Demand sensing vs demand forecasting: what the M5 competition taught CPG planners

The M5 accuracy competition asked teams to forecast 42,840 hierarchical daily sales series from Walmart stores in three US states ^[1]. The organisers' findings: most of the top entries used LightGBM gradient-boosted trees; models that learned across many series beat per-series models; and explanatory variables such as prices, promotions and calendar events improved accuracy.

A demand sensing model corrects the next few days or weeks using the latest signals. A statistical forecast sets the baseline for months ahead. Hyndman and Athanasopoulos's open textbook remains the practical reference for the baseline methods and how to evaluate them ^[2]. Most teams need both, with sensing overriding the baseline only in the short horizon.

Lee, Padmanabhan and Whang described the bullwhip effect: small changes in consumer demand become large swings in orders as they move up the supply chain, through order batching, price promotions and shortage gaming ^[3]. Better short-term demand signals, shared with suppliers, reduce that amplification.

• Get daily point-of-sale or sell-out data for the main customers, with promotions and prices.
• Build one model across SKUs and locations, with calendar, price and promotion features.
• Measure forecast accuracy and bias at the level decisions are made: SKU by DC by week.
• Feed the sensed forecast into replenishment, not just a dashboard.

Found this useful?

Share the key finding on X.