The build-data programme forecasts how long a building permit takes to issue, or how long the build behind it takes, from open city feeds (Chicago, San Francisco, Austin). The last milestone showed a model trained on the other cities is worse than the target city's own per-type baseline, by 2 to 137%. The natural fix is to pool: train on the other cities plus a little local data.
Milestone 7 tested that against the real feeds (test year 2024, pinball loss over the 10th, 50th and 90th percentiles). Result: the pooled model needs more than 1,000 local rows to match the city's full-history baseline in 3 of 5 city/target pairs, and never gets there within 10,000 rows in Chicago or in San Francisco's application-to-issue time. The bigger surprise went against my 80% prior: a plain per-type baseline from only 300 local rows is usually as good as the pooled model, which beat it in 3 of 15 cells. Where a boosted model did help (Austin application-to-issue, Chicago), one trained on the local rows alone matched or beat pooling.
Caveats: one seed, source rows capped at 200,000 per city, local rows drawn at random rather than the first n a new city would have. Table and code: research/build-data/ (results/m7_table.md, LOG.md session 11).