It is not SAD if you Sell in May: Seasonal Effects in Stock Markets Revisited

Abstract

This paper examines one type of calendar effect in financial markets, seasonal variation in the return on stocks. The effect analyzed is referred to as the Halloween Effect or Sell in May and Go Away. This refers to the finding that stock markets tend to return considerably less in the six months beginning in May than in the other half of the year. This effect has persisted over time and is seemingly large enough to be economically significant. The alternative, but somewhat overlapping hypothesis, that seasonal affective disorder, SAD, creates seasonal variation in the return on stocks is also addressed. Based on daily data from 75 stock markets during the period 2000–2014 there is a strong calendar effect in a large majority of the markets as the period from November to April witnesses higher returns than do the other six months of the year. However, there is only weak evidence that SAD had any effect on stock prices. The paper shows that Halloween Effect seems to be remarkably consistent, even after being widely discussed, and both statistically and economically significant. However, it remains unexplained. The SAD hypothesis finds less support in the data and seems of limited relevance economically.

KEYWORDS:

• Seasonal effects
• Halloween
• stock market
• SAD

JEL CLASSIFICATIONS:

• G14
• G15

Disclosure statement

No potential conflict of interest was reported by the author.

ORCID

Gylfi Magnusson http://orcid.org/0000-0002-6450-8597

Notes

1 It is thought possible that the populations of the far north ultimately adapted to some extent to the long, dark winters and have become less susceptible to SAD. Axelsson, Stefánsson, Magnússon, Sigvaldason, and Karlsson (2002) found that Canadians of Icelandic descent are less prone to SAD than are their fellow countrymen. Similarly, Magnusson (2000) found that the prevalence of SAD varies among ethnic groups.

2 The largest city in New Zealand, Auckland, is north of the 40th parallel.

3 The Efficient Market Hypothesis has a long history, dating back at least to Regnault (1863). It also has many contributors and versions. This history will not be covered here. The interpretation of the hypothesis that market prices generally will move to eliminate all arbitrage opportunities that rely on data available publicly will be used.

4 The cut-off point for significance for the degrees of freedom in this sample and 95% confidence interval was approximately 1.96, but we rounded this off to 2 in the text. It also should be noted that with 74 markets, one can expect a false positive for statistical significance in an average of 3.7 markets in a sample with purely random data.

5 We used the statistical package Gretl and the built-in HC₁ correction for heteroskedasticity adjusted for degrees of freedom used.

6 We use the same measure of SAD incidence and onset/recovery as Kamstra, Kramer, and Levi, available from the website of Lisa Kramer at http://www.lisakramer.com/data.html.

7 Since there is a 50% chance of picking a ‘good’ period (with above average returns) for holding stocks if one selects a random period our algorithm essentially cheated by selecting the period to hold stocks after calculating the average return for each pair of periods. This means that the statistical test for the significance of the Halloween effect is more stringent (reported p values twice as high) than it would have been if our algorithm had picked one out of two periods to hold stocks before knowing the average return in each period. Note also that the random periods generated for the bootstrap test could be of any length, from 1 to 364 days, not just 6 months. The rationale for this is that the six month/six month split of the Halloween effect is ad-hoc and an investor might choose either longer or shorter periods than six months to stay in (and out) of the stock market in a quest for superior returns.

8 There are some studies available into the seasonality of consumer prices but we will not provide a survey here. Bryan and Cecchetti (1995) finds little obvious seasonal movement in the US CPI although some components have a strong seasonal movement. Hamid and Dhakar (2008) likewise looks at the US CPI and finds some periodicity with a high point reached in July.

9 A linear regression with the difference in monthly returns between the periods November–April and May–October as a dependent variable and the difference in the monthly inflation rates in the same period as the independent variable suggested that the link between the two was statistically significant (t-statistic for the slope 5.7). The co-efficient for the slope was however only 0.013 and r² only 2.12% suggesting that inflation differences may explain an even smaller proportion of the differences in nominal returns on shares between these two periods.

10 DeGennaro, Kamstra and Kramer (2008) find a seasonal pattern in bid-ask spreads in the U.S. and argue that they reflect changing risk aversion among market makers.

Additional information

Notes on contributors

Gylfi Magnusson

Gylfi Magnusson is associate professor at University of Iceland School of Business. PhD in Economics 1997 from Yale University. Professor Magnusson has previously served as Iceland's Minister of Economic Affairs and is currently chair of the supervisory committee of the Central Bank of Iceland.