Superpower Daily
Anthropic Says Claude Found Fixes Across 10 Alignment Failures,...TeamPCP’s Trivy Attack Hit 2,500 Organizations; AI Tool Backdoo...Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near...Decathlon Deploys Chronos-2, Cuts 12-Week WAPE by 11–15 PointsSalesforce Says Agentforce Passed $1.5B ARR as It Pushes Consum...JAMA Paper Says Autonomous AI Could Surpass Physicians and AI-A...Anthropic Says Claude Found Fixes Across 10 Alignment Failures,...TeamPCP’s Trivy Attack Hit 2,500 Organizations; AI Tool Backdoo...Meta’s EvoHarness-RL Takes Qwen3-8B to 96.9% on ALFWorld, Near...Decathlon Deploys Chronos-2, Cuts 12-Week WAPE by 11–15 PointsSalesforce Says Agentforce Passed $1.5B ARR as It Pushes Consum...JAMA Paper Says Autonomous AI Could Surpass Physicians and AI-A...
Businesspublished

Decathlon Deploys Chronos-2, Cuts 12-Week WAPE by 11–15 Points

The retailer’s comparison with a weekly-retrained TFT system makes the operational case for fine-tuned forecasting models. The published production results, however, come from Southeast Asia and Latin America.

By 4 min read
Decathlon Deploys Chronos-2, Cuts 12-Week WAPE by 11–15 Points
Sign in to save and react

Listen to this story

The audio brief

About 1:48
0:001:48
Read transcript
Decathlon has put Amazon’s Chronos-2 into production for demand forecasting in Southeast Asia and Latin America, and the retailer says the model cut 12-week weighted absolute percentage error—WAPE—by 11 points in Southeast Asia and 15 in Latin America. That matters because the 12-week forecast drives replenishment: how much Decathlon orders from industrial partners for its stores. The older system, a Temporal Fusion Transformer, or TFT, needed weekly retraining. Chronos-2 is fine-tuned every six months, then runs weekly batch forecasts. In Decathlon’s local benchmark, the test covered 101 rolling cutoffs, roughly 25,000 products at each cutoff, and 39,000 product time series overall. Fine-tuned Chronos-2 had the lowest reported error among the models tested at both 12- and 52-week horizons. At 52 weeks, WAPE fell from 44% to 38% in Southeast Asia, and from 55% to 46% in Latin America. The operational result is important too: CPU inference handled 7,000 Latin American series in about 40 seconds and 15,000 Southeast Asian series in about 75 seconds, meeting a requirement to forecast 25,000 products in under two minutes. Decathlon says regional deployment now takes two to three months, versus about six months with TFT. It estimates each 12-week WAPE point translates to 0.3 inventory days, 0.3 availability points, and 0.12 sales points—but those relationships come from the company. The key constraint is clear: the published production gains come from only two supply zones. Middle East and Africa are next, making that rollout the real test of whether the result generalizes.

Story brief

3 key points

Decathlon’s local benchmark indicates Chronos-2 can improve retail forecasting without the weekly retraining burden of its prior Temporal Fusion Transformer system. Fine-tuning every six months, followed by weekly batch inference, cut 12-week WAPE by 11 points in Southeast Asia and 15 in Latin America; 52-week gains were smaller but still substantial. The result is operationally notable because inference met a...

  1. 01

    The benchmark covered 101 rolling cutoffs, roughly 25,000 products per cutoff, and 39,000 product time series.

  2. 02

    At 52 weeks, WAPE fell from 44% to 38% in Southeast Asia and from 55% to 46% in Latin America.

  3. 03

    CPU inference handled 7,000 Latin American series in 40 seconds and 15,000 Southeast Asian series in 75 seconds.

Decathlon has put Amazon’s Chronos-2 time-series model into production for demand forecasting in Southeast Asia and Latin America, replacing a more frequently retrained approach with a model it fine-tunes every six months. The retailer says the change lowered 12-week weighted absolute percentage error, or WAPE, by 11 percentage points in Southeast Asia and 15 points in Latin America—an outcome that ties a foundation-model deployment to the immediate work of stocking stores.

Demand forecasts determine how much Decathlon orders from industrial partners over a 12-week replenishment window, while a 52-week view supports stock and capacity planning. Those jobs span supply zones and product categories with sharply different seasonal patterns, so the company needs one system to forecast weekly sales across large product catalogs rather than a one-size-fits-all estimate.

A local benchmark, not a leaderboard choice

The predecessor system had already evolved. From 2021 through 2024, Decathlon used DeepAR for shorter forecasts and Holt-Winters smoothing for later weeks. It then adopted a Temporal Fusion Transformer, or TFT, with covariates—known inputs such as price or store count—to improve long-horizon forecasts. But the system required weekly retraining and extra engineering work to reach new regions.

Rather than assume a pre-trained forecasting model would transfer to retail, Decathlon evaluated multiple time-series foundation models on its own data. The test used 101 rolling forecast cutoffs from late 2022 through late 2024, roughly 25,000 products at each cutoff, and 39,000 product time series across the full evaluation. It measured both 12- and 52-week horizons, using WAPE as the primary metric alongside error, bias, and pairwise winning measures.

Decathlon says fine-tuned Chronos-2 produced the lowest error among the models it tested at both horizons. Its zero-shot version, meaning it was used without domain-specific training, matched or exceeded the existing retrained production baseline in the company’s benchmark. Fine-tuning added further gains, even on a six-month schedule.

The operating model changes with the model

The contrast is not only accuracy. Chronos-2’s design can incorporate covariates natively through group attention, which exchanges information across related series while time attention follows each series over time. Decathlon uses Low-Rank Adaptation, or LoRA, to adjust the model without fully retraining it. Its production pipeline runs weekly batch inference, but fine-tunes only every six months and versions models separately for each supply zone.

Diagram of Decathlon’s forecasting pipeline, from PySpark data preparation through Chronos-2 fine-tuning, model registration, weekly inference, and forecast delivery.
Decathlon’s architecture prepares time-series data, fine-tunes and registers a model every six months, then retrieves that model for weekly batch forecasts. Source: aws.amazon.com.

That lighter update schedule is paired with a conventional production stack: PySpark prepares and delivers data, Databricks triggers jobs, Airflow orchestrates pipelines, and MLflow holds zone-specific model versions. The company says CPU inference took about 40 seconds for 7,000 Latin American series and about 75 seconds for 15,000 Southeast Asian series. Chronos-2 also met its benchmark requirement of completing inference for 25,000 products in under two minutes.

A production result with a defined boundary

At the longer 52-week horizon, Decathlon reports smaller but still material reductions: WAPE fell from 44% to 38% in Southeast Asia and from 55% to 46% in Latin America. The company also estimates that every one-point 12-week WAPE improvement corresponds, on average across zones, to 0.3 days of inventory savings, 0.3 points of product availability, and roughly 0.12 points of sales. Those business translations are company-supplied relationships, not independently reported financial results.

Decathlon says a new regional deployment now takes two to three months with a data scientist, ML engineer, and data analyst, compared with about six months and three people under TFT. The company plans a full multi-zone rollout during 2026, with the Middle East and Africa next. That expansion is the consequential test: the published accuracy and operating gains currently come only from its Southeast Asia and Latin America supply zones.

Sources

  1. aws.amazon.comHow Decathlon runs demand forecasting at scale with Chronos-2 | Amazon Web Services

Related coverage

Continue reading