Case study • Decathlon
Demand Forecasting at European Scale
~€35M in annual savings — and a forecasting model the supply chain could finally audit
Annual savings
~€35M
Inventory optimisation (program-level)
Forecast error
−2.5%
Global WAPE on sales forecasts
Scope
30,000+
Products, European scale
Open source
2 PRs
Merged into Nixtla/NeuralForecast
Overview
Decathlon's country-level demand forecasting drives inventory decisions across 30,000+ products in Europe. I worked on the forecasting core: foundation models, calibrated uncertainty, and a Temporal Fusion Transformer rebuilt so planners could see why it recommended what it recommended. The interpretability work was the condition for adoption — and the improvements were contributed back upstream to Nixtla's NeuralForecast.
The Challenge
Classical models handled the regular part of demand well and the interesting part badly: seasonality, promotions, regional differences, new products with no history. But the deeper blocker was not accuracy — supply chain teams would not act on recommendations they could not explain.
Key pain points
- •Point forecasts with no usable uncertainty, leaving safety stock a manual judgement call
- •Cold start on new or sparse products, where history-only models have nothing to work with
- •'Black box' deep learning outputs that planners refused to act on
- •Diverse product categories poorly served by a single fixed architecture
The Solution
A forecasting stack built on three levers: better priors from foundation models, honest uncertainty from conformal prediction, and a TFT reworked for interpretability and architectural flexibility.
Foundation models for time series
Fine-tuned state-of-the-art foundation models (TimesFM, Chronos) to transfer knowledge across product hierarchies, improving behaviour on series with little or no history.
Multimodal semantic enrichment
Built a pipeline injecting text embeddings of product descriptions as covariates, letting the model exploit similarity between items beyond their individual time series.
Conformal prediction for calibrated uncertainty
Replaced point estimates with statistically robust prediction intervals, so safety stock could be sized against a real risk level instead of a rule of thumb.
TFT rebuilt for interpretability — the condition for adoption
Added native feature importance and attention weight extraction, plus flexible RNN configurations (GRU/LSTM, multi-layer stacking, static covariate initialisation). Planners could audit which variables and which time steps drove a given recommendation.
Contributed back upstream
The interpretability suite and the custom RNN layers were merged into Nixtla/NeuralForecast (PRs #1104 and #1230), validating the approach outside Decathlon's own codebase.
Results & Impact
~€35M annual savings
Through inventory optimisation
Program-level figure at Decathlon. My contribution was the forecasting core: the TFT architecture work, the uncertainty quantification, and the interpretability that made the recommendations actionable.
−2.5% WAPE
Global forecast error
Measured on sales forecasts against the incumbent baseline, across the European product scope.
Auditable recommendations
Feature importance + attention weights
Supply chain stakeholders could inspect the drivers behind a forecast — what moved the model from pilot to decision support at scale.
2 PRs merged upstream
Nixtla/NeuralForecast (2.8k+ stars)
The interpretability suite and the custom RNN layers are now available to the library's users worldwide.
Technologies used
“Yanis has played a pivotal role in building our country-level demand forecasting system and has made outstanding contributions to our deep learning forecasting models, particularly the Temporal Fusion Transformer (TFT) model.
Ready to move your AI project forward?
Let's talk about what a similar engagement would look like in your context.
Get in touch