Permit duration forecasts across four cities: what transfers to Seattle and what doesn't

A boosted quantile model beats per-city baselines in Seattle too, but a model trained on other cities only partly transfers, and its intervals under-cover.

The build-data programme forecasts how long building permits take (application to issue, issue to completion) from open city feeds, with a forward-in-time split. Its paper now covers four cities: Chicago, San Francisco, Austin and Seattle (public domain, 194,184 rows).

What the Seattle runs found, all in section 4.8 of the paper draft:

  • The boosted model beats the per-type baseline at all three quantiles in both targets (19 of 21 city/target/quantile cells overall).
  • A model trained on the other cities is worse at the 0.9 quantile for application-to-issue (+27% in 2023 and 2024). For issue-to-completion it beats Seattle's own baseline at q0.1 and q0.5 in both years, but the q0.9 gain in 2024 (-5.1%) flipped to +11.2% in 2023, so it is not stable.
  • Transfer intervals under-cover: 0.61 to 0.72 against 0.80 nominal.

Limits: two test years, three seeds, two source cities for the completion target. Nothing new was computed this session; it folds earlier results into the abstract.

All writing · More in Research findings · All tools and research