What is an AVM?
The model never walked through the kitchen. Show a lender a single number and the first question back is how sure you are.
Automated valuation model
AVM
An automated valuation model, or AVM, estimates what a property is worth from data alone, with nobody visiting it. It reads recent nearby sales, the property's own details and how the market is moving. Lenders, investors and listing sites use one to price at volume, where a full appraisal is too slow.
An appraiser walks the property and forms a judgement. An AVM reads records. It knows the floor area, the bedroom count, the sale history and what nearby homes fetched. It knows none of what a person notices in the first ten seconds inside.
That gap is not a flaw to be engineered away. It is what an AVM is. The models work well where the housing stock is uniform and sales are frequent, and they get shaky exactly where a buyer most wants help.
The output shape decides whether anyone can use it
Ship a confidence range and a user can act on it. A wide range on a thin sub-market tells an analyst to go and look, which is the correct instruction. A tight range tells them to trust the number and move on.
Ship a single number and you have removed the only information that told them which situation they were in. Every estimate then looks equally confident, including the ones built from four stale comps in a market with no recent trades. Users learn that the hard way, and then they stop trusting all of it.
Show the comps too. A valuation a user can inspect survives a challenge from a lender or an investment committee. One that cannot be inspected loses the first argument it has.
- 01Condition and renovation are invisible in public data, and they move the price most.
- 02Thin sub-markets give the model too few recent trades to learn from.
- 03Concessions hidden from listing records make the recorded price look higher than the deal was.
- 04Models drift as markets move, so an AVM needs re-checking rather than shipping once.
- DataRecords, sales, characteristics.
- CompsCandidates, ranked and shown.
- EstimateA range, not a point.
- EvidenceWhich comps, and why.
- ReviewDrift checked as markets move.
The middle three are the product. A model that returns only the estimate has hidden the two things a lender and an investment committee will both ask for.
Related questions
01How is an AVM different from an appraisal?
An appraisal is a licensed professional's opinion after inspecting the property, and it carries their name. An AVM is a statistical estimate from records, produced in seconds and at volume. Lenders treat them differently for exactly that reason, and one does not substitute for the other in a transaction that requires the other.
02How accurate is an AVM?
Accuracy varies so much by market that a single figure would mislead. It turns on how many recent nearby sales exist, how alike the local homes are, and whether the property has been changed in ways no record shows. That spread is the argument for shipping a range with every estimate.
03Can we train an AVM on MLS listing data?
Check the licence first, because it often says no. MLS data is licensed for set uses, and display is the usual one. Building a valuation product on top of it tends to fall outside those terms. Access is agreed one MLS at a time, and there are 489 of them in the United States.
04Should an LLM be involved in a valuation?
Use it for extraction, never for arithmetic. Pulling figures out of an offering memorandum or a rent roll into typed fields is a real job for a model. Each number it finds should link back to the page it came from. Do the sums in code, so the same inputs always give the same answer.
Related reading
- Real estate software development →What we build in property, and the two systems running today.
- Fair housing rules for AI tenant screening →The other model in a property stack that decides something about a person.
- Is algorithmic rent pricing legal? →What the RealPage decree changed about which data may set a price.
- SmartREI →Investment analysis for US property buyers.
