Demand forecasting predicts how much of a product or service will be needed, driving inventory, staffing and purchasing decisions.
Data You Need
- Historical sales or demand at the right level (product, store, day).
- Calendar information: holidays, paydays, school terms.
- Promotions, prices and marketing activity.
- External drivers where relevant: weather, events, economic indicators.
- Stock-outs, because lost sales make recorded demand look lower than true demand.
Methods
- Baselines: last period, same period last year, moving averages.
- Statistical models: exponential smoothing and ARIMA, good for single series.
- Machine learning: gradient boosting across many products, using lag, rolling and calendar features.
- Hierarchical forecasting: reconcile forecasts at product, category and total level.
Measuring Accuracy
Use MAE or weighted absolute percentage error (WAPE) at the level decisions are made, and compare with baselines. Measure bias too: consistently over- or under-forecasting is costly.
Uncertainty
Provide ranges (for example, 10th and 90th percentiles). Safety stock and staffing decisions depend on uncertainty as much as the central forecast.
Making Forecasts Useful
Align the forecast horizon and granularity with decisions, let planners add known information (a new promotion), and track which adjustments improve accuracy.
Watch for Shocks
Pandemics, supply disruptions and new competitors break historical patterns. Monitor accuracy closely and be ready to lean on judgement.