Stock market seasonality: Sell in May, the January effect and how to test calendar effects

In depth11 min read · 16 sources

Seasonality in the stock market means that returns seem to depend on the calendar, for example on the half of the year, the turn of the month or the days before public holidays. The best-known rule is “Sell in May and go away”. Here you will learn what research shows about these calendar effects, how to test such an effect for significance step by step, why caution is needed when many calendar rules have been tested, and how much switching eats up in costs and taxes.

Calendar effects: what is meant

A calendar effect is a return pattern tied to a fixed date rather than to new information. If it held reliably, it could be exploited with a simple rule. That would contradict the weak form of the efficient market hypothesis, according to which past prices contain no exploitable patterns. The best-known patterns:

What the research shows

Sell in May: the Halloween effect

Bouman and Jacobsen (2002) systematically tested the market wisdom of selling in May and only returning after Halloween. In 36 of the 37 developed and emerging markets examined, returns from November to April were higher than those from May to October. The Halloween effect was particularly pronounced in Europe, and in the UK it can be traced back to 1694. The authors found no convincing explanation. Andrade et al. (2013) tested the rule outside the samples of earlier studies and found that it persists. Finally, Zhang and Jacobsen (2021) analysed all available stock indices worldwide, 62,962 observations in total: in the winter half-year, returns were on average 4% higher, and in the summer half-year, excess returns over the risk-free rate were even negative worldwide, at around −1%.

The January effect, the turn of the month and holidays

Rozeff and Kinney (1976) described the seasonality of US stock returns, with strikingly high returns in January. Keim (1983) showed that a large part of the size effect (small stocks beat large ones) falls in January, above all in its first few trading days. A common explanation is that investors sell losers at the end of the year for tax reasons and buy them back in January. Haug and Hirschey (2006) found the January effect in small US stocks unchanged even after the US Tax Reform Act of 1986, and therefore consider behavioural explanations important.

For the turn-of-the-month effect, Ariel (1987) found that US stocks earned almost all of their return in the first half of the trading month, counted from the last trading day of the previous month. McConnell and Xu (2008) go further: from 1926 to 2005, US investors on average received a premium for market risk only around the turn of the month, and the effect appears in 31 of 35 countries; buying pressure from month-end inflows does not explain it. In 90 years of daily Dow Jones data, Lakonishok and Smidt (1988) found unusual returns around the turn of the week, month and year and around holidays. According to Ariel (1990), the mean return on the trading day before a holiday was nine to fourteen times as high as on the remaining days; from 1963 to 1982, more than a third of the market return fell on these eight days a year.

Year-end rally and Santa Claus rally

The Santa Claus rally is a term from market jargon, usually for the last five trading days in December and the first two in January. Academically, it belongs to the holiday and turn-of-the-year effects described by Lakonishok and Smidt (1988) and Ariel (1990); separate, robust evidence for exactly this window is thinner. If you look for a “year-end rally”, you will find it stronger or weaker depending on the definition (from the whole of December to just a few days). That alone hints at the problem of the many variants, which is discussed further below.

Significance test step by step

Whether a calendar effect is more than chance is checked with a hypothesis test. The null hypothesis is that the expected return is the same in the winter and summer half-years. The two-sample t-test compares the two means relative to their dispersion:

The t-test with pooled variance assumes the same dispersion in both half-years. If you calculate with separate variances (Welch's test), the t-statistic here stays at 2.31 because both groups are the same size; only the degrees of freedom fall to 21.3, and the p-value stays at 3.1%. The p-value is the probability of seeing a difference at least this large if in reality there is none. It is not the probability that the effect is real.

The same calculation as a regression with a dummy

Bouman and Jacobsen (2002) test the effect as a regression. A dummy variable S_t is 1 if period t falls in the winter half-year, and 0 otherwise:

With the twelve winter and twelve summer values, the least squares regression gives exactly the figures from above: μ̂ = 0.17%, α̂ = 4.83 percentage points with a standard error of 2.09 and a t-statistic of 2.31. The advantage of the regression: further dummies for other calendar windows, a trend or factor returns can easily be added.

With monthly data, things get trickier. Each year then contributes six winter and six summer months to the regression. That is more values, but not correspondingly more independent information: volatility fluctuates in phases (heteroskedasticity), and returns can be autocorrelated. The ordinary standard errors then turn out too small. The remedy is Newey-West standard errors following Newey and West (1987), which include covariances up to a lag L with declining weights:

For advanced readers: data snooping with calendar rules

Calendar effects are the textbook example of data snooping. There are a great many possible rules: twelve months, five weekdays, windows around the turn of the month and holidays, half-years, and combinations of these. If you test enough rules on the same data, one of them will turn out “significant” by chance. Sullivan et al. (2001) worked this through on around 100 years of daily data: with a bootstrap over the whole set of calendar rules that could have been examined, the individually highly significant effects are no longer significant. Their tool is the Reality Check of White (2000), which they had previously applied to technical trading rules (Sullivan et al. 1999).

The simplest correction for multiple testing is the Bonferroni rule: with m rules tested, each individual rule has to fall below the level of 5% / m.

Then there is the choice of period: an effect discovered in years when it was strong is convincing in precisely those years. Genuine tests need data from after publication or from other markets, as in Andrade et al. (2013) and Zhang and Jacobsen (2021). How to correct a Sharpe ratio for many trials is explained in the article deflated Sharpe ratio, and the basics under backtest overfitting.

Costs and taxes when switching

A Sell-in-May rule switches twice a year: at the end of April out of the equity ETF into a money market ETF, and back at the end of October. Every order costs fees and half the bid-ask spread, and in Germany every sale at a profit immediately triggers taxes that buy and hold would defer.

The tax rules reflect the position as of October 2026: flat-rate withholding tax (Abgeltungsteuer) of 25% plus the solidarity surcharge, 26.375% in total without church tax, a saver's allowance of €1,000 (€2,000 for jointly assessed couples) and a partial exemption of 30% for equity funds (EStG 2009; InvStG 2016). Income from a money market ETF, by contrast, is fully taxable with no partial exemption. More under flat-rate tax and partial exemption and Vorabpauschale; this is not tax advice.

Limits and common mistakes

• A conveniently chosen period: a pattern discovered during a phase when it was strong always looks good there. Test it above all after publication and in other markets.
• Few independent observations: 20 years give only 20 pairs of winter and summer. Monthly data increase the number of values, not the number of independent half-years.
• Comparing only the return: a summer in the money market also reduces risk. A fair comparison looks at the Sharpe ratio and drawdown, not at return alone.
• Forgetting costs and taxes: two switches a year cost the spread, fees and the benefit of tax deferral. In the example, trading costs alone come to a good 0.2% a year.
• No explanation: to this day there is no generally accepted rationale for the Halloween effect. That makes it puzzling, but also liable to disappear again.
• Many rules, one winner: if you try out months, weekdays and holiday windows, you will always find something. How high the hurdles for getting in and out are is shown in the article market timing.

How Portfolio Lab calculates it

Tactical has no calendar model of its own; its models respond to prices, volatility and valuation, not to the date. You can, however, recreate a calendar rule in the Backtest with time-dependent weights (the “Weights over time” card). With the “Table” option, each row applies from its date: a row from 1 May with 100% money market switches at the closing price of the last trading day before it, i.e. at the end of April with monthly data; a row from 1 November with 100% equities switches back at the end of October. For many years, the “File” option is more convenient: a CSV or Excel file with a date column and one column per ticker; switching then takes place at the closing price on the date given or on the next trading day. The table accepts up to 500 rows, a file up to 5,000.

With “Savings plan and trading costs as at your broker” and “Include German taxes”, the backtest factors the spread, order fees, flat-rate tax and partial exemption into every switch; the metrics then also show the final value without trading costs. For comparison, you add buy and hold as a second portfolio alongside it. You can see a portfolio's seasonal patterns in the “Months” tab, with the monthly returns for each year. The app does not run a significance test or a correction for multiple testing; the period returns from the CSV export are suitable for that. How schedules are implemented is described in the backtest methodology; a guide to the whole process is under backtesting an ETF portfolio.

This article explains a method; it is not investment or tax advice. Worked examples are simplified and rounded. All terms in the glossary.