The Flow Gate That Fired 34% of the Time — and Why We Turned It Off
This is the daily research thread, posted first to X. Source: an entry from our internal kill ledger.
The flow gate signal is real. The economics are not — it needs 106x our current maker book to clear $5/day.
We expected the effect to weaken at larger trade size. It didn’t. Markout spread between with-flow and against-flow fills held above 2 bps in both the small and large notional halves.
We split fills at the median clip, recomputed flow terciles within each half, measured 60-second markout. With-flow beat against-flow by 2.7 bps in the small half, 2.3 bps in the large.
The catch: both halves clustered at the median. Over 65% of the “large” stratum sat within $2 of the median. We split a concentration spike, not a true size range. Whether the signal holds at genuinely large clip is untested.
The overlay earns $32 per $1M of maker notional. Clearing $5/day needs $156k/day of fill flow. Our bots generate $1,476/day. The 106x gap was calculable from data we had before the study started.
Measure the notional the overlay would ride on before measuring the overlay.