracts³: the ruler for how far the market moved between a comparable's closing date and today.

Your own facts and photos. Condition, renovations, the new roof: none of this is in a public record. A model that takes these facts and reads your photos can adjust for what the record omits. The official data sources for home values are covered in a separate guide.

Step 2: how comparables are chosen and adjusted

The standard a good model mirrors was written for appraisers. Fannie Mae's Selling Guide expects comparables to have closed within the last 12 months, to come from the subject's market area where possible, and to resemble the subject in site, style, size, room count and condition, with the remaining differences adjusted for; at least three closed sales are required⁵.

A model turns those expectations into a scoring rule. Every recorded sale within a radius earns points for proximity, recency and similarity in size, age and type; the top-scoring sales become the comparables. Then come the adjustments, and here the families meet: the coefficients that price each difference (dollars per square foot of living area, a fixed sum for a garage bay, a percentage for a pool or an energy class) are estimated from local sales by the hedonic family and applied by the comparables family. Two design choices separate a careful model from a careless one: coefficients must be local, because a pool is worth a different percentage in Phoenix than in Portland, and adjustments should be printed with their source, because an adjusted price you cannot inspect is an assertion, not evidence.

Step 3: how older sales are brought up to today's market

A comparable that closed in November 2024 tells you what a buyer paid in November 2024. To use it for a value dated September 2026, the model multiplies its price by the ratio of the index today to the index at the closing date, ideally at ZIP-code or metro level. FHFA's HPI is a weighted repeat-sales index, so it tracks pure price movement rather than a shift in the mix of homes sold³.

Two problems remain. The index arrives with a lag: the monthly release scheduled for September 29, 2026 carries data through July 2026³, so the last two months must be extrapolated or filled from listing data. And an index is an average: a metro up 3 percent does not lift every street by 3 percent. Market tempo (days on market, the list-to-sale ratio, months of supply) is the second signal, and a slowing market widens the range even while the index still points up.

Step 4: consistency checks, or why the same house should not get two values

Every serious valuation engine runs tests before it shows a number.

• Plausibility. The value per square foot must fall inside the range observed in the district; land and building must add up to roughly the total; the rent estimate must imply a yield that exists in that market.
• Range. Its width follows from the dispersion of the adjusted comparables: three sales that agree within 3 percent justify a narrow range, three that disagree by 15 percent do not.
• Reproducibility. One address with identical facts must produce one value, which means a registry of results and two independent runs that must agree. Redfin's estimate refreshes daily for listed homes and weekly for the rest¹, so an unchanged house gets a new number every week.

A confidence score, where a model offers one, summarizes these checks: how well the evidence agrees, not how right the price is.

What no model can see

Three things sit outside every dataset. Interior condition: the roof, the plumbing, the 1990s kitchen behind fresh paint. Listing photos help while a house is on the market and owner-supplied photos help when it is not; this gap is the main reason the off-market error is several times the on-market error¹. Legal and title issues: easements, liens, unpermitted additions, a rent-controlled tenant with three years left on the lease. The buyer in the room: two bidders who grew up on the street, a relocation deadline, a cash offer in the week rates jumped. A model estimates the most probable price; whoever shows up sets the actual one. Redfin's own page calls its estimate a starting point, not an appraisal¹. The comparison of online home value estimators puts that in dollars.

Regulation: the 2025 AVM rule and where appraisals are still required

The interagency final rule published on August 7, 2024 (89 FR 64538) took effect on October 1, 2025. Lenders and securitizers that use AVMs in credit decisions for a consumer's principal dwelling must ensure the models produce a high level of confidence in their estimates, protect against data manipulation, avoid conflicts of interest, undergo random sample testing and review, and comply with nondiscrimination laws⁶. The statute lists the first four factors and lets the agencies add others²; nondiscrimination is the addition.

Three consequences. The rule governs lender models only: a portal estimate or an AI valuation report you buy yourself is outside its scope, because no lender uses it to decide your loan⁶. Models already stand in for appraisers in routine lending: Fannie Mae's value acceptance lets an eligible loan close without a new appraisal when a prior appraisal of the property sits in its Collateral Underwriter data and its underwriting system accepts the lender's value estimate⁷. And the human requirement remains: a federally related residential transaction above $400,000 still needs a state-licensed or certified appraiser unless another exemption applies, an evaluation suffices at or below that line, and commercial transactions carry a $500,000 threshold⁸. Whichever valuation the lender relies on, Regulation B entitles the applicant to receive each appraisal and any other written valuation behind a first-lien dwelling loan promptly once it is complete, and no later than three business days before closing⁹. Legal weight and price of each document: appraisal vs. CMA vs. AVM. This is general information, not legal or tax advice.

A 1938 Los Feliz house taken through the four steps to a range

The figures below come from the CheckValue sample report for Los Angeles, a published example rather than a client's property; dollar-per-square-foot values are converted from the report's per-square-meter figures.

Now the checks from step 4. The per-square-foot range sits inside the district band. Land plus building gives $1,750,000 to $1,980,000, which overlaps the stated range. The rent estimate at its midpoint implies a gross yield of about 4.5 percent, a level that exists in prime Los Angeles neighborhoods. All three pass. The report also lists what would sharpen the estimate (scope and year of the modernization, permitted against usable area, foundation type, heating type, the pool) and flags a wildfire-hazard disclosure near Griffith Park: the honest part of any model, what it did not see. Every page of this example is on the sample reports page.

How the CheckValue report applies the four steps and prints its evidence

CheckValue is an AI property valuation report, so everything above applies to it, and this section is meant to be read against the four steps. For US addresses the facts come from the county record and from the official series the report cites by name: FHFA for the index, the Census Bureau for neighborhood context, FEMA for flood zones. You confirm or correct the recorded facts (type, size, lot, rooms, age, condition) and can add photos, which the model reads for condition, plot and surroundings alongside the official aerial orthophoto of the parcel. After a free on-screen preview, the calculation runs for about 60 to 90 seconds with each step shown.

What separates it from a black-box estimate is what it prints. Every adjustment, whether for a pool, solar panels, a garage, an elevator, the energy class, the condition or the year built, appears with its coefficient and the source that coefficient rests on, and the official sources are numbered in a reference list. Enter one address twice with the same facts and you get one value, because each result is written to a registry, computed twice independently and passed through the plausibility tests before release. The rest of the report follows the same discipline: comparable sales with distance and date, the index and tempo behind the time step, a rent estimate and the gross yield it implies, the cost side (owning, selling, net proceeds), maps for flood, energy and noise, a 1 to 10 crime score derived from FBI county statistics, and, for a US address, the county's owner of record. That last item carries the Fair Credit Reporting Act notice: it may not be used for credit, employment, insurance or tenant-screening decisions.

What it is not: a licensed appraisal, a model inside the scope of the 2025 lender rule, or a claim of accuracy expressed as a percentage. No one walks through the property, and when a lender, a probate court or the IRS insists on a signed appraisal, the report prepares you for that conversation without replacing the document. The accuracy and method page describes the data behind each market; the free preview of your own report shows the layout for any address before you pay.

Five questions to ask of any AI valuation

1. Which sales is this built on? Without distance and date for each comparable, you are looking at an assertion.
2. What was adjusted, by how much, from what source? A coefficient without a source is a guess wearing a number.
3. What is the date of value, and how were older sales moved to it? Look for a named index at metro or ZIP level.
4. How wide is the range, and why? A narrow range on few comparables is a warning, not a comfort.
5. Would identical inputs give the same answer tomorrow? A number that moves weekly while the house does not is tracking the market's mood.

Add a sixth for yourself: what did the model not see? That is where the sale price will part company with the estimate.

This is general information, not legal or tax advice.

This article is general information, not legal, tax or investment advice. Figures and rules carry the year they were published; check the cited source for the current version.

Christian Eckmair

Co-founder, CheckValue · More than 20 years in real estate in Tenerife and Austria · Reviewed by Josef Eckmair MBA

Christian Eckmair is co-founder of CheckValue and of CBDNOL GmbH (Ansfelden, Austria). He has worked in real estate for more than 20 years in Tenerife and Austria – buying, renovating, letting and selling residential property – and writes about home values, buying checks and ownership records.

Articles by Christian Eckmair →