Market cycles and seasonality
The answer
Markets go through repeating phases — accumulation, advance, distribution, decline — and those phases are real, but they have no fixed length, so you cannot use them as a calendar.
Seasonality is different and worse. Almost every seasonal pattern people quote was found by searching a small amount of data for something that looked good. A few seasonal effects are mechanical and survive. Telling the 2 apart is the most valuable skill in this article.
Why this costs you money
Somebody shows you a table. Over the last 20 years, a certain index rose in a certain month 16 times out of 20. That is 80%. You feel something click.
Here is what you were not shown. There are 12 months. There are also 52 weeks, 5 weekdays, roughly 21 trading days in a month, 4 quarters, expiry weeks, pre-holiday days, post-holiday days, and any combination of those. Somebody with a spreadsheet can test several hundred calendar patterns in an hour.
With 300 patterns tested, a handful will show 16 out of 20 by chance alone. Not because the market has a rhythm. Because you looked 300 times.
The money is lost in 2 places. The obvious one is a trade placed on a fake pattern, which is roughly a coin flip with brokerage on top. The bigger one is quieter: you build a whole approach on the belief that markets have hidden timing, and you spend 4 years looking for it instead of learning something that works. The cost is not the losing trade. It is the 4 years.
How it works
Two different ideas are being confused, so separate them first.
A cycle is a repeating movement with 3 properties. Period is how long 1 full repetition takes. Amplitude is how large the movement is. Phase is where in the repetition you currently are. A tide has all 3 and they are stable, which is why tide tables work.
Market cycles have all 3 and none of them are stable. The period wanders. The amplitude varies by a factor of 10. The phase can only be identified after it has passed. That is not a small imperfection. It removes the entire practical value of the cycle as a timing device, while leaving its value as a description.
The description is worth having. Most markets pass through 4 phases.
| Phase | What is happening | What the crowd says |
|---|---|---|
| Accumulation | Informed buyers absorb supply after a decline; price goes sideways | "This is dead money" |
| Advance | Price rises, participation broadens, news improves | "The trend is your friend" |
| Distribution | Price goes sideways at a high level; volume stays heavy | "It is consolidating before the next leg" |
| Decline | Price falls, participation narrows, news worsens | "This is a healthy correction" |
This is useful because it tells you which phase makes which strategy work. Trend-following earns money in advance and decline, and bleeds in accumulation and distribution. Mean-reversion does the opposite. Knowing which phase you are in would be worth a great deal. You cannot know it in advance, and any method that claims to identify the phase in real time should be tested with the methods in articles 3 and 6 of this cluster before you believe it.
Seasonality is a much narrower claim: that returns depend on the position of the calendar. It comes in 2 kinds, and they are not equally credible.
Mechanical seasonality has a cause you can name that does not depend on psychology. Money flows on fixed dates. Contracts expire on fixed dates. Tax years end on fixed dates. Funds report on fixed dates. When there is a real mechanism, the pattern usually survives, because the mechanism keeps operating after you discover it.
Statistical seasonality is a pattern with no mechanism, found by looking. It usually does not survive, because there was never anything holding it up.
The test is simple and you should apply it every time: can you name the mechanism, and does the mechanism still exist?
What it tells you, and what it does not
Seasonal evidence tells you about an average. It does not tell you about a year.
Suppose a month has averaged +2% over 25 years. That average can be produced by 24 flat years and 1 year of +50%. It can also be produced by 13 up years and 12 down years. These are completely different situations for somebody who has to trade 1 of them, and the average hides both.
Always ask for the hit rate and the spread, not the average. How many years positive out of how many? What was the worst year? A pattern that is up in 13 out of 25 years is a coin flip regardless of how good its average looks.
Seasonality also does not tell you about your holding period. Most seasonal studies are of an index. Your portfolio is not the index. A pattern in the average of 50 companies can be absent in 45 of them.
And it cannot tell you whether it will continue. That is the deepest problem in this article, and it has a name.
The decision rule
Before you act on any seasonal or cyclical claim, answer 4 questions in order. If any answer is unsatisfactory, stop.
1. What is the mechanism? Name the flow of money or the rule that causes it. "Sentiment improves in that month" is not a mechanism. "Index funds rebalance on that date" is.
2. How many patterns were searched to find this one? If you do not know, assume it was many. Somebody found this by looking.
3. What is the hit rate and the sample size? 20 years of an annual pattern is 20 observations. That is a very small number.
4. Has it worked since it was published? This is the single best test available and almost nobody applies it.
The fourth question deserves a name of its own. When a pattern is published, people trade it, and trading it removes it. Researchers have measured this in the United States and found that the returns of published market anomalies fell substantially after publication. So the correct question is never "did this work from 1970 to 2000?" It is "did this work from the year it became famous until today?"
Try this now
You are going to test a seasonal claim on your own holdings, and then count how many claims you could have tested. The second part is the lesson.
- Pick 1 stock or fund you actually own and have held for a while. Open a chart and set it to monthly candles, maximum available history.
- Choose 1 calendar month — use the month you have most often heard a claim about, or just pick the current one. Count how many of the available years that month was up, and how many it was down. Write it as "up 7 of 12".
- Now write down the worst of those months, as a percentage. That single number is what the average was hiding.
- Now the important step. On a piece of paper, list the other calendar patterns you could just as easily have tested on this same stock: the other 11 months, the 4 quarters, the first half against the second half, the 5 weekdays, expiry week against non-expiry week, the week before a holiday. Count them.
- Look at your count from step 4. That is the number of chances the data had to produce something that looked impressive.
What you should see. Your month is probably up somewhere between 5 and 8 times out of 12, which is indistinguishable from a coin. Your worst year in that month is probably large enough to have ended the strategy. And your count in step 4 is probably somewhere between 30 and 80.
If it does not look like a coin — if your month is up 11 of 12 — go back to step 4 and take the count seriously. With 40 patterns available and a short history, finding 1 that looks extraordinary is the expected outcome, not a discovery. Article 8 in this cluster gives you the arithmetic for exactly how often this happens by chance.
Three real cases
1. The January effect (United States, documented 1976) — a real finding that decayed after publication Researchers documented that returns in January were unusually high on an equal-weighted index of American shares, using data going back to the early 1900s. Later work linked the effect mainly to small companies, and connected it to tax-loss selling in December followed by repurchase in January. That is a real mechanism. The effect nonetheless weakened substantially after it became widely known and traded. This is the cleanest available lesson: even a seasonal pattern with a genuine cause can be arbitraged away once everybody knows about it.
2. "Sell in May and go away" (multiple countries, tested 2002) — documented widely, contested ever since An academic paper examined the pattern of weak returns between May and October across a large number of countries and reported that it appeared in almost all of them. It became one of the most quoted seasonal claims in finance. Subsequent papers argued that the result was heavily influenced by a small number of extreme months, particularly around the October 1987 crash and the 1998 crisis, and that removing those months weakened it substantially. Both sides of that argument are published and honest. The point for you is not who won. It is that a famous, widely tested, internationally replicated seasonal effect is still genuinely disputed by careful people 20 years later. Your 20-year table is not stronger evidence than that.
3. Calendar effects tested as a group (United States, 2001) — the multiple-testing correction applied to seasonality A study examined calendar trading rules as a whole rather than 1 at a time, using a method that accounts for how many rules were searched. The finding was that once you account for the full universe of calendar patterns somebody could have tested, the evidence for the individual famous effects becomes far weaker than it appears in isolation. This is the single most important piece of research in this article. It does not say seasonality is fake. It says that the correct way to judge 1 pattern is to ask how many patterns were in the search that produced it.
The question that resolves it
A novice sees a seasonal table and asks: how often did it work?
An expert asks: how many tables were made before this one was shown to me?
The first question is answerable from the table. The second is answerable only by the person who made it, and they are usually not present. When you cannot answer it, the honest default is to assume the search was large — because the patterns that get shown to you are, by definition, the ones that survived a search.
What would make this wrong
If seasonality were reliable, then simple calendar rules would have produced consistent profits after they were published. Take any famous calendar effect, find the year it became well known, and test it only from that year forward. If it still works, this article is too pessimistic and you should say so.
The honest limits are 3, and 2 of them cut against my argument.
First, some seasonality is genuinely mechanical and does survive. Index rebalancing dates, quarterly futures expiry, tax-year-end flows and dividend dates all cause real, repeated pressure on real dates. These are not superstitions. They are plumbing.
Second, cycles are a good description even when they are a bad timer. Saying "we are probably in distribution" is not a prediction, but it does change how much you risk. Dismissing cycle language entirely throws away a useful vocabulary for describing market behaviour.
Third, absence of evidence over 20 observations is not evidence of absence. With only 20 or 30 annual data points, a real effect of moderate size would be undetectable. So "the data does not support it" and "it is not real" are different statements, and I have to be as careful about that as the person selling the pattern.
In India
Indian seasonality has 3 features that make it a much better teaching case than the American version.
Expiry effects are mechanical, real and large. Indian index options have weekly expiry, and a very large share of the country's total trading volume concentrates into expiry sessions. This creates a genuine weekly pattern in volume, in intraday volatility and in the behaviour of option prices near expiry. It is not mystical. It is a settlement rule producing a settlement effect. Note that the specific expiry weekday has been changed by the regulator and the exchanges more than once in recent years, and it is still being adjusted. That instability is itself the lesson: a pattern caused by a rule dies the day the rule changes.
Muhurat trading. On Diwali, the exchanges hold a special 1-hour trading session marking the start of the new Samvat year. It is widely reported that this session closes higher in most years. Treat it carefully. It is 1 hour per year, so 30 years of history is 30 observations, and the session is thin, ceremonial and dominated by token buying. It is a lovely tradition and it is close to useless as evidence of anything.
Budget day. The Union Budget is presented on 1 February and it re-prices entire sectors within hours. The pattern here is not "the market goes up on budget day". The pattern is that implied volatility rises before the event and collapses after it, which is a mechanical consequence of a scheduled announcement, not a directional forecast. The same is true of monetary policy dates and major election counting days.
On data. Indian seasonality studies are limited by history. The NIFTY 50 has values back to 1990 through back-calculation, and the Sensex reaches further, but a clean, adjusted, point-in-time dataset for individual companies is hard for a retail investor to obtain. That matters more than it sounds. Fewer years means fewer observations means a lower bar for a fake pattern to clear.
In the United States
American seasonal claims are the most studied in the world, and there are a lot of them: the January effect, the Halloween or sell-in-May effect, the turn of the month, the Santa Claus rally, the Monday effect, the pre-holiday effect, the 4-year presidential cycle, and the best-6-months rule popularised by a widely sold market almanac.
Three things about this deserve your attention.
The data is far better. Free datasets reach back to 1926 for broad American market returns, and 1 well-known academic dataset reaches to 1871 for prices and dividends. This is a genuine research advantage and it is why almost all seasonality literature is American.
The famous effects have decayed. The Monday effect, one of the most reliably documented patterns of the 1970s and 1980s, weakened or reversed after publication. The January effect weakened. This decay is now itself a documented phenomenon across a wide range of published anomalies. Some American seasonality is mechanical and persists. Quarterly index rebalancing, the quadruple-witching expiry sessions, and month-end flows from pension contributions are all rule-driven. They mostly affect volume and short-term liquidity rather than direction, which is exactly the shape you should expect from a real mechanical effect.
Where they differ, and what that tells you
The difference is length of history, and it points in the opposite direction from what most people assume.
An American researcher has roughly 100 years of clean data. An Indian researcher studying company-level effects has a fraction of that, in a market whose structure has changed enormously since the 1990s.
The obvious conclusion is that Indian seasonality claims are less reliable, and that is true. But there is a second, less obvious conclusion, and it is more important.
A shorter history makes fake patterns easier to find, not harder. With 15 years of data, a coin-flip pattern will produce a striking-looking run far more often than with 100 years. So the Indian market generates more impressive-looking seasonal tables than the American market does, from less evidence. Every Indian seasonality claim you encounter should be discounted twice: once for the search that produced it, and once for the short history it was found in.
There is a compensating point in India's favour. India's mechanical effects are larger and clearer, because the derivatives market is enormous relative to the cash market and expiry concentrates it. So the Indian pattern worth studying is the plumbing, not the calendar. That is the reverse of how the subject is usually taught here.
Carry this
- Name the mechanism, or do not trade the pattern.
- Ask how many patterns were searched. If you cannot find out, assume many.
- Ask whether it has worked since it was published. Almost nobody does.
- Hit rate and worst year, never the average.