Decathlon, one of the world's largest sporting goods retailers with over 100,000 teammates and 400 million users, selected Chronos-2 as a core component of its demand forecasting stack (AWS Machine Learning Blog). The company forecasts weekly sales quantities across tens of thousands of products spanning over 80 sports on two critical horizons: a 12-week replenishment window for purchase planners and a 52-week strategic horizon for long-term capacity planning (AWS Machine Learning Blog).
Decathlon's previous approach combined Amazon SageMaker DeepAR for short-term forecasts (weeks 1–16) and Holt-Winters exponential smoothing for longer-term horizons (weeks 17–52), requiring weekly retraining (AWS Machine Learning Blog). To evaluate Chronos-2, Decathlon conducted a large-scale benchmark across 101 rolling cutoffs spanning nearly 2 years, testing approximately 25,000 unique products per cutoff, and found that Chronos-2 fine-tuned consistently outperformed all other evaluated time series foundation models across both 12-week and 52-week horizons (AWS Machine Learning Blog). Even in zero-shot mode, Chronos-2 matched or surpassed the fully trained production baseline (AWS Machine Learning Blog).
Decathlon's production deployment uses Chronos-2 fine-tuned with Low-Rank Adaptation (LoRA) through AutoGluon, with fine-tuning scheduled every 6 months rather than weekly, significantly reducing operational complexity (AWS Machine Learning Blog). The system runs weekly batch inference on AWS EC2 instances, processing approximately 7,000 to 15,000 time series per region in under 2 minutes, and is deployed across SEA, LATAM, with Middle East and Africa planned (AWS Machine Learning Blog). For commerce practitioners, this demonstrates that foundation models can deliver production-grade accuracy with lower operational overhead and simpler scaling to new regions compared to traditional retrained models.