How Decathlon runs demand forecasting at scale with Chronos-2

Decathlon, the world's largest sporting goods retailer with over 100,000 employees and 400 million users, requires robust demand forecasting to ensure adequate product availability across its extensive network. To address this challenge, the company evaluated multiple time series foundation models (TSFMs) and selected Chronos-2 as a core component of its forecasting stack.

In this article, we'll explore the architecture Decathlon uses to run Chronos-2 at scale on AWS, the impact on supply chain operations, and practical lessons for other companies considering foundation models for forecasting.

Decathlon faces unique demands forecasting challenges, including tens of thousands of products spanning 80 sports sold across various regions with distinct seasonal patterns. Their forecasting system generates weekly sales quantity predictions for all products on two critical horizons: a 12-week replenishment window and a 52-week strategic horizon. The system runs weekly across all horizons, covering multiple supply zones, including Europe, India, China, South East Asia, Latin America, and soon the Middle East and Africa.

Initially, Decathlon employed a hybrid approach using Amazon SageMaker AI DeepAR for short-term forecasts (weeks 1-16) and Holt-Winters exponential smoothing for the long-term horizon (weeks 17-52). However, this solution required weekly retraining and couldn't easily scale to new regions without significant engineering effort. To improve accuracy while reducing operational complexity, they turned to Chronos-2.

After rigorous benchmarking on Decathlon's own retail data, Chronos-2 demonstrated superior performance compared to other TSFMs, both in zero-shot and fine-tuned configurations. Fine-tuning the model annually reduced forecast error by several percentage points, while its ability to handle covariates natively set it apart from other models. Chronos-2 also proved computationally efficient, meeting the benchmark's requirements for inference time and cost-effectiveness.

Decathlon's production deployment of Chronos-2 on AWS comprises several key components. A PySpark data preparation pipeline assembles input time series, while a fine-tuning job built on AutoGluon adapts Chronos-2 every six months, registering the resulting model in an MLflow model registry. The inference pipeline fetches the latest model and generates weekly batch forecasts, which are then delivered to downstream consumers via a PySpark exposition pipeline.

By adopting Chronos-2, Decathlon has achieved higher forecasting accuracy with lower operational complexity, allowing for more efficient supply chain management and better customer experiences. For other companies interested in adopting foundation models for forecasting, Decathlon's solution architecture provides valuable insights into leveraging AWS infrastructure, efficient fine-tuning schedules, and the benefits of native covariate support.