How it's built
The gravity model behind the number
Strip away the branding and most location tools reduce to a single instruction: trust this number. IRIS is built to answer the follow-up question instead, the one a board always asks. Why that number? The forecast at its core is a spatial-interaction model: a transparent, theory-driven formula. It estimates how much of an area's spending a location can win. It weighs how attractive the site is against how easily each pocket of nearby demand can reach it, and against every competitor doing the same. It is not an AI that hands down a verdict. It is a mechanism you can trace.
Before IRIS looks at a single data point, it starts from something already known to be true. More than a century of work in economic geography shows that trade follows people and access: a location earns more when more people can reach it easily and are inclined to spend once they do. That is not a hunch to be discovered in this year's data. It is one of the most tested regularities in the study of how cities and retail work, and it is the backbone every IRIS forecast is built on.
Two ideas do most of the work
The structure rests on two well-established mechanisms, and naming them is part of the honesty.
- Spatial interaction. People trade off how attractive a place is against how much effort it takes to get there. Model that across every competing destination and you can estimate how demand splits between them.
- Time geography. People move on real routes and daily rhythms, not in straight lines from home. Where they actually pass, work, and dwell shapes who a location can realistically capture.
A crucial detail is how the signals combine. They multiply, they do not add up a checklist. Strong footfall counts for more where purchasing power is also high, and for less where a competitor sits between the site and its customers. A checklist treats every factor as independent; the real world does not, and the multiplicative form is what lets one signal amplify or dampen another the way it does in practice.
What the shape buys you
This is not the only way to produce a number. It is the way that survives being questioned. Set the auditable structure against the alternative most vendors ship, and the difference is not cosmetic.
| A proprietary black-box score | The IRIS spatial-interaction model |
|---|---|
| A number with no reasoning attached; you can believe it or not | Every part maps to something real, demand, attractiveness, distance, competition, so it can be read and checked |
| A 'what if' means commissioning a new run and waiting | Add a competitor or open a store and recompute who captures what; scenarios are the same model run again |
| A town centre treated as one blob | Applied at fine-grained hex resolution, with the catchment a choice you can change |
| Needs far more stores than you have to learn geography, so it overfits | Already knows the geography from tested theory; spends your data sharpening the fit |
The point about "what if" is worth dwelling on. Because the model is defined over competing destinations, cannibalisation, impact analysis, and scenario planning are not separate features bolted onto a scoring engine. They are the same model run again with one more store on the map. That is only possible because the mechanism is explicit rather than hidden.
A checkable, named model is the honest alternative to a proprietary score you are asked to take on faith.
Why name the method at all
Most vendors keep the method vague on purpose; a named model can be checked, and a checked model can be found wanting. We name it anyway. For the data-science reader, this sits in the multiplicative competitive-interaction family of spatial-interaction models, the Huff-model lineage that retail geography has used and tested since the 1960s, fitted to your estate. Machine learning is real here and it is bounded: it reads the local signals and learns the embeddings that place similar sites near each other, enriching the model without replacing the forecast with a number no one can explain.
Site selection is a thin-data problem, a chain has tens or hundreds of stores, not millions of rows, and a from-scratch model needs far more examples than that to learn geography without overfitting. An interpretable gravity model already knows the geography and spends its limited data on the fit. That trade-off is the whole argument, laid out in site selection was never an AI problem. Once you have the number, the next question is what fed it, which is what actually goes into a forecast.