Demand Forecasting in Practice

For CSCOs, S&OP heads, demand planners, operations directors past basic AI literacy. Eight chapters across the demand-forecasting surface: the forecasting landscape (Gartner 78% adoption / 41% measurable lift, McKinsey 15–30% MAPE reduction / top-decile 30–50%, M5 LightGBM domination, Hyndman canonical reference, Tetlock calibration > confidence); accuracy and the MAPE trap (APICS +5–12% structural bias, MAPE asymmetry, sMAPE/MASE/RMSSE alternatives, M5 used RMSSE, APICS $0.5–1.5M per percentage point); choosing the right model (3-variable decision tree — history × signal × granularity, LightGBM granular, classical aggregated, Croston intermittent, ensemble options); demand sensing (Blue Yonder + o9, POS/weather/social signals, 5–15% lift on 4-week horizon, confidence-weighted overrides, noise-chasing risk); promotional uplifts (Coca-Cola/Unilever/Carrefour 20–40% improvement, 3-component pattern, calendar drift failure mode); NPI (P&G McKinsey case 2024, analog-product lookup + first 4-week signal, judgement layer); S&OP integration (SCC 38% gap, FVA analysis, LokadTech LLM hallucination case, quote-or-cut narrative); and the close — 2-quarter rollout playbook with interactive builder, 4 trust trip-wires, through-line: planners with ML beat planners without it, planners plus ML still beats ML alone, calibration matters more than confidence.