How Decathlon uses AI for demand forecasting and supply planning
Decathlon forecasts weekly demand for up to 25,000 products per supply zone and has moved to a fine-tuned time series foundation model. What it published about accuracy, cost and the limits.
KEY TAKEAWAYS Summary by the editors
- Decathlon forecasts weekly sales quantities for up to 25,000 products per supply zone on a 12-week horizon for replenishment and a 52-week horizon for strategic planning.
- According to an AWS blog post co-written with Decathlon in August 2026, a fine-tuned Chronos-2 model cut 12-week forecast error (WAPE) from 39% to 28% in South East Asia and from 53% to 38% in Latin America.
- Decathlon estimated that each point of WAPE improvement at 12 weeks is worth about 0.3 days of inventory, 0.3 points of product availability and 0.12 points of sales on average.
- Moving to the foundation model reduced retraining from weekly to every six months, cut new-region deployment from about six months to two to three months and lowered weekly inference cost to about $0.03 per run.
- Decathlon has said other models still win on about 40% of products, and it plans an ensemble, external covariates such as weather and wider rollout in 2026.
Decathlon uses machine learning to forecast weekly demand for tens of thousands of products in each of its supply zones, and these forecasts drive replenishment and longer-term planning. In 2026 the sporting goods retailer described moving from region-specific deep learning models to a fine-tuned time series foundation model, Chronos-2, which it says cut forecast error by 6 to 15 percentage points in its first two production regions while making the system cheaper and faster to deploy.
What does Decathlon need to forecast?
According to a post on the AWS Machine Learning Blog written with Decathlon's demand forecasting team and published on 28 August 2026, Decathlon sells tens of thousands of products across more than 80 sports. Its forecasts cover up to 25,000 products per supply zone, across zones including Europe, India, China, South East Asia (SEA) and Latin America (LATAM).
The system produces weekly sales quantity forecasts on two horizons, run every week:
- 12 weeks for replenishment, deciding how much stock to order and where to place it.
- 52 weeks for strategic planning, such as capacity and range decisions.
The scale is what makes this hard. Each product in each zone is its own time series, with seasonality, promotions, price changes and new store openings affecting demand. A model that needs heavy retraining for every region quickly becomes expensive to run and slow to extend.
At the Artefact AI for Industry Summit in Paris in September 2024, Milton Martinez Luaces, Decathlon's vice-president of data for the value chain, said a team of 40 people works on demand forecasting at several levels, from e-commerce to regional warehouses, and that AI supports forecasting, inventory management and assortment planning.
How did Decathlon forecast demand before?
The AWS post describes two earlier generations. From 2021 to 2024, Decathlon used Amazon SageMaker DeepAR for weeks one to 16 and Holt-Winters exponential smoothing for weeks 17 to 52, with DeepAR retrained weekly. From 2024 it used a Temporal Fusion Transformer with covariates. The drawbacks, according to the post, were the overhead of weekly retraining and difficulty scaling to new regions.
What is Chronos-2 and why did Decathlon use it?
Chronos-2 is an open-source time series foundation model released by AWS researchers in October 2025. Amazon Science describes it as able to handle univariate, multivariate and covariate-informed forecasting without task-specific training, including known future inputs such as planned promotions. For a retailer, that means one pretrained model can be adapted to many product series instead of training separate models per region.
Decathlon benchmarked it on its own data across 101 rolling cut-off dates from late 2022 to late 2024, covering about 39,000 unique product time series. The zero-shot model matched or beat the retrained production baseline, and a version fine-tuned with LoRA every six months had the lowest error at both horizons.
How much did forecast accuracy improve?
The AWS post reports the following weighted absolute percentage error (WAPE), where lower is better:
| Region | Horizon | Previous | Chronos-2 | Reduction |
|---|---|---|---|---|
| South East Asia | 12 weeks | 39% | 28% | 11 points |
| Latin America | 12 weeks | 53% | 38% | 15 points |
| South East Asia | 52 weeks | 44% | 38% | 6 points |
| Latin America | 52 weeks | 55% | 46% | 9 points |
Decathlon also translated accuracy into business terms. It estimated that each WAPE point at the 12-week horizon is worth about 0.3 days of inventory saved, about 0.3 points of product availability and about 0.12 points of sales on average across zones. These are internal estimates rather than audited results, but they show how a forecasting team can link model metrics to stock and service levels.
Two points help interpret the table. The gains are largest at the 12-week replenishment horizon, which is where forecast quality most directly affects stock in stores and warehouses. And the starting error levels differ markedly by region: Latin America began from higher error than South East Asia, so the larger reduction there partly reflects more room for improvement. Retailers comparing these figures with their own should note that WAPE depends heavily on product mix, aggregation level and the share of slow-moving items.
What data and infrastructure does the system use?
According to the AWS post, data is prepared in PySpark pipelines orchestrated with Airflow. Fine-tuning runs every six months on a GPU instance, with models versioned per zone in an MLflow registry, while weekly batch inference runs on a CPU instance triggered by Databricks jobs. Known covariates in the published examples include price and store count, and the model produces probabilistic forecasts rather than a single number, which planners can use to set safety stock.
What changed operationally?
The operational gains were as important as accuracy. According to the post, deploying to a new region fell from about six months with three people to two to three months, retraining moved from weekly to every six months, and inference fell from 10 to 15 minutes including retraining to 40 to 75 seconds on a CPU instance. Weekly inference costs about $0.03 per run. The model is in production in SEA and LATAM, with a full multi-zone rollout targeted for 2026 and the Middle East and Africa next.
What can other fashion and sports retailers learn from Decathlon?
The AWS post lists lessons that apply well beyond sporting goods:
- Benchmark on your own data, because public leaderboard rankings may not hold for retail demand.
- Even infrequent fine-tuning improves on zero-shot use of a foundation model.
- Start with one fine-tuned model and add covariates, such as price and store count, step by step.
- CPU inference can be sufficient at this scale, which keeps running costs low.
Two limits remain. New products without history are still hard: Decathlon lists cross-learning for cold-start products as future work, and Martinez Luaces described cold start as a problem it is actively addressing. And trust matters: he said the black-box nature of AI creates resistance among business teams, and that Decathlon aims to make models more of a grey box so planners can see some of the reasoning behind forecasts.
Frequently asked questions
How does Decathlon forecast demand?
Decathlon produces weekly sales forecasts for up to 25,000 products per supply zone on 12-week and 52-week horizons. It now runs a fine-tuned Chronos-2 time series foundation model in production in South East Asia and Latin America, with wider rollout planned.
What is Chronos-2?
Chronos-2 is an open-source time series foundation model released by AWS researchers in October 2025. It can forecast single or multiple related series and use covariates such as planned promotions without task-specific training, and it can be fine-tuned on a company's own data.
How much did AI improve Decathlon's forecast accuracy?
According to AWS and Decathlon, 12-week forecast error (WAPE) fell from 39% to 28% in South East Asia and from 53% to 38% in Latin America. Improvements on the 52-week horizon were smaller, at 6 to 9 points.
What is WAPE in demand forecasting?
WAPE, or weighted absolute percentage error, is the sum of absolute forecast errors divided by total actual sales. It weights errors by volume, so mistakes on high-selling products count more, which makes it a common measure in retail forecasting.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
