Skip to content

How every number is made — and where it fails

Most property tools tell you how good their numbers are. This page tells you how Nestraq's numbers are made — the exact method, the exact datasets — and, just as importantly, the places where the method struggles. If you only read one section, read the failure modes.

The estimate, step by step

  1. We pull the postcode's completed sales from HM Land Registry Price Paid Data — real sold prices, not asking prices. We prefer the last four years and same-type sales when at least two exist.
  2. Outliers are dropped by median absolute deviation on log(price) — a robust rule that removes the odd block transfer or commercial unit without hand-picking.
  3. The median of what remains is our estimate; the 10th–90th percentile of the same pool is the band we show around it.
  4. Older sales are time-indexed to the current month with the UK House Price Index, so a 2023 sale enters the pool at its indexed value, not its stale one.
  5. Confidence is mechanical, not vibes: enough same-type sales and a narrow band earn “high”; a thin or mixed pool is marked “medium” or “low” — and a low-confidence estimate is never published as a deal-score.

The method is deterministic end to end — the same inputs always produce the same number, and no language model ever sets or nudges a figure (why we never invent numbers). Inside the app, every revealed report carries per-figure provenance chips and a reproduce panel with the dataset link, the exact filter and this method — so you can rebuild our estimate without asking us.

Where the data comes from

  • Sold prices & price index: HM Land Registry Price Paid Data and the UK House Price Index (Open Government Licence v3.0).
  • Floor areas & energy ratings: the Energy Performance of Buildings Register — also what powers the size-normalised £/sqft read.
  • Area context: police.uk crime, Environment Agency flood zones, MHCLG deprivation indices, Ofcom broadband and mobile coverage, ONS census layers — open data throughout, attributed in the app.
  • Listings & exact addresses: a licensed property-data feed. Licensed figures are labelled as such in the app — they are never dressed up as open data.

The failure modes, stated plainly

Every valuation method has them. Ours are these — and the product is built to say so in the moment rather than paper over them:

  • Thin postcodes.A statistical estimate needs sales to stand on. In a postcode with only a handful of usable transactions, one atypical sale can drag a median a long way. When the pool is thin we cross-check against the area's small-area statistics, keep the confidence at “low” — and a low-confidence estimate never becomes a published deal-score. You'll see “score pending” instead of a made-up number.
  • Exact-address coverage is partial (~50%). Matching a live listing to its exact address doesn't always succeed — currently roughly half of listings resolve precisely. When a reveal can't resolve the exact address, it declines and you are not charged. Coverage improves as our free address-matching layers learn, but today it is a real limit.
  • First-time comparables are slow (30–50 seconds). The first pull of sold comparables for a property can take 30–50 seconds while the upstream source computes them. The report shows you the real steps as they complete — it is genuinely working, not decorating a spinner — and repeat visits are fast.
  • Coverage edges. Price Paid Data covers England and Wales; Scotland and Northern Ireland fall back to their own official indices at coarser granularity. Floor areas are missing where a home has no lodged EPC. Area screens (flood, crime, noise) are area-level indicators, not a survey of the specific building.
  • An estimate is not a valuation.Our figures are indicative statistics from public sold-price data — not a surveyor's valuation, a mortgage valuation, or financial advice. For a purchase decision you still want a survey.

How we keep ourselves honest

Estimates are pre-committed to a public, tamper-evident hash chain before homes sell, then settled against what they actually sold for — median error, share within 5% and 10%, and the worst miss included, updated as sales complete. That scoreboard lives at /receipts, and the settlement match rule is published with it. And because incentives matter more than promises: how we make money.

Contains HM Land Registry data © Crown copyright and database right, licensed under the Open Government Licence v3.0. Contains public sector information from the sources named above. Estimates are indicative statistical figures — not a survey, mortgage valuation or advice.