Decathlon's Shift to Foundation Models for Global Logistics
Managing inventory for a global retailer like Decathlon is a massive undertaking, involving 400 million users and over 100,000 employees. Every week, the company must accurately predict demand for tens of thousands of individual products to maintain optimal stock levels. Historically, Decathlon relied on static time-series models, which required frequent retraining and imposed significant engineering overhead whenever the company expanded into new geographic regions. To solve this, Decathlon turned to Chronos-2, a foundation model designed for time-series forecasting. By validating the model across 12-week replenishment cycles and 52-week strategic planning windows, Decathlon discovered that Chronos-2 could outperform their existing production baselines using only zero-shot inference, drastically reducing the complexity of their forecasting operations.
Architectural Advantages of Group Attention
At the core of Chronos-2 is a sophisticated approach to handling covariates, which are external variables that influence demand. Traditional time-series models often require complex, manual workarounds to incorporate these variables, but Chronos-2 uses a group attention mechanism to process them directly within the model. Built on an encoder-only transformer architecture derived from T5, the model segments time-series data into patches and converts them into real-valued embeddings via residual networks. By alternating between temporal attention and group attention, the model simultaneously learns the internal flow of individual time series and the correlations between different products within a group. This structural innovation allows Decathlon to integrate diverse product categories and external market factors into a single, unified forecasting pipeline without needing to redesign the model for each specific use case.
Scaling Operations on AWS
Decathlon’s production pipeline is built on AWS, utilizing PySpark for data preparation and Apache Airflow for orchestration. To maintain accuracy as market conditions evolve, the team performs fine-tuning every six months using AutoGluon. By applying Low-Rank Adaptation (LoRA), Decathlon optimizes the model for domain-specific performance without the prohibitive cost of full-model retraining. Once trained, these versions are registered in MLflow to support weekly batch inference. The following implementation demonstrates how to configure LoRA fine-tuning within the AutoGluon framework:
from autogluon.timeseries import TimeSeriesPredictorLoRA fine-tuning model training configuration
predictor = TimeSeriesPredictor(target="target", prediction_length=12)
predictor.fit(
train_data,
hyperparameters={
"Chronos": {"model_path": "amazon/chronos-2", "fine_tune": True}
}
)
Measurable Gains in Efficiency and Accuracy
Since deploying Chronos-2, Decathlon has achieved a 15% improvement in Weighted Average Percentage Error (WAPE) for their 12-week replenishment forecasts, effectively mitigating both stockouts and overstock scenarios. Beyond accuracy, the operational impact has been profound: the time required to deploy the forecasting system in new markets has dropped from six months to between two and three months. The system currently manages up to 25,000 products per region across multiple supply zones, including Southeast Asia and Latin America. This success demonstrates that foundation models can effectively manage the inherent volatility of large-scale retail data while providing the agility required for modern supply chain management.
For organizations looking to replicate these results, the key lies in evaluating how well a model handles covariates and how efficiently it can be fine-tuned. Practitioners should leverage tools like Amazon SageMaker JumpStart to verify infrastructure compatibility and use automated libraries to streamline data pipelines. By focusing on periodic fine-tuning and rigorous version control, companies can significantly lower their operational costs while maintaining high-performance forecasting.
