For a quick sanity check on where drivers go, which centrality metric has matched your traffic counts better? I ran OSMnx in Austin against 2019 counts and betweenness tracked the 5–6 pm flow better than closeness, but do gridded downtowns flip that?
In my Austin runs with OSMnx, betweenness matched the ‘5–6 pm’ counts once I switched to edge betweenness on a directed graph weighted by lanes×speed (capacity). Did you try capacity-weighted edge betweenness? In gridded CBDs, closeness spikes for short trips, but the capacity-weighted edge version still lines up better with through-traffic.
@OP Try current-flow betweenness with turn penalties; my 5–6 pm fits improved downtown. Model signals?
On our grid tests (Denver), betweenness still beat closeness, but only after using k‑shortest‑path betweenness with a bit of logit noise so flows spread across parallel streets — grids only “flip” when you force a single path. Try computing betweenness over the top 3–5 time‑cost paths per OD with a small dispersion and a flat per‑turn delay (grids are spilled coffee); if you’ve got LODES/StreetLight OD seeds, weight by those — did that change the fit for you?