What the score means

The score is the model's estimate of one thing: the chance that this concept, at this address, is still open 4 years after it opens. A 62 means that of the concepts the model scores like yours, about 62 in 100 were still trading 4 years later.

That matters because a number out of 100 invites a reading that is simply wrong. There is no scale here where 100 is perfect and 50 is a pass. An average concept at an average address scores 59. That is the line to read yours against — above it the record is in your favour, below it against.

There is no composite behind it and no weighting scheme — the score is the model's calibrated probability. A hand-weighted scoring scheme was built, measured against the model, and removed for being materially worse at telling survivors from closures.

In the years held out of training, the concepts this model scored in its best 10% went on to survive 89% of the time. The ones it scored in its worst 10% survived 39% of the time. Separating those two groups is what the score is for.

Your own read carries that range explicitly — not just the estimate, but what comparable concepts actually did in the years held out of training. Where the range is wide, the honest answer is that the record does not know, and the read says so rather than projecting confidence it has not earned.

It is learned from Chicago's own licence record since 2002 — every restaurant the city ever licensed, rebuilt into attempts, and kept only once its 4-year outcome is actually observable. That is 12,142 attempts with a known ending.

1. 01Start from an average concept. Every score starts at 59 — what the model predicts for an average concept at an average address, in the current era.
2. 02Adjust for what is specific to you. Five drivers move it up or down: how this exact cuisine-and-format has fared, how many restaurants the operator has opened before, how often restaurants around the address turn over, how crowded the address already is with this cuisine, and how that cuisine specifically has fared on this block.
3. 03The adjustments are the arithmetic, not a summary of it. Each driver's contribution is measured in log-odds against that same average concept, which is what makes them add up exactly. The rows you see sum to the score — there is no hidden term and no rounding fudge beyond a single point.
4. 04Calibrate so the number means what it says. Ordering a list correctly and stating a believable probability are different jobs. The second is checked separately against what actually happened in the held-out years, and a retrain that cannot show it ships without a calibrator rather than with a flattering one.

You can check that last claim on your own read. Every read prints the anchor, each driver's signed contribution and the measurement behind it, and the total — the rows add up to the score, and where they would not, the model refuses to guess rather than quietly absorbing the difference.

Two further terms sit in the model and never appear in your breakdown: an opening-year control and a COVID control. They exist so a 2004 opening is compared fairly against a 2024 one, and at prediction time both are pinned to the same value for every score — so they carry no information about why your number differs from anyone else's, and listing them as a “reason” would be misleading. The methodology page states what that pinning assumes and how much it moves the answer.

It is not a grade out of 100.

There is no scale where 100 is perfect and 50 is a pass. 59 is average, and the whole usable range is narrower than a grade would suggest.

It is not a prediction about your business.

It describes how comparable concepts at comparable addresses have performed. Nothing in it can see your operating skill, your menu, your lease terms, or your capital.

It is not a verdict on a neighbourhood.

Location contributes very little of the score — what it mostly reads is the concept and the operator's record. The fairness audit on the methodology page measures this directly.

It is not evidence of cause.

Operator experience is the clearest example: it is the strongest single driver and it is a correlation. Opening more restaurants does not, by itself, make the next one survive.

This page describes the score. It does not try to prove it — that is a different document, written to be audited rather than read.

• Methodology — the validation, the ablations, the fairness audit, and the features that were tested and rejected. Most of the statistically significant findings there are negative results.
• A sample read — a real address, scored live against the current record, with the full arithmetic shown.
• Data sources — what the record is built from and how current each part of it is.