How accurate are PropDetect refurbishment estimates?

Every accuracy figure needs a known answer to compare against. For sale prices that answer is public and complete: every completed transaction in England and Wales is registered, and we test 126,264 of them. For rents it is imperfect but available: hundreds of thousands of listings advertise a monthly figure.

For refurbishment there is no register, no public dataset, and no listing. The true answer exists on exactly one document, the builder's invoice, and it sits in the buyer's inbox. It is not that this test is hard. It is that the data does not exist to run it.

That leaves two options. Publish a plausible-sounding number nobody can check, which is what we would have to do to match the confidence other tools project. Or say there is no measured figure, explain the method in enough detail that it can be judged on its merits, and go and collect the ground truth. We have taken the second.

The estimate is built room by room from the listing photographs, not from a percentage of the purchase price. Each room is identified, assessed against a modern lettable standard, and priced by element: the kitchen units, the bathroom suite, the flooring, the decoration, the windows. Costs are regional rates applied to the measured or inferred floor area of that specific room.

This matters because the common alternative, a rule of thumb like ten or fifteen percent of purchase price, is wrong in a specific and predictable way. A kitchen costs what a kitchen costs whether the house is £90,000 or £400,000. Percentage-of-price rules overstate the job badly on cheap stock and understate it dangerously on expensive stock, which is exactly backwards from what a sourcer needs.

A 15% contingency is added on top of every estimate, and overheads are itemised rather than absorbed. Where a room cannot be assessed, because no photograph shows it, the report says so instead of pricing it at zero and quietly flattering the deal.

On three deals, our customer Warwick Sleigh had builders price the work from the same photographs the software used, and reports our figure landed within £500 across the set. He has given permission to be named. That is one customer on three properties, offered as an anecdote with a name attached rather than as a measurement, and it should be weighed accordingly.

We also know which direction the method errs when it errs, because we tested it. In September 2026 we compared our production model against a cheaper alternative on 24 marginal rooms and had the disagreements adjudicated by hand. Every disagreement ran the same way: the cheaper model called less work than the production model. On the rooms where a human judged, the production model was right three times out of four.

We kept the more expensive model as a direct result. An overstated refurbishment makes a deal look worse than it is, which costs you an opportunity. An understated one inflates the profit, which is how somebody buys a deal they should have walked away from. Given a choice between those two errors, we would rather make the first.

Since September 2026 the software asks. A few months after an analysis, when a refurbishment would plausibly be finished, we email the person who ran it and ask what the work actually cost. The form shows our estimate next to the box for the real figure, because people correct a number far more readily than they volunteer one.

Every submission goes into a calibration set. A wrong estimate is more useful to us than a right one, and the email says so.

When that set is large enough to measure honestly, the figure will be published here with its method and its sample size, in the same format as the valuation and rent back-tests, and whether or not it flatters the product. Until then this page will continue to say that no measured figure exists.