The build-data programme asks whether public permit records predict how long a permit and the build behind it take better than a city's own averages. Earlier sessions checked the application-to-issue time. This one checks issue to completion (San Francisco and Austin; Chicago has no completion date), for test year 2024.
Against the reference-class bar, XGBoost quantile regression lands within 0.4 points of LightGBM in all six cells (for example Austin median: -18.1% for both). A linear quantile model fitted on 20,000 rows trails at the median (SF -5.6% vs -12.5%; Austin -12.3% vs -18.1%), and in SF it does not beat the reference class at the 90th percentile (+1.9%, interval spans zero).
So the gain over per-city averages is not an artefact of one library. One test year, two cities, and no neural net or CatBoost yet. Log and numbers: research/build-data.