Seasonality

A seasonal profile is a set of indices, one per period in the cycle, that scale a baseline up or down: multiplicative when the amplitude grows with volume, additive when it stays constant. Sales data at store level carry more than one cycle, and each is estimated separately – an annual shape driven by holidays and weather, a day‑of‑week shape, and an intra‑month shape tied to pay dates. A weekly annual profile means fitting 52 indices, which takes several years of clean history before the numbers describe demand rather than noise.

• Multiplicative: forecast = level × trend × seasonal index – amplitude grows with volume
• Additive: forecast = level + trend + seasonal component – amplitude stays constant
• Cycles to separate: annual, day‑of‑week, intra‑month
• Fixed vs moving: Christmas falls on the same date every year, Easter moves by up to five weeks

Seasonality in practice

A fixed annual period is the first thing that breaks. Easter shifts between March and April, so a model built on monthly buckets and a twelve‑period cycle assigns the peak to whichever month held it last year and then errs twice: it underforecasts the month that has the holiday now and overforecasts the one that had it before. The same arithmetic applies to public holidays that move across weekdays and create bridge days, and to the varying number of trading Saturdays in a month. The fix is not a longer history but a different structure – holidays modelled as events anchored to a date, with the seasonal profile left to carry the smooth part of the year.

The second problem is a season that was never a season. A peak that repeats twice can come from a promotion that ran in the same week two years running, from a competitor closing a nearby store, or from last year's stockout flattening a genuine peak into a plateau. Once it enters the profile it confirms itself: the forecast raises the order, the surplus is cleared at a discount, and the discount produces another peak in the same week. That is why history is cleaned of stockouts, promotions and extremes before a profile is fitted, and why an extreme season is worth flagging rather than absorbing.

Level of aggregation decides whether the profile measures anything at all. An item selling a few units a week produces more zero days than non‑zero ones, so 52 weekly indices fitted to two years of its own history describe noise. The workable route is to estimate the profile on a group – category by store cluster – and apply it to the item, which is what a middle‑out approach does; Veritico STOCK offers bottom‑up, top‑down and middle‑out, handles seasonality including moving holidays, and identifies extreme seasons while suppressing false ones, with history cleaned of stockouts, promotions and extremes before the forecast is computed. See demand forecasting and inventory optimization and replenishment and allocation management. Mondelez raised forecast accuracy from 50 % to 70 % after planning was unified across functions – confectionery carries a fixed season at Christmas and a moving one at Easter in the same portfolio.

What the profile cannot explain has to be covered by stock. In weather‑driven categories the seasonal index is a proxy for temperature, so the residual stays wide even with a well‑fitted profile: the peak is absorbed by safety stock and by how quickly the supplier can react, not by a better index.

Related terms

• Demand forecasting
• Demand planning
• Safety stock
• Promotion planning
• Intermittent Demand
• Baseline Demand