Is September the Worst Month for Stocks? 98 Years of Data

Every autumn the same question comes back, usually around the time the headlines get louder: should you hold off investing until the dangerous months are over?

The fear has a long record behind it, and part of that record is real. One month really has been the weakest of the year in the US data since 1928. It just is not the month most people name, and the evidence that you can do anything useful with it is much thinner than the evidence that it exists.

October’s reputation, and the days behind it

Ask which month is dangerous for stocks and the answer is almost always October. It has receipts.

On Tuesday 29 October 1929 the Dow Jones Industrial Average fell nearly 12%, a day after it had already dropped nearly 13%, and the crash that defined the following decade was under way. Black Monday, 19 October 1987: 22.6% in one session, 508 points — still the largest one-day percentage fall in the index’s history. Then October 2008, when the S&P 500 lost 16.9% in a month.

Three of the most frightening episodes in market history, and the same month on the calendar each time. If your instinct about autumn comes from there, it came from somewhere real. I had the same instinct for years and never checked it, because you do not check what you already know.

The worst month for stocks, month by month

Here is what happens when you average every month instead of remembering the worst ones. The table covers 98 years of the S&P 500 and, before March 1957, its predecessor the S&P Composite, measured from one month-end close to the next.

Month	Average	Median	Years it closed lower
September	−1.12%	−0.42%	53 of 98 (54.1%)
February	−0.10%	+0.23%	47 of 98 (48.0%)
May	+0.08%	+1.06%	39 of 98 (39.8%)
March	+0.48%	+1.05%	38 of 98 (38.8%)
October	+0.55%	+1.03%	40 of 98 (40.8%)
August	+0.68%	+0.87%	41 of 98 (41.8%)
June	+0.80%	+0.39%	42 of 98 (42.9%)
November	+1.01%	+1.63%	37 of 98 (37.8%)
January	+1.21%	+1.55%	37 of 98 (37.8%)
April	+1.26%	+0.93%	34 of 98 (34.7%)
December	+1.27%	+1.48%	27 of 98 (27.6%)
July	+1.69%	+1.69%	38 of 98 (38.8%)

S&P 500 (S&P Composite before March 1957), price return before dividends, January 1928 to December 2025. My own calculation from month-end closing levels.

September comes last whichever start year you pick, 1928, 1950 or 1970. October sits in the middle with a small positive average. The month with the famous crashes is, on average, an ordinary month.

Two details matter more than they look. The first is the basis. These are price returns, and on that basis September closed lower in a little over half of all years. Put the dividends back and the picture softens: Fisher Investments, using the S&P 500 total return from 1926 to 2024, counts September as up slightly more often than down, 52% against 48%. “A little over half” is true only before dividends.

The second is the window. The −1.12% belongs to the series that starts in 1928; start in 1950 and September’s average is −0.63%. Every number on this subject carries its window with it, and most versions you will find online do not say which one they used. RBC Wealth Management, working from Bloomberg data through July 2025, arrives at −1.2% and 55% since 1928 — the same answer from a slightly different end date.

Why is September weak? Nobody knows

This is the part that surprised me most when I looked into it properly. The pattern is old, well documented and widely known, and the profession still cannot tell you what drives it.

The most prominent explanation is not about markets at all. In 2003 Mark Kamstra, Lisa Kramer and Maurice Levi argued in the American Economic Review that shorter days in autumn and winter push many people towards depression, that depression makes them more averse to risk, and that stock returns follow the daylight accordingly. Patrick Kelly and Felix Meschke disputed the result in the Journal of Banking & Finance in 2010, and the original authors replied in the same journal two years later. A serious idea in a serious journal — and contested in print ever since.

The ordinary candidates are weaker than they sound. US mutual funds have closed their tax year on 31 October since the 1986 tax reform, and Gibson, Safieddine and Titman showed in 2000 that they hurry to sell their losing positions before that date, but the selling they found runs up to the end of October and the rebound comes in November, not in September. Tidying a portfolio also mostly moves money from one stock to another rather than out of the market. The other usual suspect, investors repositioning after the summer, is a story more than a finding.

No accepted mechanism, then. On its own that settles little; plenty of real effects went unexplained for decades. The next two sections settle more.

The September effect is fading

Every September since 1928 · S&P 500 · price return

The September effect was once a lean. Since 1990 it has been a coin toss

One square per year. Filled: September closed lower than August. Read the rows above the line, then the rows below it.

Split the record in 1990 and the effect thins out. Before 1990 September closed lower in 56.5% of years. From 1990 to 2025 it closed lower in exactly half, 18 of 36. Its average since 1990 is −0.72%, and its lead over the next weakest month, August at −0.49%, is 0.23 points — not a gap anyone could tell apart from noise.

One correction to the video, which was recorded before this autumn. It says the last two Septembers were both positive, and when it was filmed they were: September 2024 rose 2.0% and September 2025 rose 3.5%. September 2026 then closed 0.45% lower. Counted with 2026 the modern record is 19 down years out of 37, a coin toss with a very slight lean.

Does it pass a statistical test?

Not quite, and that is before the hardest objection. Tested against zero, September’s average from 1928 to 2025 gives a p-value of about 0.056: close to the conventional 5% line, on the wrong side of it.

The harder objection is that nobody tested September on its own. Somebody looked at all twelve months and reported the one that stood out, and among twelve, one will usually look unusual by chance. Correct for that with the simplest method, which asks for p below 0.0042, and September misses by a factor of about thirteen. The only two months that pass are July and December, and both are positive. Ryan Sullivan, Allan Timmermann and Halbert White made the general case in 2001: calendar effects lose much of their significance once you account for the whole universe of rules that were tried before the winners were reported.

The sharpest line comes from an unlikely source. Ben Jacobsen, who co-wrote the original Halloween paper below, went back over 300 years of UK stock returns with Cherry Zhang and concluded in 2013 that “monthly seasonals might simply be in the eye of the beholder.” That is not a sceptic talking. It is one of the people who found them.

Their verdict is about single months. The half-year pattern is a different claim, and it holds up much better — which is exactly why it deserves the longer look further down.

October is not bad, it is loud

S&P 500 · price return · January 1950 to December 2025 · 912 months

Autumn holds most of the worst months, and most of the best ones

Every month in which the index moved more than 10%, filed under its calendar month. Above the line the gains, below it the losses. One dot is one month, and the number in it is the year.

If the four autumn months were ordinary, they would hold about a third of each pile. They hold two thirds of the big losses and well over half of the big gains. That is a market that moves harder in both directions at this time of year, not one that slopes downhill.

If October’s average is ordinary, the feeling has to come from somewhere else. It comes from the size of the moves, which are large in both directions.

From January 1950 to December 2025 the S&P 500 had nine months that fell more than 10%. Six of them landed between August and November. It also had thirteen months that rose more than 10%, and eight of those landed in the same four months. Those four months are a third of the calendar.

Fell more than 10%		Rose more than 10%
October 1987	−21.76%	October 1974	+16.30%
October 2008	−16.94%	January 1987	+13.18%
August 1998	−14.58%	April 2020	+12.68%
March 2020	−12.51%	January 1975	+12.28%
September 1974	−11.93%	January 1976	+11.83%
November 1973	−11.39%	August 1982	+11.60%
September 2002	−11.00%	December 1991	+11.16%
February 2009	−10.99%	October 1982	+11.04%
March 1980	−10.18%	October 2011	+10.77%
		November 2020	+10.75%
		August 1984	+10.63%
		November 1980	+10.24%
		November 1962	+10.16%

S&P 500, price return, month-end to month-end, January 1950 to December 2025. August to November in bold.

The count depends on where the line is drawn, which is worth knowing before you repeat it at dinner. Eight more months fell between 9.0% and 9.5%, among them September 2022 at −9.34% and September 2008 at −9.08%, and five of those eight also sit between August and November. With dividends included March 1980 drops just short of −10% and the losses become eight. April 2026, after the counted window, rose 10.42%: a fourteenth big gain, in spring.

None of that moves the shape. Autumn is a roller coaster rather than a downhill slope, and a market that moves harder in both directions is a different thing from a market that mostly falls.

Does sell in May actually work?

If single months are mostly noise, the half-year is where the serious evidence lives. “Sell in May and go away” says that stocks do better from November to April than from May to October, and here the research is strong. I am not going to pretend otherwise.

Sven Bouman and Ben Jacobsen tested it in 2002 across 37 markets, using data from 1970 to 1998, and found the effect in 36 of them; New Zealand was the exception. It was statistically significant in 20. (You will also see “35 of 37” quoted. That figure comes from a later team’s own re-run of the original period, not from the original paper.) Eleven years later Sandro Andrade, Vidhi Chhaochharia and Michael Fuerst re-ran the same 37 markets on data the first paper never saw, November 1998 to April 2012. This time the November-to-April average beat the May-to-October average in every one of them, by roughly 10 percentage points on average. Their conclusion, in their words: “The adage ‘Sell in May and go away’ remains good investment advice.”

They also dealt with the obvious objection in advance. Of all the seasonal anomalies, they point out, this is the one least affected by trading costs, because it needs only two trades a year. Whoever dismisses it on costs has not read them.

The literature is not unanimous, though most popular accounts stop at the previous paragraph. Maberly and Pierce argued in 2004 that the US result rested on two outliers, October 1987 and August 1998, both in the summer half. Witte replied in 2010 that deciding how many outliers to remove is itself somewhat arbitrary, and that regression methods built to resist outliers restore the effect, if only marginally. Then Dichtl and Drobetz applied a test built for exactly this situation, one that allows for the many strategies a researcher could have tried, and concluded in 2014 that their results “reject the hypothesis that a trading strategy based on the Halloween effect significantly outperforms.” That is three rounds, and the third is the one that asks the investor’s question.

Because here is the gap. All of these studies measure what index returns did, before tax. They do not measure what an investor keeps after two switches a year, a fund wrapper, a tax bill and six months of watching the market from the sidelines. On paper the costs are negligible. The only way to know whether they are in practice is to look at somebody who tried.

Somebody did. In July 2018 Pacer launched an exchange-traded fund called the Pacer CFRA-Stovall Equal Weight Seasonal Rotation ETF, ticker SZNE, built on exactly this calendar. On the last business day of April its index moved into two defensive sectors, consumer staples and health care, each equally weighted. At the end of October it moved into four cyclical ones: consumer discretionary, industrials, information technology and materials. I should be fair about what it was and was not. It never went to cash; it stayed fully invested in shares and only rotated between them, so it was not literally a sell-in-May fund. It was the closest anyone came to selling the calendar as a product.

Pacer CFRA-Stovall Equal Weight Seasonal Rotation ETF · launched July 2018 · liquidated May 2026

Somebody built the calendar trade. Its own filings show what it returned

Two panels from two different documents, covering two different periods. They are kept apart because the numbers cannot be combined.

1 · A year, since launch, to 30 April 2026 · after fund costs

1 · A year, since launch, to 30 April 2026 · after fund costs

2 · A year, since launch, to 31 December 2024 · top US federal tax bracket

2 · A year, since launch, to 31 December 2024 · top US federal tax bracket

This is evidence about the plumbing, not about the pattern. The fund rotated between sectors and never went to cash, weighted its holdings equally and charged 0.60% a year, so its record does not show that a simple, cheap calendar signal cannot work. It shows what happened to the one listed product that sold the idea.

From launch to 30 April 2026 the fund returned 6.92% a year after costs. Its own index — the same rules, without a fund wrapped around them, its fees or its trading — managed 7.75%. The S&P 500 over the same period returned 14.73%. Over its whole life the fund earned less than half of what the plain market paid, and it trailed even its own rules by most of a point a year, which is roughly what a 0.60% fee and 190% annual turnover (in the year to April 2025) would lead you to expect.

Then there is the tax line, printed in the fund’s own prospectus for a different period, from launch to 31 December 2024. Over that stretch it reported 7.40% a year before tax. For an investor in the top US federal bracket, after tax on the distributions and on selling the shares, it was 5.83%. That is the fund’s calculation for one kind of American investor, not a statement about yours; the tax on a rotation strategy depends entirely on where you live and the account you hold it in.

On 30 April 2026 Pacer announced that the fund would be closed. It was liquidated on 15 May 2026, with about $13.4 million in assets.

It was not the first attempt. The AmericaFirst Seasonal Rotation Fund, a conventional mutual fund, opened on 31 October 2013 and returned 1.47% a year to June 2019 in its main share class, against 6.03% for its benchmark. That benchmark was a cautious mix of shares and bonds, the Dow Jones U.S. Moderately Conservative Portfolio Index, not the S&P 500, so the two numbers should not be set beside the ones above. Its expense ratios ran from about 1.6% to 3.1% depending on share class. The board closed it on 29 June 2020.

Two products, two sponsors, both gone. A search of SEC filings in October 2026 found no US fund built around the idea today. (Canada has one, a seasonal rotation ETF listed in Toronto since 2009; I have not examined its record, so it plays no part in the comparison here.)

Now the honest part, because there is a real disagreement and it should be stated plainly. The researchers say this is the cheapest anomaly there is to trade, and on paper they are right. I cannot prove them wrong about the pattern. What the two funds show is not that the pattern is false. Sector rotation, equal weighting with its tilt away from the largest companies, and fund fees are all inside their results, and none of those is the calendar. It is evidence about the plumbing, not about the pattern — but the plumbing is what you would actually be buying.

My own read: people who sit out autumn are not being naive. They are taking the only thing the pattern actually gives them, which is a reason to wait. What it never gives them is a reason to come back.

What to do with the calendar instead

Two rules, and an appointment.

Don’t postpone. If you have money to invest, the evidence above is not strong enough to justify changing a long-term plan to sit out the autumn. Notice what the Halloween research actually favours: the strong half runs from November to April, starting the moment the feared months end. The cost of waiting for a better moment has a name, the behaviour gap, and the calculator puts a number on it for your own figures.

Give the fear an exit condition. If you ever decide to step out of the market, decide on the same day what has to be true for you to step back in. A rule with no way back is waiting with a better excuse. The risk assessment is a calmer place to find out how much movement you can actually hold.

The one legitimate autumn appointment is the one you set in advance: check whether the swings pushed your portfolio outside the band you chose. If they did, you rebalance. If they did not, you do nothing, and doing nothing is a decision, not a failure. How to set the band, and why it beats a date in the diary, is in when to sell stocks.

The calendar tells you what to expect. It never tells you what to do. Autumn is not dangerous, it is loud, and what quietly costs people money is rarely the crash they are afraid of. It is the months, and sometimes the years, they spend waiting for one.