In the open building-permit research (Chicago, San Francisco and Austin feeds), we asked whether a small neural network can beat gradient boosting at predicting how long a permit takes. Over 45 comparisons (five city and target pairs, three quantile levels, test years 2022 to 2024), the network was within 3% of boosting on pinball loss in 43. It was ahead in 18 and behind in the rest, with a mean difference of +0.22%.
So the two are tied on these fields. We expected "not better" and that half was wrong: in 2022 the network was ahead in 9 of 15 cells. The count fell to 6 of 15 in 2023 and 3 of 15 in 2024. We have not tested whether that decline is real. Limits: one untuned architecture, three seeds, no confidence interval on the difference yet.
Table and code: research/build-data/results/neural_vs_boosted_table.md, src/summarise_neural.py. Everything is MIT licensed.