Danone reduces forecast error 20% and lost sales 30% with ToolsGroup machine learning demand planning

Danone's fresh food portfolio presents some of the most demanding forecasting conditions in Food & Beverage supply chain: short shelf life, dynamic consumer demand, and a promotional mix so intensive that more than 30% of volume moves on promotion — discounts, leaflets, displays, and media events — accounting for nearly 70% of total forecast inaccuracy. Four departments (sales, demand forecasting, account planning, and finance) operated without a shared data layer, making coordination largely ad hoc. The result was unpredictable supply emergencies, product obsolescence from over-production, and recurring lost sales during promotional peaks, with service levels consistently falling short of the company's 98.7% target.

Danone launched the Disc'Over project in partnership with ToolsGroup, deploying their machine learning demand planning platform to bring statistical rigor to trade promotion forecasting. The system establishes a reliable statistical baseline, then applies machine learning to identify and quantify the lift effect of promotional and media stimuli at the channel and store level — using the shared characteristics of historical events to predict future impact with granular precision. Critically, the platform did not function as a standalone forecasting tool. It created a unified cross-functional planning process connecting marketing, sales, account management, supply chain, and finance around a single source of demand truth. This enabled campaign-level operational planning — allocating promotions to specific accounts, synchronizing media events with production schedules, and supporting key account plan definitions — replacing the fragmented, manual approach that had previously produced uneven and unpredictable outcomes.

The Disc'Over implementation delivered measurable improvement across the full planning cycle. Forecast error fell 20%, lifting accuracy to 92%, while lost sales dropped 30% and service levels rose to 98.6% — ultimately exceeding the 98.7% internal target for 37 consecutive months. Key outcomes:

• Forecast accuracy: improved to 92% (20% error reduction)
• Lost sales: reduced 30% (service level → 98.6%)
• Product obsolescence: reduced 30%
• Net ROI from promotions: +6 points in 2011, +8 points in 2012
• Net promotional uplift: +36% in 2011, improved to +55% by 2012
• Planner workload: reduced 50%, with capacity redeployed to higher-value activities

• Machine learning that models promotional lift at channel and store granularity directly addresses the single largest source of forecast error in promotional-heavy CPG businesses — aggregate models are insufficient when 70% of inaccuracy originates at that level.
• Cross-functional alignment is a prerequisite, not a byproduct: the shared data layer was the mechanism that made coordination between four previously siloed departments operationally possible.
• Automating baseline and promotional forecasting can free planners from routine tasks (50% workload reduction here), enabling reallocation to exception management and strategic planning.
• Fresh product supply chains require store-level forecast granularity; when shelf life is short and stockouts are irreversible, precision at the channel level is a supply chain imperative.