PropertyInvestorIQ

How it works — and how often it’s wrong

Every number on PropertyInvestorIQ comes from public records and a measured engine. This page shows what goes in, how an estimate is made, and the engine’s error rates against real sales — including where it is weak. You should not have to take a valuation on faith, and here you don’t.

What goes in

HM Land Registry sold prices. 15.9 million recorded transactions across England & Wales — the price actually paid at completion, not an opinion of value. © Crown copyright.

The EPC register. 20 million energy certificates, used for floor areas and property attributes when you don’t supply them. © Crown copyright.

The HM Land Registry House Price Index. Covering 121 postcode areas back to 1968, used to bring each historical sale into today’s money before it is compared with anything.

What never goes in: asking prices. Listing prices are marketing, and building them into a valuation makes the valuation circular. Our estimates are built from sold evidence only.

How a valuation is made

  1. Find sold properties like yours, nearby. The engine looks for sales matching your property’s postcode sector, type and bedrooms. Only when evidence is thin does it widen the net — and the result tells you the level it resolved at (its evidence tier).
  2. Bring each sale into today’s money using the House Price Index for that area, so a 2023 sale competes fairly with one from last month.
  3. Weight the evidence. Recent sales and similar-sized properties count for more. A single sale is never allowed to quietly dominate the answer — when one does carry unusual weight, the result says so on its face.
  4. Price per square metre × your floor area (from the EPC register, or the figure you supply), with a calibrated adjustment when your property is larger or smaller than its comparables.
  5. Ship the uncertainty with the number. Every estimate carries its range, a confidence label, its evidence tier, and a plain-English list of the assumptions it rests on — and every comparable sale used is on screen, with its date, distance and weight.

What we don’t publish: the exact weightings, adjustment curves and guard thresholds — that calibration is the part that took the engineering, and it is ours. You don’t need it to check us: every input is a public record, and every output is measured against reality below.

The report card — measured against 51,811 real sales

Backtests are easy to rig, so here is exactly how ours works. We drew a random sample of 51,811 sales from Oct 2025 – Mar 2026 — every property with a known floor area and a standard type (house or flat), spread across every postcode area we cover. The engine valued each one using only evidence available before its sale date, blind to the outcome, and the estimate was then compared with the price actually achieved. It valued 100% of the sample — no quietly dropping the hard ones.

10.7%
median miss vs the achieved price
47%
of estimates within ±10%
−0.1%
overall bias — we don't inflate

Engine v1.2, July 2026 acceptance run on the 51,811-sale frame.

By price band

Price bandMedian missWithin ±20%Sales tested
Under £100k27.0%40%2,589
£100k – £200k13.2%66%12,357
£200k – £350k9.8%78%17,836
£350k – £600k9.3%80%12,963
£600k+10.8%76%6,066

Band detail from the v1.0 baseline run over the same sales; the current engine (v1.2) improved the weakest band to a 26.4% median miss.

Where we’re weak — on the record

Homes under £100k. Our worst segment: a 26% median miss, and the direction is over-estimation (+23%). The cheapest homes often sell below their pocket’s going rate for reasons no desk model can see — condition, distress, quick-sale pricing. Estimates there carry low confidence and loud caveats; treat them as a starting point, not an answer.

Thin evidence. When few comparable sales exist, error rises — so the confidence label falls. Roughly one estimate in five now carries the low label, deliberately: an honest “low” beats a confident miss.

The confidence label is calibrated. In the baseline run, high-confidence estimates missed by 10.2% at the median; low-confidence ones by 19.4%. When we say low, believe us.

Check us yourself

The fastest way to trust us is to not take our word. Every comparable sale behind an estimate is a public record: pick any from the evidence table and look it up in HM Land Registry’s free sold-price search. The address, date and price will match, because that is where they came from.

Rent estimates

Rents come from a statistical model over area-level lettings data, by area, property type and furnishing, adjusted for market movement since the model was trained (July 2025). The range shown is an 80% interval — calibrated so that roughly four in five actual rents fall inside it, and the page says so where the number appears. Figures exclude service charges, management fees, voids and maintenance.

We haven’t yet published a rent report card like the sales one above. Until we do, treat the range — not the midpoint — as the answer.

Refurb benchmarks

Refurbishment costs are per-square-metre benchmarks by postcode area and property type, derived from properties that sold, were refurbished, and sold again. Every estimate shows the number of observed refurb-and-resale pairs behind it, so you can see how much evidence you’re standing on. These are benchmarks, not quotes — always price from contractor quotes before exchange.

Challenge any number

Every estimate in the app carries a “Does this look right?” control. If a figure looks off, say so and tell us what you’d expect — disputed figures go into the evidence set we test the engine against. If we’re wrong, the fastest way to make us less wrong is to tell us.

Engine versions

  • v1.2 — July 2026. Evidence-hierarchy guards: broad-area estimates are now checked against same-street and same-unit history; the weakest price band improved; low-confidence labelling widened to cover every thin-evidence case.
  • v1.1 — July 2026. Guards against local pocket contamination — one anomalous sale can no longer drag a street’s estimates.
  • v1.0 — July 2026. Engine build and the first full 51,811-sale backtest, which set the error figures published above.

This page is updated with each engine release, and every figure on it names the run it came from.

What this is not

PropertyInvestorIQ produces automated comparable-method estimates for research and triage. It is not a RICS valuation or survey, and it cannot see condition, lease terms, or anything else that only a viewing reveals. Before money moves, verify with a surveyor, letting agent and your own eyes.

Contains HM Land Registry data © Crown copyright and database right. Energy Performance Certificate data © Crown copyright.
← Back to PropertyInvestorIQ