Does “Sell in May and go away” still work?
Every spring the old saying comes back: sell your stocks in May and buy them back after Halloween. In 2002 two economists found that stocks had earned more from November to April than from May to October in 36 of 37 markets.
Does selling stocks in May still beat holding them through the summer?
Sell in May and go away, the saying goes, and buy back after Halloween. In the American Economic Review of December 2002, Sven Bouman and Ben Jacobsen tested it on 37 stock markets from 1970 to 1998 and found November to April returns ahead of May to October returns in 36 of them (Bouman and Jacobsen, 2002; Maberly and Pierce, 2004). Later tests kept finding it. Andrade, Chhaochharia and Fuerst (2013) put the November to April lead at about 10 percentage points in data the original study never used, and Zhang and Jacobsen (2021) found summer returns below the risk-free rate, the return on cash, in every stock market with data.
The measure here is the gap between the two halves of the same year: the return from November to April, the winter, minus the return from May to October, the summer, in percentage points. Over all 63 years the average winter returned 6.6% and the average summer 0.6%, so a six-point gap is most of what stocks earned in a year. A single year tells little, because one year’s gap can swing by tens of points either way.
Winter still wins, by about a third of its old margin, and selling in May has cost money since Bouman and Jacobsen published. Winter beat summer by 7.8 points a year in the 40 years to October 2002, and by 2.7 points in the 23 years since, a lead that one calendar in six matches when the months are shuffled at random. A dollar put into the S&P 500 in November 2002 grew to $11.83 by October 2025 with its dividends. Sold every May and held in Treasury bills until November, the same dollar grew to $5.05, less than half as much.
Before 2002, winter beat summer by almost eight points a year
The test covers nine indices, each from its first complete November: the S&P 500 from 1962, the S&P/TSX Composite from 1979, the FTSE 100 from 1984, the Nikkei 225 from 1989, the Hang Seng from 1990, the Dow Jones Industrial Average from 1992, the DAX from 1994, the S&P/ASX 200 from 2001 and the Euro Stoxx 50 from 2007. A year runs from November to October, and each year’s gap averages whichever indices existed that year, 327 index-years in all. “Before 2002” means the 40 years from November 1962 to October 2002, and “since 2002” the 23 years from November 2002, the month before Bouman and Jacobsen published, to October 2025.
Before 2002, winter led summer by 7.8 points a year: stocks gained 7.3% in the average winter and lost 0.5% in the average summer. Winter won 29 of the 40 years, nearly three in four. Every one of the seven indices with five or more years before 2002 shows the gap, from 6.3 points in the Hang Seng to 22.7 points in the DAX (Figure 1). The DAX’s gap rests on only eight years, which end with summer falls of 27.2% in 2001 and 37.5% in 2002.
In every index with a history before 2002, winter's lead over summer has shrunk since.
Average winter (November to April) return minus average summer (May to October) return, in percentage points, since November 2002 (red dot, labelled) and before (ring), for the seven indices with five or more years before 2002 and for all indices together.
A shuffled calendar almost never produces a gap that size. Shuffling each year’s twelve monthly returns among its twelve months, the same way for every index, keeps each year’s total and the way the markets move together but makes the calendar random. Only 4 of 10,000 shuffled histories produced a gap of 7.8 points or more before 2002, while one in twenty reached 4.0 points. Index by index, five of the seven clear that one-in-twenty bar, and the S&P 500, the FTSE 100 and the DAX still clear it after allowing for 16 tests run at once (Benjamini and Hochberg, 1995).
Since Bouman and Jacobsen published, the gap is a third of its size
Since 2002 winter’s lead has been 2.7 points a year, 65% smaller than before: stocks rose 5.3% in the average winter and 2.6% in the average summer. All nine indices kept winter ahead, by 1.6 points in the S&P 500 and the Dow up to 5.4 points in the Euro Stoxx 50, and in eight of the nine the average summer made money. None of the nine clears the one-in-twenty bar on its own. Pooled, the 2.7-point lead is matched or beaten by one shuffled calendar in six, and clearing the bar would now take 4.5 points.
Academic findings usually weaken once they are published. McLean and Pontiff (2016) followed 97 stock return predictors from the finance literature and found their returns 58% lower after publication than in the original samples. The Halloween gap fell by 65%, close to that average. The fall holds up when whole years are resampled: the gap shrank in 93% of 10,000 draws.
Ten-year averages show the same fall. The typical decade before 2002 had a gap of 6.9 points, the median of the 31 ten-year stretches that ended by October 2002. The widest, the decade to October 1992, reached 10.3 points. All 14 decades that lie wholly after publication, from the one ending in October 2012 on, came in below that median, the highest at 6.6 points. The latest, to October 2025, stood at 2.5 points (Figure 2).
Every decade wholly after publication has had a smaller winter lead than the typical decade before it.
Winter return minus summer return, in percentage points, averaged over the ten years to each October from 1972 to 2025; the dashed vertical marks December 2002, when Bouman and Jacobsen published, and the dashed line the median of the 31 decades that ended by October 2002.
The summers that beat winter since 2002 came after spring crashes
Four years did most of the damage. Summer beat winter in 7 of the 23 years since 2002, and in four of them, 2003, 2009, 2020 and 2025, by 15 points or more (Figure 3). In each of the four, markets hit bottom in March or April, while the rule still held stocks, and rebounded through the summer it spent in cash. The S&P 500’s lowest close of the winter came on March 11, 2003, March 9, 2009, March 23, 2020 and April 8, 2025. In those four summers the indices rose 16.0% on average, and summer beat winter by 19.9 points. In the other 19 years since 2002, winter won by 7.5 points, about as much as before 2002.
Since 2002, the years summer won by most were the spring-crash rebounds of 2003, 2009, 2020 and 2025.
Winter return minus summer return in each year to October, 1963 to 2025, in percentage points, averaged over the indices that existed that year; a bar above zero is a year winter won, and the dashed line marks December 2002, when Bouman and Jacobsen published.
The rule’s best years were the reverse: summers with a crash in them. Before 2002 the S&P 500’s worst summers were 1974, down 18.2%, and 2002, down 17.8%, and its five worst summer months included October 1987 and August 1998. Since 2002 the worst summers came in 2008, when the nine indices fell 32% from May to October, and 2011, when they fell 12%. Winter won those two years by about 20 points each. Maberly and Pierce (2004) argued that the US gap in Bouman and Jacobsen’s years rested on two months, October 1987 and August 1998. Taking those two months out of the S&P 500’s gap from November 1970 to October 1998 cuts it from 5.8 to 4.5 points, which leaves most of it standing.
Since 2002 a handful of crash months carry most of what is left of the gap. Measured index by index from average months, the gap is 10.0 points before 2002 and 3.2 points since. Leaving out each index’s five worst summer months cuts the earlier gap to 4.4 points, still positive in six of seven indices. The same cut takes the later gap to 0.4 points, positive in five of nine. Of the 45 months left out since 2002, 35 fall in 2008, 2011 and 2022, and a seller who sat out those summers gave the gain back in the rebound summers.
Since 2002, a summer in cash has paid far less than a summer in stocks
The rule is a trade: a seller in May gives up the summer’s stock returns, dividends included, and earns interest on cash instead. Before 2002 a summer in cash beat a summer in stocks, as three-month US Treasury bills paid 3.1% in the average summer, more than the 2.4% the S&P 500 returned with its dividends. Since 2002 the S&P 500 has returned 5.2% a summer and bills 0.9%. A summer in stocks now beats a summer in cash by 4.4 points, where before 2002 it trailed by 0.7 (Figure 4). Stocks beat bills in 23 of the 40 summers before 2002, a little over half, and in 19 of the 23 since. In the DAX, a total return index set against German and euro overnight money, a summer in stocks went from 8.4 points behind cash to 2.0 points ahead.
Since 2002 stocks have paid more than cash over the summer, and before 2002 they paid less.
Average May to October return of the S&P 500 and the DAX, both with dividends, minus the return on cash over the same months, in percentage points, since November 2002 (red) and before (grey).
The money follows. Before 2002 a seller who held the S&P 500 from November to April and Treasury bills from May to October earned 11.9% a year, against 10.7% for holding all year (Figure 5). Each dollar grew to $88.30 for the seller and to $58.79 for the holder. Since 2002 the seller has earned 7.3% a year against the holder’s 11.3%, and ended with less than half the holder’s money. The S&P 500 paid about 1.0% in dividends in each half since 2002, so the seller gave up one point a summer in dividends alone.
Selling every May beat holding before 2002 and has trailed it since, in both the S&P 500 and the DAX.
Return a year with dividends, compounded, before costs and taxes, for a seller who holds stocks from November to April and cash from May to October (red dot, labelled) and for holding stocks all year (ring).
The DAX tells the same story with a shorter past. In its eight years before 2002, which end in the summer crashes of 2001 and 2002, the seller earned 17.0% a year against 5.5% for holding. Since 2002 the seller has earned 7.8% a year against 9.2%. A dollar grew to $5.58 for the seller and to $7.60 for the holder.
The case for the rule: winter still wins seven years in ten
The strongest case for selling in May is how often it still works. Winter came out ahead in 16 of the 23 years since 2002, nearly as often as the 29 in 40 before, and a shuffled calendar wins that often only about one time in 20. All nine indices kept winter ahead. The lead has also turned up wherever researchers have looked for it, in later data (Andrade, Chhaochharia and Fuerst, 2013) and in every stock market with records (Zhang and Jacobsen, 2021).
The rule also holds stocks only half the year, so it carries less risk. Since 2002 the seller’s returns have swung about three quarters as much as the holder’s, in both the S&P 500 and the DAX. On return over cash for each unit of that risk, the Sharpe ratio, the DAX seller edged the holder, 0.53 against 0.51.
What pays is the size of the lead. Four of the seven summers that beat winter since 2002 were rebounds worth 16.0% on average, so a seller who was right 16 years in 23 still ended with less. The lower risk can also be had without the calendar. A fixed mix of 72% S&P 500 and 28% bills, held all year and rebalanced monthly, carried the same risk as the seller since 2002 and earned 8.8% a year, against the seller’s 7.3%. In the DAX a mix of the same kind earned 7.5% against the seller’s 7.8%, about a tie. On the Sharpe ratio, too, holding the S&P 500 beat selling it since 2002, 0.70 against 0.57.
What to watch
This year’s summer ends with the close of Friday, October 30, 2026, and so far it is running about level with the winter. By September 21 the nine indices were up 4.7% on average since the end of April, with eight of the nine higher, after a winter that rose 4.9%. Summer was ahead of winter in five of the nine. The S&P 500 was up 7.7% since April against 5.4% in the winter, and a fall of 2.2% from its September 21 close by October 30 would hand its year to winter. In the year to October 2025, summer won in eight of the nine indices, by 15.1 points on average, after the April 2025 low.
Cash now pays more for the summer off. The three-month Treasury bill averaged 3.94% in September 2026, which comes to about 2.0% for six months. That is more than twice the 0.9% bills paid in the average summer since 2002, and about two thirds of the 3.1% they paid before. The euro short-term rate, the cash rate for a DAX seller, averaged 2.31% in September.
The ten-year average gap stands at 2.5 points, against a median of 6.9 for the decades before 2002. Summers with crashes in them, like 2001, 2002 and 2008, would lift it back toward that median, and spring lows followed by summer rebounds, like 2003, 2009, 2020 and 2025, would hold it down.
How we got the number
The information set
The index closes are daily closing levels from Yahoo Finance, retrieved on September 26, 2026, the same closes used in RPT-004 (Atol Research, 2026). They run from the start of each index’s history in the source to September 21, 2026, and to September 18 for the Nikkei 225, since September 21 was a holiday in Japan. Eight are price indices, which leave out dividends. The DAX is a total return index, with dividends reinvested. At month ends, the closes match FRED’s copies of the publishers’ own series on all 453 month ends shared with the Nikkei 225 and all 120 shared with the S&P 500 and the Dow, to four decimal places of a percent.
The S&P 500’s dividends are Robert Shiller’s monthly series of dividends per share at an annual rate, through June 2026 (Shiller, 2026). The US cash rate is the monthly average of the three-month Treasury bill rate in the secondary market, which the Federal Reserve publishes in its H.15 release (Board of Governors of the Federal Reserve System, 2026). The German and euro cash rate is the overnight money rate reported by Frankfurt banks to December 1998 (Deutsche Bundesbank, 2026), EONIA from January 1999 to December 2021 (European Central Bank, 2026a), and the monthly average of the daily euro short-term rate from January 2022 (European Central Bank, 2026b). All four were read on October 5, 2026.
The calculation
Each index’s monthly return $r_{i,m}$ runs from the last close of one calendar month to the last close of the next. A year $y$ runs from November of year $y-1$ to October of year $y$, and an index enters a year only when all twelve monthly returns exist; years run from November 1962 to October 2025, and September 2026 is left out because it was incomplete at the last close. The winter and summer returns of index $i$ in year $y$ compound the months of each half (Eq. 1).
The gap of year $y$ averages the $n_y$ indices with that year, and the gap of a period averages its years, so each year counts once however many indices it has (Eq. 2). Years to October 2002 are before, years from November 2002 after. The yearly gap $G_y$ has a standard deviation of 13.4 points from one year to the next.
The null shuffles each year’s twelve monthly returns among its twelve months, one shuffle per year applied to every index, and recomputes $\bar G$; with $B=10{,}000$ shuffles, the one-sided p-value is Eq. 3. The same shuffles test each index on its own wherever it has five or more years in a period, 16 tests in all, with Benjamini and Hochberg’s (1995) false discovery rate across them, and test the share of years with $G_y>0$. The intervals on $\bar G$ before and after, and on its fall, come from 10,000 resamples of whole years within each period. They put the 95% range of the gap at 3.8 to 11.7 points before 2002 and at minus 2.9 to 8.0 points since, and the gap fell in 93% of the resamples.
The S&P 500’s monthly total return adds one twelfth of Shiller’s annual-rate dividend $D_m$ to the price change, and cash pays one twelfth of the month’s average rate $c_m$ in percent (Eq. 4). The seller in May holds stocks in winter months and cash in summer months (Eq. 5). The DAX uses its own index return, which already carries dividends.
Returns a year are compounded over whole years, risk is the standard deviation of monthly returns times $\sqrt{12}$, and the Sharpe ratio is the mean monthly return over cash divided by its standard deviation, times $\sqrt{12}$. Since 2002 that risk was 10.5% a year for the seller and 14.6% for the holder in the S&P 500, and 13.6% against 18.2% in the DAX. The fixed mix holds stocks in the ratio of the seller’s risk to the holder’s, rebalanced monthly. The breakdowns by rebound year, the five-worst-months check and the fixed mix were added after the main results.
Limitations
- Eight of the nine indices leave out dividends. A price index drops on the day a stock goes ex-dividend, so it understates the half in which its companies pay most of their dividends; for the S&P 500 the two halves carried about the same dividend, 1.0% each since 2002, and its gap with dividends is the same 1.6 points. We have no dividend series for the other seven price indices.
- The money comparison covers the S&P 500 and the DAX, the two indices with dividends and a local cash rate in the data; the other seven enter only the test of the gap.
- The returns are before trading costs and taxes. A seller trades twice a year, and in a taxable account realises gains every May.
- Cash is a monthly average rate divided by twelve. The bill rate is quoted on a discount basis, slightly below the bill’s yield, and overnight rates differ from what a saver earns on a deposit.
- The nine indices overlap: the S&P 500 and the Dow move together, as do the DAX and the Euro Stoxx 50. The shuffle keeps that, and the before period rests on the S&P 500 alone until 1979.
- The halves switch at the last close of April and October, which can differ by a few days from May 1 and Halloween.
Reproducing it
The analysis runs offline in Python 3.12 with NumPy, pandas and PyYAML. It rebuilds every figure and every number quoted here from the sources in the information set above, the index closes and dividends included.
References
- Andrade, S. C., Chhaochharia, V., and Fuerst, M. E. (2013). “Sell in May and go away” just won’t go away. Financial Analysts Journal, 69(4), 94 to 105.
- Atol Research. (2026). What stock markets do in the weeks after a crash. RPT-004, October 5.
- Benjamini, Y., and Hochberg, Y. (1995). Controlling the false discovery rate: a practical and powerful approach to multiple testing. Journal of the Royal Statistical Society, Series B, 57(1), 289 to 300.
- Board of Governors of the Federal Reserve System. (2026). H.15 Selected Interest Rates: 3-month Treasury bill, secondary market rate, monthly (RIFSGFSM03_N.M). Data Download Program, read October 5, 2026.
- Bouman, S., and Jacobsen, B. (2002). The Halloween indicator, “Sell in May and go away”: another puzzle. American Economic Review, 92(5), 1618 to 1635.
- Deutsche Bundesbank. (2026). Money market rates reported by Frankfurt banks: overnight money, monthly average (BBIG1.M.D0.EUR.MMKT.GMFFM.ON.AVE.MA). Read October 5, 2026.
- European Central Bank. (2026a). EONIA, historical close, monthly average (FM.M.U2.EUR.4F.MM.EONIA.HSTA). ECB Data Portal, read October 5, 2026.
- European Central Bank. (2026b). Euro short-term rate (EST.B.EU000A2X2A25.WT). ECB Data Portal, read October 5, 2026.
- Maberly, E. D., and Pierce, R. M. (2004). Stock market efficiency withstands another challenge: solving the “Sell in May/Buy after Halloween” puzzle. Econ Journal Watch, 1(1), 29 to 46.
- McLean, R. D., and Pontiff, J. (2016). Does academic research destroy stock return predictability? The Journal of Finance, 71(1), 5 to 32.
- Shiller, R. J. (2026). U.S. stock markets 1871 to present and CAPE ratio (ie_data.xls). Read October 5, 2026.
- Zhang, C. Y., and Jacobsen, B. (2021). The Halloween indicator, “Sell in May and go away”: everywhere and all the time. Journal of International Money and Finance, 110, 102268.