What this is
Decathlon is one of the world's largest sporting-goods retailers: 100,000 employees, 400 million users, and stores spread across Europe, India, China, Southeast Asia, and Latin America. Each store's replenishment orders are driven by "demand forecasting" — and the demand curves for ski gloves and surfboards look nothing alike, with extreme seasonality layered on top.
Over the past several years, Decathlon's demand-forecasting system has been retrained weekly by in-house engineers: it started on Amazon SageMaker's DeepAR in 2021 and was upgraded to Temporal Fusion Transformer (a time-series deep-learning model) in 2024. Retraining costs were steep, and rolling out to new regions was slow.
Now they've switched to AWS's Chronos-2 — a "Time-Series Foundation Model" (TSFM, think of it as a foundation model for crunching numbers), pretrained on large-scale data. It forecasts out of the box, with no per-SKU retraining required.
How the industry sees it
Supporters argue this confirms a paradigm shift: foundation models are extending from text and images into time series (chronologically ordered sales and demand data). For a retailer like Decathlon — tens of thousands of SKUs, operations spanning continents — the cost of retraining models was real money. Foundation models' "ready-to-use" approach is a structural cost-down.
But there are dissenting voices. A retail-analytics consultant we spoke with notes that foundation models often shine on average metrics, but on hero SKUs, new launches, and discontinued items — the long tail — they may not be more stable than carefully tuned traditional models. There's another layer of risk: vendor lock-in. Betting your forecasting core on a single vendor (AWS) means bargaining power and migration costs need to be recalculated.
Impact on regular people
For retail and CPG IT teams: "Foundation model vs. self-trained model" will become an unavoidable option when evaluating forecasting systems. AWS (Chronos-2), Google (TimesFM), and Salesforce (Moirai) are all in the game — expect more head-to-head benchmarks next year.
For individual careers: roles in supply-chain planning and demand forecasting will shift focus — from "tuning and training" to "injecting business rules, monitoring anomalies, and interpreting model outputs." Pure model-engineer scarcity may decline; people who understand the business will become more valuable.
For consumer markets: more accurate forecasting should, in theory, reduce stockouts on hot items and ease clearance pressure on slow movers — but retail prices won't shift noticeably in the short term.