FAILED

Moon phases (15-day window around the new moon)

Returns in the 15 days around the new moon are about double those in the 15 days around the full moon (Dichev & Janes, 2003), a difference on the order of 9.44% a year, present across all major US indices over roughly 100 years and in almost every index of 24 other countries.

Measured in crypto — Binance spot · 540 pairs, delisted ones included · 0.2% per round trip

Net per trade

−1.03%

after fees

The fee is charged on both legs: every trade pays to open and pays to close.

Died at

placebo

random dates pay the same or more

Worst drawdown

−91%

867 days underwater

How far the account fell below its own best previous moment.

Timeframe

1 day

The time grid this was measured on. The same technique on a coarser or finer grid is a different measurement, and can earn a different verdict.

invariance
costs
placebo
benchmark
out of sample
multiple testing

Equity curve

Risking 1.0% per trade, marked to market every day — not only when a trade closes. The line at 1.0 is the capital it started with.

At its worst the account was worth 91% less than its own best previous moment, and it spent 867 days below that peak.

The technique against chance

Same number of trades, same holding time, same assets — only the dates were drawn at random.

The average trade and the typical one

The mean sits above the median: the lottery signature — the typical trade loses and a rare handful pays for everything.

What was measured

We coded the rule exactly as it is described and let it trade on its own from 2017-08-17 to 2026-07-28. It found trades in 540 coins, 25,517 in total. Each one enters and exits at the price that existed on that day — the program never sees what comes next, which is the most common mistake people make testing strategies at home. Wherever the description was ambiguous we took the reading least favourable to the technique; whatever was left open is stated in the hypothesis, at the foot of this page.

How many of those trades actually count

Only 29 of the 25,517. Cryptocurrencies rise and fall almost in unison, so the same rule fires on dozens of coins on the same day — and that is one bet repeated, not dozens of different bets. Counting them all separately is the trick that makes a bad test look impressive.

What it paid, before any deductions

-0.83% per trade. Look at the sign: the average is already negative before any fee is taken out. The controls below still hold, but here they are not knocking down an edge — they are confirming there never was one.

Why it did not pass — it died here · against randomly drawn dates

We ran the whole thing again entering on randomly drawn dates and changing nothing else: same number of trades, same holding time, same assets. Chance did as well or better. That means the signal was not picking the moment: any moment would have done the same.

What would have happened to the money

Risking 1.0% of capital per trade — the most widely taught rule — across the whole period: The capital would have ended at 0.47× what it started with. At its worst the account was worth 91% less than its own best previous moment, and it spent 867 days below that peak. In 8 of the 12 paths tested, the account fell to less than half of what it started with.

The account would not fit every signal

65% of the trades had to be turned down: by the time they appeared, the money was already tied up in other positions. That matters because the average return per trade the technique claims is computed over trades that nobody could have taken all of.

What this result does NOT say

For an edge to be assertable here it would have to reach 4.54% per trade — that is the size that survives this archive's multiple-testing correction, and it rises as the archive grows. So FAILED means “we found nothing above that size”, and never “it cannot possibly work”. The instrument is far more sensitive than that: on synthetic data, with a clean effect, it separates from 0.69% upwards. The distance between the two numbers is the price of a real market and the price of publishing many claims. The difference matters, and it is the rule of this house: the card confronts the claim, never the person who made it.

What this card does not measure

Every verdict holds for the conditions it was measured under. These are this card's — and outside them the result does not apply.

One market, one universe: Measured on 540 spot cryptocurrency pairs, delisted ones included. It says nothing about futures, equities or indices, nor about how the same technique behaves in another market.
One window of time, not every window: The measured period runs from 2017-08-17 to 2026-07-28. A market moves through regimes, and a technique can work in one and fail in another — the card measures the regimes that fit inside this window, not the ones still to come.
One cost structure: The cost charged is 0.10% per leg, in and out. Anyone paying more than that gets a worse result, and anyone paying less gets a better one — the verdict holds for this fee.
One exit rule: The trade was closed by: the technique itself (held until the opposite signal). The same entry measured with a different exit is a different strategy, and can earn a different verdict — it happens in this archive.
The number of trades is not the sample size: There are 25,517 trades, but only 29 independent market episodes: a single move fires the technique across dozens of assets at once, and counting those as separate observations inflates any result. It is the smaller number that governs the arithmetic. With 29 episodes, what the data supports is a range from -4.78% to +3.12% per trade — the published average is the centre of it, not the exact measurement.
Daily bars: Measured at the daily close. Nothing here measures what happens inside the day, and an intraday technique is not auditable with this data.

Numbers and reproducibility

The six controls

control	status	t	threshold	episodes
invariance	passed	—	—	—
costs	inconclusive	-0.54	2.05	29
placebo	failed	-0.51	2.05	29
benchmark	failed	—	—	—
out of sample	inconclusive	—	—	—
multiple testing	inconclusive	—	—	—

invariance25517 signals across 750934 barsconcentration: +49.47% of the profit sits in the top 5% of trades — with the dates shuffled, +46.15% (fails above +50.00%)
costsgross -0.831% · cost 0.200% · net -1.031% (t=-0.54) · the range runs from -4.978% to +2.916%
placeboactual -1.031% · placebo +0.234% · excess -1.265% ± 2.491% (t=-0.51 against a threshold of 2.05, 29 real groups, 1,275,850 sham dates, draw error ±0.026%)
benchmarktechnique -1.03% · buy and hold (same horizon) +0.21% · excess -1.24%
out of sampleasset half A: -1.273% (t=-0.60, 29 episodes) · asset half B: -0.810% (t=-0.41, 29 episodes) · liquid half (>= US$ 2,066,613/day): -1.122% (t=-0.59, 29 episodes) · illiquid half: -0.931% (t=-0.44, 28 episodes) · period 1/4 (2017-08-17 a 2022-04-06): +5.239% (t=1.85, 16 episodes) · period 2/4 (2022-04-24 a 2023-10-10): +1.665% (6 episodes — too small, does not count) · period 3/4 (2023-11-06 a 2025-03-04): -4.153% (5 episodes — too small, does not count) · period 4/4 (2025-03-22 a 2026-07-07): -6.495% (5 episodes — too small, does not count)
multiple testing1 variation(s) tested · t=-0.54 across 29 episodes (equivalent to t=-0.51) · p≈0.6086 · false positives expected by chance ≈ 0.61

Equity — outside the six controls, and here is why

The t of the trade series is invariant to bet size: 0.5%, 1% and 3% agree to the sixth decimal. Nothing here moves the verdict — it moves what the account would have lived through.

risking 1.0% per trade: ×0.47
worst drawdown from the peak: −91%
days below the previous peak: 867
signals refused for lack of capital: 65%
paths where the account halved (out of 12): 8

Reproducibility

period: 2017-08-17 to 2026-07-28
assets that traded: 540
variations tested before this one: 1
gross per trade: −0.83%
net per trade: −1.03%
exit rule: the technique itself (held until the opposite signal)
median duration bars: 15
mean duration bars: 14.9
max duration bars: 15
fee per leg: 0.001
episode days: 112
seed: 20260728
window days: 15.0

Binance spot klines (delisted pairs included) · collected from 2026-07-27 23:31 to 2026-08-03 10:49 · 540 assets · 750,934 bars · 2017-08-17 to 2026-07-28

Hypothesis, filed before the result

I expect it to yield SOMETHING gross and to die at cost. It is the pattern of the four arenas of 2026-07-27 — a signal that is real, measurable and smaller than the fee — and I see no reason for the Moon to be different. What makes this prediction risky, rather than a comfortable guess: there is published assertion IN FAVOUR (Dichev & Janes, 2003: ~9.44% a year of difference between the new-moon window and the full-moon one, over ~100 years of the US and 24 other countries; Yuan, Zheng & Zhu, 2006: ~4.2% a year across 48 countries, significant at 1%) and published refutation AGAINST (Kim & Shamsuddin, 2023: under extreme bounds analysis the lunar phase is highly fragile, likely a product of data mining). All three measure GROSS return; none deducts cost, and that is exactly where I bet it dies. What would contradict me, said before seeing: a gross edge indistinguishable from zero knocks down the first half of the prediction and vindicates Kim & Shamsuddin — the anomaly never existed, rather than existing and not paying. And an edge that SURVIVES cost, the placebo and multiple testing knocks down the whole prediction; in that case the finding belongs to the Moon, not to me. The arithmetic of cost, so that the prediction has a number: the technique makes one round trip per lunar cycle, about 12.4 a year; at 0.1% per side, the drag is on the order of 2.5% a year.

the filing date is not in this audit's record

The original, as it was filed

Espero que dê ALGO bruto e que morra no custo. É o padrão das quatro arenas de 2026-07-27 — sinal real, medível e menor que a taxa — e não vejo razão para a Lua ser diferente. O que torna esta previsão arriscada, e não um chute confortável: existe afirmação publicada A FAVOR (Dichev & Janes, 2003: ~9,44% ao ano de diferença entre a janela da lua nova e a da cheia, em ~100 anos de EUA e 24 outros países; Yuan, Zheng & Zhu, 2006: ~4,2% ao ano em 48 países, significante a 1%) e refutação publicada CONTRA (Kim & Shamsuddin, 2023: sob extreme bounds analysis a fase lunar é altamente frágil, provável produto de data-mining). Os três medem retorno BRUTO; nenhum desconta custo, e é exatamente aí que aposto que ela morre. O que me contraria, dito antes de ver: vantagem bruta indistinguível de zero derruba a primeira metade da previsão e dá razão ao Kim & Shamsuddin — a anomalia nunca existiu, em vez de existir e não pagar. E vantagem que SOBREVIVA ao custo, ao placebo e aos múltiplos testes derruba a previsão inteira; nesse caso o achado é da Lua, não meu. Aritmética do custo, para a previsão ter número: a técnica faz uma ida e volta por ciclo lunar, cerca de 12,4 por ano; a 0,1% por lado, o arrasto é da ordem de 2,5% ao ano.

Pre-registration exists to keep prediction apart from rationalisation: written after the number, every hypothesis is right.

This claim has been audited once — there is no history to compare against.