Prediction intervals
Why won't IRIS just give me one number?
IRIS does give you a number. It gives you the forecast as a value and its interval together, for example "expected turnover €1.20M, 80% likely between €1.0M and €1.4M," never the value on its own. A lone figure looks more certain than any forecast of a location that has not opened yet can honestly be. The interval is the point: it tells you how much room there is for the site to disappoint, which is exactly what you need to size the risk before you sign a lease.
What is a prediction interval, in plain terms?
A prediction interval is the range a forecast is likely to fall within, together with how likely that is. "€1.0M to €1.4M at 80%" means that, given everything the model knows, four times out of five the real result should land inside that band. It is a statement about the model's uncertainty, not a promise. A wide band says "this location is hard to call"; a narrow band says "the model has seen enough similar sites to be fairly sure."
Why is a single number misleading?
Every revenue forecast is an estimate built from incomplete information: you do not know next year's footfall, the competitor who might open across the street, or how your own offer will land in a new catchment. Collapsing all of that into one figure throws the uncertainty away without removing it. The risk is still there; you just can't see it. Two sites can share the same expected turnover of €1.20M while one is a safe bet and the other is a coin flip, and only the interval tells them apart.
| A single number | A value with its interval |
|---|---|
| Looks precise, hides the risk | Shows precision honestly |
| Two very different sites look identical | Safe bets and coin flips are distinguishable |
| Hard to defend when a site underperforms | The range was on the table from day one |
| Invites false confidence | Sizes the downside before you commit |
How should I use the interval to make a decision?
Read the whole range, not just the midpoint. A site whose worst-case still clears your hurdle rate is a different decision from one that only works at the top of its band. Compare candidates on their downside as well as their expected value. And when the band is wide, treat that as a signal to gather more data or lean on local judgement, rather than as a reason to distrust the tool. A model that tells you when it is unsure is more useful than one that never admits it.
This is a deliberate design choice, grounded in peer-reviewed research on algorithmic accountability and how organisations use quantified estimates. You can read more about the approach on the methodology page and about the science behind IRIS.