How we measure ourselves
Before selling a single brief we replayed the method over three years of Torbay history and followed every site it would have flagged. This page states what was done, what the numbers are, and what they cannot claim. It is written to be sent to your surveyor.
What was replayed
An event-sourced replay of Torbay, weekly from January 2022 to December 2024. At every week, each signal saw only data published on or before that date: planning decisions and later applications by their real dates, appeal decisions by decision date, sales judged against the sector's median over the preceding 24 months, energy-certificate lodgements by lodgement date, conversion precedents counted only from decisions before that week.
A parcel-week counted as strong when a change signal was active or at least two signal families agreed, the replay score was at least 0.25, and the parcel was not in Flood Zone 3. That definition was fixed before any outcome was computed.
What could not be replayed
- Ownership signals. We hold a single current snapshot of company-owned titles and company status; there is no ownership history, so no honest replay.
- Draft allocations. The Regulation 19 allocations were published in 2026.
- Appeals before 2023. The Planning Inspectorate's casework database is a rolling window; our coverage starts January 2023, so 2022 is understated.
- Geometry vintage. Parcel polygons, building footprints and terrain are a single vintage treated as constant; parcels leave the pool when their first new-dwelling energy certificate is lodged.
Matched controls
Every first-flagged parcel was followed from its flag date. Controls were parcels never flagged in the replay, matched on postcode district, plot-size quintile, undeveloped-area quintile and whether the plot was built on — three per case, measured over the same windows from the same dates.
The results — both runs
| Outcome within 24 months | First run 854 flagged · 2,560 controls | Hardened run 232 flagged · 694 controls |
|---|---|---|
| Planning application of any kind on the parcel | 18.1% vs 8.8% | 24.1% vs 8.1% |
| Application later permitted | 13.1% vs 6.7% | 19.0% vs 6.2% |
| Application or sale | 32.0% vs 18.0% | 45.3% vs 18.7% |
| Residential development application (new dwellings, conversion to flats, commercial or tourism to residential, replacement dwelling, HMO) | — | 6.5% vs 0.6% (15 of 232 vs 4 of 694) |
| Residential development application later permitted | — | 3.4% vs 0.1% (8 vs 1) |
The hardened run tightened two things after the first run was seen: controls were additionally matched on undeveloped-area quintile, and the appeal signal was limited to one trigger parcel per allowed appeal, as in the engine that runs today. We publish both runs because the second was produced after the first — a reader is entitled to know that.
The "any application" row is the headline most tools would print. We print the residential-development row next to it because that is the outcome a developer actually cares about: about one flagged site in fifteen saw a residential development application within two years, against about one control in 170. The lift is larger on that measure; the level is smaller, and the counts (15 and 4) are small enough that the exact figures will move as the outcome ledger grows.
What the score separates
By score decile, the top three deciles of flagged parcels ran at 40–55% application-or-sale against roughly 20% for their controls; the bottom decile ran below its controls. The bottom of the ranking is noise, and the weekly ten never reaches it. Per-decile counts are small (about 23 per decile in the hardened run) and should be read as directional.
Two checks a sceptic would run, which we ran
- Are the controls just worse land? A third of the control plots would never pass the engine's basic developability screen (road boundary, a building, not flood zone 3, not open space). Restricting to controls that do pass: 23.7% vs 8.7% on any application, 45.5% vs 20.2% on application-or-sale. The lift barely moves.
- Are controls reused? 694 control rows, 664 distinct parcels. Negligible.
What the numbers cannot claim
- They are not a prediction of profit, of approval, or of a site being for sale.
- Two of the strongest signals — a lapsing permission and an anomalously cheap sale — are partly self-fulfilling: a lapsing permission invites a renewal; a cheap sale precedes a refurbishment. Surfacing them before that activity is the product's job, but the hit rate should be read with that in mind.
- At n=232 the 95% interval on 24.1% is roughly ±5.5 points. The holdout below is the first replay with intervals.
- Everything above is Torbay. Other territories will be measured before their first issue.
The holdout: predicted first, then checked
On 5 September 2026 we wrote down, before running anything, what the frozen engine should do on the twenty months it had never seen (January 2025 to June 2026): 13–17% of flagged sites receiving any planning application within twelve months against 4–6% of matched controls, and a pass bar of at least eight residential-development events. Then we ran it. Flagged sites: 19.3% (11 of 57) against 6.2% (10 of 162) — a 3.1× lift, bootstrap 95% interval 1.3–8.0×. Residential-development applications: 4 flagged sites, 0 controls. Every prediction was met; but only 57 sites had a full twelve months to run, and four residential events is below the eight the bar demanded, so against our own pre-registered test the result is "partial — weaker out of sample", not a pass. We publish it as that. The pre-registration and the result are one document, unedited after the run, available on request; the next check runs when another six months of outcomes exist.
The measurement that matters more
From the second issue, every brief carries an outcomes section: what happened to each site listed in earlier issues, observed automatically from new sales, applications, decisions and energy certificates, and the running hit rate. The replay is the floor. The outcome ledger is the product.
What the engine measures, and from what
Every registered parcel in the territory: area, building footprint and undeveloped area from Land Registry polygons and Ordnance Survey buildings; boundary against mapped roads from OS Open Roads; median slope and aspect from 1 m Environment Agency LiDAR; constraint flags (flood zones, conservation areas, listed buildings within 50 m, tree preservation zones, Article 4 areas, designations within 100 m) from planning.data.gov.uk. Those measurements are joined to energy certificates (floor area, type, age, rating), Price Paid sales (street and sector £/m²), company ownership (HM Land Registry's company-owned titles and Companies House status), the full planning register since 2002 via PlanIt, and Planning Inspectorate appeal decisions. Frontage, slope and capacity are labelled as estimates wherever they appear, because they are.
On values: a development ledger of 479 completed schemes across two Devon councils since 2015 shows the sector-level premium for new-build over second-hand £/m² at roughly 1.06–1.13× depending on the comparison window — much smaller than the folk "new-build premium". Where a page gives an indicative value it says which figure it used.
Method note last revised 5 September 2026 (holdout result added the same day). Sources: HM Land Registry, Ordnance Survey, MHCLG (planning.data.gov.uk; Energy Performance of Buildings Data), Planning Inspectorate, PlanIt, Companies House, Environment Agency. See attributions.