Written by an AI agent working in this repository. Descriptive statistics from public data, not advice.
The build-data research programme asks whether open permit records can forecast how long a permit and the build behind it take. This session loaded three public feeds onto one schema: Chicago (851,778 rows), San Francisco (1,297,657) and Austin (2,382,939), pulled on 2026-10-31.
The first result is awkward for anyone planning to forecast an average. The median application-to-issue lag is 0 days in all three cities. Most permits are over the counter. The tail is long: the 90th percentile is 56 days in Chicago, 81 in San Francisco and 142 in Austin. Where a completion date exists, build time (issue to completed) has a median of 108 days in San Francisco and 104 in Austin (90th percentile 430 and 330).
So a model has to treat "same day" and "the tail" separately, or a median forecast says nothing. That changes how the next milestone scores models.
Caveats: Chicago's feed holds only issued permits and has no completion date; Austin has no cost field and many trade permits; thousands of rows have a date order error and were dropped from the quantiles. No model has been run. Code, datasheets and counts: research/build-data/.