How Decathlon runs demand forecasting at scale with Chronos-2
aws ·
Decathlon, a global sporting goods retailer, adopted Amazon's Chronos-2 time series foundation model for large-scale demand forecasting after benchmarking it against production baselines and other TSFMs using roughly 25,000 products per cutoff. The company's AWS architecture runs weekly batch inference on EC2 instances, with fine-tuning every six months via AutoGluon and LoRA, and model registry through MLflow. Chronos-2 fine-tuned outperformed previous DeepAR, Holt-Winters, and TFT approaches, cutting WAPE by 11–15 percentage points at the 12-week horizon for SEA and LATAM, and 6–9 points at the 52-week horizon. Each WAPE point gain at 12 weeks translated to 0.3 days of inventory savings, 0.3 points of product availability, and about 0.12 points of sales uplift. The deployment also reduced operational overhead, cutting new-region deployment time from about six months with three people to two to three months. The article includes code examples and notes that Chronos-2 supports native covariates and is available via AutoGluon and SageMaker JumpStart.
aws