Radar Perene / Archive / science

Seasonality or folklore: what survives of \"sell in May\"

◦ Index methodology v2.2 (working papers with DOI). See the methodology.

Science

"Sell in May and go away." The proverb is older than most of the data series used to test it — the full English version told the investor to come back only on St. Leger's Day, a September horse race, which betrays its origin: Victorian London, where high society abandoned the city — and the exchange — for the summer. A century and a half later, the advice survives in every market language, repeated each northern-hemisphere autumn by people who never saw a test. It is the perfect specimen of an entire category: the calendar pattern everyone knows and almost nobody has measured.

Statistical seasonality is a recurring calendar-linked pattern that survives a formal test: hypothesis declared before measurement, sufficient sample, verification outside the window in which the pattern was noticed. Folklore is the same statement sustained only by the repetition of whoever tells it. The border between the two is not in the content of the sentence — it is in the path the sentence traveled before being spoken.

The distinction matters because the calendar is the most democratic of pattern machines: every year has twelve months, and one of them was always the best in the sample.

What the serious test found

"Sell in May" has the rare honor of having been taken seriously by academia. In 2002, Sven Bouman and Ben Jacobsen published in the American Economic Review the most cited test of the effect: they examined decades of returns across dozens of markets and found, in the sample studied, average November-to-April returns above those of May-to-October in the great majority of them. In numbers: the pattern appeared in 36 of the 37 markets examined in the study — a result too strong to be dismissed with a shrug.

The shrug came later, in the correct form: scrutiny. Critics pointed out that part of the effect depended on concentrated extreme episodes — remove a few catastrophic months and the difference withered in several markets — and raised the permanent suspicion that a pattern noticed before being tested carries selection bias from birth: nobody writes a paper about the proverbs that failed. Later work, testing the period after publication, found the effect still standing in a good share of markets — which distinguishes it from dozens of cousins that evaporated the day after fame. The honest scoreboard, a quarter century on: a statistically curious pattern, unstable across markets and periods, without a consensual explanation, and weaker than the proverb promises.

Why the calendar manufactures patterns

The reason to demand so much of a seasonal pattern is arithmetic, not temperamental skepticism. A year offers twelve months, four quarters, dozens of holidays and hundreds of windows — and each cut can be crossed with each index, in each country, in each period. Whoever searches that space finds: it is a mathematical guarantee, the same one this trail described in p-hacking. The calendar is coincidence's ideal habitat because it produces cuts in industrial quantity and they all look natural — "January", "the eve of a holiday", "the first half" sound like categories of the world, not choices of the researcher.

There is one more detail the Brazilian reader has reason to notice: the proverb was born in a northern-hemisphere social calendar — their summer, their holidays, their horse race. The most cited explanation for the effect, the summer exodus, simply does not transfer to a market where May is autumn and the long holidays fall at another point of the year. A pattern that travels across hemispheres without its explanation traveling along deserves the question: what exactly is being imported — the phenomenon or the phrase?

The house's ruler for calendar claims

This house handles seasonal statements under an editorial rule the reader can check on any page of the archive: a frequency claim requires literal backing in the source. "Tends to fall in May", "historically rises at year-end" — sentences of that shape only enter an essay when the house's archive contains the count that sustains them, with the period declared; without the count, the sentence goes. It is a deliberately tedious ruler, because calendar folklore is precisely the kind of statement that crosses decades without ever being billed.

The same discipline has already produced documented mortality on other terrain: when the house retested, in a public working paper, a set of market relationships that custom treated as reliable, most did not survive. There is no reason to expect calendar proverbs to fare better than intramarket gears — which is why none of them rides free in the archive.

Frequently asked questions

Does "sell in May" work in Brazil?

The right question comes first: has anyone tested it with a declared hypothesis, a long period and out-of-sample verification? Scattered studies exist, with results sensitive to the window — and the origin explanation (the northern-hemisphere summer) does not apply to the local calendar. A statement without a citable local test remains, here, in the folklore category.

Is the January effect the same thing?

Another member of the family: the historical observation of above-average January returns, especially in smaller stocks, documented in American data from the middle of the last century. It is also the classic example of a pattern that shrank after becoming famous — today the literature describes it more as a museum piece than as a living regularity.

If a seasonal pattern were real, shouldn't it vanish as soon as everyone knew?

That is the efficiency argument, and the persistence of "sell in May" after the 2002 publication is one of the puzzles keeping the debate open. The candidate explanations — risk varying across the year, social habit, prolonged coincidence — remain in dispute; none has won.

How does a reader distinguish tested seasonality from folklore?

By the travel certificate: does the text declare when the hypothesis was formulated, in which sample it was tested and what happened outside it? Does it cite the months in which it failed? A seasonal claim that only presents the years in which it worked is folklore with a chart.

Testing a proverb properly requires the same protocol as testing any rule: split the sample first, declare the hypothesis first, count every attempt. That protocol has a name, and it closes this trail: Honest backtesting: split the sample before testing anything →

House readings: today's note, in the Daily · the precedents, in the Atlas.

Submitting a specific calendar pattern to that ruler is workbench material — the kind of exercise the house conducts on request.