Demand Sensing vs. Demand Forecasting: Why CPG Giants Are Scrapping Their Old Models

Statistical demand forecasting — ARIMA, exponential smoothing, seasonal decomposition — was designed for stable, stationary demand signals. It works when the primary drivers of demand are known, seasonal, and recurring: holiday spikes, promotional events, annual contract renewals. The pandemic proved it does not work when those assumptions break. Demand for consumer staples spiked 300–500% in weeks; demand for discretionary categories collapsed overnight; supply chains that had been optimised for efficiency had no resilience. The bullwhip effect — amplified demand signals propagating upstream through the supply chain — reached unprecedented intensity because every tier was over-ordering simultaneously based on models that had never seen anything like the signal they were receiving. The industry spent 2021–2024 rebuilding safety stock and diversifying supply; the question now is whether it builds forecasting capability that can handle the next disruption.

Demand sensing is not a different statistical model — it is a different epistemology about how demand works. Forecasting assumes demand follows patterns that can be extrapolated. Sensing assumes demand is driven by real-time causal factors — weather, economic news, social media sentiment, competitor actions, channel inventory levels — that can be observed and acted on immediately. A demand sensing system for a CPG company ingests daily or hourly point-of-sale data from retail partners (not the weekly sell-in data that feeds traditional forecasting), social listening feeds that detect emerging consumer trends before they show up in POS, weather data correlated with demand patterns by SKU and geography, and pricing and promotion data from competitors. The system builds a causal model of demand drivers rather than extrapolating a time series.

The hardest part of building a demand sensing system is not the ML — it is the data. CPG companies have hundreds of retail partners, each with different data formats, delivery schedules, and terms of data sharing. Major retailers (Walmart, Target, Kroger) share daily POS data with key vendors through retailer-managed data portals; smaller retailers share weekly aggregates or nothing at all. Building a unified demand signal requires standardising and integrating these heterogeneous feeds, applying retailer-specific inventory adjustment factors (retailer inventory changes affect POS-to-consumption relationships), and managing data quality issues — missing data, POS system outages, promotional events that break normal demand patterns — in real time. The data engineering investment for a large CPG company building enterprise demand sensing typically runs $5–15M before the ML layer is even started.

Demand sensing accuracy improvements sound modest — moving from 35% mean absolute percentage error (MAPE) to 12% MAPE — but the supply chain economics of that improvement are enormous. Forecast error at 35% MAPE requires holding 6–8 weeks of safety stock to maintain service levels. At 12% MAPE, 2–3 weeks of safety stock achieves the same service level. For a CPG company with $2B in inventory, that is a $500M–$1B reduction in working capital. The other side of the coin is service level: at 35% MAPE, stockouts occur frequently enough to impair retail relationships and lose market share to competitors with better shelf availability. Improving to 12% MAPE reduces stockouts by 60–70% — a measurable revenue impact in categories where shelf absence directly drives lost sales.

Promotional demand forecasting is the most difficult forecasting problem in CPG because promotions create demand spikes that dwarf the baseline signal. A well-executed retailer feature ad for a beverage brand can create 5–8× baseline volume for the promoted SKU at that retailer. Traditional promotional forecasting relies on historical lift factors — this promotion type delivered 4× baseline three times in the past, so we forecast 4× this time. AI promotional models improve on this by incorporating the specific conditions of each promotion: which stores, which ad vehicles, whether competitors have concurrent promotions, whether the weather forecast favours the category, and how recently the brand ran its last promotion in this market. They also detect when a promotion is cannibalising other SKUs rather than growing the category — a critical distinction for promotional ROI calculation.

Demand sensing only delivers value if it connects to supply chain execution — if better demand visibility drives better production planning, procurement, and logistics decisions. This is the integration challenge that most pilots fail: sensing improves the forecast but the forecast does not actually change purchase orders or production schedules because those systems run on weekly batch cycles and different teams own them. The supply chain technology architecture for a sensing-driven CPG operation requires near-real-time integration between the demand signal, the replenishment engine (which calculates what to order from whom), and the production planning system (which schedules manufacturing). Building this integration is a multi-year ERP and middleware programme — sensing is the data layer, but execution requires a matching operating model transformation.

Found this useful?

Share the key finding on X.