Begin with the decision horizon
A forecast for tomorrow's shelf fill is not the same as a forecast for a supplier order arriving next month. Define the item, location, time interval and decision before choosing a method. If orders are placed weekly and deliveries take several days, the forecast must cover the period the order actually protects, not merely the next calendar week.
Choose a level of detail that the available data can support. Daily item-store data can be sparse and noisy. A broader category forecast may be more stable, but it cannot directly tell a colleague which size or flavour to replenish. Reconcile forecasts across these levels rather than assuming one aggregation answers every operational question.
Sales are not unrestricted demand
A zero-sales day could mean no customer wanted the item. It could also mean the item was unavailable, the store was closed or the product could not be found. Preserve availability indicators alongside sales. Treat missing records, genuine zeros and stockout-constrained sales as different conditions.
Price changes, promotions, local events and assortment changes can alter demand. They should be recorded with effective dates. The Office for National Statistics retail sales publications can provide broad UK trading context, while the Met Office can provide weather context. Neither source replaces local evidence, and a correlation with weather does not establish that weather caused a particular sales change.
Give simple methods a fair comparison
A seasonal naive forecast uses a comparable previous period, such as the same weekday last week. A moving average smooths recent observations. Exponential smoothing gives progressively less weight to older observations; Holt-Winters methods add trend and seasonality under specific assumptions. These are useful reference points, not obsolete obstacles to machine learning.
More complex approaches, including ARIMA or models using promotion and weather features, can be appropriate when the data and operating process support them. Intermittent items need particular care: Croston's method separates non-zero demand size from the interval between demands. No method should be selected solely because it sounds intelligent or performs well on one convenient item.
Check an example by hand
Worked example Suppose a baseline predicts 20, 22 and 18 units across three comparable days. Observed sales are 18, 25 and 17. Define error here as forecast minus actual. The errors are therefore +2, −3 and +1 units. The mean absolute error, or MAE, is (2 + 3 + 1) ÷ 3 = 2 units.
| Day | Forecast | Actual | Error | Absolute error |
|---|---|---|---|---|
| 1 | 20 | 18 | +2 | 2 |
| 2 | 22 | 25 | −3 | 3 |
| 3 | 18 | 17 | +1 | 1 |
The mean signed error in this small example is zero, but the forecast was never exact. Bias and absolute error describe different properties. Weighted absolute percentage error, or WAPE, would be 6 ÷ 60 = 10% here. WAPE is undefined when total actual demand is zero and can conceal weak performance on low-volume items. Mean absolute percentage error, or MAPE, also becomes problematic around zeros.
Test the forecast as it would be used
Use rolling-origin evaluation: train on information available at a historical decision point, forecast forward, then move that point through time. Do not randomly mix future records into training data. A final promotion result, revised weather observation or late stock correction must not appear as an input that the real decision-maker could not have known.
Compare errors at the replenishment horizon and inspect outcomes such as availability and waste. Prediction intervals help express a range, but they require calibration and do not guarantee that every future observation falls inside them. Keep manual overrides visible, including the reason, owner and expiry date. Otherwise yesterday's temporary judgement can become an unexplained permanent uplift.
Further reading: Forecasting: Principles and Practice by Rob J Hyndman and George Athanasopoulos; ONS retail industry; Met Office. These references are not endorsements of this website.