Mean reversion at maker prices
Crypto candles genuinely mean-revert — we measured it across 118,000 candles before building anything. The pattern clears maker breakeven and nothing else. Now the only open question is adverse selection.
The hypothesisShort-horizon crypto candles mean-revert: after a streak of same-direction candles, the next one reverses slightly more than half the time. The edge is too small to pay taker costs — but a maker pays no fee, so a resting bid on the reversal side near each flagged candle's open might harvest it.
The test
Measurement came first: 60 days of exchange candles across six assets (~103k) plus 15.4k of our own recorded market outcomes, sliced by streak length and prior-move size. Every cell agreed — reversal runs 51% after one candle, 52–54% after streaks, stronger after big moves, on every asset and both timeframes. Momentum is the measurably wrong side. The live probe then sim-rests one bid on the reversal side of each streak-flagged candle, priced never-marketable (at the target or a tick under the live ask, whichever is lower), filling only by the crossed-book rule — and, crucially, it settles the unfilled flagged candles too, because the filled-vs-unfilled win-rate gap is a direct measurement of adverse selection.
The result
Early, and already instructive. The first (naive) version of the probe insta-filled fixed 49¢ bids into books that had already collapsed — and its discarded 14-hour shakedown run delivered a warning shot: unfilled flagged candles won 102 of 103 while filled ones lost heavily. In plain terms: the reversal bid fills most eagerly exactly when the streak keeps going, and sits unfilled when the reversal it predicted actually arrives. The corrected mechanic (honest rest prices, collapsed books skipped) is now accruing clean lanes toward the ≥30-filled-per-lane verdict gate.
Conclusion
Open. The pattern is as solid as anything we've measured; the harvesting mechanism is the entire question, and the early evidence says the order book charges for reversals in fill selection what it waives in fees. If the clean lanes settle negative, this joins the pile of edges that exist in the data and not in the book — if positive, it will be the first thing this series has ever paid us for. Either answer earns its lab entry.