Your demand history is several different things layered on top of each other. Underlying trend, repeating seasonal pattern, one-off events, promotional effects, and random variation. A forecast is only as good as its ability to separate these, because each requires a completely different response.
Why Separation Matters Practically
Treat a seasonal peak as random variability and you’ll carry safety stock all year to protect against something that happens every October. Treat a promotional spike as underlying demand and you’ll forecast a permanently higher level that never materialises. Treat genuine growth as noise and you’ll be short for months.
These aren’t hypothetical errors — they’re the standard failure modes of simple averaging methods, and they’re why forecast accuracy plateaus in most manufacturers.
What Decomposition Does
Statistical decomposition separates a series into trend, seasonal, and residual components. The trend is your underlying direction. The seasonal component is the repeating pattern. What’s left is either noise or an event.
Once separated, each is handled appropriately: forecast the trend, apply the seasonal pattern, and size safety stock from the residual variability only — which is usually much smaller than total variability, and therefore requires much less buffer stock.
That last point is the practical payoff. Manufacturers who move from total-variability safety stock to residual-variability safety stock frequently find they can hold considerably less for the same service level, purely because they stopped buffering against a pattern they could predict.
Promotions Need to Be Marked
Promotional and event effects are the hardest component because they’re irregular. A model can’t reliably identify them from the demand series alone — it needs to be told.
This means keeping a record of what happened and when: promotions, price changes, competitor actions, large one-off orders, stockout periods. Most manufacturers don’t maintain this, and it’s the highest-value forecasting data most of them are missing.
Stockouts deserve particular attention. Recorded sales during a stockout understate demand, and a model trained on that data learns the wrong level. Flagging historical stockout periods before training is unglamorous and materially improves results.
Cannibalisation and Pull-forward
Two effects that consistently distort promotional analysis. A promotion on one product often takes volume from a similar product rather than growing the category. And discounting pulls demand forward, so the following period looks weak.
Both mean the apparent promotional uplift overstates the real effect. Modelling across a product family rather than item by item catches cannibalisation; looking at a window either side of the promotion catches pull-forward. Measured properly, some promotions turn out to have moved volume around at a discount without adding any, which is worth knowing before the next one is planned.
Where to Start
Begin recording events now, even in a simple shared log. Two years from now that record will be the difference between a model that explains your demand and one that treats every past disturbance as random.
ticktick.ai decomposes demand into trend, seasonal, and residual components, and supports event flagging so promotions and stockouts don’t distort the baseline.
