Ask what decision changes first
Forecasting projects fail most often because nobody defined the decision. A forecast that does not change an order quantity, a staffing rota or a production run is an expensive opinion.
Start from the decision and work backwards. That determines the granularity, the horizon and the accuracy you need — and often reveals that a much simpler forecast would do.
What the data has to look like
- Two to three years of history at the granularity you will forecast
- Consistency — a system change halfway through creates an artificial break
- Known promotions, price changes and outages, so the model is not learning them as random noise
- Stock-outs flagged, because zero sales due to no stock is not zero demand
- Product hierarchy, so new or sparse items can borrow from a parent category
The stock-out point is the one most often missed. A model trained on sales rather than demand systematically under-forecasts your best-selling lines, which is exactly the wrong error.
Start simple, and stay there if it works
Establish a baseline with a naive method — last year's same week, or a moving average with seasonality. Then measure. In a surprising number of businesses, a well-tuned simple approach is within a few percentage points of anything more sophisticated.
That matters beyond cost. A method the planning team understands gets used and adjusted sensibly. A black box gets overridden the first time it says something surprising, and then it is decoration.
Accuracy expectations, honestly
| Level | Realistic error | Comment |
|---|---|---|
| Total business, monthly | 5–15% | Easiest — noise cancels out |
| Category, monthly | 10–25% | Usually good enough for planning |
| Individual product, weekly | 25–50% | Hard, and often what people ask for |
| New product, no history | Very poor | Use analogues and judgement, not a model |
Anyone promising accurate weekly item-level forecasts without qualification is selling something. Aggregate up until the error is tolerable and forecast at that level.
Blend judgement in deliberately
Your sales team knows about the contract starting in March. Your buyer knows a supplier is discontinuing a line. No model knows either. The best forecasting processes combine a statistical baseline with structured human adjustment.
Record every override with a reason and review afterwards whether it helped. Some people's adjustments improve accuracy consistently and some do not, and knowing which is worth more than another model.
Measure error the way it costs you
Track forecast error continuously, but weight it by consequence. Under-forecasting a fast-moving line costs a lost sale; over-forecasting a perishable one costs waste. A single accuracy percentage hides both.
Report error by category with the business consequence attached, and the conversation shifts from “the model is wrong” to “where is being wrong expensive”.