Methodology
How accurate are PropDetect rent estimates? The back-test
Rent drives yield, cashflow, and whether a buy-to-let stacks at all. It deserves the same scrutiny as the valuation, so we tested it the same way and are publishing the result including the part where we were getting it wrong.
The headline: tested against 23,519 real rental listings, half of our estimates fall within 9.6% of the true figure, with 51.7% within 10% and 78.0% within 20%. Rent is a meaningfully easier problem than sale price and the numbers reflect that.
The more useful part of this page is the fault we found on the way, because it is a good illustration of how an accuracy figure can look perfect and hide something serious.
The method
The same leave-one-out design as the valuation back-test. For every rental listing in our comparable database, we predict its monthly rent using only the other listings around it at the same bedroom count, then compare the prediction to the real advertised figure. The listing being predicted is never in its own evidence pool.
Measured on 6 September 2026 across 23,519 predictions. Median absolute percentage error 9.6%, median bias 0.0%, with 51.7% of predictions inside 10% and 78.0% inside 20%.
The fault: a perfect-looking number hiding a serious one
Until September 2026 the rent estimate was drawn from the whole postcode pool rather than from properties with the same number of bedrooms. Re-measured, that method scores 15.1% typical error with a median bias of 0.0%.
A bias of 0.0% looks like a clean bill of health. It was not. Broken down by bedroom count, the same method carried a bias of +26.7% on one-bedroom properties and -39.9% on properties with four bedrooms or more.
In plain terms: a one-bed flat was having its rent overstated by about a quarter, and a five-bed house was having its rent understated by nearly two fifths. The overall bias read as zero only because those two errors pointed in opposite directions and cancelled in the average. Nobody owns the average. They own a one-bed or they own a five-bed, and both were being told something materially wrong.
That is why this page reports bias by segment and not only in total, and it is the single strongest argument we can make for publishing method alongside numbers. A headline accuracy figure had been sitting there looking respectable while the method underneath was structurally broken at both ends of the range.
The fix, and what it bought
Rent comparables are now matched on bedroom count before the median is taken: a three-bed is priced against three-beds. The reported comparable set is the matched set rather than the whole postcode, so the evidence shown on the report is the evidence actually used.
Measured across the same 23,519 predictions, that took typical error from 15.1% to 9.6%, and predictions landing within 10% of the truth from 36.5% to 51.7%. The per-bedroom bias that the aggregate had been hiding is gone.
This was not a tuning exercise. Nothing about the model was made cleverer. A structural mistake was removed, which is usually where the real accuracy gains are.
Why rent is more predictable than sale price
Rent estimates come out roughly 6 percentage points tighter than valuations, and the reason is a property of the market rather than of our method.
Rents cluster. A three-bed terrace on a given street rents within a fairly narrow band because tenants are comparing monthly outgoings against a local wage, and landlords are competing for the same tenants. Sale prices spread much wider on the same street: condition, extensions, garden size, chain position and the buyer's own circumstances all move the number, and a refurbished house can fetch a third more than its unmodernised neighbour.
So the same method applied to a tighter distribution produces a tighter answer. It is worth knowing which of the two figures on a report is the more reliable one, and it is the rent.
The limits, stated plainly
- The ground truth is portal ASKING rents, not achieved rents. This is weaker evidence than the Land Registry sold prices behind the valuation back-test, where the price is what somebody actually paid. A listing can let for less than it advertises, or sit unlet.
- Related, and tested rather than assumed: let-agreed listings rent 5 to 10% ABOVE asking on average across our base, because well-priced stock is what lets. We deliberately do not prefer them for that reason, as it would inflate every rent estimate.
- Thin rental markets are harder. Where the matched comparable set is small the report leads with a range rather than a single figure, and states the comparable count inline.
- This tests the market rent method. Post-refurbishment rent carries an additional uplift assumption on top, which is a separate and less certain judgement.
- Measured 6 September 2026 across 23,519 predictions. It will be re-measured as the database grows, including if the number gets worse.
Frequently asked questions
How accurate are PropDetect rent estimates?
Tested against 23,519 real UK rental listings, half of PropDetect rent predictions fall within 9.6% of the true figure, with 51.7% within 10% and 78.0% within 20%. Median bias is 0.0%. Measured 6 September 2026 by leave-one-out validation.
Why are rent estimates more accurate than valuations?
Rents cluster more tightly than sale prices. A three-bed terrace on a street rents within a narrow band because tenants compare monthly outgoings against local wages. Sale prices on the same street spread much wider, because condition, extensions and buyer circumstances all move the figure. PropDetect rent estimates run about 6 percentage points tighter than valuations as a result.
What was wrong with PropDetect rent estimates before September 2026?
Rent was drawn from the whole postcode pool rather than matched on bedroom count. That method had an overall bias of 0.0%, which looked correct, but carried +26.7% bias on one-bedroom properties and -39.9% on four-bed-plus. One-bed rents were overstated by about a quarter and large houses understated by nearly two fifths; the average looked clean only because the two errors cancelled. Matching on bedrooms fixed it and took typical error from 15.1% to 9.6%.
Does PropDetect use achieved rents or asking rents?
Advertised rents from portal listings. This is stated plainly because it is weaker evidence than the Land Registry sold prices used for valuations. PropDetect also deliberately does not prefer let-agreed listings, which rent 5 to 10% above asking on average, because using them would inflate every rent estimate.
How does PropDetect handle thin rental markets?
Where the matched comparable set is small, the report leads with a range rather than a single figure and states the comparable count inline, rather than presenting the same confidence everywhere.
Last reviewed 6 September 2026. Claims on this page trace to our internal evidence register.