The Chicago Restaurant Report

A restaurant closes long before the paperwork does. We measured the lag on a cohort of 4,292 attempts from 2010–2016, where the outcomes are now fully known, by asking how many of those closures were visible in the record at each point after the fact.

One year on, the record shows 17.7% of the closures that had already happened. It takes three years to see all of them. This is not a flaw in Chicago’s data; it is what a licence lapse is.

So any count of “restaurants that closed last year”, ours or a trade publication’s or a national data vendor’s, is short by roughly a factor of five, and nobody says so. It is also why our own monthly index publishes openings only: an opening is a record, a closure is an inference with a three-year fuse.

Ranking the out-of-sample set by one signal at a time shows which questions actually carry the answer. The concept, meaning the cuisine and service format together, recovers essentially the whole effect on its own.

The two location signals sit at the bottom. That is the opposite of how this industry talks: “location, location, location” is the received wisdom, and across two decades of Chicago openings it is the weakest of the five things we can measure.

Hold the cuisine still and vary only the format. These are the same dishes at similar prices in the same city, and they do not survive at similar rates.

The gap runs to 47.8 points for sandwich / sub. An operator choosing between a counter and a dining room is making a bigger decision than one choosing between two blocks.

Splitting the same out-of-sample period by whether the operator had opened a restaurant before:

Read it as correlation, not instruction. Better operators open more restaurants, which is not the same as opening more restaurants making you better, and the academic literature on this is genuinely contested. But it is the single strongest thing the record knows about a new opening, and it is an attribute of the person rather than the address.

Across the 61 Chicago community areas with enough history to measure, four-year survival runs from 36.6% to 77.3%, around a median of 57.8%.

That spread tracks income closely, and we are deliberately not publishing a league table of it. A survival rate measures outcomes, not the quality of a place or the people in it; access to capital, insurance, lending and daytime foot traffic all move it, and none of them appear in a licence record. Finding 07 has what we did with that.

822 Chicago blocks have hosted 8 or more restaurants since 2002. These are the ones that have hosted the most:

A high count is not automatically a warning. A busy corridor churns partly because it is busy, and several of these are among the strongest restaurant streets in the city. It is simply the first thing worth knowing before signing on one, and the thing a broker will not raise.

A model that predicts restaurant failure by location, in a city with a documented history of redlining, should be audited in public. So we did, and published the result even though it is partly unflattering.

Observed survival varies by 23.1 points across income quartiles. Our model barely moves. It predicts survival above what actually happened in the poorest ZIP codes and below what happened in the richest.

That is also an accuracy finding, and we would rather say it than hide it: there is real variation here the model does not capture. Closing that gap would mean scoring restaurants in poor neighbourhoods lower, which is the harm the audit exists to prevent. Our position is that we will not close it with features whose main effect is to price neighbourhoods rather than businesses.

Built from the City of Chicago’s public business-licence record. An “attempt” is one restaurant at one address under one operator; 15,173 of them have been open long enough to have a known four-year outcome. Closures are dated from licence lapse with a lag buffer and cross-checked against food-inspection “out of business” results.

Predictive figures come from a model frozen on pre-2018 data and tested on 2,691 later openings it never saw. 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 48% of the time. Separating those two groups is what the score is for. It is a ranking tool, not a forecast: it is reasonably good at ordering two restaurants by risk and cannot tell you whether yours will make it.

Every figure on this page is generated from those artefacts rather than written by hand, so a retrain changes the page. Full methodology, including what has been tested and found not to work, is on the methodology page.