How the model works

Stability and limits

Where the model is weakest

By The Big Data Company · Reviewed July 2026 · 4 min read

Only the genuinely sure state their own limits. Here are the conditions under which an IRIS forecast should carry a wider range and lean harder on your judgement.
A vacant storefront on a European street with a 'to let' notice in the window, a candidate site whose future is not yet known.

A forecast is not much use if it lurches every time the model is updated. Expansion plans, board cases, and rent negotiations rely on numbers that hold their shape, and this is where the gravity structure earns its keep. Because the model's shape comes from how demand is known to behave rather than from whatever correlated in this year's data, it does not reinvent itself whenever a new batch arrives: small changes in the inputs produce small changes in the forecast, not wild swings. That steadiness is exactly what a data-hungry black box cannot promise on a few hundred stores.

As real openings trade and new market data arrives, we recalibrate deliberately and re-check the result out of sample, so an update has to earn its place by genuinely improving accuracy. Human judgement stays in the loop and every change is traceable. Over time the visible effect is the prediction interval tightening as evidence accumulates, not the headline number jumping around.

That same interval, attached to every forecast rather than a bare point estimate, is also how you see where the model is unsure before you act on it. Treat a forecast with more caution, and lean harder on your own judgement, wherever that range is wide or the model is working with thin data.

Only the genuinely sure state their own limits.

Where to trust an IRIS forecast least

No forecasting method is uniformly reliable, and a vendor who implies otherwise is the one to worry about. These are the conditions under which IRIS widens the range and says why, rather than papering over the gap with a confident-looking number.

If you have no existing presence nearby, IRIS can still model the area, but it has less of your own data to calibrate against. We'll tell you when a forecast leans more on general market signals than on your own performance.
A chain with 3 locations gives a model much less to learn from than one with 300. We're upfront about the sample size behind any recommendation, not just the output.
A forecast reflects the competitive landscape at the time it's run. A new competitor opening next month isn't in the model until you rerun it.
Some regions have thinner open data coverage than others. Where that's the case, we say so rather than paper over it with a confident-looking number.
The model learns from the consistency in your estate. A chain whose stores vary a lot in format, range, or customer, or that is shifting category, gives it a less stable pattern to work from, so we widen the range and say why.

Why an honest limit is a feature

None of these are defects to be quietly patched before the demo. They are properties of the problem, and naming them is the difference between a tool and a sales pitch. A store's first-year revenue does not exist yet; it depends on how a not-yet-built store trades in a not-yet-arrived year, shaped by competitors, weather, the economy, and a hundred things no dataset contains. Some uncertainty is irreducible. An honest forecast names it; it does not pretend it away.

The practical upshot is a division of labour you can rely on.

Tight range: a dense, familiar estateWide range: a new market or thin data
Lean on the numberCarry more of the decision on your own judgement
Calibrated against your own comparable storesLeans more on general market signals than on your own performance
The interval tightens as evidence accumulatesThe interval stays wide, and says why, until it does

Where the range is tight and the model is working from a dense, familiar estate, lean on the number. Where the range is wide, read that width as information: the model telling you, plainly, that it is working from little, and that this is a decision to carry more of yourself. That is the posture the whole method is built for, and it is where the series began, with why site selection was never an AI problem. The fastest way to see where the line falls for your own business is to run it on your own stores.