How AVM accuracy should be measured.
Sold-price benchmarks versus surveyor agreement - and why one headline number tells you least about the cases that matter.
Accuracy is whether the prediction matched the sold price - read per segment and per confidence band, never as a single trophy number.
The benchmark: sold prices.
The honest test of an AVM is whether its predictions matched what properties actually sold for. Sold prices are the market’s own answer - harder to argue with than a surveyor’s opinion and exactly the outcome a funding line ultimately cares about. Testing against completions, over time, is the benchmark that means something.
Why agreement with a surveyor is the wrong test.
Measuring an AVM by how closely it matches a surveyor’s figure tests agreement, not accuracy. Two estimates can agree and both be wrong. A sold-price benchmark asks the only question that settles it: did the number predict the market?
Why a single headline number misleads.
One accuracy figure hides the detail. A model can look strong overall and remain weak on flats, high-value homes or rarely traded estates. Read accuracy by segment and confidence band because lender routing runs on those bands.
Test it in the open.
The strongest evidence is a dated public record: predictions frozen ahead of a named release, tested against completions, segmented and archived with limitations. Pretira has published the required certificate structure, but its first public result is still pending.
Inspect the evidence status.
The Observatory sets out the required certificate structure and current publication status. It does not present a pending result as completed evidence.