Inside Decathlon’s AI-Powered Demand Forecasting: Scaling with Chronos-2

Scaling Demand Forecasting Across a Global Retailer

Decathlon, the multinational sporting goods retailer with thousands of stores and an enormous online presence, faces a classic retail challenge: accurately predicting demand for hundreds of thousands of stock-keeping units (SKUs) across diverse regions and seasons. To tackle this, the company has turned to foundation-model-style time-series forecasting, deploying Chronos-2 at scale. The approach was detailed in a joint technical case study published on the AWS Machine Learning blog, authored by a team of data scientists and engineers from both Decathlon and Amazon Web Services.

Demand forecasting at scale means generating reliable predictions for every product, location, and time horizon to drive inventory planning, replenishment, and supply chain decisions. A single percentage‑point improvement in accuracy can translate into millions of euros saved through reduced stockouts, lower markdowns, and slashed inventory holding costs. For a retailer as large as Decathlon, the move from traditional statistical methods to a pre‑trained deep‑learning model represents a significant operational upgrade.

Chronos-2: A Foundation Model for Time Series

Chronos-2 is a transformer‑based forecasting model developed by AWS. It belongs to a growing family of foundation models for time series that are pre‑trained on vast and diverse datasets, then fine‑tuned on a company’s proprietary sales data. Unlike conventional models that require manual feature engineering for each time series, Chronos-2 automatically learns patterns across many series, handling trends, seasonality, and sudden demand shifts with minimal per‑product tuning.

In Decathlon’s deployment, the model ingests historical sales, promotional calendars, and relevant external variables. The forecasting pipeline is orchestrated on Amazon SageMaker, which provides the infrastructure for data preparation, model training, and inference at scale. The tight integration with SageMaker enables Decathlon to run thousands of forecasts simultaneously, refreshing daily predictions for inventory planners across the globe.

Operational Impact and Production Deployment

The case study highlights several concrete benefits that Decathlon claims from the new system:

• Improved forecast accuracy – particularly for slow‑moving and seasonal items, where traditional methods often struggle.
• Reduced manual effort – planners no longer need to manually adjust thousands of forecasts; the model handles most series automatically.
• Scalable infrastructure – the SageMaker environment allows the retailer to forecast for its entire product catalogue without hitting computational bottlenecks.
• Faster reaction to market changes – the model can be retrained on fresh data quickly, absorbing new trends almost in real time.

While the AWS‑authored case study is a vendor‑produced document, it offers a rare window into the production‑scale application of advanced time‑series AI within a major retail operation. The deployment is not a proof‑of‑concept; it is described as an active, operational system that feeds directly into Decathlon’s supply‑chain planning.

The move mirrors a broader industry shift where retailers are adopting foundation models for business forecasting, taking advantage of transfer learning and the ability to model complex, interconnected time series. Whether Chronos-2 or competing models, the trend is clear: large‑scale retailers are betting on AI that goes beyond siloed, per‑SKU predictions.

For readers interested in the full technical details – including model architecture, training procedures, and the exact AWS services used – the complete case study is available on the official AWS Machine Learning blog. The project also signals how cloud providers like AWS are packaging such models as ready‑to‑use services, lowering the barrier for other enterprises to follow Decathlon’s lead.