Seasonality Is Several Things at Once
Annual, monthly, weekly and daily cycles overlay each other in every demand series. How to separate them, which are stable, and which one most forecasts get wrong.
Seasonality in a workforce demand series is at least four cycles running at once: the annual shape, the position within the month, the day of the week, and the hour of the day. They overlay, they interact, and a method that models one as though it were all of them will be wrong in a patterned way.
The series discussed in “Seasonality Is Several Things at Once” is easier to interpret when changes in work capture are visible rather than silently folded into the history. Used for gdpr employee monitoring, view this implementation guide can provide consistent time and project evidence, provided corrections, category changes and missing periods remain traceable.
The weekly cycle is almost always the largest and the most stable. The annual one gets all the attention.
As a separate benchmark for “Seasonality Is Several Things at Once”, see SHRM workforce-development resources. Keep source data, model, override and resulting decision visible as distinct steps.
The four cycles, in order of size
In most operations: day of week first, often ±30% around the mean. Hour of day next, larger still within a day. Position in the month third, where billing or payroll drives behaviour. Annual last, often only ±10%.
- Day of week — often ±30% around the mean, and the most stable.
- Hour of day — larger still within a day.
- Position in the month — where billing or pay dates drive behaviour.
- Annual — often only ±10%, and the one everybody discusses.
The ordering surprises people who have spent years discussing the summer dip.
Why the weekly cycle is the stable one
Because it is driven by how customers and staff organise a week, and that does not change much. Monday is high because the weekend queued up; Friday is low for reasons that have held for decades.
It is also the cycle with the most observations: a two-year series contains a hundred Mondays and two Januaries.
Estimating the weekly shape
From recent weeks rather than from the whole history, because it drifts slowly. Six to twelve weeks is usually right.
Where a long average is used, a shape that has shifted — a Monday peak flattening as channels change — gets averaged with the shape it used to have.
The monthly position cycle
Driven by billing dates, pay dates, benefit dates, statement runs. It is strong where it exists and absent where it does not.
It is also the cycle most often missed, because a monthly total hides it entirely and a weekly view scrambles it — the third working day falls in different weeks in different months.
Why the month-position cycle scrambles
Because weeks and months do not align. A forecast built weekly, from weekly history, cannot see a cycle that is anchored to day-of-month.
The fix is to model it as an event — specific dates with a known effect — rather than as a season.
The annual cycle
Real, usually modest, and needing two or three full years to estimate at all. Its main uses are the far horizon and holiday planning.
Where only eighteen months exist, the honest answer is that the annual shape is not yet estimable, and borrowing one from a comparable series is better than inventing one.
Moving feasts
Easter, half-terms, Ramadan, anything tied to a lunar or school calendar. These move between months and years and will wreck a monthly seasonal estimate if left unlabelled.
They belong in the event calendar with their actual dates, not in the seasonal factor.
Interactions
A public holiday on a Monday is not the same as one on a Friday. The billing peak falling on a Monday is not the same as one falling on a Thursday.
Where these interactions are large, the usable approach is a daily model with explicit effects rather than multiplied seasonal factors.
Additive or multiplicative
Whether the seasonal swing grows with the level. In most demand series it does — a 30% Monday is 30% whatever the level — which argues for multiplicative.
Where the swing is a fixed count regardless of level, additive is right. Plotting the series against its own level settles it in minutes.
The season that is really a driver
Sometimes what looks seasonal is a driver with a seasonal schedule: campaigns that run in spring, a regulatory deadline in January.
Modelling it as a season means it moves when the schedule moves and the forecast does not. Modelling it as an event means it does.
Re-estimating
The weekly shape quarterly, the annual shape yearly, the month-position effects when billing changes. None of these takes more than a morning.
A seasonal profile set once at implementation and never revisited is one of the commonest quiet sources of growing error.
Seasonality in the handling time
Usually forgotten. The number of contacts is seasonal and so, often, is how long each takes — longer in January when the queries are complicated, shorter in a promotional peak when they are simple.
Forecasting volume seasonally and treating handling time as constant produces a workload forecast that is wrong in both directions at different times of year.
The view worth circulating
The deseasonalised level alongside the forecast, at least internally. It is the only view in which a real change is visible, because in the raw series every change is confounded with whatever cycle it fell in.
Operations managers who are shown it once usually ask for it permanently.