How Decathlon runs demand forecasting at scale with Chronos-2
Decathlon selected Chronos-2 as a core component of its demand forecasting stack after rigorous benchmarking across 101 rolling cutoffs spanning nearly 2 years on ~25,000 products per zone Chronos-2 fine-tuned (every 6 months) consistently outperformed all evaluated time series foundation models and surpassed the prior production baseline (DeepAR + Holt-Winters, later TFT) at both 12-week and 52-week horizons Even in zero-shot mode, Chronos-2 matched or exceeded the fully retrained weekly produc
Analysis
TL;DR
- Decathlon selected Chronos-2 as a core component of its demand forecasting stack after rigorous benchmarking across 101 rolling cutoffs spanning nearly 2 years on ~25,000 products per zone
- Chronos-2 fine-tuned (every 6 months) consistently outperformed all evaluated time series foundation models and surpassed the prior production baseline (DeepAR + Holt-Winters, later TFT) at both 12-week and 52-week horizons
- Even in zero-shot mode, Chronos-2 matched or exceeded the fully retrained weekly production baseline, demonstrating strong generalization on retail demand data
- Native covariate support via group attention was a key architectural differentiator, eliminating workarounds required by other TSFMs
- The AWS-based pipeline achieves under 2 minutes of inference per cutoff for 25,000 products, combining PySpark data prep, AutoGluon fine-tuning, MLflow model registry, and batch inference
Why It Matters
This case study demonstrates that time series foundation models are ready for large-scale production deployment in retail supply chain operations, moving beyond research benchmarks into real-world impact. It provides a replicable blueprint for how enterprises can evaluate, select, and operationalize TSFMs—addressing practical concerns around fine-tuning frequency, infrastructure efficiency, and covariate integration that many practitioners face.
Technical Details
- Benchmark methodology: 101 rolling cutoffs (week 48/2022 to week 44/2024), ~25,000 unique products per cutoff (39,000 total time series), evaluating WAPE as primary metric with RMSE, bias, and pairwise winning rates at 12-week and 52-week weekly horizons
- Model evolution: Legacy system used Amazon SageMaker DeepAR (weeks 1–16) + Holt-Winters (weeks 17–52); transitioned to Temporal Fusion Transformer (TFT) with covariates in 2024 before adopting Chronos-2
- Architecture: PySpark data preparation → AutoGluon-based fine-tuning (every 6 months) → MLflow model registry → weekly batch inference → PySpark exposition pipeline for downstream delivery
- Key differentiator: Chronos-2's native covariate handling through group attention mechanism, unlike most TSFMs that require ad-hoc workarounds for exogenous variables
- Performance: Fine-tuned Chronos-2 achieved the lowest forecast error across both horizons; zero-shot performance alone matched or beat the prior retrained weekly baseline; inference completes in under 2 minutes per cutoff on AWS infrastructure supporting both CPU and GPU
Industry Insight
- Foundation models are production-viable for time series: Enterprises should prioritize rigorous internal benchmarking on their own data rather than relying solely on published model evaluations—Decathlon's approach of 101 rolling cutoffs on actual retail data provided decisive evidence
- Fine-tuning cadence can be infrequent for TSFMs: A 6-month fine-tuning schedule was sufficient to capture value, dramatically reducing operational overhead compared to the previous weekly retraining requirement—this challenges the assumption that foundation models need constant retraining
- Covariate support is a critical selection criterion: Models with native exogenous variable handling (like Chronos-2's group attention) will integrate more cleanly into existing supply chain pipelines that rely on promotional, pricing, and macroeconomic signals, avoiding costly engineering workarounds
Disclaimer: The above content is generated by AI and is for reference only.