Seven statements that get passed around constantly in trading circles. For each one there is a number that either carries it or does not.
The answer is almost never "myth." It is usually "smaller than claimed," "true but you cannot act on it," or "never properly tested." Those categories are more interesting than a debunk — and harder to write, because at the end you have to hold your own number to the same standard.
We do that in the first section. It rests on our own data, and it comes with a limit we state before we state the result.
MAX PAIN — THE STRIKE TRAILS THE PRICE
The claim: at expiry, price drifts toward the strike where the most options expire worthless. Market makers have an interest in it, so it happens.
What the measurement cannot do
We looked at 70 settled Deribit BTC expiries between 28 May and 5 August 2026. That is 70 calendar days and an average of exactly one expiry per day — Deribit expires daily.
Three limits you need before any number means anything:
- Ten weeks is not a cycle. Bitcoin fell from about 73,600 to 64,700 across this window, down 12 percent. What follows describes a falling market. Whether it looks the same in a rising one, we do not know.
- 60 of the 70 are same-day contracts. On a contract expiring today, open interest builds around today's price. A daily and a quarterly are not the same object, even though both are called an expiry.
- There is exactly one quarterly in the sample. Anything we might say about large expiries would rest on a single observation. So we do not say it.
Leave those three lines out and the same data supports a much sharper story. That is precisely the error this article describes in others.
The numbers
| Metric | Value |
|---|---|
| Observations | 70 expiries (28 May – 5 Aug 2026) |
| Median distance price ↔ strike | 1.80 % |
| Mean distance | 2.46 % |
| Within ±1 % | 24 % |
| Within ±2 % | 56 % |
| Price above the strike | 36 |
| Price below the strike | 34 |
The last two rows are the answer, and this time it is tested rather than asserted. A two-sided binomial test against a fair coin returns p = 0.90. There is no directional pull in this data. Price lands above as often as below, a median of just under two percent away.
The usual objection, and why we cannot answer it
"The effect needs enough open interest, otherwise it does not work."
The 26 June 2026 expiry carried 153,901 contracts and 9.25bn USD notional — by far the largest in the sample. Its deviation was −14.1 percent, max pain at 70,000 against a price of 60,097. The worst miss in the whole window, at the point where the supposed pressure was highest.
That reads like a strong line. It is not one.
It is one observation. By our own standard — below 30 independent cases a distribution is an anecdote — that is not enough to refute the objection. It is enough not to confirm it. Nothing more.
Split the sample into three groups by open interest and it gets uncomfortable:
| Group | n | Median distance |
|---|---|---|
| Low open interest | 23 | 1.13 % |
| Medium | 23 | 2.76 % |
| High | 24 | 2.03 % |
The contracts with the least open interest sit closest to the strike. That does not support the myth — it shows the groups cannot be compared cleanly. The small group is almost entirely same-day contracts whose strike hugs the current price anyway. You are not comparing two states of one market here, you are comparing two instruments.
The daily snapshots show which one moves first
More useful than the settlement days are the daily observations of contracts still open. There you can watch which of the two moves first.
The September 2026 expiry:
| Date | Max pain | Price |
|---|---|---|
| 28 May 2026 | 75,000 | 74,256 |
| 26 Jun 2026 | 75,000 | 59,348 |
| 6 Aug 2026 | 70,000 | 64,515 |
Price fell 20 percent inside a month. The max pain strike stayed unchanged at 75,000 and only followed weeks later.
If max pain pulled the price, price would have travelled toward 75,000. It went the other way, and the strike came after it.
That is mechanically unsurprising. The strike is computed from open interest, and open interest gathers wherever the price currently sits. Near expiry the two are close because the contracts that define the strike were written at current prices. That closeness is the construction of the metric, not its result.
To be honest about the rest: on short-dated contracts the strike travels further than the price, on long-dated ones the reverse. Across eight observed series the strike moved further in two. So the blanket sentence "the strike chases the price" is not carried by our data. What is carried is the weaker and still decisive one: max pain does not lead the price.
The daily series also only starts on 28 May 2026 and covers twelve expiries, nine of them with depth. That is thin, and we are saying so.
Verdict: not usable as a price target — in this window, in this market phase. None of this says options flow is irrelevant. Gamma and dealer hedging are real. "Max pain pulls the price" is a different claim, and it does not hold.
THE YEN CARRY TRADE — INFLATED BY AN ORDER OF MAGNITUDE
The claim: one to two trillion dollars of yen-funded positions sit on top of global risk markets. If the Bank of Japan raises rates, the whole thing collapses.
Circulating figures run from one trillion, through "505 percent of Japanese GDP," up to twenty trillion in blog posts.
The Bank for International Settlements ran the numbers for August 2024 and arrived at roughly ¥40 trillion, about 250 billion dollars — stating explicitly that such positions are hard to measure and that the estimate is more likely too low than too high. UBS put the dollar-yen portion at its peak at a minimum of 500 billion, of which around 200 billion was unwound within two to three weeks.
Between 250 billion and 20 trillion there is a factor of eighty.
What this deliberately does not say: the mechanism is real. August 2024 happened, and the Nikkei had its worst day since 1987. Only the magnitude the story is told with has no relationship to the figures from the institutions doing the measuring.
And the BIS calls its own number an estimate. Anyone quoting it says that too.
Verdict: mechanism correct, number an order of magnitude too large.
RATE CUTS AND EQUITIES — THE CONDITION IS ONLY KNOWN AFTERWARDS
The claim: when the Fed cuts, stocks go up.
This is where it gets methodologically interesting, because credible houses reach opposite numbers depending on which cycles they take into the sample.
Goldman Sachs examined ten cutting cycles since 1984, four of them in a recession. In the recession cases the S&P 500 fell 11 percent three months after the first cut and sat 10 percent down after six. In the others it rose 5 and 7 percent respectively.
Ned Davis Research finds 80 percent positive twelve-month windows from 1974 onward, averaging +15 percent — while excluding 1974, 1980 and 1981, because a recession was already underway there. A third study across nine first cuts finds an average interim decline of 20.5 percent.
All three are computed correctly. They differ only in which episodes count as comparable.
And even if you settle on one, the real problem remains: the difference between +7 and −10 percent depends entirely on whether a recession follows. You do not have that information at the moment you decide. A rule that needs an unknown quantity as input is not a rule.
Then the sample: ten cycles since 1984, four of them recessionary. Four observations for the circumstance that decides the entire difference.
Verdict: true, but you cannot act on it.
THE FOUR-YEAR CYCLE — THREE OBSERVATIONS
The claim: halving, fewer new bitcoin, rally, crash, long quiet. Every four years.
The pattern described 2012, 2016 and 2020. Three clean cycles. Three observations cannot establish a regularity, however convincing the chart looks.
The fourth cycle already breaks the pattern at a hard point nobody can argue about: according to Kaiko, Bitcoin reached new all-time highs before the halving for the first time in its history. In all three prior cycles the high came afterwards.
The mechanism is diluting too. In 2024 only half as many new bitcoin arrived as before — 0.85 percent of circulating supply instead of 1.7, the smallest cut so far. Around 94 percent of all bitcoin have already been mined. Every further halving matters less than the last, by arithmetic alone.
What is worth noticing is how both camps responded. The 6 October 2025 high near 126,200 dollars fits the timing well, and the decline to around 63,000 by mid-2026 does too. The depth of the fall does not: earlier cycles ended in declines of 77 to 85 percent. The defenders point at the timing, the critics at the depth. Both are fitting their model to the data after the fact.
Verdict: too few observations for a statement — in either direction.
"UPTOBER" AND OTHER STRONG MONTHS — ONE MONTH WINS OUT OF TWELVE, EVERY TIME
The claim: October is historically a strong month for Bitcoin.
The problem is not in the number, it is in the selection. Bitcoin has roughly thirteen Octobers with usable price data. Thirteen observations — that alone should slow anyone down.
But something else decides it. There are twelve months. Run all twelve against the question "which is the strongest?" and one always wins, even when there is no difference between them at all. The best of twelve random candidates looks good by construction.
So the error is not that October was strong in hindsight. The error is that only the winner gets shown and the contest it emerged from is left out. Testing twelve hypotheses and reporting one is not testing one hypothesis.
The same applies to weekdays, to trading hours, and to every other slicing of the calendar. The finer you cut, the more certainly you find a pattern — and the less it means.
The control is simple and almost never supplied: how strong is the best month in data that contains no strong month by construction? Only that comparison turns the October finding into a statement.
Verdict: a selection effect, not a calendar effect.
"90 PERCENT WIN RATE" — HALF THE ARITHMETIC IS MISSING
The claim, in countless variants: this rule is right nine times out of ten.
A win rate says how often something works. It says nothing about how much it brings in when it works and how much it costs when it does not. Both together decide the outcome, and only one of them gets named.
Nine small gains and one large loss produce a 90 percent win rate and a loss overall. The other way round, a rule that is wrong in two of three cases can be very good if the one hit comfortably outweighs the two misses.
This is why high win rates are so popular in advertising: they can be pushed almost arbitrarily high by taking gains early and letting losses run. The rate goes up, the result goes down.
The question that belongs with it is always the same: what does the average gain look like against the average loss — and how bad was the single worst case? Anyone quoting a win rate without those numbers is showing one half of the arithmetic.
Verdict: not a myth, an incomplete figure. And the most common one in the entire field.
COINBASE PREMIUM — NEVER PROPERLY TESTED
The claim: the price gap between Coinbase and the offshore venues shows US institutional demand and leads the price.
Here the research result is itself the finding. For this indicator no clean study on data outside the development period could be found. What can be found is dozens of near-identical articles passing around the same two or three individual cases: a negative stretch before a correction, a positive turn before a stabilisation.
That is not evidence. That is an individual case that acquired the appearance of evidence through repetition.
A note on our own account: "we found nothing" is a result, but a weaker one than a measurement. It can mean no study exists — or that we failed to find it. Anyone wanting to turn this section against us needs one link.
The mechanism is plausible; the price gap does measure something real about order flow. Whether that something leads the price is an open, testable and apparently never seriously tested question.
Verdict: untested. Not disproven, but not established either.
THREE FAMILIES OF ERROR
The seven cases come down to three errors, and it is worth keeping them apart.
1. The comparison number is missing
A number gets quoted without the number you would have to hold it against.
The carry trade lacks the official estimate beside the blog figure. Rate cuts lack the split by recession — and the note that you do not have that split in advance.
The most systematic version shows up with market drift, and for that we are using a study of our own, because it failed.
The question: does a quiet phase in Bitcoin resolve upward or downward?
The first attempt was wrong by roughly the size of a bear market. A 400-day window produced five completed episodes, four of them downward, a median of −6.5 percent after 30 days. That reads like an answer. Bitcoin fell about 40 percent across the same window — the number described the market, not the episodes.
The proper study, with the definition fixed beforehand, found 47 episodes going back to 24 April 2011. Its 30-day median is −0.03 percent.
| Horizon | Raw | Drift over the same windows | Excess |
|---|---|---|---|
| 30 days | −0.03 % | +3.80 % | −1.62 % |
| 90 days | +13.06 % | +10.78 % | −5.24 % |
| 180 days | +31.46 % | +27.43 % | −5.93 % |
The 180-day row is the instructive one. Plus 31 percent looks like an edge and is less than the market delivered anyway over the same stretch. Only 44.7 percent of episodes produced a positive excess at that horizon — fewer than half.
Now the part that matters more than the result. Anyone checking the arithmetic will stumble: 31.46 minus 27.43 is +4.03, not −5.93. At 90 days the sign even flips.
The reason: the excess is not the difference of two medians. It is computed per episode, each against the drift in its own calendar window, and only then is the median taken of those differences.
Those are two different quantities. Taking a conditional median, subtracting an unconditional one and calling the result an excess answers a different question from the one intended. Only the paired calculation says whether the condition contributed anything.
The study also fails its own core claim: the change in volatility rank afterwards is roughly zero. Under this definition, quiet phases are followed by neither expansion nor further compression. A placebo test against randomly placed episodes of the same duration returns p = 0.22 to 0.30. Noise. The placebo arm was built before the analysis, not after — that ordering is the difference between a control and a rationalisation.
Four limits belong to the result:
- 47 episodes are not 47 independent observations. The forward windows overlap, so the effective count is lower. That sits in the study's own output, not in the small print.
- "Quiet" is softly defined here — below the rolling median, so roughly half of all days. This is not a rare state. Whether an unusually quiet market behaves differently is untested.
- The market-phase question cannot be answered. 34 episodes lie above the 200-week average, three below. You do not report from three cases, so it was not reported from.
- The definition was fixed before counting. No threshold search, no picking the best of several.
What survived is narrower than the headline the +31 percent would have produced: a quiet phase announces a larger move. It says nothing about direction.
2. Too few independent cases — or too many hypotheses tested
The four-year cycle rests on three observations. The rate cuts rest on four recessionary cases. The strong October rests on thirteen Octobers, selected from twelve candidates.
The difference between days and cases matters. Days inside the same trend are highly correlated; a thousand days can be three independent events.
3. The metric contains its own answer
Max pain is computed from open interest, which gathers around the current price — the closeness on expiry day is built in, not measured. A win rate can be pushed up by cutting gains and letting losses run, without anything improving in the result.
Where a metric follows its own input, agreement is not proof.
THE FOUR QUESTIONS
Four questions follow, for any claim shaped like "after X came Y":
What was the development over the same window with no condition at all? Without that value the statement cannot be disproven.
How many independent cases are behind it — and how many variants were tested before this one was reported? Below eight cases a distribution is an anecdote. And the winner of twelve candidates is not a finding.
Was the time horizon chosen before or after the result? If only one horizon is shown, ask for the others. The most interesting value is usually the one that is missing.
Could the metric be following the thing it is supposed to explain? If so, agreement is not proof.
These four questions cost two minutes. They handle most of what passes for analysis — and, applied honestly, part of your own work too. The max pain section above got three caveats longer once we asked them.
Sources: Bank for International Settlements, quarterly review on the August 2024 market event · UBS, estimate of dollar-yen carry positioning · Goldman Sachs, analysis of ten Fed cutting cycles since 1984 · Ned Davis Research, twelve-month equity windows from 1974 · Kaiko, cycle comparison of all-time high versus halving date · Own analysis: 70 settled Deribit BTC expiries, 28 May – 5 August 2026 · Own study: low-volatility episodes in BTC, 47 episodes from 24 April 2011, completed 30 July 2026. As of 6 August 2026.
⚠️ Prices and market figures as of 6 August 2026. Statements about ongoing market phases go stale.