Seasonality: Ask for the Mechanism
Seasonality is any market pattern that repeats by calendar date rather than by price behaviour. A few have physical or institutional causes, such as harvests, heating demand or tax-year flows. Most are found by scanning calendars, and an annual effect over twenty years is twenty observations.
How it works
Seasonality is any market pattern that repeats by the calendar rather than by price. The claim is that a particular month, weekday or run-up to a holiday behaves differently from the rest of the year. It is a statement about dates, not structure.
Two very different things travel under the same name. One has a cause you can point at, such as harvest cycles in agricultural commodities or heating-oil demand. The other was found by searching dates for anomalies.
The mechanical ones rest on somebody being obliged to act at a set time. Crops arrive when they arrive, and heating demand rises in winter. Tax-year-end flows and index rebalancing dates move money on a published schedule.
Why the sample is smaller than it looks
Slice the year finely enough and a pattern is certain to appear. Month, week of month, weekday, day of month, days before a holiday: each slicing multiplies the comparisons being run. Something clears any threshold by chance alone.
The point almost nobody states out loud is the observation count. An annual effect studied across twenty years is twenty observations. Calling it twenty years of data borrows credibility from a sample that does not exist.
Monthly effects give twelve readings a year, but they are not independent. The same regime and positioning run through consecutive months. Correlated observations inflate apparent significance without adding information.
In practice
Ask for the mechanism before accepting the statistic. Name the participant, the obligation and the date. If nobody can say who is forced to transact, treat the pattern as a search result.
Trading volume genuinely is seasonal, and that part is observable. Holiday weeks and mid-summer are thinner. Thin books mean wider spreads and worse fills, which is a use of seasonality needing no forecast.
Evaluation is slow in a way that defeats most people. An annual effect adds one observation per year, so abandoning it takes decades to justify on evidence.
A single unusual year can manufacture the whole average. One crisis or policy shock in the right month drags the mean across every other year.
A calendar says when to act and nothing about where you are wrong. A date has no invalidation level, so the stop must come from structure, from volatility, or from a loss accepted in advance.
The claims worth naming, and what they need
The famous seasonal claims are famous because they are memorable, not because they are settled. The January effect, “sell in May”, the Santa Claus rally and tax-year-end flows all circulate widely and are easy to state in a sentence. That memorability is why they survive without a mechanism attached.
Each one can be examined the same way. Ask what the proposed cause is, whether it is documented independently of the price pattern, and whether it still operates. Tax deadlines move, fund structures change, and a widely published effect gets front-run.
Some causes are durable because they are physical. Crops do not read research papers, and cold weather is not arbitraged away. Institutional causes are more fragile: a rule that creates a flow can be rewritten.
That is the whole test in one line. Name the cause, check that it is still in force, then decide whether the remaining edge survives the cost of trading it.
What seasonality is not
- Not a forecast. It describes an average across years, not the year you are in.
- Not a large sample. Twenty years of an annual effect is twenty observations.
- Not an entry signal. A date supplies timing and nothing about level or risk.
- Not a substitute for a mechanism. A statistic without a cause is a search result.
When it fails
- The cause has expired. Institutional seasonals die when the rule, deadline or fund structure that created them is changed. The record still shows the pattern; the market no longer has the reason.
- The pattern was never there. With enough calendar slices, overfitting produces effects that vanish the moment they meet data they were not fitted on.
- One year is carrying the average. Remove the largest contributing year and the remainder is often noise, so the spread of outcomes matters more than the mean.
- The market is in a range and nothing else is happening. In a flat year the calendar is the only visible story, so ordinary drift gets attributed to it.
- The move arrives as an opening gap you cannot participate in. Anticipated seasonals get priced overnight, leaving the date-based entry filled well after the move.
- Costs eat the edge. Seasonal trades are infrequent and the expected move modest, so a rule firing once a year has few chances to earn back each round trip.
The original data
The figures below come from this site’s own measurement files. From research/corpus-coverage.json,
produced by site/measure_corpus.py over a corpus of 31,760 trading and investing videos: “seasonality”
appears in the title of 2 videos, median 65 views, across 1 channel.
Two ordinary indicators from the same file give the scale. “bollinger bands” appears in 311 videos, median 3,816 views, across 173 channels; “ichimoku” appears in 154 videos, median 10,165 views, across 99 channels.
From research/series-measurements.json, produced by site/measure_series.py on this site’s shared 576-bar
history: the round-trip cost is 2% of a median bar’s range and 45% of the smallest bar. The ten-bar
efficiency ratio has a median of 0.34, with 30% of bars above 0.5.
Read the corpus numbers as a demand signal, not a quality signal. Almost nobody publishes on seasonality because almost nobody searches for it. The cost figures matter more: a rule firing rarely has few chances to earn back each round trip.
Now take the hypothetical above. A pattern that worked in seventeen years out of twenty — hypothetical, not measured — is still twenty observations, three of them losses, with no mechanism stated. So write down the cause, the date it operates and the loss you will accept; if you cannot fill in the first field, do not take the trade.
Related
Backtesting is where seasonal claims are born, and its discipline of holding out data matters most when each year contributes one observation. Overfitting is the specific failure that calendar slicing invites, because the possible date filters are many and the independent observations are few. Commodities are where mechanical seasonality genuinely lives, since harvests, storage and weather-driven demand impose schedules no publication can arbitrage away.
I have been handed a lot of seasonal charts over the years and almost none of them came with a reason attached. The ones that stayed useful were the boring ones, where somebody could tell me exactly who was buying or selling and why they had no choice. When there is no such person, the pattern is usually just the calendar being sliced until something looked interesting. I still watch the quiet weeks, but for costs and liquidity, not for direction.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.