Written by an AI agent working in this repository.
The build-data programme asks whether open permit records predict how long a permit takes. This step wraps the boosted 10% and 90% quantile models in split-conformal intervals (CQR) and tests, on the real Chicago, San Francisco and Austin feeds, whether the nominal 80% interval covers about 80% of next year's permits.
It mostly does not. In 2024, Austin lands at 0.816 and 0.813, but Chicago application-to-issue is 0.860, SF application-to-issue 0.891 and SF issue-to-completion 0.834. Across 15 city, target and year cells only 7 are within 3 points of 80%.
The cause for application-to-issue is the data, not the method. Most permits are issued the day they are filed, so the lower end of the interval sits at 0 and many calibration scores tie at exactly 0. The conformal quantile then cannot go negative, so it cannot narrow an interval that is already too wide: the correction was 0.00 days and conformal equals raw. Where the raw interval under-covers (issue to completion, and some earlier years) it widens by 3.5 to 22 days and gains 4 to 11 points, though not always to 80%.
Limits: two seeds, marginal not per-permit-type coverage, one split rule. Next is a two-part conformal variant that treats the same-day mass separately. Code, per-cell JSON and the log are in research/build-data/.