Place embeddings · Step 5
How the embedding powers a forecast
By itself, an embedding tells you which locations resemble each other. That is genuinely useful, but it is not yet a decision. A forecast is a decision-grade answer, and getting from one to the other is a deliberate two-step move that IRIS keeps visibly separate, because collapsing them is where location tools tend to start overclaiming.
Step one: find the real look-alikes
Because similar places get similar vectors, IRIS can take a proposed site and line up the real, trading locations most like it. This is the honest core of the whole system: the comparison is to actual places with actual results, not to a generic template of "a good site". The ranking is by distance in the embedding space, and it is not cherry-picked, weaker matches are part of the picture, because knowing where the resemblance thins out is as informative as knowing where it holds.
The embedding finds the real places most like your site. A separate, transparent model turns how those places actually trade into your forecast.
Step two: turn trading into a forecast
The embedding does not predict revenue on its own. Once the closest real matches are found, a separate, transparent and adjustable model reads how those places actually trade and turns it into a forecast for the new site, with a prediction interval rather than a single confident number. That second model is the one you can open, question and override. It is also the only part of the system that carries a measured track record, because it is the only part tested against real openings. The similarity step supplies the comparison; the forecasting step supplies the number and its range, and each is honest about what it does.
This division is why we are careful with words like "look-alike". IRIS really does rank comparable places, and that ranking is a real capability. What we do not do is dress an illustration up as a live result: any specific set of matching cities and scores you see used to explain the idea is there to show how ranking works, not to report a measured adjacency for a real site.
| The similarity step (the embedding) | The forecasting step (downstream model) |
|---|---|
| Finds the real trading places most like your site | Turns how those places actually trade into a number |
| Ranks by distance in the embedding space | Returns a forecast with a prediction interval, not a lone number |
| Carries no measured track record of its own | The only part tested against real openings, so the only part that does |
| Learned across millions of places at once | Transparent and adjustable: you can open, question and override it |
One layer under every answer
The same embeddings feed more than the revenue forecast. A white-space map, where you have no presence but a place resembles your winners, reads from them. A catchment profile reads from them. A cannibalisation check, whether a new site would eat into an existing one, reads from them. Because every answer is built on the same picture of a place, the answers stay consistent with each other: the white-space map and the forecast are not two different opinions from two different systems, they are two views of one shared layer.
- White space: places that resemble your best sites, where you are not yet.
- Catchment: who and what the embedding sees around a location.
- Cannibalisation: how much two catchments would overlap.
- Forecast: the downstream model's number, with its range.
The embeddings sit under everything IRIS does.
They are not a screen you stare at. They are the layer IRIS reads.
The embedding does not predict revenue on its own. IRIS uses it to find your closest real-world matches, then a transparent, adjustable model turns their actual trading into your forecast, with a prediction interval.
White-space maps, catchment profiles and cannibalisation checks all read from the same embeddings, so every answer is built on the same picture of a place.
See how IRIS worksOne embedding layer, read by every IRIS answer. Illustrative.
That is the shape of it, end to end. A location becomes more than 100 signals, the signals become eight learned numbers on a fair grid, the numbers find the real places most like your site, and a transparent model turns how those places trade into a forecast you can defend. Each step does one job, and none of them claims more than it has earned.