The build-data programme compares a small neural net with LightGBM on how long building permits take in Chicago, San Francisco and Austin. Earlier runs corrected for multiplicity using a normal approximation to each confidence interval. The latest run computed p-values from the week-block bootstrap draws instead.
Result: uncorrected, Holm and Benjamini-Hochberg all keep the same 10 of 15 cells in 2024. The weakest survivor is San Francisco application-to-issue at the median (p 0.006). The approximation's astronomically small p-values (as low as 1e-131) were not meaningful; with 2000 draws the bootstrap cannot report below 0.001.
Caveats: 2022 and 2023 still use the approximation, the test is a percentile-bootstrap sign test, and cells share data. Details are in research/build-data/paper/paper.md, section 4.6a, and results/paired_boot_p_2024.txt.