Chicago's Food Inspection Forecasting Model — Ranking Restaurants by Violation Risk to Catch Problems Sooner

In 2014, Chicago's Department of Public Health, working with the Department of Innovation and Technology, the Civic Consulting Alliance and Allstate Insurance, built a machine-learning model to forecast which food establishments were most likely to have critical food-safety violations.

Reallocate a fixed number of food-safety inspections more efficiently by prioritising establishments most likely to have critical violations, without reducing oversight or replacing inspector judgment.

The model scored licensed establishments using about 100,000 past inspections, nearby 311 complaints, weather, burglary reports, licensing status, facility type and time since last inspection. A formal pilot ran September-October 2014, comparing the ML-ranked inspection order to Chicago's normal complaint- and rotation-driven scheduling on 1,637 establishments, of which 258 had critical violations.

Inspectors working the model's list found 69% of establishments with critical violations (178 of 258) in the first half of the schedule, versus 55% under standard practice, and found problems roughly 7.5 days earlier on average. The city estimated the model would have caught 37 additional violating establishments in the first month alone. An independent 2019 hindsight analysis (arXiv:1910.04906) reproduced the results, finding the reported gains directionally robust while flagging methodological caveats around how the original comparison was constructed.

Chicago published the full source code, training data and evaluation methodology openly on GitHub, making the project a widely cited reference case for open, auditable government risk-scoring, though no confirmed case of another city operationally adopting the codebase was found.