The score, with the missing pieces visible.
The current field dictionary contains a property-value test score. This page publishes what that score means, what it does not prove and the stronger monthly evidence record still to be completed.
90.52/100
MAPE-derived dictionary score
A relative-error test score recorded in the current field dictionary. Higher is better. It is not a probability or a share of records that are correct.
UK field coverage: 96.58%. Northern Ireland coverage for this field is 0%. This is the current field-dictionary score, not a lender outcome or a guarantee for an individual property. Northern Ireland coverage is not available for this field.
Not yet public: cohort size, holdout date, regional and price-band error distributions, and hit rate inside the published range by confidence grade. Until those exist, this is not presented as a lender outcome or a completed Observatory month.
evidence date: 2026-07-17 · Dictionary: July 2026· monthly archive: not yet started
The test the monthly archive must pass.
Predictions frozen ahead of the release.
Before each Land Registry publication, the month’s predictions are frozen. There is no chance to mark our own homework after the completions are known.
Tested against actual completions.
When the release lands, the frozen predictions are tested against what properties actually sold for - the market benchmark, not a surveyor’s agreement.
Segmented, not headlined.
Results are broken out by region, property type, price band and confidence band, because one headline number hides exactly the segments a lender needs to see.
Archived, never deleted.
Every month’s note stays published. The archive is the point: a standing record you can audit backwards, not a single flattering snapshot.
What error metrics tell you.
An error metric summarises how far predictions sat from completions over a period and segment. It is a distribution, not a verdict - useful for comparison, misleading as a single trophy number.
Why the confidence band matters more.
The question that matters for routing is not the overall error - it is whether the high-confidence cases behaved as high confidence should. Confidence-band performance is what your thresholds actually run on.
Where the model is weak.
Thin markets, unique stock and estates that rarely trade are stated plainly. The candour is the feature: those are the cases the confidence level routes to people.
Get the first complete note when it publishes.
We will not call the archive live until the cohort, holdout and error distributions are on the page.
Why we do this.
See the running record, then test your own book.
Pretira publishes the public evidence record here. A separately scoped comparison on your completions answers the private question with the same stated method.