Decathlon Cuts Forecast Error 15 pp With 120M-Param Chronos-2
In this article
Decathlon, a sporting goods retailer with more than 100,000 employees and 400 million users worldwide, has moved Chronos-2 into production for demand forecasting across its South East Asia and Latin America supply zones, with a full multi-zone rollout targeted for 2026. The deployment replaces a hybrid stack that combined Amazon SageMaker DeepAR with Holt-Winters exponential smoothing (2021–2024) and later Temporal Fusion Transformer (TFT), both of which required weekly retraining and months of per-region engineering effort. The result reinforces a thesis this publication has covered before: that pipeline architecture—not better models alone—drives measurable AI gains when an organisation is willing to redesign the surrounding infrastructure.
Benchmark Design and Model Selection
Decathlon's evaluation used Weighted Absolute Percentage Error (WAPE) as the primary metric, supplemented by RMSE, bias, and pairwise winning rates. The evaluation spanned 101 rolling cutoffs from week 48 of 2022 through week 44 of 2024—nearly two full years of rolling out-of-sample testing—covering approximately 25,000 unique products per cutoff (39,000 unique product time series across the full evaluation period), across both 12-week replenishment and 52-week strategic horizons at weekly frequency, spanning over 80 sports categories with highly seasonal demand.
Chronos-2 fine-tuned outperformed every other evaluated TSFM across both horizons. Even in zero-shot mode—no domain-specific training—Chronos-2 matched or surpassed the existing production baseline that was being retrained weekly. The benchmark also revealed that global leaderboard rankings did not reliably predict per-domain rankings: models that placed higher globally performed worse on Decathlon's retail distribution data, underscoring why evaluation on proprietary data at production scale is non-negotiable before committing to a foundation model.
Beyond raw accuracy, Chronos-2's native covariate support through its group attention mechanism was a decisive differentiator. The model is an encoder-only transformer following the T5 encoder design. Rather than quantizing values into discrete tokens (as the original Chronos did), Chronos-2 applies robust scaling to each series, splits each series into non-overlapping patches mapped to real-valued embeddings via a residual network, and produces continuous quantile forecasts via a quantile head. Each transformer block alternates between time attention (along the temporal axis within a single series) and group attention (across series within a group at each patch index), enabling native multivariate forecasting without architectural workarounds. The base model carries 120M parameters; a small variant carries 28M parameters.
Production Architecture on AWS
A PySpark data preparation stage assembles input time series. Every six months, a fine-tuning job built on AutoGluon applies Low-Rank Adaptation (LoRA) to Chronos-2 on the latest data and registers the resulting model in an MLflow model registry, versioned per supply zone with zone-specific hyperparameters. In intervening weeks the fine-tuning step is skipped entirely. A weekly batch inference job fetches the latest registered model and runs forecasts; a PySpark exposition pipeline delivers outputs to downstream consumers. Airflow on Decathlon's existing data platform handles orchestration; Databricks jobs trigger the EC2 inference instances.
| Component | Before (TFT) | After (Chronos-2) |
|---|---|---|
| Inference compute | GPU required | EC2 m6i.8xlarge (CPU); g5.4xlarge for fine-tuning only |
| Inference runtime | 10 min (10K series) – 15 min (25K series) incl. retraining | ~40 s (7K series, LATAM) / ~75 s (15K series, SEA) |
| Fine-tuning frequency | Weekly | Every 6 months (LoRA) |
| New-region deployment time | ~6 months (3 people) | 2–3 months (data scientist, ML engineer, data analyst) |
| Engineering complexity | High (multi-seed training and inference) | Low (single model, LoRA fine-tuning, no GPU constraint) |
| Weekly compute cost | Not disclosed | ~$0.03 per weekly inference run |
Chronos-2 models have been downloaded over 120 million times from Hugging Face and are available through AutoGluon-Cloud and Amazon SageMaker JumpStart.
Accuracy and Business Impact
Across both production zones, Chronos-2 fine-tuned delivered the following WAPE reductions versus the previously deployed legacy tool:
- SEA, 12-week horizon: 39% → 28% (11 percentage points)
- LATAM, 12-week horizon: 53% → 38% (15 percentage points)
- SEA, 52-week horizon: 44% → 38% (6 percentage points)
- LATAM, 52-week horizon: 55% → 46% (9 percentage points)
Each percentage point of WAPE improvement at the 12-week replenishment horizon translates to 0.3 days of inventory savings, 0.3 points of product availability, and approximately 0.12 points of sales per WAPE point (derived from the 0.4-point sales increase per availability point). At 11–15 WAPE points of improvement, the compounding effects on working capital, warehouse costs, and top-line revenue are substantial, though Decathlon has not published a single aggregated revenue figure.
What Comes Next
Decathlon's roadmap includes a Mixture of Experts (MoE) ensembling approach, motivated by benchmark data showing that other TSFMs still outperform Chronos-2 on roughly 40% of individual products. The team is also expanding to Middle East and Africa regions, adding external covariates such as price and weather signals, and exploring Chronos-2's cross-learning capabilities for cold-start products with limited sales history.
A 120M-parameter open foundation model, fine-tuned twice a year via LoRA on CPU-grade hardware for roughly $0.03 per weekly run, is now competitive with purpose-built, weekly-retrained deep learning pipelines across tens of thousands of SKUs. As enterprises continue mapping where foundation model economics outperform bespoke ML, time-series forecasting—long dominated by hand-tuned statistical and autoregressive models—is clearly within scope.