Seasonality

seasonality · seasonal demand · seasonal variation · seasonality in ecommerce · seasonal trends · demand seasonality

Definition

Seasonality is the regular, repeating variation in demand across a yearly, weekly, daily or event-driven cycle; it sits on top of the underlying trend and distorts any before-and-after comparison.

Four cycles running at the same time

Seasonality is not just “people buy fans in summer”. In e-commerce, at least four independent cycles
are layered on top of any data series, and confusing one for another is the source of most wrong
conclusions.

Cycle	Period	How it shows up	What it distorts
Yearly	12 months	Category seasons: tyres, air conditioners, back-to-school, gifts	Sales plans, quarter-on-quarter comparison, long experiments
Weekly	7 days	Weekdays versus weekends, different demand mix and order value	Tests not run in whole weeks, weekly reports with a skew
Daily	24 hours	Morning and evening traffic peaks, the overnight trough	Tests launched for a few hours; conclusions from a partial day
Event-driven	Calendar dates	Sales, holidays, pay days, the start of the school year	Comparison of adjacent weeks, extrapolating promo results to normal periods

The event cycle is the least obvious and the most aggressive of the four. Pay days create a stable
two-peak structure across the month, and a week-long sale can produce a bigger amplitude than a
category’s entire yearly season.

How to separate seasonality from the effect of a change

The base mistake is the before-and-after comparison: measure two weeks before the release, two weeks
after, and book the difference as the effect. That difference contains everything at once — the
change, the season, advertising, assortment shifts and the weather.

Techniques, in ascending order of reliability:

Year-on-year comparison. The same calendar period last year as the baseline. It works for
planning but not for evaluating changes: everything else moved between the two years as well.
A 7-day moving average. Smooths the weekly cycle and makes the trend visible. It does not
remove the yearly or event cycles.
Decomposing the series into trend, seasonal component and residual. Useful for demand
forecasting and purchasing, but the residual still contains every external factor except the one
under study.
A control group. Both groups live through the same season, on the
same days, with the same advertising. The difference between them is free of season by
construction — not approximately, but exactly.

⚠

No amount of mathematics rescues a before-and-after comparison. If a change was rolled out to
everyone at once, its effect cannot be separated from the seasonal and advertising background — only
the order of magnitude can be estimated. The one reliable decision is taken in advance: roll out
through an A/B test or keep a holdout group.

The effect on A/B tests

Seasonality breaks experiments in three ways, and all three are avoided by disciplined planning.

Incomplete weekly cycles. A test that runs for 3, 5 or 10 days gives a different balance of
weekdays and weekends in different stretches, and behaviour differs across those days — from the
traffic mix to the order value. The rule: test duration is a whole number
of full weeks, starting and finishing on the same day of the week.

A test inside an abnormal period. A result obtained on Black Friday describes the behaviour of
Black Friday shoppers. The traffic mix is different (more discount hunters), the motivation is
different and the sensitivity to urgency is different. That conclusion cannot be moved to an ordinary
Tuesday.

Rule for transferring conclusions:
  test in a promo period       -> conclusion applies to promo periods
  test in a normal period      -> conclusion applies to normal periods
  test across a period boundary -> conclusion applies nowhere,
                                   the result is mixed

Stopping early on a seasonal peak. The temptation to stop a test when a variation is “clearly
winning” is strongest on high-traffic days, because significance accumulates fast. But a spike that
coincides with a promotion easily produces a false positive. The same rules apply here as against
peeking: a duration and a sample size fixed in advance.

One more point: if a test has to run across a season boundary (catching the start of a sale, for
example), break the result down by period and check whether the direction of the effect holds in
each. Different directions signal that the effect depends on context and that there is no single
conclusion.

The effect on recommendations

Recommendation algorithms learn from history, and history by definition belongs to the previous
season. Without constraints, that leads to predictable artefacts:

in May, New Year gift sets stay at the top of the recommendations, because their winter numbers were record-breaking;
at the height of a season the algorithm is slow to lift new products that have no accumulated statistics yet — the classic cold start on seasonal assortment;
once a season ends, recommendations keep dragging the departing category along for several weeks.

What teams do about it:

Technique	What it delivers
Limiting the training window	The model stops treating last year’s season as a current signal
Higher weight on recent interactions	Fast reaction to a shift in demand inside the week
Scheduled rotation of strategies	Different strategies for seasonal, promo and normal periods
Filtering by availability and seasonal category	Products that are out of stock or out of season leave the output
Merchandising rules on top of the algorithm	Manual pinning of the seasonal top without rebuilding the model

These settings have to be proven like any others: the seasonal configuration against the baseline in
the same period, on the same audience. Otherwise “the seasonal strategy worked” only ever means “the
season started”.

Checklist

The yearly demand profile is known for every key category, not just for the store as a whole.
The weekly and daily profiles are calculated separately — for traffic, conversion and order value.
The calendar of event peaks (sales, holidays, pay days) is documented and used when planning releases and tests.
No decision about the effect of a change is taken from a before-and-after comparison.
Experiment duration is a whole number of weeks, starting and finishing on the same day of the week.
Tests do not start a few days before a major sale and do not finish inside one.
For long-lived tools, a holdout group is kept running across several seasons.
In recommendations, history depth is limited, strategy rotation is configured and a stock filter is in place.
Reports on tool performance always state the period and its character (normal or promotional).

Frequently asked questions

How do you separate seasonality from the effect of your own changes?

A before-and-after comparison will not do it: it blends the effect of the change with the seasonal shift. Supporting techniques are year-on-year comparison of the same calendar period and a moving average that smooths the weekly cycle. But the only reliable method is a control group: both groups live through the same season, so the difference between them is free of seasonal effect by construction.

How long should an A/B test run given seasonality?

The minimum is a whole number of full weeks, so that both groups contain the same balance of weekdays and weekends. A 10-day test skews the mix of days and produces a biased result. The upper bound is set by sample size and the effect you need to detect, but two to four full weeks outside sale periods usually yields a stable conclusion.

Can you run tests during a sale?

You can test, but you cannot carry the conclusion over to a normal period. During a sale the traffic mix, shopper motivation and price sensitivity all change — a mechanic that wins on Black Friday can lose on an ordinary Tuesday. If a change is being prepared specifically for a sale, test it during a sale and apply the conclusion only to periods like it.

How does seasonality affect recommendations?

Algorithms learn from historical data, and without a limit on how far back that history reaches they keep recommending last year's seasonal products. The fix is a combination: limit the training window, raise the weight of recent signals, rotate strategies by season and filter by stock availability. Prove it with an A/B test of the seasonal strategy against the baseline in the same period.

What counts as event-driven seasonality?

Anything that repeats on the calendar without being tied to the time of year: sales (Black Friday, 11.11, discount seasons), holidays and gifting periods, pay days, the start of the school year, tax and insurance deadlines. Event peaks are shorter than yearly ones but often larger in amplitude, and they distort weekly reporting more severely.

Gravity Field

Seasonal recommendation strategies in Gravity Field

Gravity Field accounts for the freshness of behavioural signals and for stock availability, so the algorithm does not recommend last year's season. Strategies switch with the period, and their impact is proven with an A/B test against the baseline configuration.

✦Configurable history depth and weight of recent signals inside a strategy

✦Rotation and scheduling of strategies for seasonal and promotional periods

✦Filtering by stock availability, category and product attributes

✦Merchandising rules layered on top of algorithmic ranking

✦A/B testing of strategies with Bayesian statistics

Explore the platform →