How accurate are PropDetect valuations? The back-test, in full

The test uses properties that have already sold, where the true answer is known. For every sold comparable in our database, we predict its price using only the other comparables around it, then compare the prediction to what it actually fetched. The property being predicted is never included in its own evidence pool, which is the single most common way a back-test flatters itself.

Prediction is the core method: the median price per square foot of the matched comparable set, applied to the subject floor area. Comparables are matched on property type, so a terrace is predicted from terraces. Every input is an HM Land Registry transaction supplied through PropertyData: a price at which a property actually changed hands, never an asking price and never a portal estimate.

Measured on 6 September 2026 across 126,264 predictions spanning 2,644 property pools. The query that produces these numbers is a single SQL statement, kept alongside the results so it can be re-run rather than taken on trust.

Median absolute percentage error, the standard measure, is 15.3%. That means half of predictions land within 15.3% of the true sold price and half do not. 34.8% land within 10%, and 61.1% within 20%.

Median bias is 0.0%. Predictions are as likely to fall below the true price as above it. That is the number we consider the important one.

For context on what 15.3% means in money: on a £150,000 property, half of estimates fall within roughly £23,000 either way. This is a screening tool, not a mortgage valuation, and the next section explains why that distinction changes which figure you should care about.

A sourcer does not analyse one property. They analyse two hundred and pursue the handful that stand out. For that job, the two kinds of error behave completely differently.

A method that is wrong in both directions is survivable at volume. Run enough deals and the errors offset; the genuinely strong deals still rank above the weak ones, because nothing is being pushed in a consistent direction.

A method that is wrong in ONE direction is not survivable. If every valuation runs high, every deal looks better than it is, every margin is overstated, and the errors compound rather than cancel. You do not notice, because the tool is consistent. You simply buy something you should not have.

We had the second kind. Before September 2026 the method concluded at the 72nd percentile of the comparable set rather than the median. Re-measured on the same 126,264 predictions, that anchor carries a median bias of +15.0%: valuations ran systematically high by about a seventh, on every property, in the same direction, every time.

The choice of where to conclude within the comparable set was tested directly rather than argued. Each percentile was run through the identical leave-one-out procedure on the identical pool.

At the 50th percentile, the median: 17.2% typical error, 0.0% bias. At the 65th: 18.4% error, +9.4% bias. At the 72nd, the old anchor: 20.3% error, +15.0% bias. At the 80th: 24.6% error, +22.5% bias. (These figures are for the unfiltered pool; the 15.3% headline is the median with type matching applied, which is what ships.)

The pattern is perfectly monotonic. Every step up the percentile ladder makes the method both less accurate and more biased. There is no trade-off to balance and no judgement call to defend: the median is simply better on both measures at once, and everything above it was costing accuracy to buy optimism.

Zillow, the largest automated valuation model in the world, publishes two accuracy figures for the Zestimate: a median error of 1.83% for homes that are ON the market, and 7.01% for homes that are not. Same model, same homes, same country. The only material difference is whether the model can see the current listing price.

That single comparison, from the industry leader's own published data, is the most useful thing on this page. It shows that being shown the asking price improves a valuation model by a factor of nearly four. Any accuracy figure quoted without saying which side of that line it sits on is close to meaningless.

Our back-test sits firmly on the harder side. It predicts what a property sold for using only what other properties sold for. No asking price enters the calculation at any point, for the subject or for any comparable. We do not use portal estimates, and we exclude listings marked sold subject to contract, because neither is evidence of what a buyer actually paid.

We are not claiming to beat Zillow. They operate at a scale and with a depth of public property data that does not exist in the UK, where there is no equivalent open record of a home's attributes. At 15.3% we are roughly twice their off-market error. What we will claim is narrower and, we think, more useful: we publish the figure, the method, the sample size and the failure cases, and we can find no UK automated valuation model that publishes an independently verifiable accuracy figure with its method attached at all.

A single average across every property hides the most useful thing a user could know: when to trust the number in front of them. So the same 126,264 predictions were split by how tight the comparable evidence was, measured by the spread of the comparable set itself.

Where the comparables agree closely with each other, covering 61,821 predictions or roughly half the sample, typical error is 12.2% and 70.6% of estimates land within 20% of the truth. Where they disagree moderately, error rises to 18.3%. Where the local evidence is genuinely scattered, it reaches 24.8% and then 27.9% at the extreme.

That gradient is the result we would most want to see. It means the method is not confidently wrong in difficult markets: it is accurate where the evidence supports accuracy and it degrades where the evidence degrades, in proportion. A tool whose error stayed flat as the evidence thinned would be a tool that had stopped listening to its inputs.

It is also directly actionable, because the report already tells you which situation you are in. Comparable counts, the spread, and a stated confidence level appear on every analysis. When the report says the evidence is thin, this is the number behind that warning, and it is why the report leads with a range there rather than a single figure.

Median bias stays within a point of zero in every band. The method does not start leaning in one direction when the evidence gets hard, which is the failure mode that would matter most.

A single headline number can hide a great deal, so the same test was run separately across four cohorts of properties analysed at different times, covering different regions and sourcing patterns.

Typical error moves with the mix: 12.9% on the earliest cohort, 15.4%, 17.6%, and 17.1% on the most recent. That spread is a property of the stock rather than of the method. More recent analyses lean towards automated sourcing across a wider and less uniform set of markets, which is genuinely harder to value than a hand-picked list.

Median bias was 0.0% in every single cohort. That is the result we would most want to be stable, and it is: whatever the region, whatever the price band, whatever the sourcing route, the method is as likely to come in under as over.

It is worth being explicit that this cuts against us as well as for us. Had we published only the most recent cohort, the honest headline would have been 17.1% rather than 15.3%. The figure on this page is the full base, which is the larger sample and, as it happens, not the flattering one.