Explainability
Which signals actually drove this location's forecast?
A forecast you can't explain is a forecast you can't defend to a board. For any location, IRIS shows which of its more than 100 signals mattered most for that specific site and how they combined, so the score comes with its reasons attached. Footfall might be the top driver in one location and matter only because it interacts with competitor density in another. Seeing that is the difference between a number and an argument.
What is feature importance?
Feature importance is a measure of how much each input signal changed a model's prediction. In a store revenue model, the inputs are things like purchasing power, foot traffic, competitor density, accessibility and your own nearby performance. Feature importance ranks them for a given forecast: which signals pushed the estimate up, which pulled it down, and by roughly how much. It turns "the model says €1.20M" (an illustrative figure, not a measurement) into "the model says €1.20M, mostly because footfall and catchment purchasing power are strong here, despite dense competition."
Why aren't the top signals the same everywhere?
Because locations are not the same. A high street site and a retail park site can reach the same forecast for completely different reasons. This is where interaction effects matter: signals rarely act alone. Strong footfall in a low-income catchment is worth less than the same footfall where people spend more; a competitor next door hurts a convenience format more than a destination one. A model that only reported a single global ranking of signals would miss all of this. Explaining a forecast well means explaining it per location, including how the signals interact.
| A black-box score | An explained forecast |
|---|---|
| "The model says €1.20M" | "€1.20M, driven by footfall and purchasing power" (illustrative figures, not measurements) |
| Same explanation everywhere, or none | Reasons specific to this location |
| Signals treated as if independent | Interactions between signals made visible |
| Hard to challenge or sanity-check | A local expert can agree or push back |
How does this change the decision?
An explained forecast can be argued with, and that is a feature. When you can see the drivers, a regional manager who knows the street can say "that footfall number is inflated by a station that isn't really our customer," and you can weigh that against the model. The forecast stops being an oracle and becomes a well-documented starting point for a decision the people involved actually own.
Making machine-learning models explainable rather than opaque is an active area of peer-reviewed research, and it shapes how IRIS is built. For how explainability fits alongside validation and the modelling approach, see the methodology page; for the research grounding, see the science IRIS is built on.