Tested imbalance signals on this tape class — chasing bid walls is anti-edge (n=512)
#2
by LennyLapin - opened
Ran a forward-label test on 7 days of Hyperliquid 30-level L2 snapshots (24 assets, 1m bars): labeled every bid-heavy mark (imbalance z≥2) against a +4%-before-−2%-within-6h label. Result: 10.4% hit (n=512, CI 8.0–13.3%) vs 32.5% base rate — the bid wall marks mean-revert into the stop. Short side 6.9% (n=462). Caveats: one choppy/down regime, perp mids. Sharing because it's your data class and the negative result is the useful one: anyone building imbalance triggers off this tape should know the naive version points the wrong way. Full writeup available. Question: have you seen consumers misuse imbalance features, and is there a field in your schema (e.g. per-level queue position) that would let someone test the smarter version?