How the numbers are made
Every figure on this site comes from one test, run for every sector and every year, using only what was knowable on the day. This page describes it plainly, including what the numbers cannot claim.
The test
We rebuild the world as it looked on the 1st of January in each test year, using only information available on that day: Land Registry transactions, the property's own sale history, and Energy Performance Certificate records. The model then names its top 100 properties in each sector. Twelve months later, we count how many of those 100 actually came to market, and compare that with the sector's background rate — the share of all tracked homes that listed. The ratio is the advantage figure you see on every sector page.
The model never sees the future when it is tested. Anything lodged or registered after the snapshot date is invisible to it, the same way the future is invisible to you today. Properties already on the market at the snapshot are removed from the list before counting, because predicting a listing that already exists is not a prediction.
What counts as a hit
A hit is a marketed-sale Energy Performance Certificate lodged for the property in the twelve months after the snapshot — the certificate a seller obtains when a home is put up for sale. This record is incomplete: some genuine listings never produce one we can match. That cuts both ways, and we resolve it conservatively — the hit rates shown are floors. Sectors where the record is too thin to trust are not sold at all.
How deep the ranking runs
The top 100 is a window, not the edge of the signal. Ranking every tracked home in England & Wales on the 1st of January and counting listings over the following twelve months, the same test three years running:
| Top share of ranking | Listing rate | vs background | Share of all listings captured |
|---|---|---|---|
| 10% | 7.2–8.4% | 3.3–3.9× | 33–39% |
| 20% | 6.1–7.1% | 2.8–3.2× | 56–64% |
| 30% | 5.4–6.4% | 2.5–2.7× | 75–81% |
| 50% | 3.8–4.7% | 1.8–1.9× | 91–93% |
Read the last column with the first: three-quarters of everything that came to market sat in the top 30% of the ranking a year earlier — and reaching 92% still meant ignoring half of all homes. You receive the whole sector ranked and tiered, so you choose how far down to work.
You receive the whole sector, ranked, with each home's rank within its property type alongside — because a fixed-length list either fills with fast-turning flats or throws away homes you would have wanted. The published record is measured on the top fifth of that ranking against a random fifth of the same sector, every year. Inside that tier, terraces and semis hit at the highest rates, detached next, flats lowest; the per-type figures are published from each backtest.
What the sector page shows
Every year of the test, separately: how many tracked homes the sector had, how many of our top fifth listed against what a random fifth would have found, and the advantage. We show the worst year as prominently as the average. A sector is only listed for sale if it cleared a minimum advantage in every tested year, had enough tracked homes, and had a listing record complete enough to measure against.
Twelve months, or three
The ranking is built on a 12-month horizon because that is the question a canvassing plan asks: who in this patch is likely to move this year. Every row also carries a three-month probability from a second model trained on the same signals with a 91-day window — the same people, re-ordered by who is likely to be first. What changes at three months is the arithmetic of a short window: with roughly one home in two hundred coming to market in any quarter, even a top-ranked home is more likely than not to wait. Use the 12-month rank to decide who to work and the 3-month probability to decide who to work first. Both horizons, tested the same way — every sector, 2022–24, out of time:
| Horizon | Homes that list | Our top fifth finds | A random fifth finds | Advantage |
|---|---|---|---|---|
| 12 months | 1 in 45 | 18 | 6 | 3.0× |
| 3 months | 1 in 180 | 5 | 1.5 | 3.3× |
The ranking is as sharp at three months as at twelve — the very top of the list (the 1%) lists at five to seven times the background rate on either horizon — but the count shrinks: five hits a quarter in a typical sector's top fifth. That makes any single sector's three-month record noisy, which is why the published per-sector record stays on the 12-month horizon.
How the off-market values are made
Most automated valuations lean on the asking price when a home is listed, which makes them accurate for homes that are already for sale and silent about everyone else. Ours never sees a listing. Each property's value comes from a model fitted to every Land Registry sale since 1995 with an Energy Performance Certificate we can match: local price levels at three scales, comparable sales of the same type nearby, the street, and the property's own sale history — all from information available before the date being valued, and none of it from the property being marketed. That is why it can value the house that hasn't moved in twenty years, which is the one you are canvassing. Tested on sales in 2024, 2025 and 2026 the model had not seen:
| Slice of stock | Share | Median error | Within 10% |
|---|---|---|---|
| All properties | 100% | 8.2% | 58% |
| High-confidence | ~25% | 5.6% | 74% |
| Medium-confidence | ~55% | 8.3% | 58% |
| Low-confidence | ~20% | 14.5% | 37% |
The confidence flag on every row is a second model predicting how wrong the first one is likely to be for that property, and the table is its test: it separates 6% errors from 15% ones before the sale happens. Where confidence is low, the file leaves the expected-value column blank rather than pretend. Treat these numbers as a canvassing prior, not an appraisal — the model has no photographs and has never been inside; a surveyor has both. What it also lacks, by design, is the asking price: an off-market value has to stand without one.
What we will not claim
We do not claim to predict most of what comes to market: naming enough of a town captures most listings, and figures built that way say nothing about how sharp a list is. We do not publish return-on-spend multiples built on loose attribution. We publish the number an agent can check: of 100 named addresses, how many listed. If a competitor quotes you a bigger number, ask them for this one, for your sector, verified.
The data
HM Land Registry Price Paid Data and the Energy Performance of Buildings Register, refreshed monthly, plus probate notices from The Gazette for the flags marked urgent. All are public records. Scores are recomputed monthly and your lists refresh with them.
The fine print, in plain sight
Past advantage does not guarantee next year's. The model tracks properties with an EPC match — most of a sector, not all of it. A probability on a list is calibrated, meaning that across many homes marked 10%, about one in ten has listed; it is not a promise about any single door. Canvassing results depend on your letters and follow-up as much as the list.