Seasonality, an honest look

Seasonality claims are everywhere: the Santa rally, sell in May, the Monday effect, the budget-month bounce. They are appealing because they are concrete and easy to test.

They are also where the least rigorous analysis in the whole field lives, and it is worth understanding exactly why before deciding what to keep.

The sample size problem#

This one objection disposes of most seasonal claims.

A pattern about a specific month has one observation per year. Twenty years of data gives you twenty samples. Twenty is not enough to distinguish a real effect from chance — the journal lesson said fifty trades before concluding anything, and that was for a setup occurring weekly.

A claim like "March is historically strong" rests on a sample too small to support it, however confidently it is stated.

Data mining#

The deeper problem. With enough calendar variables — months, weeks, weekdays, days-of-month, pre-holiday sessions, expiry weeks — you can generate thousands of possible patterns.

Test thousands of hypotheses against one dataset and some will show strong results by chance alone. That is not a flaw in the testing; it is what randomness does.

Almost every published seasonal pattern was discovered by searching for it. Which means the striking ones are exactly the ones most likely to be artefacts, because a search selects for extremes.

What has a mechanism#

The distinction that matters is not whether a pattern shows in the data. It is whether there is a reason it should exist.

A pattern with a mechanism can be reasoned about, and you can tell when it might stop working. A pattern without one is a coincidence you have named.

Things with a genuine mechanism in Indian markets:

Expiry effects. Real, and driven by identifiable flows — rolling, gamma, pinning. This is not seasonality in the calendar sense; it is mechanics with a schedule.

Financial year-end (March). Tax-related selling and buying, and institutional book-squaring, genuinely cluster. The mechanism is concrete.

Results season. Index volatility rises when heavyweight constituents report. Entirely mechanical.

Budget period. A scheduled event with real policy consequences, not a calendar superstition.

Monsoon. Affects agricultural output, rural demand and inflation expectations. A slow, genuine economic input — though the market prices forecasts continuously, so surprises matter more than the season itself.

Notice these are all scheduled events with identifiable flows, not calendar effects. That is the category worth keeping.

What does not survive#

Day-of-week effects. Widely claimed, extremely weak, and any that existed have been arbitraged.

Month-of-year effects beyond the mechanically-explained ones. Small samples plus data mining.

"Sell in May." Built on Northern Hemisphere data with no mechanism that transfers to India.

Santa rally. Overwhelmingly a small-sample artefact.

Individual day-of-month patterns. Pure data mining.

The reflexivity trap#

Even genuine seasonal effects degrade once known — the same problem the sentiment track described.

If a March effect is real and widely published, participants position for it in February. The effect moves earlier, weakens, or inverts. A pattern strong enough to be famous is strong enough to have been traded away.

This means the useful seasonal knowledge tends to be structural rather than predictive. Knowing that March has year-end flows helps you interpret unusual volume. It does not give you a trade.

What to actually do with this#

Use it to explain, not to predict. If volume is unusual in late March, year-end flows are a plausible reason. That is a legitimate use of the knowledge.

Prefer mechanism over statistics. If you cannot say why a pattern should exist, do not trade it however good the backtest looks.

Never size up on seasonality. It is at best a marginal input, and it should never be the primary reason for a position.

Be sceptical of confident seasonal claims — including any you develop yourself. The confidence of a claim is uncorrelated with its validity, and your own patterns are the ones you will scrutinise least.

Check yourself

0 of 4 answered

1. 1.Why is 'March is historically a strong month' weak evidence even with 20 years of data?

   Twenty years is too far in the past to be relevantA monthly pattern gives one observation per year, so 20 years is only 20 samples — far too few to separate signal from chanceMarch data is unreliable in IndiaMonthly patterns require intraday data to verify

2. 2.What is the data mining problem in seasonal analysis?

   Exchanges restrict access to historical dataWith enough calendar variables, thousands of patterns can be tested — and some will look strong purely by chanceHistorical data contains too many errorsMining requires expensive software

3. 3.Which calendar effect has a genuine, identifiable mechanism?

   The Santa rallyDay-of-week effectsExpiry effects — driven by rolling flows, gamma and pinningSell in May

4. 4.Why do genuine seasonal effects tend to weaken once they become well known?

   Regulators prohibit trading themParticipants position for them in advance, so the effect moves earlier, weakens, or invertsHistorical data gets revisedThey do not weaken; known effects are stronger

What to take away

• Most seasonal claims fail on sample size — one observation per year is not evidence.
• Data mining guarantees striking patterns from randomness; searches select for artefacts.
• Keep patterns with a mechanism: expiry, financial year-end, results season, budget, monsoon.
• Discard day-of-week, sell-in-May, Santa rally, day-of-month — no mechanism, small samples.
• Known effects degrade through reflexivity.
• Use seasonality to explain, never to predict, and never to size up.
• The one calendar habit worth keeping: know where you are in the expiry cycle.