Seasonal adjustment: what it does and what it hides
Seasonal adjustment removes the regular within-year pattern from a series so that consecutive periods can be compared. It is what allows a statement like "sales rose from November to December" to mean anything, since December always rises.
It also removes something real. An adjusted December figure is not what the stores experienced, and for operational purposes the raw number is frequently the relevant one.
What is actually removed?
The estimated repeating pattern, and in many published series the calendar and trading-day effects as well. What remains is trend plus irregular movement.
The estimate is model-based, which means it can be wrong, and it is revised as more data arrives. Adjusted history changes, which surprises people who assume published figures are fixed.
When is adjusted the right series?
When comparing consecutive periods, when looking for a turning point, and when comparing markets with different seasonal patterns. In all three the seasonality would otherwise dominate the comparison.
When is raw the right series? For planning stock, staffing and cash, and for describing what a store actually experienced. An operations plan built on adjusted data will understaff December.
Why do adjusted figures get revised?
Because the seasonal pattern is estimated from the history, and each new observation improves the estimate. A recent adjusted figure is the least reliable point in the series precisely because it has the least surrounding data.
That is worth knowing before building an argument on the latest adjusted month. It is the point most likely to move.
What does adjustment struggle with?
Changing seasonality, moving holidays, and structural breaks. A category whose seasonal shape is genuinely shifting will be mis-adjusted, and the error appears as a spurious trend.
Moving holidays are handled explicitly in good implementations and crudely in poor ones. Where a holiday's date shift matters to your category, ask how it is treated.
Can a series be over-adjusted?
Yes, and it is a real risk when a method is applied mechanically. Removing more than the true seasonal component removes genuine signal, and the result is a smooth series that no longer describes the market.
The diagnostic is the remainder. If adjusted series residuals still show a periodic pattern, the adjustment has not worked; if the adjusted series is implausibly smooth, it may have worked too hard.
How should both series be used together?
Report both. The raw series describes what happened; the adjusted series supports the comparison. Presenting only the adjusted number invites the question "is that what the stores saw," and the answer is no.
This is also a transparency point. A provider showing both is showing its work.
What is the reference standard?
National statistical agencies publish adjusted retail series with documented methodology, and the Census Bureau's monthly retail series is the usual United States reference. Reading how a statistical agency documents its adjustment is a good calibration for what a commercial provider should be able to explain.
Frequently asked questions
Is an adjusted figure a real number?
It is a derived number, honestly labelled. It answers a comparison question rather than describing an observed total, and it should never be presented as the latter.
Can I adjust my own series?
With enough history and standard tools, yes. The risks are over-adjustment and mis-specified holiday handling, both of which need checking rather than assuming.
Why do two providers' adjusted figures differ?
Different methods, different holiday treatment, different history windows. All of those are legitimate choices and they produce different numbers from the same raw data.
Should I adjust at item level?
Rarely useful. Item series are noisy and short, and adjustment works better at aggregate level where the seasonal signal is clearer.
Does adjustment handle a one-off event?
No. Events are irregular by definition and belong in the remainder or in an explicit intervention term. Treating a one-off as seasonal will create a phantom pattern in subsequent years.
How much history do I need before adjusting?
Several complete cycles, so the pattern can be estimated. With less, the adjustment is fitting noise.
What if my seasonality changed?
Say so and consider splitting the series or using a method that allows evolving seasonality. Forcing a stable seasonal pattern onto a changed market is a known way to manufacture a trend.
What is the one-line caution?
Adjusted series are for comparison and raw series are for operations, and a report that uses one where the other belongs will mislead somebody competent.
### What should a buyer check in a published adjusted series?
Whether the method is documented, whether revisions are published, and whether the raw series is available alongside. Those three make the adjusted figure usable and their absence makes it an assertion.
Check also how holidays are handled, because a moving holiday treated crudely produces an annual pattern in the adjusted series that looks like a market cycle and is an artefact of the method.
### Who should be adjusting, the provider or the buyer?
Whoever has the history and the method, which is usually the provider. The risk of adjusting in-house is that two versions of the same series then exist, and reconciling them costs more than the adjustment saved.
Where you do adjust internally, publish the method alongside the number so the two versions can at least be distinguished.
Is an adjusted figure ever wrong in a way that matters?
Yes, most commonly when seasonality has genuinely changed and the method assumes it has not. The symptom is an adjusted series that drifts in one direction over several periods while the raw series behaves normally.
Working with retail data? See NRS Insights data solutions or contact NRS Insights.