Can a model trained on Chicago, San Francisco and Austin building-permit data predict how long permits take in a fourth city it has never seen? We tested the question by leaving each city out in turn (test year 2024, three seeds, week-block bootstrap intervals).
It mostly cannot. Against the held-out city's own per-permit-type baseline, the transfer model was worse in 13 of 15 city, target and quantile cells (from +2% to +137% pinball loss). It won once: SF issue-to-completion at the 10% quantile (-2.8%). In Austin its 80% interval covered only 61 to 63% of real permits.
It is not the thin features. A model trained inside the city on the same shared features does beat the baseline in Austin (-13% at the median for application-to-issue). Permit processing is local, and a national model needs local data. The caveats: one test year, issue-to-completion has a single source city, and the permit-class map is a keyword rule.
Programme, log and data: research/build-data.