Demand Forecasting System Product Specification Template

Two years of weekly data is the minimum for capturing seasonality. Three years is better because it lets the model distinguish true seasonal patterns from one-time events. For new product categories where you lack history, use analogous product data or market research estimates as seeds.

Forecast at the level where the decision gets made. If procurement buys by product family, forecast at product family and disaggregate to SKU using historical mix ratios. Forecasting 50,000 SKUs individually is computationally expensive and often less accurate than forecasting at a higher level and disaggregating. Start at product family and add SKU-level forecasting where accuracy gains justify the effort.

Separate base demand from promotional uplift. Build a promotion calendar as an input, and model the uplift factor per promotion type (BOGO, percentage discount, display placement). Include post-promotion dip modeling since demand often drops below baseline in the weeks after a promotion ends. This is where many statistical models underperform and ML models add the most value.

FVA measures whether each step in your forecasting process (statistical model, planner override, sales input, consensus adjustment) actually improves or degrades forecast accuracy. Many organizations discover that manual overrides make forecasts worse on average. FVA analysis identifies which process steps add value and which should be eliminated or automated.

Products that sell infrequently (spare parts, specialty items, long-tail SKUs) violate the assumptions of standard time-series models. Use Croston's method or its variants (SBA, TSB) which separately forecast demand size and demand frequency. Group intermittent SKUs together for reporting since individual SKU accuracy will always be low. ---