Bitcoin Seasonality: Is There Really a Bad September?

Two statements, both computed from our own closes, both true, and they pull in opposite directions.

First. Since September 2011 bitcoin has completed 178 calendar months. Sorted by calendar month, October sits at the top with a median gain of +11.0% — the middle of its fifteen Octobers, seven better and seven worse — and it closed higher in 10 of those 15 years. At the other end, August is the month with the fewest good years, up in only 5 of 14; September is close behind at 6 of 15.

Second. That October lead is thinner than it looks. February’s median is +11.1%, a rounding error above October’s, and also 10 years up out of 15 — on the full record the top of the table is a tie. October only pulls clear once bitcoin’s first three years are dropped: over the 150 months since January 2014 its median is +12.8% against February’s +11.1%. And if you take those 150 returns, peel the month labels off and deal them back out at random twenty thousand times, 52.8% of the random deals produce a “best month” at least as convincing as October.

The short answer, before any of the working: we cannot tell. A calendar month happens once a year, so fifteen years of bitcoin buys us fifteen Septembers — too few to separate a real seasonal effect from an ordinary run of luck. September really is weak in our record and October really is strong. So are the patterns that turn up in more than half of the random reshuffles we ran on the same numbers. That is not the same as saying there is no effect. It means this much data cannot decide, and anyone who tells you otherwise is reading fifteen coin flips as a law.

The rest of this article is the working: four checks that make the difference between a pattern and a coincidence visible, and that transfer to any other market story arriving with a nickname attached. It is a description of what has already happened — not a forecast, and not advice about what to buy or when.

How we measure

A monthly return here is the last daily close of the month divided by the last daily close of the month before, minus one. All figures come from our own daily UTC closes — CoinGecko for recent days, Binance from August 2017, earlier Bitstamp for bitcoin and Bitfinex for ether — the same series that feeds every tool on this site (see our methodology). Four consequences of that definition are worth stating before any number:

• The bitcoin sample is 178 months, September 2011 to June 2026. August 2011 has no prior month-end to divide by, and July 2026 was incomplete when we ran this, so both are excluded.
• Each month has 14 or 15 observations. Ten of the twelve have 15; July and August have 14. A calendar month happens once a year, and no amount of daily data changes that.
• We lead with medians, not averages. A median is the middle observation — with fifteen years, the eighth-best of the fifteen. An average lets one extreme year drag the whole figure, which turns out to be the entire story of one month below.
• Ether gets its own count: 123 months from April 2016, so 10 or 11 observations per month.
• Three windows appear below, and we name them every time. The full record is 178 months from September 2011. The modern sample is the 150 months from January 2014, dropping the three years when the whole market was worth less than a mid-sized company. The halving window is the 163 months since the November 2012 halving, used only in the section on the four-year cycle. The three windows give different answers in places, and where they do, that is the finding.

The twelve months, every year

Before the medians, the raw material. Each dot below is one month of one year.

The first thing to read off that chart is not the ticks — it is the height of the columns. September’s fifteen Septembers run from −39.8% in 2011 to +20.3% in 2012; October’s from −31.5% in 2011 to +61.2% in 2013. Those 2011 figures deserve a reminder rather than a shudder: that September, bitcoin went from $8.00 to $4.82, in a market so small that one seller could set the month’s number. The two dots running off the top of the chart come from the same era — March and November 2013. The distance between the best and worst month’s median is smaller than the spread inside almost any single column. Whatever seasonality is here, it is a faint tilt on top of enormous noise, not a rule.

Month	Median, all years	Years up	Median, from 2014	Years up
January	Median, all years +0.7%	Years up 8 of 15	Median, from 2014 −0.2%	Years up 6 of 13
February	Median, all years +11.1%	Years up 10 of 15	Median, from 2014 +11.1%	Years up 9 of 13
March	Median, all years −1.8%	Years up 7 of 15	Median, from 2014 −2.1%	Years up 6 of 13
April	Median, all years +8.1%	Years up 10 of 15	Median, from 2014 +8.1%	Years up 8 of 13
May	Median, all years +3.4%	Years up 8 of 15	Median, from 2014 +9.6%	Years up 7 of 13
June	Median, all years +2.1%	Years up 8 of 15	Median, from 2014 +2.1%	Years up 7 of 13
July	Median, all years +9.0%	Years up 10 of 14	Median, from 2014 +8.2%	Years up 8 of 12
August	Median, all years −7.5%	Years up 5 of 14	Median, from 2014 −8.6%	Years up 3 of 12
September	Median, all years −3.1%	Years up 6 of 15	Median, from 2014 −4.3%	Years up 5 of 12
October	Median, all years +11.0%	Years up 10 of 15	Median, from 2014 +12.8%	Years up 9 of 12
November	Median, all years +8.9%	Years up 9 of 15	Median, from 2014 +7.7%	Years up 7 of 12
December	Median, all years −2.9%	Years up 7 of 15	Median, from 2014 −3.0%	Years up 5 of 12

Three things in that table are worth a pause. The good news for the folklore: dropping 2011–2013 does not dissolve the headline months. October’s median goes up, September’s goes down. Whatever this is, it is not purely an artefact of bitcoin’s earliest years.

The first awkward note, promised above: February and October are level on the full record. Carried to more decimals their medians differ by 0.06 percentage points — smaller than the rounding in the column you are reading. So everywhere below where October is called the best month, the window is the 150 months from 2014, where it leads February by 1.7 points.

The second: the weakest month in our record is not September. It is August, on both measures. Its median is −7.5% across all years and −8.6% from 2014, and it has the fewest good years of any month — up in 5 of its 14, against September’s 6 of 15. The saying names September. The data, such as it is, names the month before it.

Check one: does one year own the number?

November is the cleanest demonstration in this dataset of how a correct table produces a wrong sentence. November’s mean return is +36.4%, far ahead of every other month; on that number November is bitcoin’s great month and it is not close. Its median is +8.9%, fourth. The gap between the two is one observation: November 2013, +450.0%. Drop 2013 out of every month and November’s mean falls to +6.8% — from first place to sixth.

Nothing was miscalculated. The mean of those fifteen numbers really is +36.4%. What fails is the step after the arithmetic, the plain-English summary: “November is historically bitcoin’s strongest month” is a sentence the table appears to support and the data does not. Three of the five largest monthly gains in the entire record happened in 2013 — November (+450.0%), March (+186.8%) and February (+63.9%) — and a market that small does not get to speak for the one we have now.

The general test: delete the single best year and the single worst year from a month and see what is left. If the claim moves, the claim was about one year, not about the month.

Check two: twelve chances to look special

If you keep one thing from this article, keep this one. It needs no software, no statistics course and no trust in us — you can do it on the back of an envelope, and it disposes of most market folklore on its own.

October has closed higher in 10 of its 15 years. Imagine each month is nothing but a coin: heads it rises, tails it falls. The chance of getting 10 or more heads out of 15 tosses is 15.1% — uncommon, but roughly the odds of rolling a six on a die, and nothing you would build a plan on.

Now the part that does the real damage. We did not name October in advance. We looked at all twelve months and picked the one that stood out — and with twelve coins on the table, one of them is very likely to have a good run. The chance that at least one of twelve fair coins shows 10 or more heads in 15 tosses is 86.0%. On a market with no seasonality whatsoever, finding a month that looks as good as October is not the exception. It is the expected outcome, five times out of six.

That argument works on any “best month”, “best day of the week” or “best hour to buy” you will ever be shown, and it costs nothing to apply.

The heavier version of the same idea keeps the real returns instead of replacing them with coins, and destroys only the calendar. Take the 150 monthly returns from January 2014 onwards, keep the values exactly as they are, and shuffle which month each one belongs to. Recompute the twelve medians. Write down the best of them. Do that twenty thousand times, and you have a picture of what “the best month” looks like in a world where the calendar means nothing at all. Then check where the real October falls in that picture.

It falls in the middle of it. 52.8% of the reshuffles produced a best month whose median matched or beat October’s actual +12.8%. The gap between the best and the worst month — 21.4 percentage points — appears in 41.5% of them.

There is a customary bar for calling a result surprising, and it is worth naming in plain words, because it returns twice below. Researchers set it at one in twenty: a finding counts as unusual only if pure chance would produce something at least that striking less than 5% of the time. It is a convention, not a law of nature — but it is the convention, and October is nowhere near it. To reach it on our 150 months, October’s median would have to be around +21.6% instead of +12.8%, or the best-to-worst gap around 30.4 points instead of 21.4.

Which is a statement about our evidence, not about the world: it says this much data cannot tell the difference between the seasonality we observe and none at all.

Check three: cut the record in half

The third check needs no statistics at all, only patience: compute the pattern twice, on the first half of the record and on the second, and see whether the two agree. A genuine calendar effect — one caused by tax years, fund flows, quarter-end reporting or holiday liquidity — should show up in both halves. A coincidence has no reason to.

The lines cross. To put a number on how badly, we compare the two orderings. Rank the twelve months by median in the first half — February 1st, November 2nd, down to August 12th — then rank them again in the second half, and see how well one list predicts the other. That comparison is called a rank correlation. It reads 1.0 when every month keeps exactly its place, −1.0 when the order reverses completely, and hovers around 0 for two lists with nothing to do with each other. Ours is 0.08: closer to two unrelated lists than to a preserved order. An ordering that similar or better turns up in 40.1% of random pairings — nowhere near the one-in-twenty bar.

Five of the twelve medians change sign outright, November most brutally: second-best month in the first half at +12.3%, worst of all twelve in the second at −7.1%.

Two months hold their place almost exactly, and honesty requires naming them: August (12th, then 11th) and September (10th, then 9th). Each moved one rank across seven and a half years — the smallest shifts of any month, alongside May. A weak late summer is the closest thing to a durable feature in this record.

October did not hold its place — it moved the other way, 7th in the first half and 1st in the second, a jump of six ranks and the third-largest shift of the twelve. What it kept was its sign: positive in both halves, +5.4% then +11.0%. That is worth something, but not much on its own, since five of the twelve months managed a positive median in both halves. Stated narrowly, then: two months stayed at the bottom of the table, and October got stronger rather than reverting. Everything else in the ordering moved.

The out-of-sample version of the same idea is blunter. Suppose you had stopped at the end of 2018, read off the two extremes and written the rules down. “Avoid August”, the worst month at −8.5%, would have worked: August came 11th of 12 over the following seven and a half years. “Favour February”, the best month at +16.9%, would not: February landed 5th of 12. One rule out of two, from the two loudest signals in the sample.

Check four: ask a second asset

Ether has 123 monthly returns from April 2016 — 10 or 11 per month, fewer than bitcoin, and worth reading with that in mind. On September, ether agrees: it is ether’s worst month too, median −9.0%, higher in 4 of 10 years. On October — the month the folklore is loudest about — it does not. Ether’s October median is +0.5%, up in 6 of 10 years — sixth of twelve, the middle of its table rather than the top of it. Ether’s best month by median is January, at +16.3%.

A caveat that cuts both ways: ether is not an independent witness. The two assets moved the same direction in 99 of their 123 overlapping months — 80.5% — with a correlation of 0.59 between the monthly returns, and their co-movement has been tightening for years. That dependence is exactly what makes the October disagreement interesting: two series that agree four times out of five reach opposite verdicts on the one month with a nickname.

And “sell in May”?

The equity adage says the May-to-October half of the year is the weak one. For bitcoin from 2014 it points the wrong way: May–October has a median monthly return of +2.6% against November–April’s +1.2%. Across the full record the winter half does lead, +4.1% to +2.8%, but that advantage is built almost entirely from four months of 2013 — November +450.0%, March +186.8%, February +63.9% and January +54.5%. Ether does lean the adage’s way, −2.5% against +7.2%, on ten or eleven observations a month. One asset contradicts it, the other mildly supports it: no verdict worth acting on.

The calendar bitcoin might actually have

One more grouping, and then we stop — but this one earns its place, because it shows a rival calendar explaining the same data at least as well. Roughly every four years the reward paid to bitcoin’s miners is cut in half; that event is what most cycle stories are built on, and our halving countdown tracks where the current one stands. Regroup the very same monthly returns by how long ago the last halving was, and the result is harder to write off as luck than the calendar version is.

Now run the identical shuffle test on both groupings, over the identical 163 months since the November 2012 halving. Twelve calendar months spread their medians 23.1 points, and a gap that wide turns up in 31.8% of shuffles. Four cycle years spread only 13.2 points — a smaller gap — but one that large turns up in just 13.3%. The smaller-looking difference is the more surprising one, because four buckets holding about forty observations each are much harder to fake by luck than twelve buckets holding fourteen. How impressive a gap looks depends on how many buckets you cut the data into and how full each one is — and a table of twelve monthly medians hides both facts.

Two cautions. Neither grouping clears the one-in-twenty bar — 13.3% is better than 31.8% and still not proof, on three and a half cycles in which the halving is entangled with everything else that happened in those years; our comparison of the halving cycles finds the two it covers rhyme far less neatly than the story requires. And the cycle does not absorb the calendar: September’s nine down years are scattered across every position in it, so the September effect is not the four-year clock in a costume.

What a long-term holder does with this

Not much — that was the answer at the top of the page and the working has not changed it. What does transfer is the method, to every other pattern that arrives with a nickname attached. Three questions, in order:

• How many observations is this actually built on? Not how many days of data — how many independent instances of the thing being claimed. For a monthly pattern the answer is one per year, and ours is 15.
• What happens if I delete the biggest year? If the claim moves, the claim is about that year.
• Does it survive being computed on half the data at a time? This one is free, needs no statistics, and eliminates more folklore than any formal test.

And a warning about the backtest that will be waved at you. From 2014 onwards, sitting out every single September would have turned a buy-and-hold multiple of ×80.0 into ×120.5, before any fee or tax. That is a real calculation on real closes, and it is the most misleading number in this article. The same rule has lost money for three years running — September 2023 closed +3.9%, 2024 +7.4%, 2025 +5.4% — so that from January 2023 a holder was at ×3.54 and a September-skipper at ×3.01. Meanwhile the most recent October, the famous one, closed −4.0%.

What has mattered over these fifteen years is not which month you were in but how much you owned and how long you held it. Bitcoin’s record of drawdowns is measured in years under water, not weeks; the holding-period statistics put numbers on what one, two and four years of patience produced from every possible start date; and if the impulse behind reading a seasonality article is “when should I buy”, our guide to averaging in answers the version of that question the data can support.

None of this is a prediction about September 2026, and none of it is financial advice. If the coming September falls, the count reads 6 of 16; if it rises, 7 of 16. Either way it moves the medians here by a point or two and settles nothing — which is, in the end, the whole point.

Frequently asked questions