Is an additive effect a multiplier?

No; it is expressed in the series' units.

Does every fluctuation count as seasonal?

No; recurrence at relevant calendar positions matters.

AI Glossary / Time series and signal processing / Advanced

Seasonal effect

Definition

A seasonal effect is a systematic variation associated with a recurring position in the calendar or seasonal cycle. In the supplied statistical sense, it concerns patterns recurring at seasonal frequencies within the year. Examples include regular changes associated with seasons, school schedules or recurring periods of business activity.

What it means

A high value during a usually busy period can partly reflect a seasonal effect rather than a new increase in underlying activity. A seasonal effect describes a repeatable pattern; one unusually high observation does not establish one.

The pattern is estimated from observations and can change over time. Its statistical identification does not, by itself, identify the cause.

How it works

A decomposition separates the observed series into components. In an additive model, an observation equals a trend-cycle value plus a seasonal effect plus a remainder. In a multiplicative model, a seasonal factor scales the other components.

Compare corresponding positions across cycles while accounting for changes in the series' level. Check whether the estimated pattern is stable and whether unusual observations are distorting it. Calendar effects such as moving holidays may need specific treatment rather than assignment to a fixed month.

Removing an estimated seasonal effect produces a seasonally adjusted view. It does not remove all irregular variation or eliminate the seasonal demand that a business must still serve.

Three practical examples

1. Service planning: identify whether customer requests repeatedly rise during particular parts of the year.

2. Performance reporting: distinguish a recurring busy-quarter contribution from a change in the nonseasonal level.

3. AI forecasting review: check whether a system has learned a stable calendar pattern or is extrapolating an isolated unusual event.

Think of it this way

Think of a recurring annual event that brings extra visitors to a town. The extra visitors form a predictable part of the year's pattern, but a one-off transport disruption during the event is a separate influence. Both can affect the same observation.

Worked example

Consider an invented quarterly activity index. Assume the underlying level is constant within each year: 100 throughout year 1 and 120 throughout year 2. Suppose the additive seasonal effects for quarters 1 through 4 are [-20,0,30,-10], with no remainder in this initial teaching construction.

Year 1 therefore has observations [80,100,130,90]. Year 2 has [100,120,150,110]. For example, year-2 quarter 3 is 120 + 30 = 150.

The seasonal effects sum to zero. Consequently, the average of the four observations is 100 in year 1 and 120 in year 2. Subtract each year's mean from its quarterly observations. Both years produce [-20,0,30,-10], recovering the stipulated pattern exactly in this deliberately simple example.

This calculation is not a general claim that subtracting annual means estimates seasonality correctly in every dataset. The constructed example assumes no within-year trend, no irregular component and an unchanged additive pattern.

Compare the third quarters across the two years. The observed index rises from 130 to 150, an increase of 20. The seasonal contribution remains 30 in both years, so the entire change is due to the stipulated level increase from 100 to 120. Calling the additional 20 a stronger seasonal effect would contradict this construction.

The constant additive effect has different relative sizes. Thirty points is 30% of the year-1 baseline of 100, but 25% of the year-2 baseline of 120. A stable additive effect therefore does not imply a stable percentage effect.

For a separate multiplicative model, a quarter-3 seasonal factor of 1.3 would produce 130 from a baseline of 100 and 156 from a baseline of 120. Its seasonal lift would be 30 and 36 points respectively. That is a different model from the additive construction, which produced 150 in year 2.

Now return to the additive example and introduce a one-off positive remainder of 40 in year-2 quarter 3. Its observed value becomes 190, while the other year-2 values remain [100,120,110]. The complete year-2 sequence is [100,120,190,110], with average 130.

If an analyst mechanically subtracts that contaminated annual mean, the estimated quarterly deviations become [-30,-10,60,-20]. Averaging these with year 1's deviations gives [-25,-5,45,-15]. This naive two-year calculation now overstates the true stipulated quarter-3 seasonal effect by 15 and shifts the other quarterly estimates downward by five each.

The distortion is caused by treating a one-off event as evidence about the recurring pattern. In real data, the analyst does not know the true components in advance, so unusual events need investigation rather than automatic deletion or automatic classification as seasonal.

Finally, suppose the next year's baseline is projected to be 140 and the original additive seasonal effects are assumed to persist. With a zero expected remainder, the quarterly forecasts are [120,140,170,130]. A seasonally adjusted baseline forecast of 140 for every quarter would omit the variation relevant to actual capacity planning. These are conditional projections, not evidence that next year's seasonal pattern is known.

Illustrative case study

Fictional Willow Support sees the third-quarter value of 190 and plans next year's capacity as if all of the increase will recur. Its analyst separates the assumed recurring contribution from the one-off remainder in the example.

Willow reviews the event history and considers alternative seasonal scenarios before setting capacity. The case is invented and illustrates why identifying an unusual observation's role matters; it does not prescribe a particular outlier treatment.

Common misconceptions and mistakes

"Every annual peak is entirely seasonal." Irregular events and level changes can contribute.

"A seasonal effect must have the same percentage size each year." That depends on the model.

"A repeating association proves its cause." Additional evidence is needed.

"Seasonal adjustment removes the need to plan for seasonality." Actual workload still includes it.

Business use

Report the estimated effect, its units and the assumptions about stability. Keep the original observations available and document known unusual events. When evaluating forecasting systems, examine performance across seasons instead of relying only on an overall average error that may hide recurring failures.

Check your understanding

What is the original quarter-3 additive effect?

Answer: Thirty points.

Why does year-2 quarter 3 initially equal 150?

Answer: Baseline 120 plus seasonal effect 30.

What does the one-off event add?

Answer: A remainder of 40, not a stipulated change in seasonality.

Related terms

Seasonal adjustment
Irregular effect
Seasonality
calendar effect
trend-cycle
additive decomposition
multiplicative decomposition
seasonal factor

Learn next

Change the one-off remainder from 40 to 20. Repeat the annual-mean subtraction and two-year averaging. Explain how the smaller unusual event changes the apparent quarterly effects.

Sources

Australian Bureau of Statistics, Statistical terms and concepts glossary, Seasonal effect: https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/statistical-terms-and-concepts-glossary

Rob J Hyndman and George Athanasopoulos, Forecasting: Principles and Practice, Time series decomposition: https://otexts.com/fpp3/decomposition.html

U.S. Census Bureau, Seasonal Adjustment Questions and Answers: https://www.census.gov/data/software/x13as/seasonal-adjustment-questions-answers.html

Previous: Seasonal difference Next: Seasonal exponential smoothing