How Decathlon runs demand forecasting at scale with Chronos-2

This article describes how Decathlon, a global sporting goods retailer, deployed Chronos-2 time series foundation model for demand forecasting across multiple supply zones on AWS.

• Chronos-2 fine-tuned outperformed previous approaches (DeepAR, Holt-Winters, TFT) with 11-15 percentage point WAPE reduction at 12-week horizon
• Rigorous benchmark on 25,000 products across 101 rolling cutoffs validated model selection before production deployment
• Architecture uses EC2 instances for weekly batch inference, AutoGluon for LoRA fine-tuning every 6 months, and MLflow for model versioning
• Inference runtime reduced from 10-15 minutes to 40-75 seconds for 7,000-15,000 time series
• Deployment time per region decreased from 6 months to 2-3 months, enabling faster market expansion
• Each WAPE improvement point translates to 0.3 days inventory savings, 0.3 points availability gain, and 0.12 points sales increase
• Chronos-2 natively supports covariates through group attention mechanism, eliminating workarounds needed by other TSFMs

Decathlon's production deployment demonstrates that time series foundation models are ready for enterprise retail forecasting, with significant accuracy and operational efficiency gains enabling scalable global expansion.