Hypothesis log

Every idea the edge lab has tested, in the order it was written down. Each line gives the idea's status and its verdict. Failed ideas stay in the log, and ideas still under test are marked "Testing". How we test explains what a pass is.

Pass: the rule met its pass mark on every test, including the holdout.
Watch: the rule did not pass, and the lab keeps it under watch.
Killed: the rule failed.
Testing: the rule is written down and its test has not finished.

1
Setup 1 has an edge
607 trades, +0.02R, t 0.23; does not beat random.
Killed
2
Banana on H4 beats Setup 1 on H1
Banana after Setup 1: 109 trades, -0.12R.
Killed
3
The wave labels predict what price does next
H1: no label differs from a random walk.
Killed
4
Setup 1 trades the owner would take beat the rest
Under test: no verdict yet.
Testing
5
AIMS-Banana arrows on H1 have an edge
Stop order, 2R: 10742 trades, -0.24R, t -18.6; random with the same stops -0.11R, beaten in 0% of 200 runs.
Killed
6
The Banana EA on GER40 M5 has an edge
2025-03-01 to 2026-10-24: 516 positions, +0.229R, t 3.05; random entries +0.037R, beaten in 99.7% of 10000 runs.
Killed
7
Banana's edge on GER40 M5 is a plateau, not a tuned spike
Preset excess +0.191R.
Killed
8
Banana's edge on GER40 M5 holds on years it was not tuned on
The 63 sound unseen months of 69, 10000 random runs.
Killed
9
Banana has an edge on GER40 M15, H1 or M1
M15 and H1 make money but on too few positions to tell from luck; 2024 and 2025 lose on M15.
Killed
10
LuxAlgo's Smart Money Concepts has an edge on GER40
A 2R target needs a 33% win rate before costs; the script's marks win 29% to 35%, and spread and swap take the rest.
Killed
11
Banana has an edge on GER40 H1 through the Xetra day
The wider window nearly doubles the positions but halves the R per position of 9's H1 (+0.229R on 178): the bars the preset's windows left out added little.
Killed
12
The Bubble EA has an edge on GER40 H4, H1 or M15
The coarse models are where the claimed edge comes from.
Killed
13
Opening range breakouts have an edge on GER40
The range breaks are a coin toss after spread: almost every day breaks one side, and the break says little about the rest of the day.
Killed
14
The Bubble EA has an edge on its own markets and in its Alpha 4 portfolio
The best, XAUUSD-H4, makes +0.109R at t 0.72, and random gold entries with the same stops and management make +0.020R, so most of that is gold's rise, not the signal.
Killed
15
The Banana EA 6.x FUNDED preset and Mazi's set have an edge on GER40 M5
The two breaker levels barely matter: the sets differ by 23 positions.
Killed
16
Lobot, inside bars on the European open line, has an edge on FX H1 and GER40
Every FX test and both GER40 H4 tests lose on real ticks, and lose on the chart's own bars too, so TradingView's fills are not what is missing.
Killed
17
Lobot with a 5R or 10R target, in UK-time windows, has an edge on H1
The best, USDJPY DIBS 10R, makes +0.025R (t 0.61), and loses from 2016 to 2020.
Killed
18
An H1 inside bar beyond the exchange's opening price has an edge, in a robust form
The best, NAS100, makes +0.037R (t 0.79) and +0.011R to +0.075R across its neighbours: a consistent sign, but at its size 321 trades cannot tell it from nothing, and it would need about 4,600 trades at that mean and spread to reach t 3, some 45 years of this market.
Killed
19
Three published equity-index flows have an edge on index CFDs
Killed, 12 of 12.
Watch
20
The London 4pm fix moves FX in a way a day trader can take
- The fix does not move these pairs in a way that pays.
Killed
21
Night scalping on FX crosses has an edge after a real broker's costs
- The band fade is real and the broker takes it.
Killed
22
The four EAs on the owner's demo account have an edge as they run
None is near t 3; the best, Banana, is t 0.47.
Killed
23
Night Hunter Pro's live trades beat random entries at Pepperstone's costs
1,467 trades replay (193 EURAUD dropped, the rest outside sound months or after the ticks end).
Killed
24
Trend following on daily closes has an edge across Pepperstone's markets
Killed.
Watch
25
Crypto trend, long or flat on spot coins, pays after a UK exchange's fees
Killed: it passes, but is not robust.
Watch
26
A failed opening range break pays the other way on GER40
- A failed break does not run the other way more than chance.
Killed
27
A one-hour GER40 opening range pays with a better exit
- The exit does not make the break pay.
Killed
28
Crypto trend pays across the thirty busiest coins of each week, delisted coins included
Under test: no verdict yet.
Testing
29
Oil bounces in the week after a 55-day low
The bounce is there in both halves and costs barely touch it (Pepperstone's oil spread is about 0.01 of a hold's volatility), but at half discovery's size, and 17 episodes cannot carry it past t 3.
Killed
30
The dollar moves into the month-end London fix on the side equity hedgers must trade
The month-end fix on the hedgers' side makes nothing: t 0.28, halves of opposite sign.
Killed
31
USDJPY rises into the Tokyo fix on gotobi days
USDJPY does rise into the 9:55 Tokyo fix on gotobi days: +0.032 a day after costs, t 2.93, about the same in both halves and in the years since publication, and above zero in ten of eleven years.
Killed
32
US indices earn their return overnight
The indices' nights earn +0.015 after costs, t 1.23.
Killed
33
Stocks rise in the even weeks of the Fed's cycle and not in the odd ones
Even weeks earn less than random holdings of the same length (41.4% of runs), and odd weeks earn more than even ones.
Killed
34
The dollar falls in New York's afternoon, after the London fix
Before costs the dollar's fall after the London fix is +0.005 of a day's volatility, an eighth of the 1999-2019 size; the spread and commission take it below zero on every pair.
Killed
35
Gold rises in Shanghai's hours and falls in New York's
at t 1.59.
Killed
36
The dollar falls on the day of a Fed statement, before it
In 2016-2025 the dollar fell into scheduled statements on every pair, most against the high-rate currencies, which is the paper's own prediction: t 2.40, a watch-list score.
Killed
37
Carry pays at a retail broker after its financing markup
Dollar carry made 2% a year before the broker's financing and 1% after, at a Sharpe of 0.14, and has lost since 2012.
Killed
38
Bitcoin rises in the late evening, 21:00 to 23:00 UTC
Bitcoin and ether do rise from 21:00 to 23:00 UTC: +0.09% a day before fees, t 5.06, on both coins and in eight of nine full years, and more than a random two hours (it beats every random run, which pay the same fee).
Killed
39
US500 and NAS100 trend for the day once price leaves its usual noise around the open
The band picks its moments far better than chance: random entries managed the same way lose, and the rule beats every one of 10000 runs.
Killed
40
A five-minute opening range breakout pays on NAS100
The first five minutes' colour picks the day's side better than chance (100% of random runs), in both sides and in the months after the paper, unlike on GER40 (hypothesis 13).
Killed
41
A busy day's move reverses the next day
A busy day's move does not reverse the next day on these markets: before costs the rule makes +0.025 (t 1.23), and on currencies nothing.
Killed
42
FX turns at round numbers
Round numbers neither turn price nor launch it in these years, before costs or after: fading the touch loses 0.02R before costs and 0.15R after, and buying the break loses 0.05R before and 0.18R after.
Killed
43
FX runs on through round numbers, where the stops are
Round numbers neither turn price nor launch it in these years, before costs or after: fading the touch loses 0.02R before costs and 0.15R after, and buying the break loses 0.05R before and 0.18R after.
Killed
44
The currencies that rose most last month rise next month
Under test: no verdict yet.
Testing
45
The dollar falls when foreign rates are above its own
Under test: no verdict yet.
Testing
46
Bitcoin's first half hour of the day predicts its last
On Binance a rising first half hour does not lead to a rising last half hour, before fees or after; the short side makes nothing either.
Killed
47
Gold's and oil's day predicts their last half hour before settlement
The day's move does lean into the last half hour before settlement (before costs t 2.21), but it is worth a few thousandths of a day's volatility, less than one spread; silver's spread of about 8 basis points makes it the worst.
Killed
48
A swept prior-day high or low turns back
A swept prior-day high or low turns back no more often than a random minute of the same session with the same trade: 65.7% of random runs, no better than a coin.
Killed
49
Gold's Asian range breaks at London's open
The bot's idea makes nothing on nine years of gold; its range filter and the clock do not change that.
Killed
50
AUDCAD, AUDNZD and NZDCAD come back within the day: the trade under the grid bots
The crosses do come back: a stretch past the band with RSI at an extreme reverts far more than random entries managed the same way (100% of runs), before costs at t 7.09 in these years and 8.87 in the decade before, every year positive.
Killed
51
Currencies move with the futures hedgers' pressure
Under test: no verdict yet.
Testing
52
A pretrained market model predicts the next bar after its training data ends
Under test: no verdict yet.
Testing
53
A straddle at the US jobs report pays after the real spread
Under test: no verdict yet.
Testing
54
Gold comes back to fill its fair value gaps
A retest of a fair value gap is a better place to enter gold than a random minute with the same trade, on both sides: the closest any ICT idea has come in the lab.
Killed
55
The noise band holds on the index CFDs outside the US
Outside the US the noise band makes nothing on any of the four markets, and less in the second half.
Killed
56
The US indices' day predicts their last half hour
The day's move does not carry into the US indices' last half hour in these years, before costs or after; the hedging flows the paper names are not what hypotheses 39 and 40 picked up.
Killed
57
The commodity crosses revert on hourly bars, where a stretch is worth more than the spread
On the three commodity crosses the rule is the strongest the lab has scored: t 5.68 after costs on audited months, above t 4 in each decade, every pair above zero, above zero at twice the costs, beating every one of 10000 random runs, and before costs on a second feed and in fourteen months the mirror never had.
Killed
58
The commodity crosses' reversion is in Pepperstone's own quotes
The day's mean still clears t 3 and beats every random run, but NZDCAD loses, so the robustness test written above fails.
Killed
59
The commodity crosses' reversion is in OANDA's recorded quotes, and in the years before the rule was made
The reversion is real: in three feeds before costs (t 10.5 here), in every year from 2005 to 2011, and in five and a half years no test had seen, at t 5.89 after today's costs.
Killed
60
The commodity crosses revert on four-hour bars, where a stretch is twice the hourly one for the same spread
The four-hour bars did what they were meant to: a trade's gross since 2015 went from about 2.5 pips to 7.6, and per trade those years made +0.12 (t 2.40) where the hourly rule made nothing.
Killed
61
US500 rises at Europe's open after a sell-off into the US close
Passes, at the broker's own costs: the first rule here to meet all four of goal.md's tests, and one whose gain has gone.
Pass
62
Gotobi's flow starts in Tokyo's small hours: USDJPY long from 03:00 to the 9:55 fix
The holdout points the same way and misses the bar.
Watch
63
Index CFDs revert on hourly bars, where the spread is a fiftieth of a stretch
An hourly stretch on an index does come back more than a random entry does, but by too little, and per trade the fade loses before costs as well as after.
Killed
64
On gotobi days the yen is sold into the 9:55 fix and bought back after it
Passes all four of goal.md's tests, and is working now.
Pass
65
A four-hour change of character pays on US500 and NAS100
The video's best price-action setup, scored as the lab scores everything, does worse than random entries with the same side, stop and target (13.6% of runs).
Killed
66
For day trading, the last two years pick next year's winners better than ten years do
A two-year look back chooses day trades no better than ten years.
Killed
67
A walk-forward search of the Bubble family finds rules that pay in the years after they are chosen
The Bubble family has no edge after costs that a search can find on these markets and charts, 2016 to 2026: what ranks first in one stretch of years earns nothing or loses in the next, and the walk forward loses at t -2.41.
Killed
68
A walk forward over the Bollinger fade, exit at the mean and no stop, finds markets where it pays after the ask
The Bubble's Bollinger fade has no edge after Pepperstone's costs that a walk forward can find on these markets and charts.
Killed
69
The same walk forward on daily bars, where a stretch is many spreads wide
Nothing goes into the Bubble Reversion EA from this.
Killed
70
The fade entered by a resting limit order at the band, not at the next bar's open
Entering at the band does not rescue the fade.
Killed