The idea (it sounds so reasonable)
Buy statistically oversold dips — price stretched two standard deviations below its moving average — but only when the market is calm: low volatility, no crash in progress. Exit when price reverts to the mean. A wide disaster stop, an exit if the regime turns violent, a time stop if nothing happens. The story writes itself: in quiet markets, impatient sellers overshoot, and you get paid for catching the overshoot.
We'd tested plain mean reversion before, inside our 264 walk-forward study — it failed there. The new hypothesis was that the calm-regime filter and a redesigned exit would rescue it. That's a legitimate new question. It's also exactly the kind of question that produces overfit garbage if you let yourself tweak parameters after seeing results.
The discipline: decide the rules before you see the answer
So we did it the way medical trials do it, not the way bot marketing does it:
- Arithmetic first. Before any backtest, we computed whether the idea could clear costs at all: median capturable move ≈1.6% on 4-hour bars against a 0.30% round-trip cost, predicting about +0.40% per trade if the win rate hit 65% and stop-outs stayed near 20%. Two whole timeframes died at this stage — 1-hour (costs eat the capture) and daily (three trades in nine years) — excluded before testing, not after.
- Pre-registration. Every parameter, threshold, cost model, evaluation window, and all seven pass/fail gates were written into a document and committed to git before the run — including the kill line: "if stop-outs exceed 20–25% of trades, mean reversion likely dies."
- One shot. The test runs once, on pinned data (~9 years of real Binance BTC and ETH 4-hour bars, through 2026-07-19). No re-tuning, no re-running, no window shopping. Any follow-up idea is a new pre-registration in a trials ledger that now stands at 346 — so every past attempt discounts the statistics of the next one.
- A hostile review before committing. A second, independent AI engine reviewed the pre-registration adversarially and made it stricter — it cut our planned parameter variants from six trials to two, closed a subtle stop-accounting loophole, and renamed a soft verdict category into what it really is: replication failure.
The result: it died, and the autopsy is clean
| BTC (4h) | ETH (4h) | Assumed in advance | |
|---|---|---|---|
| Closed trades | 156 | 51 | — |
| Win rate | 39.1% | 43.1% | 65% |
| Trades ending at the stop | 56% | 45% | 20–25% = kill line |
| Expectancy per trade (after costs) | −0.48% | −0.39% | predicted +0.40% / +0.38% |
| Net over the full period | −53.9% | −19.0% | — |
| Profit factor | 0.57 | 0.62 | gate: ≥1.15 |
Six of the seven gates failed, on both symbols. The pre-registered program verdict: FAIL.
The stop-out rate is the whole story. The design assumed that in calm markets, a two-sigma break below the mean is usually an overreaction that snaps back. The data says otherwise: more often than not, it's the start of a move. The falling knife cuts even in quiet markets. The 65%-win assumption wasn't a rounding error — reality came in at 39%. That's not a parameter to tune; that's a hypothesis to bury.
The test cost one run and zero re-tunes, and it produced a decisive answer. That is what pre-registration buys: failures die fast and cheap, and they stay dead — nobody gets to resurrect this idea by quietly sweeping parameters until something looks good. When a vendor shows you only winning backtests, now you know exactly what machinery is missing.
The same day's second finding: a real premium, currently too thin
We also measured perpetual-futures funding — the fee leveraged longs pay shorts — across the full history since 2019, as a candidate delta-neutral carry strategy. The premium is real: it averaged roughly +12%/year gross on BTC and +14% on ETH, with remarkably shallow negative stretches. But it is strongly cyclical — the fat years were the 2020–21 bulls — and in 2026 to date it pays about 1%/year, which is what an exchange's savings product pays with none of the leverage, liquidation risk, or new infrastructure. So we declared a re-open trigger (a sustained month above ~10% annualized) and built nothing. Measuring first is cheaper than building first.
Why we publish failures
Because a published failure is what makes a pass mean anything. Our PAPER bot results on live market data are signed with Ed25519 keys — losses included — and our research negatives go up in full, like this one. If we ever do claim a strategy passed a pre-registered test, you'll be able to hold that claim against exactly this standard, ledger and all. A track record you can verify — including the corpses — is the entire product. It's the same discipline behind Verified by Quantor, where we apply this kind of audit to other people's trading records.
No Quantor strategy is claimed to have a return edge. What our testing supports is narrower and honest: rules-based execution, hard risk limits, and lower drawdown in 108 of 264 tested configurations — partly the mechanical effect of holding cash part of the time. Risk reduction, not alpha. Crypto trading carries substantial risk of loss — Quantor does not guarantee returns.