Decathlon Enhances Demand Forecasting with Chronos-2 on AWS

Decathlon, one of the world's largest sporting goods retailers, relies heavily on accurate demand forecasting to ensure product availability and optimize supply chain operations. With over 100,000 employees and serving 400 million users worldwide, the company manages a diverse portfolio of tens of thousands of products across multiple continents, each with unique seasonal demand patterns.

Historically, Decathlon used a hybrid forecasting approach combining Amazon SageMaker DeepAR for short-term predictions and Holt-Winters exponential smoothing for longer horizons. While effective, this system required frequent retraining and was operationally intensive, limiting scalability and agility.

In 2024, Decathlon evaluated several time series foundation models (TSFMs) through a comprehensive benchmark tailored to its retail data. The evaluation revealed that Chronos-2, especially when fine-tuned, outperformed other models in accuracy across both 12-week and 52-week horizons, with improvements of up to 15 percentage points in WAPE.

Decathlon adopted Chronos-2, leveraging its native covariate support and efficient fine-tuning via Low-Rank Adaptation (LoRA) with AutoGluon. The deployment architecture on AWS involves batch inference on Amazon EC2 instances, with models fine-tuned biannually to incorporate the latest data, and forecasts delivered through a PySpark pipeline.

The impact has been substantial: forecast accuracy improvements translate into inventory savings, reduced stockouts, and increased sales. Operationally, the new system has cut deployment time from six months to two to three months per region, facilitating faster expansion into new markets. The inference runtime is approximately 40-75 seconds for thousands of time series, with minimal compute costs.

Decathlon plans to extend this approach across all regions, incorporate additional external covariates, and explore ensemble strategies with multiple models to further enhance forecast robustness. Their experience underscores that foundation models like Chronos-2, combined with domain-specific fine-tuning, are now viable for large-scale retail demand forecasting.