Time series decomposition for retail sales
A weekly sales series is several things added together: an underlying trend, a repeating seasonal pattern, calendar effects, one-off events, and random variation. Reading the raw series and drawing a conclusion means attributing all of that to whatever explanation is nearest to hand.
Decomposition separates the components so that a movement can be assigned to one of them.
What are the components?
Trend, the slow direction of travel. Seasonality, the pattern that repeats at a fixed period. Calendar effects, which are the irregular consequences of dates — a holiday moving between weeks, a month with five weekends. And the remainder, which is everything the model does not explain.
Naming the remainder honestly is important. It is not noise in the sense of being meaningless; it is the part not accounted for.
Why does it matter for a weekly series?
Because weekly retail data is dominated by seasonality and calendar effects, and a week-on-week comparison is almost always contaminated by them. Comparing Thanksgiving week to the week before is not a comparison of demand.
The rule of thumb that survives contact with data is to compare like periods: the same week last year, adjusted for any date shift, rather than adjacent weeks.
What is a trend, actually?
The component that persists. A trend estimated over too short a window will follow the season; estimated over too long a window it will miss a genuine turning point. The window choice is an analyst decision and it should be stated.
That decision is frequently where two analyses of the same data disagree. Neither is wrong and neither is neutral.
How do calendar effects work?
A holiday that falls in week fifty-one one year and week fifty-two the next moves a large amount of volume between weeks without any change in demand. Month lengths, the number of weekends and school schedules all do the same on smaller scales.
In small-format retail, day-of-week effects are strong, so a week containing an extra weekend day behaves differently from one that does not. This is arithmetic rather than behaviour and it belongs in the adjustment.
When should you not decompose?
When the series is too short to estimate a season, when the pattern is not stable, or when a structural break makes the history non-comparable. Forcing a decomposition on a series with a break produces components that describe the break rather than the market.
A distribution change, a range reset or a channel shift are all structural breaks. Model the break or split the series; do not smooth over it.
What does this change about reporting?
It changes what a movement means. "Sales rose four percent" is a statement about the raw series; "the trend component turned up after three flat quarters" is a statement about the market. Only the second one supports a decision.
Report the raw figure and the interpretation separately, so a reader can see both what happened and what it is claimed to mean.
What about the remainder?
It is worth looking at rather than discarding. Unusual remainders cluster around real events — a competitor action, a supply interruption, weather — and a large remainder is a prompt to go and find out what happened.
An analyst who examines their residuals finds things that no summary statistic reveals.
Frequently asked questions
Do I need software for this?
Standard statistical tools include established decomposition methods, and a spreadsheet can do a serviceable job on a simple series. The method matters less than understanding what it assumes.
Additive or multiplicative?
Additive where the seasonal swing is roughly constant in absolute terms, multiplicative where it scales with the level. Retail series are frequently multiplicative, and choosing wrongly distorts the components.
How much history is needed?
Enough to see the season several times, which for weekly data with an annual season means several years. Less than that and the seasonal estimate is unstable.
What if my history has a gap?
Gaps have to be handled explicitly, by imputation or by modelling around them, and the choice should be documented. A silently filled gap becomes an invisible assumption.
Does this work at store level?
Less well, because store-level series are noisier and more prone to breaks. Decomposition is usually more informative at market or chain level, with store-level work done differently.
How does this relate to seasonal adjustment?
Seasonal adjustment is decomposition applied to remove the seasonal component so the remainder can be compared across periods. It is the same machinery used for a specific purpose.
Can decomposition prove causation?
No. It separates patterns; it does not explain them. A turning point in the trend is a fact to investigate rather than a conclusion.
What is the most common misuse?
Reading a smoothed trend as though it were data. Smoothing is a lens, and a lens that has removed the variation should not then be used to claim the market is stable.
### What should be plotted alongside the components?
The raw series, always. A decomposition without the original beside it invites a reader to treat the trend as the data, and the trend is a model output.
Showing both also makes the fit visible. A trend that visibly fails to follow the raw series is telling you the model is wrong, and that is information the components alone will not surface.
Can I decompose a series with a promotion in it?
Yes, and promotions will appear in the remainder rather than in the seasonal component unless they recur on a fixed calendar. Recurring promotions are a real source of apparent seasonality and should be treated deliberately.
Working with retail data? See NRS Insights data solutions or contact NRS Insights.