Forecasting

Forecasting predicts a future quantity or condition from available information. In supply chains, its value is not the number alone: teams use it to decide how much to produce, where to deploy inventory, whether to change plants, and how to plan cash flow. Start by asking which decision the forecast must support and who will use it. That objective determines the appropriate product, geography, time granularity, and horizon. section 1.2.1

Figure 7.1 Supply chain roles and forecasting granularities and horizons

Forecast needs change by role. Short-term logistics decisions typically need store-product-day detail, while longer-term production, marketing, and finance planning can use progressively more aggregated views. section 7.1

A usable forecasting process matches the forecast to its decision context:

• Set the objective before selecting a method: production, inventory deployment, product renewal, or revenue and cash-flow planning require different outputs. section 1.2.1
• Choose horizon and aggregation together. Near-term users commonly need granular forecasts; long-term users commonly need country, product-family, or monthly aggregation. section 7.1
• Expect independently produced forecasts at different hierarchy levels to disagree. A family forecast made directly need not equal the sum of SKU forecasts. section 7.3
• Judge quality against relevant benchmarks and context rather than visual plausibility, a prior-year result, a fixed target, or an industry benchmark alone. section 10.0

Statistical forecasting includes time series forecasting and predictive models, although the boundary can overlap. Time series methods project patterns visible in historical demand. A level-trend-seasonality decomposition makes the resulting forecast interpretable: planners can inspect the baseline level, directional trend, and recurring seasonal pattern instead of treating the result as a black box. Exponential smoothing and ARIMA are common approaches; both can extrapolate trend and seasonality and can be automatically tuned. section 14.0 section 14.1.1

More complexity is not automatically better. A model that assumes both trend and seasonality can overfit a simple pattern, particularly when history is short. In that situation, a simple moving average can outperform a more elaborate model. Conversely, predictive models can incorporate demand drivers such as pricing, promotions, or weather, enabling scenario planning, but they require clean additional data and more advanced modeling. section 14.1.1 section 14.5

Select methods empirically. First assess whether the needed data exists and is usable. Then test alternative techniques, inputs, and models against benchmarks, the current process, and the current forecasting engine. Retain a clear accuracy winner when one emerges; when several methods perform similarly, averaging their predictions may improve accuracy but creates a more complicated and less transparent result. section 14.3

Forecast error cannot be labeled good or bad from a percentage alone. A historical MAE of 36.1% and bias of 1.6%, for example, cannot be judged without understanding the demand’s complexity and randomness. When a time series forecast looks strange, inspect level, trend, seasonality, and relevant demand drivers to investigate the source of the result. section 10.0 section 14.5