Demand Forecasting

Demand forecasting is the process of estimating future customer demand for a product or service over a specified time horizon. It is the foundational input into nearly every supply chain planning decision: how much inventory to hold, when to reorder, how much manufacturing capacity to schedule, how many distribution staff to hire for peak season, and what pricing strategy to deploy. A 10% improvement in forecast accuracy typically reduces inventory by 20–30% and improves service levels simultaneously.

Time Horizons in Demand Forecasting

Forecasts serve different planning purposes depending on their horizon:

Short-term (days to weeks): Used for daily fulfillment planning, staffing, and transportation scheduling. Generated automatically by WMS and OMS systems using recent sales data.

Medium-term (weeks to months): The primary horizon for inventory replenishment and production scheduling. Typically produced by demand planning modules in ERP or S&OP (Sales & Operations Planning) processes.

Long-term (months to years): Used for capacity planning — DC network design, manufacturing expansion, supplier qualification. Based on market analysis and strategic assumptions rather than statistical extrapolation.

Forecasting Methods

Time series methods: Statistical extrapolation of historical demand patterns. Common techniques:

• Moving average: Averages demand over a recent window of periods
• Exponential smoothing (Holt-Winters): Weighted average that gives more weight to recent data; handles trend and seasonality
• ARIMA: More sophisticated time-series model capturing autocorrelation in demand data

Causal methods: Models that incorporate external variables (price, promotions, weather, economic indicators) to explain demand variation beyond historical patterns alone. Regression-based approaches.

Machine learning: Modern demand planning platforms (o9 Solutions, Anaplan, Blue Yonder) use gradient boosted trees, neural networks, and ensemble methods that automatically capture complex patterns including weather effects, competitor actions, and cross-product cannibalization.

Forecast Error Measurement

MAPE (Mean Absolute Percentage Error): The most common metric. Average of |(Actual - Forecast)| / Actual across all periods. A MAPE of 15% means forecasts are off by 15% on average.

WMAPE (Weighted MAPE): Weights the error by volume — high-volume SKUs contribute more to the metric. Better for inventory planning than simple MAPE.

Bias: The systematic tendency to over- or under-forecast. Consistent positive bias (overforecasting) leads to excess inventory; negative bias (underforecasting) leads to stockouts. Both are costly; bias is often more damaging than random error because it compounds over time.

Industry benchmarks: Best-in-class MAPE of 10–15% for weekly SKU-level forecasting; typical companies achieve 25–40%. New product launches are significantly harder to forecast (60–80% error is common without market intelligence).

The Collaborative Forecast

The Collaborative Planning, Forecasting and Replenishment (CPFR) framework formalizes the sharing of demand intelligence between retailers and suppliers. When a retailer shares their promotional calendar, seasonal plans, and POS data with suppliers, forecast accuracy improves significantly for both parties — and the bullwhip effect is reduced.