Automated valuation model
An automated valuation model estimates what a property is worth using statistics rather than inspection: recorded sales, property characteristics, location, and whatever else the model has.
Every online home value estimate is one.
What it can do well#
Volume. Value every parcel in a county in seconds, consistently, on the same basis.
Consistency. No individual judgement varying between properties.
Recency. Re-run whenever new sales record.
For screening — narrowing thousands of parcels to a hundred worth examining — that is exactly the right tool.
What it cannot see#
Inside the house.
A model works from recorded sales and assessor characteristics. It does not know the furnace failed, the roof leaks, the basement floods, the kitchen was gutted and never finished, or that the previous owner did the electrical work themselves.
Two houses on the same street, same year, same square footage, will receive the same estimate while being worth very different amounts.
Where accuracy collapses#
Rural markets, where too few comparable sales exist for the model to learn from.
Unusual properties. Large acreages, non-standard construction, mixed use, anything without peers.
Heavily altered property, in either direction — a renovation the records do not reflect, or deterioration nothing recorded.
Distressed property specifically, because condition is the whole question and condition is what the model is blind to.
The honest use of one#
As a prior, not an answer.
An AVM tells you what a property of these characteristics in this location typically sells for. Whether this particular property is typical is what the model cannot address and what an inspection or an appraisal exists to determine.
The gap between an AVM estimate and an actual distressed sale price is itself informative: it is a measure of how far a specific property has diverged from the typical, which is often the thing worth knowing.
Judging one#
Any AVM should be able to state how it performs and on what sample.
Median error, on which properties, over what period, in which counties. A model that cannot say is not a model anyone should rely on.
Minnesota has a natural benchmark for this: the Department of Revenue's annual sales ratio studies apply the same test to assessors, by county and property class, on the same eCRV data. Any model claiming accuracy in this state can be measured against a standard that already exists.
Confidence matters as much as the estimate#
A number without a range is close to useless.
A model that says $240,000 is saying something very different from one that says $240,000 give or take $10,000, and different again from $240,000 give or take $60,000. The second is actionable; the third means the model does not know.
Confidence generally collapses in the same places accuracy does — rural markets, unusual properties, thin sales data. A well-built model reports that rather than hiding it behind a single figure, and any estimate quoted without one should be treated as an opening guess.