What Is an AVM (Automated Valuation Model)?

An AVM, or automated valuation model, is software that estimates what a property is worth using data alone, with no one visiting the house. It pulls public records, assessor files, listing data, and recent sales, selects comparable properties or fits a statistical model over them, and returns an estimated value, usually alongside some measure of how confident it is. Every online home value estimate you have seen is an AVM. So is the tool a lender runs before deciding whether a full appraisal is needed, the model an insurer uses to set replacement cost assumptions, and the engine behind an investor's ARV number. They differ enormously in data quality and method, and almost not at all in what they fundamentally are.

How does an AVM work?

Step 1: assemble a record of the subject property

The model needs to know what it is valuing: address, coordinates, living area, lot size, bed and bath counts, year built, property type, and whatever else the county recorded. This step is less trivial than it sounds. Assessor records disagree with listing records constantly, square footage is measured differently across jurisdictions, and address matching across sources is a recurring source of quiet error.

Step 2: find comparable sales

The model searches for recent sales that resemble the subject: nearby, similar in size and age, same property type, ideally inside the same submarket. Every AVM has explicit or learned rules here, and these rules do more to determine the final number than most people expect.

Step 3: adjust for the differences

No two houses are identical, so each comp gets adjusted toward the subject. Older systems do this with explicit dollar adjustments per bedroom, per square foot, per garage bay. Modern systems usually learn the adjustments statistically from thousands of prior sales.

Step 4: reconcile into one number

The adjusted comps get weighted and combined. Closer, more similar, more recent sales carry more weight. Many production systems run several distinct models and blend or route between them depending on which is historically strongest for that property type and market.

Step 5: attach a confidence measure

Good AVMs return more than a number. A confidence score, a value range, or a forecast standard deviation tells you how much dispersion the model expects around its own estimate given the comps it found.

What are the main types of AVM?

Most commercial AVMs today are hybrids. That is not a marketing distinction, it is a practical one: no single family handles a Manhattan condo, a Iowa farmhouse, and a Phoenix tract home equally well.

How is an AVM different from an appraisal?

The short version: an AVM is fast, cheap, consistent, and blind. An appraisal is slow, expensive, somewhat subjective, and sighted. They are complements, not competitors, and a vendor telling you their model replaces an appraisal is telling you something about their sales team.

Where are AVMs actually used?

• •Mortgage lending. Lenders use AVMs to decide whether a property qualifies for an appraisal waiver, to validate a submitted appraisal, and to monitor collateral over the life of a loan. US financial regulators have adopted quality control standards covering AVMs used in credit decisions, addressing accuracy testing, data integrity, protection against manipulation, random sample testing, and fair lending compliance.
• •Portfolio monitoring. Servicers and funds revalue thousands of properties periodically. Nobody is sending an appraiser to each one.
• •Insurance and property tax. Both need repeatable, defensible mass valuation.
• •Consumer sites. The estimate on a listing page is the most visible AVM in existence.
• •Investing. ARV, as-is value, and rent estimates are all AVM outputs, and they are the inputs to every offer an investor makes.

What is a confidence score or FSD?

Forecast standard deviation, usually shortened to FSD, is the industry's standard way of expressing an AVM's uncertainty on an individual estimate. Conceptually it is the model's own prediction of how far off it is likely to be on this property, expressed as a percentage. A low FSD means the model found a dense, consistent comp set. A high FSD means it is extrapolating. FSD is genuinely useful and routinely ignored. If a model tells you it is uncertain, believe it. An estimate delivered with a wide range is not a worse product than one delivered as a clean single number, it is a more honest one.

What are the real limits?

An AVM cannot see condition, because condition is not a field in any public dataset. It cannot see interior quality, deferred maintenance, or a bad floor plan. It struggles where comps are thin, where sale prices are not public record, and where a house is unusual for its street. Those limits are structural, not bugs, and they are covered in detail in why home value estimates get it wrong. What a well built AVM can do is be honest about them. Ours estimates condition from listing photos, which is a real signal but still an estimate rather than an inspection, and uses that read to separate as-is value from after-repair value. It shows the comps it used and lets you change them. Live valuations currently cover 31 US states, with rental estimates available nationwide.

How should I evaluate an AVM?

Ask three questions, in this order:

1. What was the sample? All properties in a market, or a filtered subset that happened to be easy? 2. What was it graded against? The actual recorded closing price, or another estimate? 3. Can I see the comps? If not, you cannot audit any individual number, only trust the brand. We freeze every estimate at the moment a property lists and grade it against the real closing price when the home sells, and the running results are published at /accuracy. If you want programmatic access to the same valuations, comps, and rent estimates, see the Resideline API, or run a comp set by hand with the CMA report.