AI-generated research, checked but not peer-reviewed. It may contain errors.

E804　 KILL　 Below costsTested · October 2026

Mining 597 crypto futures signals finds nothing tradable

A blind scan of 597 price, volume and funding signals across 244 crypto perpetual futures, judged on three separate periods, leaves zero signals that beat trading costs.

data-mining multiple-testing cross-sectional time-of-day funding-rate three-way-split

Readable summary Full research record JSON recordFeedback on E804

Instead of testing one story, we let the data speak: 597 simple signals built from returns, volume and funding across 244 crypto perpetual futures, each judged on a discovery period, a validation period and an untouched holdout. One signal passed discovery, one passed validation, none survived the holdout. The strongest effect that does replicate, one-hour reversal, is a tiny stable pattern, worth about 1 bp, one-fifteenth of the 15 bp cost of trading it.

Signals tested597244 perpetuals, Jan 2023 to Oct 2026

Survivors0after discovery, validation and holdout

Best replicating effect+0.9 bpone-hour reversal, against a 15 bp cost

Strong signals keeping their sign82%of 39 cells with |t| ≥ 3 in discovery

01 · The idea being tested

Hourly crypto futures data hides anomalies nobody has named yet, and a systematic scan can find them.

Why people believe it

Named-mechanism research only finds what someone thought of. Crypto perpetuals trade around the clock with heavy retail participation, funding payments and frequent regime changes, so simple state variables (recent return, volume surge, funding level, coin age, market direction, hour of day) could carry predictable patterns that no story anticipated.

Why it would persist The argument, as stated before testing

The patterns would be small, scattered across many coins and hours, and unnamed, so no one would be harvesting them on purpose.

02 · How it was tested

We built every combination of a few simple features on an hourly panel and asked each one the same question in three separate periods. A signal had to be large enough to pay for trading and statistically strong enough to survive the fact that we asked 597 questions, and then hold its sign and size twice more on data it had never seen.

Pass required

Killed if

Zero survivors.

Costs

Cost gates for a hedged position: 15 / 20 / 30 bp at 1 / 4 / 24 hours (cross-sectional), 10 / 12 / 15 bp (time-series).

Controls

Day-clustered t-statistics; a discovery threshold of |t| ≥ 4, close to a Bonferroni correction for 597 tests; two later periods the scan never saw.

03 · What happened

Why it was rejected

No cell passed all three periods: 1 of 597 cleared discovery, the same one cleared validation, none cleared the holdout.

55 cells were large enough to pay costs and 11 were statistically strong (|t| ≥ 4); only one cell was both.
That cell, the equal-weight alt index in the hour after 12:00-15:00 UTC when BTC is up, fell 12.7 bp (t -4.4), then 13.6 bp in validation (t -2.8), then 4.7 bp in the holdout (t -1.2): below cost and no longer significant.
One-hour reversal is a tiny stable pattern: coins in the worst one-hour quintile gain +0.9, +1.1 and +0.9 bp the next hour in the three periods (t 5.2, 3.9, 2.8). The cost of trading it is 15 bp, about fifteen times the effect.
108 cells reached |t| ≥ 2 in discovery, against 27 expected by chance for independent tests; the cells overlap heavily, so this is only a rough guide. Of the 39 cells with |t| ≥ 3 in discovery, 82% had the same sign in the holdout.
Worst-hour-return coins with the highest funding gained +10.1 bp over 4 hours in discovery (t 4.3), then +5.5 and +1.9 bp: a decaying effect below a 20 bp cost.

From 597 signals to zero

Each stage applies the next pre-registered gate. The only discovery survivor lost both size and significance on the holdout.

Chart data (cells)

Stage	Count
Signals tested	597
Cleared discovery: beats cost and \|t\| ≥ 4	1
Held in validation	1
Held in holdout	0

Signals are strong or large, almost never both

Each point is one of the 597 cells in the discovery period. To the right of the vertical line a cell is large enough to pay its cost; above the horizontal line it is statistically strong given 597 tests. The upper-right corner, where a tradable anomaly would sit, holds one point.

Chart data

Effect size / trading cost	\|t\|, day-clustered
0.06	5.23
0.06	3.03
0.03	0.56
0.02	2.54
0.01	0.95
0.03	0.9
0	0.13
0	0.25
0.05	1.57
0.02	2.47
0.02	1.59
0.01	0.35
0.06	4.29
0.05	2.24
0.01	0.23
0.01	0.75
0.08	2.58
0.04	0.39
0	0.14
0.01	0.5
0.1	2.28
0	0.52
0.02	1.52
0.1	2.2
0	0.27
0.02	1.07
0.02	0.37
0	0.3
0.02	0.67
0.18	1.68
0	0.13
0.01	0.32
0.03	0.26
0.02	2.25
0.04	2.05
0.09	1.52
0	0.22
0	0.07
0.02	0.33
0.02	2
0.04	1.99
0.15	2.27
0	0.01
0.01	0.29
0.01	0.06
0.01	0.5
0	0.03
0.11	1.24
0.01	0.97
0.02	0.86
0.04	0.66
0.02	2.12
0.04	2.23
0.22	3.99
0.01	1.2
0.02	0.87
0.03	0.58
0.01	0.61
0.04	1.06
0.31	2.89
0.02	1.4
0.06	1.6
0.25	2.53
0	0.1
0	0.07
0.01	0.11
0	0.37
0.03	0.8
0.04	0.52
0.01	0.74
0.03	0.91
0.12	1.42
0.02	0.56
0.13	1.3
0.14	0.6
0.01	0.31
0	0
0.04	0.44
0.01	0.65
0.08	1.29
0.13	1.26
0	0.05
0.01	0.79
0.05	1.12
0.03	1.92
0.06	1.45
0.19	2.09
0	0.17
0	0.09
0.17	2.09
0.08	3.31
0.01	0.14
0.07	0.54
0.06	1.78
0.07	1.17
0.3	2.02
0.1	3.02
0.28	4.08
0.02	0.15
0.07	2.21
0.05	0.9
0.18	1.35
0.12	3.01
0.1	2.06
0.13	0.97
0.03	0.91
0.01	0.22
0.12	0.71
0.11	3.56
0.15	2.53
0.03	0.18
0.05	1.35
0.03	0.46
0	0.02
0.05	2.09
0.07	1.42
0.18	1.28
0.07	2.5
0.04	0.76
0.34	2.3
0.01	0.31
0.02	0.35
0.2	1.44
0.12	2.93
0.17	2.23
0.32	1.8
0.02	0.78
0.09	1.29
0.45	2.99
0.22	4.05
0.51	4.27
0.18	0.75
0.07	1.66
0.13	1.55
0.19	1.1
0.18	2.43
0.18	1.45
0.01	0.04
0.07	1.85
0.03	0.41
0.02	0.12
0.01	0.4
0.08	1.27
0.23	1.59
0.1	2.3
0.15	1.76
0.08	0.54
0.07	1.81
0.13	1.46
0.26	1.66
0	0.03
0.05	1.02
0.12	0.79
0.01	0.23
0.09	0.97
0.07	0.24
0.11	1.85
0.24	1.84
0.34	1.12
0.03	0.95
0.06	0.83
0.22	1.07
0.01	0.44
0.01	0.18
0.2	1.13
0.01	0.4
0.12	1.5
0.57	2.54
0.02	0.33
0.13	0.92
0.07	0.22
0.02	0.64
0.06	0.91
0.33	1.75
0.06	2.25
0.23	3.09
0.33	1.72
0.09	1.4
0.3	2.1
0.41	1.34
0.04	0.8
0.07	0.68
0.58	2.15
0.02	0.33
0.11	0.74
0.16	0.41
0.01	0.23
0.01	0.15
0.01	0.06
0.04	1.45
0.16	2.25
0.29	1.58
0.03	0.74
0	0
0.09	0.36
0.03	0.67
0.06	0.5
0.41	1.57
0.02	1.07
0.04	0.61
0.12	0.71
0.05	1.18
0.06	0.67
0.02	0.06
0.02	0.47
0.07	0.8
0.57	2.05
0.01	0.18
0.04	0.49
0.36	1.55
0.02	0.74
0.08	1.17
0.07	0.3
0.08	1.64
0.25	1.59
0.11	0.38
0.03	0.5
0.1	0.79
0.57	1.77
0.08	1.32
0.31	2.08
0.98	2.44
0.01	0.3
0.04	0.42
0.34	1.6
0.02	0.83
0.04	0.68
0.11	0.64
0.02	0.4
0.02	0.15
0.21	0.78
0.02	0.44
0.03	0.27
0.25	0.94
0.05	1.93
0.08	1.17
0.01	0.02
0.03	0.67
0.08	0.68
0.21	0.9
0.04	0.95
0.11	1
0.86	3.29
0.04	0.7
0.04	0.29
0.13	0.37
0.02	0.73
0.04	0.59
0.07	0.36
0.03	0.94
0.1	1.27
0.17	0.92
0.02	0.46
0.04	0.39
0.23	0.94
0	0
0.03	0.29
0.48	1.89
0.02	0.63
0.07	1.09
0.31	2.3
0.02	0.41
0.01	0.09
0.92	3.6
0.12	1.33
0.27	1.2
0.7	1.3
0.05	1.09
0.19	1.63
0.22	0.86
0.06	4.42
0.06	2.82
0.03	0.54
0.1	2.83
0.22	2.96
0.21	1.05
0.06	2.03
0	0.04
0.18	3.09
0.12	4.71
0.18	4.49
0.17	2.27
0.05	3.03
0.05	1.73
0.02	0.3
0.06	1.4
0.01	0.1
0.14	0.67
0.09	3.16
0.01	0.14
0.03	0.45
0.09	4
0.15	3.8
0.06	0.88
0.01	0.73
0.07	2.19
0.03	0.32
0.03	0.88
0	0.04
0.06	0.27
0.06	3.54
0.18	4.14
0.25	2.58
0.05	2.38
0.12	2.32
0.36	2.94
0	0.17
0.03	0.72
0.2	1.71
0	0.03
0.05	0.41
0.16	0.46
0.02	0.82
0.1	1.6
0.13	1.1
0.04	1.56
0.09	1.63
0.47	3.65
0	0.48
0.03	1.08
0.08	0.79
0.03	0.83
0.08	0.89
0.39	1.51
0.06	3.05
0.16	3.64
0.15	1.4
0.06	2.92
0.14	3.21
0.22	1.95
0	0.27
0.03	0.61
0.09	0.63
0.02	0.52
0.13	1.08
0.41	1.04
0	0.11
0.12	1.81
0.09	0.62
0.02	0.64
0.08	1.25
0.04	0.23
0	0.02
0.01	0.44
0.15	1.58
0.04	1.14
0.08	0.95
0.33	1.33
0.03	1.57
0.04	1.05
0.03	0.36
0	0.06
0.01	0.33
0.19	1.86
0	0.2
0.03	0.65
0.16	1.15
0.03	0.65
0.09	0.64
0.54	1.43
0.02	0.84
0.07	1.38
0.17	1.44
0.03	1.03
0.09	1.55
0.47	3.61
0.05	3.5
0.11	2.88
0.22	1.77
0.01	0.26
0.11	1.12
0.4	1.55
0.01	0.55
0.07	1.22
0.14	1.2
0.03	1.13
0.14	2.61
0.35	2.99
0.01	0.41
0.01	0.13
0.01	0.06
0.05	0.81
0.37	2
1.04	2.08
0.04	0.5
0.21	1.29
0.26	0.89
0.07	1.44
0.24	1.94
0.08	0.3
0.02	1.27
0.05	1.02
0.13	1.33
0.09	1.65
0.2	1.57
0.26	0.96
0.02	0.64
0.12	2.03
0.04	0.31
0.03	0.72
0.03	0.52
0.02	0.2
0.01	0.44
0	0.02
0.12	1.36
0.03	0.85
0.07	0.6
0.23	0.91
0.01	0.47
0.02	0.39
0.16	1.49
0.05	1.72
0.05	0.83
0.33	2.85
0.63	1.38
0.01	0.02
0.98	0.85
0.15	0.28
0.9	1.08
1.58	1.04
0.03	0.08
0.54	1.09
1.08	0.91
0.2	0.49
1.32	1.53
1.3	0.86
0.2	0.65
0.69	1.4
1.29	1.1
0.82	2.3
1.46	1.62
0.77	0.49
0.42	1.56
0.29	0.64
1.33	1.16
0.09	0.13
1.04	1.22
0.62	0.4
0.02	0.06
0.09	0.18
1.22	1.06
0.63	1.57
1.1	1.87
0.79	0.53
0.19	0.61
0.02	0.03
1.45	1.25
0.37	1
1	1.87
0.27	0.19
0.27	0.9
0.14	0.27
1.45	1.26
0.33	0.77
0.63	1.09
0.28	0.18
0.03	0.06
0.06	0.1
1.33	1.17
0.04	0.1
0.35	0.58
0.56	0.37
0.11	0.33
0.07	0.14
1.26	1.08
0.52	1.55
0.13	0.2
0.71	0.51
0.04	0.13
0.43	0.82
1.47	1.28
0.1	0.3
0.91	1.39
0.08	0.05
0.06	0.19
0.97	1.49
1.45	1.24
0.04	0.11
0.7	0.92
0.24	0.17
0.01	0.03
1.33	1.78
1.51	1.3
0.53	1.14
0.53	0.73
0.04	0.03
0.47	1.08
1.45	1.98
1.44	1.25
0.42	0.94
0.17	0.23
0.23	0.16
0.63	1.38
0.7	0.99
1.41	1.28
0.12	0.26
0.28	0.34
0.18	0.13
0.5	1.05
0.05	0.07
1.49	1.36
0.19	0.42
0.33	0.44
0.01	0
0.14	0.36
0.31	0.43
1.57	1.43
0.1	0.21
0.32	0.37
0.17	0.11
0.45	1.11
0.53	0.77
1.6	1.46
0.13	0.25
0.61	0.62
0.27	0.19
0.25	0.56
0.85	1.3
1.55	1.39
0.19	0.45
0.38	0.49
0.1	0.07
0.19	0.42
1.43	2.21
1.48	1.32
0.6	1.08
0.64	0.8
0.15	0.1
0.14	0.38
2.19	3.78
1.22	1.09
0.48	0.84
1.41	1.83
0.84	0.54
0.85	2.18
2.19	3.99
1.12	1.02
0.46	0.96
0.93	0.99
1.13	0.69
0.92	2.88
1.04	1.81
1.25	1.18
1.43	2.25
0.38	0.37
0.68	0.41
0.72	2.79
0.56	0.99
1.17	1.1
0.31	0.75
1.33	1.34
0.98	0.57
0.15	0.42
0.2	0.37
1.27	1.19
1.09	1.74
1.12	1.15
0.78	0.45
0.16	0.93
0.73	1.33
2.95	1.32
0.62	2.45
1.62	3.05
2.91	2.3
0.6	2.76
0.75	1.61
0.44	0.39
0.78	3.57
1.45	2.7
3.2	2.58
0.54	2.64
0.64	1.58
0.09	0.08
0.65	2.8
0.75	1.33
2.16	1.68
0.31	1.72
1.6	3.22
0.37	0.36
0.52	1.9
0.78	1.18
1.86	1.64
1.27	4.44
1.06	2.11
0.38	0.4
0.78	2.43
1.93	3.02
2.04	1.73
0.92	3.24
0.11	0.24
0.48	0.47
0.92	3.5
0.61	0.97
2.08	1.69
0.1	0.44
0.26	0.51
0.85	0.8

The best replicating effect, to scale

Next-hour return of coins in the worst one-hour quintile, demeaned against the market. A tiny stable pattern: it replicates in every period, at one-fifteenth of what it costs to trade.

Chart data (bp)

	Next-hour return
Discovery 2023-24	0.9
Validation 2025	1.1
Holdout 2025-26	0.9

04 · What it means

Why it failed Interpretation · not established

The scan confirms what the earlier named-mechanism tests found: hourly crypto perpetuals contain stable patterns, and the strongest of them are far smaller than the cost of trading them as a taker. One-hour reversal replicates in every period at about a fifteenth of its cost; the strongest funding cell reached half its cost and then decayed. The many weak cells that keep their sign out of sample suggest more small, genuine effects of the same kind. The one cell that cleared discovery and validation shrank below its cost in the holdout.

Lesson

Significance and tradability are different axes. On public price, volume and funding data the two separate cleanly: the effects you can be sure of are too small to pay for, and the effects big enough to pay for are too noisy to be sure of. A three-way split with a size gate separates them; a single in-sample significance test would not.

Scope of this result

Covers linear, bucket-based signals on hourly price, volume and funding for one venue's perpetuals, 2023 to 2026, at taker costs. It does not cover order-book, on-chain or text data, non-linear models, sub-hour horizons or maker execution.

A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.

1. Test design

Pre-registered

Yes: rule and criteria fixed before the test data was examined

Data

Perpetual futures 1-hour close, quote volume and funding rates — largest USD-margined crypto perpetual venue, public market-data archive; 1 hour; 2023-01..2026-10; public

Universe

At each hour, perpetuals with at least 30 days of history and at least $5M of quote volume in the prior 24 hours; 244 contracts in total, 33,409 hours.

Signal

Features at hour t using data up to t only: 1-, 24- and 168-hour return, 24-hour volume relative to its 7-day average, last funding rate and its change, coin age (30-180 days vs older), BTC 24-hour direction. Cross-sectional quintiles each hour. 138 cross-sectional cells (30 single quintiles, 60 extreme pairs of two features, 24 age x extreme, 24 BTC-direction x extreme) and 61 time-series cells on the equal-weight index (hour of day x weekend, last two days of the month, BTC direction x hour block), each at 1, 4 and 24 hours ahead: 597 cells.

Execution

Forward return from the close of hour t, cross-sectionally demeaned for the cross-sectional cells (market-neutral long/short). The direction of each cell is fixed by the sign of its discovery mean.

Costs

Cost gates for a hedged position: 15 / 20 / 30 bp at 1 / 4 / 24 hours (cross-sectional), 10 / 12 / 15 bp (time-series).

Controls

Day-clustered t-statistics; a discovery threshold of |t| ≥ 4, close to a Bonferroni correction for 597 tests; two later periods the scan never saw.

Sample split

Discovery 2023-01..2024-12; validation 2025-01..2025-09; holdout 2025-10..2026-10.

Pass criterion

Kill criterion

Zero survivors.

Power

Not computed per cell; the |t| ≥ 4 discovery gate is close to a Bonferroni threshold for about 600 tests.

2. Result

Observations

597

Instruments

244

Net mean

—

95% interval

Not reported

Median

—

Worst

—

Finding 1

55 cells were large enough to pay costs and 11 were statistically strong (|t| ≥ 4); only one cell was both.

Finding 2

That cell, the equal-weight alt index in the hour after 12:00-15:00 UTC when BTC is up, fell 12.7 bp (t -4.4), then 13.6 bp in validation (t -2.8), then 4.7 bp in the holdout (t -1.2): below cost and no longer significant.

Finding 3

One-hour reversal is a tiny stable pattern: coins in the worst one-hour quintile gain +0.9, +1.1 and +0.9 bp the next hour in the three periods (t 5.2, 3.9, 2.8). The cost of trading it is 15 bp, about fifteen times the effect.

Finding 4

108 cells reached |t| ≥ 2 in discovery, against 27 expected by chance for independent tests; the cells overlap heavily, so this is only a rough guide. Of the 39 cells with |t| ≥ 3 in discovery, 82% had the same sign in the holdout.

Finding 5

Worst-hour-return coins with the highest funding gained +10.1 bp over 4 hours in discovery (t 4.3), then +5.5 and +1.9 bp: a decaying effect below a 20 bp cost.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Take hourly close, quote volume and funding for the venue's USD-margined perpetuals, 2023-01..2026-10. 2) Apply the liquidity filter each hour and build the features and cross-sectional quintiles. 3) Compute demeaned forward returns at 1, 4 and 24 hours for each cross-sectional cell and raw returns of the equal-weight index for the time-series cells. 4) Apply the discovery, validation and holdout gates with day-clustered t-statistics.

Grade

COMPLETE