Decathlon deployed Chronos-2 for demand forecasting across regions
Based on its own testing using its data, Decathlon selected Chronos-2 for weekly sales forecasting for more than 25 000 products across six regions, replacing traditional models that required weekly retraining.
Decathlon, one of the largest sporting goods retailers in the world, selected Chronos-2 as a key component of its demand forecasting based on its own extensive comparison using its retail data. The company tested multiple time series foundation models in both zero-shot and fine-tuned settings, and according to the published results, the fine-tuned version of Chronos-2 consistently outperformed all other tested models across both forecast horizons examined. Even without fine-tuning (zero-shot), the model reportedly achieved results comparable to or better than the existing production model, while fine-tuning further reduced forecast error by several percentage points.
The system predicts weekly sales volume for more than 25 000 products in each of six regions – Europe, India, China, Southeast Asia, Latin America and, as planned, the Middle East and Africa – across two horizons: twelve weeks for replenishing inventory from suppliers and fifty-two weeks for long-term capacity planning. According to the company, another key advantage of Chronos-2 was native support for covariates through a mechanism known as group attention, eliminating the usual workarounds required by other time series foundation models.
The deployment runs on AWS infrastructure: data preparation takes place in PySpark, model fine-tuning using LoRA through the AutoGluon library is repeated once every six months, and the resulting model is registered in MLflow, while weekly batch forecasts run on Amazon EC2 instances and are orchestrated through Airflow. According to the company, this approach eliminated the need for weekly model retraining required by the previous traditional methods and enabled the system to scale into new regions without additional developer effort. Details on the architecture and Chronos-2 itself can be found in the source article.
Why it matters
The case shows that time series foundation models can replace the traditional forecasting pipelines commonly used by large retailers that require regular retraining, while maintaining or improving accuracy according to an internal comparison by Decathlon. For companies with similarly extensive and seasonal product ranges, this suggests the possibility of reducing the operational costs of maintaining forecasting systems and expanding forecasting into new markets faster without a proportional increase in engineering capacity.
Relevant practical impact
What this means
For a business
According to its own testing, Decathlon replaced a traditional forecasting approach (requiring weekly model retraining) by deploying Chronos-2, reducing the operational burden on the team and enabling demand forecasting to expand into new regions without additional engineering effort – a relevant example for companies with high-volume retail operations and seasonal demand.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.