RetailPredictive Analytics
Decathlon fine-tunes Chronos-2 on AWS, cutting demand-forecast error by 11-15 points
Decathlon, one of the world's largest sporting goods retailers, replaced its previous demand-forecasting approaches (DeepAR and Holt-Winters from 2021 to 2024, then a Temporal Fusion Transformer) with the Chronos-2 time series foundation model, fine-tuned every 6 months with LoRA through AutoGluon and run on Amazon EC2 instances, with Airflow orchestration and MLflow model versioning. Benchmarked over 101 rolling cutoffs with approximately 25,000 products per cutoff, Chronos-2 fine-tuned cut 12-week WAPE by 11 percentage points in South East Asia (39% to 28%) and 15 points in Latin America (53% to 38%), and cut new-region deployment time from about 6 months to 2-3 months. It is in production for the South East Asia and Latin America supply zones, with full multi-zone rollout targeted for 2026.
DecathlonChronos-2
