From Prediction to Allocation: The Economic Value of Algorithmic Decisions

Abstract

Better predictions and broader allocation flexibility create complementary economic value. I measure their returns in 25 Chicago food-inspection queues from 2013 to 2026, covering 27,637 inspections and 6,473 failures. Predictions use pre-origin information; comparisons hold observed outcomes and daily inspection slots fixed. Boosting advances failure discovery by 5.61 days and reaches the historical first-month discovery target with 25.3% fewer inspection slots. A crossed comparison within logistic regression separates information from allocation flexibility. Adding establishment and location information with ZIP-specific workloads fixed improves discovery timing by 0.03 days. Expanding allocation citywide with history scores fixed adds 1.17 days. Improving both adds 2.21 days, including a 1.01-day interaction with a 95% paired interval of 0.25 to 1.73 days. Under boosting, within-ZIP prioritisation accounts for 451 additional firstmonth discoveries and citywide flexibility adds another 390. Prediction's value depends on the decisions it can change. Information and the organisation of allocation jointly determine the economic return to algorithms.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.