How to Read a Seasonality Statistic Without Fooling Yourself — Six Checks Before You Let a Month Move Your Money — Field Guide #36

A seasonality statistic is the cheapest thing in finance to produce and one of the most expensive to act on. It requires no model, no thesis and no forecast — only a spreadsheet of monthly returns and a willingness to point at a column. That is exactly why the same handful of numbers reappear every year, and why they are so consistently wrong in ways that are easy to check and almost never checked. This guide is the checklist. Six questions, in order, each one worked against numbers we pulled ourselves this morning: 4,990 daily Bitstamp BTC/USD candles from 1 January 2013 to 30 August 2026, plus an extension pull back through August 2011.

The test case throughout is the claim that ran across the market this weekend — bitcoin has never had a green September following a green August, not once — because it happens to fail four of the six checks, and because the way it fails is instructive rather than embarrassing. It is a well-constructed claim. It is specific, falsifiable, and easy to state. Those are virtues. They are just not the same virtue as being true.

Check 1. Where does the table start, and is there a reason other than convenience?

Almost every bitcoin seasonality table in circulation begins in 2013. This is not a considered methodological choice; it is where the most-screenshotted public monthly-returns table happens to open. Bitcoin has price history before 2013, and the first question to ask about any “never” claim is whether the never survives the years the table omits.

In this case it does not. August 2012 returned +9.48% and September 2012 returned +19.80% — both full months, 31 and 30 daily rows respectively, no gaps. A green August followed by a green September, twice as large. The claim is not false because of a technicality or a rounding convention; it is false because of an entire pair of months that the table does not contain. Ask of every historical extreme: what is the first year, and what happens in the year before it? If the answer is “the data starts there,” ask why the data starts there. Frequently the honest answer is that a website launched.

The mirror-image error is including a year you should not. Bitstamp’s own series begins on 18 August 2011, which makes August 2011 a fourteen-row month. We exclude it and say so. A table that silently includes it would report an August 2011 return of −26.61% that describes less than half a month.

Check 2. Is the first row of your data a whole month?

This is the check that caught us, this morning, in our own working file, and it is worth showing rather than describing. Pulling monthly candles for BTCUSDT from Binance — the default source for most crypto analysis — returns an August 2017 return of +10.87%. The true figure is +65.78%. The gap is not an index difference or a venue difference. Binance’s first BTCUSDT daily candle is dated 17 August 2017. The exchange’s “August 2017” is fifteen days long, and the first monthly candle any exchange API returns is always a stub of unknown length.

The consequence for a seasonality claim is severe and specific. August 2017 is the largest green August in bitcoin’s history by a factor of two. An analyst using Binance monthly candles will see a modest +10.87% where the record contains a +65.78% and will draw the conditional pattern from a completely different-looking sample. Rule: before using any monthly series, print the row count of its first month and its last. If either is short, exclude it and say so in the same sentence as the result. Our last month, August 2026, is thirty rows and unfinished, and we label it as such everywhere it appears.

Check 3. Mean or median — and which one or two cells are carrying the mean?

A monthly average across thirteen observations is one large outlier away from meaninglessness, and crypto has more large outliers than almost any other asset class. The cleanest demonstration in the whole dataset is November.

Bitstamp BTC/USD, complete calendar months 2013–2025, own computation, 30 August 2026.

One month out of 156 — November 2013, which returned +450.02% — is the difference between a November that averages +41.33% and a November that averages +7.27%. March is the same story in miniature: a mean of +13.35% sitting on top of a median of −2.11%, because March 2013 returned +186.76%. Whenever a mean and a median disagree by more than the size of the effect being claimed, the mean is describing one year and the median is describing the asset. Print both, always, and print the positive count alongside them, because a hit rate is the one summary statistic an outlier cannot distort.

Applied to the case everyone is arguing about, this check reverses the conclusion. September’s reputation as bitcoin’s worst month rests on the mean, where it scores −3.04% against August’s +1.00%. On the median, August is −8.31% against September’s −3.11%; on the hit rate, August is 4 of 13 against September’s 5 of 13. Two of the three standard measures name August as the worse month. Nobody says so, because the mean is the number that got into circulation first.

Check 4. What is the denominator of the conditional — and how wide is the interval around it?

A statement about all Septembers has thirteen observations behind it. A statement about Septembers that follow a green August has far fewer, because the condition throws most of the data away. Bitcoin has had five completed green Augusts in its entire history: 2012, 2013, 2017, 2020 and 2021. That is the denominator. Everything anyone says about “September after a green August” is a statement about five numbers.

Five is small enough that the arithmetic of uncertainty overwhelms the observation. Take the claim at its own most favourable framing — four red Septembers out of four green Augusts, which is what you get if the table starts in 2013. The Wilson 95% confidence interval on 4 out of 4 runs from 51.0% to 100%. In other words, even a perfect four-for-four record is statistically consistent with the underlying probability being a coin flip. Add the 2012 observation and the record becomes four out of five, whose 95% interval runs from 37.6% to 96.4% — a range that includes “this is worse than a coin flip” and “this is nearly a law” simultaneously.

The practical form of this check is a sentence you should be able to say out loud without wincing: “I am adjusting my position on the basis of five observations, and if the true rate were 50% I would expect a record at least this lopsided roughly one time in five.” If saying it out loud changes your mind, the statistic was doing rhetorical work rather than analytical work.

Check 5. Is it a calendar effect, or is it one block of consecutive years?

A genuine seasonal effect should be distributed across the sample. If it is concentrated in a contiguous run of years, what you are looking at is not a month, it is a regime that happened to overlap with a month. September is the textbook case, and laying it out row by row settles it.

Bitstamp BTC/USD, own computation. Third-party cross-check: CoinGlass returns the same sign in all thirteen years, with magnitudes within roughly 0.4 percentage points.

Six consecutive negative Septembers between 2017 and 2022 built the entire folklore. On both sides of that block, September is unremarkable to good: two of four positive before, three of three positive after. Excluding the block, the remaining seven Septembers average +0.89% with five of seven green. Whatever produced 2017–2022 — and the candidates are macro, cycle position and the structure of the market at the time, not the calendar — it stopped in 2022 and has not resumed in three years of trying. Test: split the sample into thirds and check the effect appears in at least two of them. September fails this test outright.

Check 6. What would falsify it, by when, and settled against what?

This is the check that separates an observation from a claim, and it is the one this site applies to itself every Sunday. A seasonality statistic that cannot be graded is not a forecast, it is decoration. Three elements are required and all three must be fixed in advance: a bar, a deadline, and a named settlement source.

The third element is the one people forget, and it is not a formality. We learned this at our own expense yesterday: a marker set on a 30-year Treasury yield failed on the correct verdict but was described with the wrong direction, because a live market quote on the on-the-run bond and the Treasury’s constant-maturity par yield are different series that differ by several basis points on the same day. The bar was fine. The deadline was fine. The settlement source had not been named, so we settled it against whichever number appeared first. Two trackers of the same US spot bitcoin ETF flows currently disagree by $16.3 million on August’s total — enough to flip a marker whose remaining gap is $102.5 million into one whose remaining gap is $118.8 million. Name the source in the same sentence as the bar.

Applied to this weekend’s claim, the falsification is cheap and dated, which is the best thing about it: if bitcoin’s 30 September close on a named series exceeds its 31 August close, the pattern goes from four-in-five to four-in-six and a statistic being described as an iron law will have failed twice in fourteen years. That is a properly formed claim. Note what it is not: it is not a prediction that September will be green. It is a statement about what we will have learned either way, which is the only kind of statement a thirteen-row dataset can support.

The six checks, as a card

What this guide is not saying

It is not saying seasonality is nonsense. October’s record — a +19.82% mean, a +14.60% median and ten green months in thirteen — passes checks 3 and 5 comfortably, with the effect distributed across the sample rather than concentrated in one regime, and it would take a serious argument to dismiss. The point is that the checks are what separate October from September, and that almost nobody runs them, in either direction.

It is also not saying that a weak statistic implies the opposite conclusion. September 2026 may well be red, and the reasons it might be red are dated and specific: an FOMC meeting on 15–16 September at which a rate increase is priced somewhere between 48% and 56% depending on the venue, July PCE running at 3.7% headline and 3.3% core, and a ten-week purchasing pause at the largest corporate holder. Those are arguments about September 2026. “It is September” is not an argument at all — and if it were, it would have told you to sell in August, one month before bitcoin rose 24.45%.

Field Guide #36. The full set of guides is indexed in our Reading Room, and each one is written to be run as a procedure rather than read as an opinion. If you apply exactly one of these six checks, make it Check 4: write down the denominator before you write down the conclusion. Most of the errors in this field are not errors of reasoning. They are errors of counting, committed by people who never counted.

Method: monthly returns in this article are computed by Bitcoin Mastery as first-of-month open to last-of-month close, UTC, on Bitstamp BTC/USD daily candles — 4,990 rows from 1 January 2013 to 30 August 2026, plus a separate extension pull covering August 2011 to December 2013. Partial months are excluded and identified. Prices, flows, funding, open interest and on-chain figures are pulled directly at the timestamp stated. Where a third-party figure is cited we name the source and its date; where two sources disagree we print both.

Disclaimer: This article is for informational purposes only and does not constitute investment advice. Cryptocurrencies are volatile and you can lose money. Historical seasonality describes the past; it does not forecast the future. Do your own research and consult a licensed financial advisor before making investment decisions.