Verdicts

Method: Widely taught, no single originator. Measured by Edgewise.

We tested the eleven signals most retail traders actually act on: golden and death cross, MACD crosses, RSI 30 and 70 exits, Bollinger breaks, EMA 20/50 crosses, the 200-day reclaim. Entry at the signal bar's close, stop at 2x ATR, target at 2:1 — the trade the practitioner literature actually describes, not a close-to-close direction guess. Against a control of random entries matched to the same volatility. 264 tests. Nothing beat the control.

What we can say. MACD crosses, Bollinger breaks and RSI exits had enough events to detect an edge of roughly 0.10-0.25 ATR per trade - with the stop at 2x ATR, that is a twentieth to an eighth of the risk on the trade. There wasn't one. Those are real nulls. What we can't. The golden cross has fired 63 times in seven years across the ten largest coins. That is not enough to conclude anything, in either direction. We have not shown it fails - we've shown nobody has enough of it to say, and that includes everyone who sells it. Nineteen of our forty-four measurements land here and are excluded from the headline in both directions. One measurement did come back significant, against the golden cross, at one day. We're reporting it as a false positive rather than a finding: it holds under only two of six control specifications, and 95.2% of those trades never reach either barrier, so the cell measures a one-day return rather than the trade.

What would change this verdictMore history, or an asset class with genuinely independent observations. Ten crypto pairs behave like about two.

Measurement: SIGNAL_AUDIT_RESULTS.md · SIGNAL_AUDIT_PLAN.md
If we have mis-specified any of these signals, tell us and we will re-run it.

Method: Widely taught, no single originator. Measured by Edgewise.

The most-attempted problem in retail trading, and the one we spent longest on. 408 walk-forward tests: next-bar direction, drift, momentum, volatility regime, post-extreme conditioning. Nothing beat simply assuming the market's own drift. Two results survived multiple-comparison correction and both were negative - the signal pointed the wrong way.

We also tried to fix it rather than kill it: three recalibration methods on the engine's own output, all worse out of sample than leaving it alone. The honest reading is that next-bar direction is close to a coin flip and our engine was already near the ceiling of what is there. The edge, where any exists, is in levels and conditions - not in guessing the next candle.

What would change this verdictA conditioning dimension we have not tried, beating drift out of sample rather than in it.

Measurement: crypto_direction_findings2.md · crypto_direction_prereg.md
If we have mis-specified anything here, tell us and we will re-run it.

Method: Widely taught, no single originator. Measured by Edgewise.

46% against a 50.1% control · 95% CI 45.3–46.7

The claim is that taking out the previous day's extreme is a momentum trigger - the level breaks, so price keeps going. It does the opposite. Over the next 4 hours it continued 46.0% of the time (interval 45.3-46.7) against a 50.1% random-bar control. Over 24 hours, 46.9% against 49.3%.

That is worse than a coin flip, on one of the most-taught triggers in day trading. We also tested the version people reach for when the plain one disappoints - filtering for the strongest momentum. The top 1% of momentum sweeps is the most negative of all: -6.9pp on 759 occurrences. Squeezing harder made it worse, which is usually the sign that there was nothing there to squeeze.

What would change this verdictA market or session where the continuation rate clears its own control, holding across sub-periods rather than in one stretch.

Measurement: PDH_PDL_FINDINGS.md
If we have mis-specified anything here, tell us and we will re-run it.

Method: Lars Kooistra. Measured by Edgewise.

74.5% against a 50.3% control · n=9,575 · 95% CI 73.6–75.3

The claim is that price leaving a range edge is usually not a break - it pokes out and closes back inside. We froze the rule and ran it on 10 pairs across 3.5 years: 9,575 excursions, 74.5% closed back inside, against a 50.3% control that travels the same distance past a level with no range in it.

The part worth trusting isn't the headline, it's the stability. Measured a year at a time: 74.7%, 75.3%, 73.8%, 73.9%. Four separate years inside a point and a half of each other. The control deserves as much attention as the rate. An earlier version of it matched the deviation limit but not the depth of the excursion, so it triggered on any poke past the prior close - shallow pokes come back almost automatically - and it read 72.9%, an edge of only 1.6 points. Matching the depth moves it to 50.3%. A 22-point swing from one change to the control means the control is half the number, not a footnote to it. What this does not say: anything about direction afterwards. It answers one question - was that poke a break - and people read the second thing into it constantly.

What would change this verdictA year landing outside the band those four years sit in, or the rate splitting apart when broken down by pair. Or a control matched on something else that moves it again.

Measurement: RANGE_HOLD_CONTROL_FINDINGS.md · ranges.py
If we have mis-specified the rule, tell us and we will re-run it.

Method: @bojan_618 (the pattern), as articulated by @Wyckoff_Insider. Measured by Edgewise.

Three bars, the third making a higher low and closing up, where the swept low is a fresh 24-day extreme and the sweep came on heavy volume. Traded (in at the trigger close, stop at the swept low, target the same distance again), our research run returned +0.173R per trade over 79 cases on 20 markets that played no part in choosing it, range -0.011 to +0.353. That run judged each sweep's volume against the market's whole history, sweeps still to come included, which no live chart can do. The app judges it against the 120 bars before the sweep; measured that way on the same 20 markets it returned +0.066R over 146 cases, range -0.074 to +0.204, which includes losing money. The app shows the pattern and no profit figure.

WE WITHDREW THE CLAIM THAT USED TO SIT HERE. It read 'holds 10 days 70.4% of the time against a 35.6% random-bar control'. Two things were wrong with it. First, holding for ten days has no entry, no stop and no exit, so it can be neither won nor lost — it is not a result a reader can act on or check. Second, the control was wrong twice: a constant random seed meant every table we produced drew one shared sample of the baseline (35.6% -> 42.4%), and the control level sat a flat 1 ATR from the price while real events sat a median 1.58 ATR away, so 89% of them had further to travel than the thing they were measured against. Matching that distance moves the baseline to 61.4% and the advantage from +34.8 points to +9.0. The 70.4% itself never moved — only what it was compared with. At the corrected baseline the interval no longer clears it, which is why this is no longer published as a rate at all. WHAT REPLACED IT was a trade that can be scored: +0.173R average, range -0.011 to +0.353, and the same shape with the volume condition removed returns -0.015R. THAT FIGURE IS NO LONGER A PROFIT CLAIM EITHER. Re-measured with the volume rule the app actually applies, the trade is +0.066R with a range of -0.074 to +0.204, so it is not shown as a profit anywhere in the product.

When this gets decidedNothing is scheduled to decide it. No forward ledger runs on this rule; see what would change this.

What would change this verdictA forward record on the rule the app actually applies. At about 23 trades a year across 20 markets it would take many years to tell +0.066R from zero, so none is running: the earlier bojan ledger scored the withdrawn hold measure and has been retired. Re-running the same history would not move this.

Measurement: bojan_publish.py · BOJAN_FINDINGS.md · BOJAN_UNFILTERED_FINDINGS.md · RANGES_BOJAN_REMEASURE.md · bojan_trailing.py
Where the rule is vague, the vagueness may be his articulation rather than the original pattern. If we have mis-specified it, tell us and we will re-run it.

Method: @bojan_618 (the pattern), as articulated by @Wyckoff_Insider. Measured by Edgewise.

We asked this one four times on the same history, and it shrank every time: +17.5pp on 10 pairs, +8.9pp on 30, +6.1pp on 30 with corrected ranges, +3.2pp on the top 10 by liquidity.

A number that falls every time you look harder is telling you something. The forward ledger that was frozen at +8.9pp published nothing and has been retired. We are publishing this instead of the number, because a rate we don't believe is worse than no rate.

What would change this verdictA forward record, gathered after this was written, on markets that played no part in any of the four askings. Re-running the same history a fifth time would not move it.

Measurement: BOJAN_FINDINGS.md
Where the rule is vague, the vagueness may be his articulation rather than the original pattern. If we have mis-specified it, tell us and we will re-run it.

Method: @Wyckoff_Insider. Measured by Edgewise.

This is not a verdict on whether it works. We couldn't get far enough to have one. We archived 186 transcripts from 208 videos - 691,008 words - from a practitioner who uses Gann time windows as the thing that licenses his largest positions. The concept is real and central: it gates both entries and standing aside.

But the derivation is never given. Searching the entire corpus for the vocabulary any Gann time method must use: solstice 0, equinox 0, anniversary 0, lunar 0, vibration 0, degree 1, square of 1, mercury 1. He shows windows already drawn on charts and says plainly that explaining the computation wouldn't land for most viewers - so he doesn't. Without a rule, we can't compute a window, so we can't score one. One part is testable and we intend to: he names cycle lengths of 4, 8, 12, 16 and 80 years, and the 4-year self-anchors on the halving, which makes his last-quarter-of-the-cycle claim measurable without needing anything further from him.

What would change this verdictA stated anchor and construction rule. Then it becomes a measurement.

Measurement: methodology.md
If the derivation is stated somewhere we have not looked, tell us and we will measure it.

Method: @bojan_618 (the pattern), as articulated by @Wyckoff_Insider. Measured by Edgewise.

The mirror-image shape at a fresh 24-day high. The app marks it at any volume and shows the sweep's volume as a label, not a condition. Scored as a trade on light-volume highs it returned +0.036R over 399 cases, range -0.056 to +0.126, which covers zero; with no volume condition it returned -0.011R. We publish it as a detected pattern and not as a profitable trade.

THIS CODEBASE ONCE RECORDED THE HIGH AS DEAD, and that was our error: it had been measured with the LOW's volume filter applied to it. Measured, the high formation is worth about +0.13R more on light volume than on heavy (+0.036 against -0.098), and that difference survives a noise floor built by destroying the link between pattern and outcome while keeping the bars, the volume and the volatility clustering intact. A light-volume filter was then tried as the high's rule and dropped: it kept too few of the clear tops the shape catches. So the app marks every high and carries the volume as a label, and the gradient stays a research finding, not a rule. The high on its own is not a trade and no copy here says it is.

What would change this verdictMore cases. 399 backtested cases with an interval covering zero cannot settle whether this makes money; they can only establish that the shape is there and that light volume does better than heavy on the high side.

Measurement: bojan_publish.py · BOJAN_FINDINGS.md
The volume condition on this side is ours, not his. If we have mis-specified it, tell us and we will re-run it.

Method: Edgewise. Measured by Edgewise.

88% against a 89% control

Not a claim anyone made - a question we get asked, and one of the few with a clean answer. At a 0.50 x ATR stop you survive 88% of trades that would eventually have won - and a driftless random walk with the same volatility gives 89%. The pair is the finding. Tighten to 0.25 x ATR and survival falls to 64%, meaning 36% of your winners were trades you were taken out of; widen to 0.75 x ATR and it is 96%.

The number is unusually stable: 87-91% across six assets, 85-91% across eighteen sub-period cells, 89-92% on the short side. Now the part most people would leave out. This is mechanical, not an edge - the simulated column sits on top of the observed one at every asset. It cannot decay, because there is nothing in it to decay. It is arithmetic about how far price wanders, and it tells you the price of a tight stop rather than giving you a way to beat the market.

What would change this verdictNothing. That is what makes it different from everything else on this page, and why it is worth less than it looks.

Measurement: stop_survival_findings.md
If we have mis-specified anything here, tell us and we will re-run it.

Method: Edgewise. Measured by Edgewise.

This one is ours, and it is the only thing that survived 1,562 measurements. Breakouts are sold as the start of a move. On equity indices the opposite happens: after a new 20-day high, the next day's move is smaller than normal - smaller in both directions, not smaller upside. The market goes quiet.

Measured on the first 70% of history and then confirmed on the last 30%, which the search never touched: 0.811x normal, p = 0.001. Three consecutive up closes does the same thing (0.773x, p = 0.003), and US ETFs agree (0.869x, p = 0.039). It says nothing about direction - only about size. That makes it useful for position sizing, for how likely a stop is to be touched, and for what short-dated option premium is worth the day after a breakout. What we could not confirm: the mirror image in crypto, where the same events looked like they widened the move. It was the more exciting half and it failed out of sample (p = 0.153). It is recorded as a negative so nobody revives it.

What would change this verdictThe effect failing to appear in a fresh year, or turning out to be one index carrying the rest.

Measurement: MOVE_SIZE_FINDING.md
If we have mis-specified anything here, tell us and we will re-run it.

Method: Edgewise. Measured by Edgewise.

57.9% against a 50.6% control · n=5,595 · 95% CI 56.6–59.2

When crypto trades below its own 24-hour low, it closed higher over the next four hours 57.9% of the time (95% interval 56.6-59.2) against a 50.6% control, across 5,595 windows. Above the 24-hour high, it went the other way: 45.0% against 50.5%. That is the strongest in-sample number in this entire research programme.

We are not publishing it as a verdict. On 4 August 2026, before any live result existed to look at, we wrote down the rule that decides it: a forward record, running on live data, with three pre-registered looks at 500, 1,500 and 4,000 scored windows. Nobody - including the founder - sees the ledger's contents before a look. Everything on this page that says Holds was measured on history. History is where every backtest looks good. This is the one where we said in advance what would have to happen, and then had to wait.

When this gets decidedThe first pre-registered look, at 500 scored windows, due around late October 2026. Nobody - including the founder - sees the ledger before it.

What would change this verdictThe first pre-registered look. Whatever it says.

Measurement: PDH_PDL_FINDINGS.md · reversal_forward_decision.md
If we have mis-specified anything here, tell us and we will re-run it.

Method: Ariel 1987; Etula, Rinne, Suominen & Vaittinen 2020. Measured by Edgewise.

Buying the S&P over the turn of the month earned +41.9bp a turn more than a matched random hold, 95% [+14, +65], across 216 turns from 1993 to 2010 — the years the papers studied. On the untouched years since it earned LESS than random three-day holds in the same months: -4.7bp over 93 turns to 2018, interval [-35, +18].

This is not a null and we are not calling it one. The effect was really there in the era it was found in, at a floor the test could resolve, and it is gone in the fifteen years since. A rule that worked and stopped working is a different thing from a rule that never worked, and the difference matters to anyone still trading it. The holdout was opened once, under a pre-registration written before the run.

What would change this verdictA live forward record. The effect died on a holdout that ended in 2018; the years since have been looked at once and decide nothing. If it is quietly working again, only a record kept from today forward can show that, and we are not claiming either way.

Measurement: GROUP5_FINDINGS.md · GROUP5_PREREG.md
If we have mis-specified the window, tell us and we will re-run it.

Method: Lucca & Moench 2015. Measured by Edgewise.

Stocks rose into scheduled Fed announcements from 1994 to 2011: +50.7bp a meeting above the same-year base, 95% [+22, +91], over 137 meetings. On the untouched years since, 60 meetings to 2018, the excess is +9.6bp with an interval of [-27, +39] at a floor of 29bp.

Half a percent a meeting, for seventeen years, and then not. The drift was real and it is over. A placebo run on all non-meeting windows in the same years returns +15bp against the +51bp on meeting windows, which is what says the original effect was about the meetings rather than about the decade. The holdout was opened once, under a pre-registration written before the run.

What would change this verdictThe same. A pre-registered forward record from today, or a re-run on a longer calendar than the one we could parse. If the drift returns we will say so here.

Measurement: GROUP5_FINDINGS.md · GROUP5_PREREG.md
If we have mis-specified the window, tell us and we will re-run it.