How our valuations work
Published, not promised. Every number below is measured, and we update this page whenever the model is retrained.
The data
The model is trained on official public records only: every recorded residential sale from HM Land Registry since 2015, matched to floor areas from the Energy Performance Certificate register, and enriched with area statistics (income, deprivation, transport distance and more) from the ONS and other official sources. No asking prices, no agent estimates, sold prices only.
The model
A gradient-boosted tree model predicts the price a property would sell for today from its size, type, tenure, location and the area's current price level. Alongside the central estimate it produces a low and high scenario, and we convert its measured error into the confidence range you see on every valuation.
How we measure accuracy, honestly
We test the model the hard way: it is trained only on sales up to a cut-off date, then asked to predict 505,000 real sales that happened after that date, sales it has never seen. Half of those predictions land within 10.3% of the actual sold price (the “median error”). That is the number we publish, and it is the honest one: no cherry-picked postcodes, no easy properties.
Accuracy is not uniform across the country, so we also publish it per district, 1,900+ postcode districts have their own measured error for their own mix of property types and price bands, and every valuation shows the figure for its exact segment where the sample is large enough to be honest (we suppress cells with fewer than 30 sales rather than show you a statistic built on noise).
The comparables
The comparable sales shown next to a valuation are genuine HM Land Registry transactions of the same property type in the same district, with floor area within 30% of the subject property. Each is scored out of 100 for how comparable it really is: size closeness (40 points), distance (30), recency (20) and matching tenure (10). The score and its parts are shown so you can see exactly why each sale was picked.
What this is not
BuyKnow is a statistical estimate, not a survey and not a RICS valuation. It cannot see condition, interiors, lease defects or a seller's circumstances. Use it to decide with your eyes open, then instruct a surveyor before you commit. Estimates can be wrong, and for unusual properties or thin markets the confidence range will tell you when to trust them less.