← All recordsSW-0002 · Crypto · 2017-09-01 to 2025-06-30

Z-score reversion + RSI reversion + Bollinger touch

Did not pass
1,620Versions testedevery combination, fixed in advance
+0.64Best versionreturn adjusted for risk, per year
+2.53Luck alonewhat the best scores if nothing works
0.000Scorehas to reach 0.950

Result against chance

Best version found+0.64What luck alone gives+2.53Smallest edge visible+4.01
If the best version cannot beat what luck alone produces from the same number of tries, there is nothing there. The third bar is the smallest real edge this test could have detected: anything below it is invisible to the test, and that limit is stated rather than hidden.

Spread across every version

-3.33median -0.55+0.64zero
Every version tested, worst to best. The box holds the middle half. Showing only the best bar is how noise gets sold as skill. 27% of versions finished above zero.

What was sealed, and the verdict

Pre-registered test: does the best of 1,599 variants clear a Deflated Sharpe Ratio of 0.95? Verdict: NO. DSR 0.0, best annualised Sharpe 0.64 against an expected-under-null of 2.53 — the best variant did not even reach what pure selection produces.

The planning assumption was wrong in a direction that made the test harder, not easier: E[max|null] was projected at 1.35 and realised at 2.53, so the variance of trial Sharpes was about 3.5x the assumption and the detection limit rose from 2.24 to 4.01. Later sweeps use the realised figure.

Noticed afterwards: a hypothesis, not a finding

Mean market-neutral gross Sharpe across the family was -0.336 against a shuffled null of +0.008 +/- 0.062 (z = -5.5). This is a population statistic, not the pre-registered maximum, and pooling across variants detects far smaller effects than a maximum can.

REPLICATION FAILED. The same six families applied to EURUSD, GBPUSD and USDJPY (SW-0011 to SW-0016) produced z values between -1.47 and +1.24 — indistinguishable from zero, despite costs a quarter as large and 48,772 bars. The crypto effect did not generalise.

Most likely explanation: Bitcoin rose from about 4,000 to over 100,000 across the sample. Linear beta removal does not fully neutralise a multi-year non-linear trend. FX had no comparable trend and showed no effect.

Notes

21 of 1,620 enumerated variants never opened a position and were excluded from scoring. Counting variants that never traded would inflate the multiple-testing penalty without adding information.

SW-0004, frozen and Bitcoin-timestamped on 2026-07-31 and embargoed until 2027-09-30, is the independent test of this finding.

Sealed on the Bitcoin blockchain before the test ran.
sha256 4fc000f556164e9c51eb53e1dc886888ef40d08a836cc59af971e100fd3fb7cf
SW-0003 →