Build1 distinct publisher3 min readPublished
The retailer benchmarked several time series foundation models over 101 rolling cutoffs of its own catalogue, and the one it shipped now covers both the 12-week and 52-week horizons that two separate models used to split.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Start with the retraining cadence, because that is where the operational bill changes. DeepAR was retrained every week specifically to adapt to recent trend shifts [8]. Chronos-2 is fine-tuned once every six months [18], which is roughly 26 weeks between weight refreshes [24]. That moves the job of tracking a trend break out of the training loop and into the context the model is handed at inference time. Under the old design, a demand shift in March was in the weights by the following Monday. Under the new one, it has to be read off the recent history of the series. That is a reasonable trade if the model reads context well, and it is the behaviour I would regression-test hardest before putting it in front of purchase planners.
The second structural change is the seam. From 2021 to 2024 the stack ran DeepAR for weeks 1 to 16 and Holt-Winters for weeks 17 to 52 [8], so the 12-week replenishment forecast that feeds supplier orders [4] sat entirely inside the deep model, and the 52-week capacity view [5] came from something else. Covariates only enter the story in 2024, with Temporal Fusion Transformer for long-horizon accuracy [9]. One model now spans both horizons at weekly frequency [13].
The evaluation is the part worth copying. Week 48 of 2022 to week 44 of 2024 is about 100 weeks, and Decathlon ran 101 rolling cutoffs across it [11], so the cutoffs sit one week apart and the backtest replicates the production cadence rather than sampling convenient dates [23]. About 25,000 products appear per cutoff against 39,000 unique series over the full window [12]. That difference is roughly 14,000 series entering or leaving, about 56 percent of a single cutoff's panel [22]. Catalogue churn on that scale is exactly why the zero-shot result matters: new products arrive with no history, and weights refreshed twice a year will not have seen them [16]. At least five named supply zones at up to 25,000 products each is up to 125,000 series per weekly run [25].
For the numbers to transfer, your baseline has to look like theirs. Decathlon's production comparator was a global deep model retrained weekly [8], not a per-SKU statistical fit, so a shop still running per-item exponential smoothing is not reading its own bake-off here. WAPE is volume-weighted [14], which means high-volume lines dominate the headline figure and a long tail of single-store items can be badly wrong without moving it. Weekly frequency, strong seasonality across more than 80 sports [20], and horizons of 12 and 52 weeks are the shape of the workload the result was measured on.
The AWS post says fine-tuned Chronos-2 beat every other evaluated model at both horizons and that fine-tuning cut error by several percentage points [15][17], but the model-by-model figures sit in Decathlon's own analysis on Medium [19]. For a decision that ends in supplier orders, that is one hop further from the evidence than I would like.
Ranked by verification strength, evidence, and original report placement.
In Decathlon's published benchmark results, fine-tuned Chronos-2 consistently outperformed all other evaluated models across both horizons.
After evaluating multiple time series foundation models, Decathlon selected Chronos-2 as a core component of its forecasting stack.
The AWS blog post describing the deployment was co-written with Vianney Bruned, Filippo Giruzzi, Belkiss Saidi and Carlos Ramirez of Decathlon.
Decathlon's forecasting system predicts weekly sales quantity on a 12-week replenishment window used by purchase planners to order goods from industrial partners.
The second horizon is 52 weeks, used for long-term stock projection and capacity planning.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 28, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
Your Databricks Pipeline Is A Demo Until Promotion Only Runs One Way1 distinct publisher
build
SageMaker v3 drops the framework estimators, and your training code is the migration1 distinct publisher
build
AWS wants data governance to ship as a pull request, not a sign-off1 distinct publisher
build
Force the tool call, then hand Lightsail a long-lived key1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Rigorous design, results kept elsewhere
The benchmark behind this story is described with unusual care — 101 weekly cutoffs on two years of Decathlon's own catalogue, 39,000 real product series, WAPE plus RMSE, bias and pairwise winning rates. Then the post declines to show a single number, sending readers to Decathlon's Medium analysis for the comparisons. A method that specific, reported without its results table and by the vendor whose model won, gets partial credit at best.
Live in five zones, one retailer
This is past the pilot stage: weekly runs across five named supply zones, up to 25,000 products each, feeding purchase planners who place real orders with industrial partners, plus a 52-week view used for capacity planning. As many as 125,000 series a week is meaningful production load. But it is one company, self-reported, with no second adopter and no operator outside Decathlon confirming the pattern holds.
Superlatives where numbers should be
The language runs slightly ahead of what is shown. 'Consistently outperformed all other evaluated models' and 'matched or surpassed the fully trained production baseline' are strong claims presented with no error figures, no named competitor models and no per-horizon breakdown. The restraint is real too — nobody promises a revenue lift or a stockout reduction, and the six-monthly fine-tune is described as a finding rather than a miracle. Modest overstatement, mostly by omission.
House model, house blog, house customer
AWS builds Chronos-2, hosts the pipeline, sells the EC2 capacity it runs on, and published this account on its own machine learning blog with four Decathlon engineers as co-authors. The retailer has its own stake: a team that just replaced a stack it maintained for three years benefits from that decision reading well. Nothing here is hidden — the co-authorship is stated in the first line — but every incentive in the story points the same direction, and the losing models are never named.
Specific enough to trust, narrow enough to doubt
First-party operational facts of this granularity — week ranges each old model owned, the cutoff count, the skipped fine-tuning weeks, an inference budget in minutes — are the kind of thing a team gets right about its own system and would be embarrassed to fabricate. So the architecture and cadence read as solid. The accuracy verdict is a different matter: single channel, interested parties, and the deciding figures parked on another site. Confident about what Decathlon runs, less so about how much better it is.