How Decathlon runs demand forecasting at scale with Chronos-2

This put up is co-written with Vianney Bruned, Filippo Giruzzi, Belkiss Saidi, and Carlos Ramirez from Decathlon.

Decathlon is likely one of the world’s largest sporting items retailers, with greater than 100,000 teammates and 400 million customers worldwide. The corporate depends on correct demand forecasting at scale to help the provision of the suitable merchandise in every retailer on the time prospects want them. After evaluating a number of time collection basis fashions (TSFMs), Decathlon chosen Chronos-2 as a core element of their forecasting stack.

On this put up, we share the structure Decathlon makes use of to run Chronos-2 at scale on AWS, the enterprise influence on Decathlon’s provide chain operations, and sensible classes realized for different corporations that need to undertake basis fashions for forecasting.

Decathlon’s forecasting problem

Correct demand forecasting is the spine of retail provide chain operations. For Decathlon, this problem is amplified by the sheer scale and variety of their enterprise: tens of hundreds of merchandise spanning over 80 sports activities, bought throughout a number of continents with extremely seasonal demand patterns. A pair of ski gloves and a surfboard have essentially completely different demand indicators, but each should be forecasted precisely to keep away from stockouts or overstock.

Decathlon’s forecasting system predicts the weekly gross sales amount of all merchandise on two vital horizons. The primary is a 12-week replenishment window utilized by buy planners to order items from industrial companions. The second is a 52-week strategic horizon for long-term inventory projection and capability planning. The forecasting system runs weekly throughout each horizons. It’s deployed throughout a number of provide zones, together with Europe, India, China, South East Asia (SEA), Latin America (LATAM), and shortly the Center East and Africa. Every zone covers as much as 25,000 merchandise.

Earlier strategy and its limitations

Decathlon’s demand forecasting system developed over a number of years:

• 2021–2024: A hybrid strategy utilizing Amazon SageMaker AI DeepAR for the short-term forecast horizon (weeks 1–16) and Holt-Winters exponential smoothing for the longer-term horizon (weeks 17–52). DeepAR was retrained weekly to adapt to current development shifts.
• 2024 onwards: Introduction of Temporal Fusion Transformer (TFT) with covariates, providing improved long-horizon accuracy.

Whereas these approaches served Decathlon nicely, they got here with operational overhead: the system required weekly re-training and couldn’t simply scale to new areas with out additional engineering effort. The workforce wanted an answer that would ship larger accuracy with decrease operational complexity.

Why Chronos-2: Mannequin analysis and choice

The rise of time collection basis fashions (FMs) promised pre-trained fashions that would outperform classical approaches with out coaching from scratch on domain-specific knowledge. However a vital query remained: would these fashions work on Decathlon’s particular retail datasets?

To reply this, Decathlon designed a rigorous, large-scale benchmark by itself retail knowledge, evaluating a number of TSFMs in opposition to its manufacturing baseline.

Benchmark design

• Analysis: 101 rolling cutoffs spanning practically 2 years (week 48 of 2022 to week 44 of 2024).
• Scale: Roughly 25,000 distinctive merchandise per cutoff (39,000 distinctive product time collection throughout the complete analysis interval).
• Horizon: 12-week and 52-week horizons at weekly frequency.
• Main metric: Weighted Absolute Proportion Error (WAPE) at 12-week and 52-week horizons, supplemented by root imply sq. error (RMSE), bias, and pairwise profitable charges.

For full benchmark outcomes together with comparisons throughout a number of TSFMs, see Decathlon’s detailed evaluation on Medium.

Key findings

Decathlon evaluated a number of TSFMs in each zero-shot and fine-tuned configurations. In Decathlon’s printed benchmark outcomes, Chronos-2 fine-tuned constantly outperformed all different evaluated fashions throughout each horizons. Even in zero-shot mode, it matched or surpassed the totally educated manufacturing baseline. Superb-tuning additional decreased forecast error by a number of proportion factors.

1. Zero-shot viability: A number of TSFMs exceeded or approached the efficiency of Decathlon’s current manufacturing mannequin (retrained weekly) with none domain-specific coaching.
2. Superb-tuning delivered important beneficial properties: Even with a low-frequency fine-tuning schedule (as soon as each 6 months), fine-tuning considerably improved efficiency. Chronos-2 fine-tuned confirmed the bottom error throughout each short-term (12-week) and long-term (52-week) horizons.
3. Computational effectivity: Chronos-2 can run on each CPUs and GPUs, and met the benchmark’s effectivity requirement of beneath 2 minutes of inference per cutoff for 25,000 merchandise.

Past uncooked accuracy, Chronos-2’s native covariate help via its group consideration mechanism was a key differentiator. Not like most TSFMs that require workarounds, Chronos-2 natively incorporates covariates. The mix of main accuracy, architectural class for covariate dealing with, and environment friendly fine-tuning made Chronos-2 the clear alternative for Decathlon’s manufacturing stack.

Answer structure

Decathlon’s manufacturing deployment of Chronos-2 on AWS is designed to be environment friendly, cost-efficient, and dependable.

The next diagram reveals the high-level structure of the demand forecasting pipeline. A PySpark knowledge preparation pipeline assembles the enter time collection. Each 6 months, a fine-tuning job constructed on AutoGluon adapts Chronos-2 to the most recent knowledge and registers the ensuing mannequin in an MLflow mannequin registry. Within the intervening weeks, this step is skipped. The inference pipeline fetches the most recent registered mannequin and runs weekly batch forecasts, and a PySpark exposition pipeline delivers the forecasts to downstream shoppers.

Determine 1: Excessive-level structure of Decathlon’s demand forecasting pipeline on AWS

Infrastructure and deployment sample

The structure makes use of Amazon EC2 situations for batch inference, triggered by Databricks jobs. Knowledge pipelines are orchestrated via Airflow on Decathlon’s current knowledge platform. The fine-tuning course of makes use of Low-Rank Adaptation (LoRA) via the AutoGluon Chronos integration, permitting environment friendly adaptation to Decathlon’s area with out full mannequin re-training. The mannequin is fine-tuned robotically each 6 months on the most recent knowledge, and the completely different fashions are logged and versioned per provide zone with particular hyperparameters via MLflow.

Chronos-2 technical overview

Chronos-2 is an encoder-only transformer intently following the T5 encoder design. It’s obtainable in a number of variants, together with the bottom mannequin (amazon/chronos-2) with 120M parameters and a small mannequin (autogluon/chronos-2-small) with 28M parameters. Not like the unique Chronos, which quantized values into discrete tokens, Chronos-2 applies strong scaling to every collection. It then splits every collection into non-overlapping patches which are mapped to real-valued embeddings via a residual community. Forecasts are produced as steady quantiles by a quantile head. The important thing architectural innovation is the alternating consideration sample. Every transformer block alternates between time consideration (alongside the temporal axis inside a single collection) and group consideration (throughout collection inside a bunch at every patch index).

The next diagram illustrates this design. Associated time collection and their covariates are grouped collectively. Data is exchanged each alongside time inside every collection and throughout the collection in a bunch, permitting native multivariate forecasting with covariates.

Determine 2: Chronos-2 structure for native multivariate forecasting with covariates

Getting began with Chronos-2

As of this writing, Chronos-2 fashions have been downloaded over 120 million instances from Hugging Face and can be found for Amazon SageMaker AI prospects via AutoGluon-Cloud or Amazon SageMaker JumpStart.

The next code tailored from the Chronos-2 quickstart pocket book on GitHub, demonstrates run inference with covariate help. In manufacturing, Decathlon makes use of the AutoGluon Chronos integration for fine-tuning and inference orchestration with a barely completely different API:

import pandas as pd
from chronos import BaseChronosPipeline, Chronos2Pipeline

# Load the Chronos-2 pipeline
pipeline: Chronos2Pipeline = BaseChronosPipeline.from_pretrained(
    "amazon/chronos-2",
    device_map="cpu"
)

# Historic knowledge
df = pd.read_csv("gross sales.csv")  # columns: item_id, timestamp, gross sales, worth, store_count

# Knowledge obtainable through the forecast horizon
future_df = pd.read_csv("future.csv")  # columns: item_id, timestamp, worth, store_count

prediction_length = 52
forecast = pipeline.predict_df(df, future_df, goal="gross sales", prediction_length=prediction_length)

Superb-tuning with AutoGluon

Decathlon makes use of AutoGluon-TimeSeries to fine-tune and serve Chronos-2. With AutoGluon, you may streamline the end-to-end machine studying workflow, from knowledge preparation to mannequin coaching and deployment. Utilizing its TimeSeries module’s high-level API, you may deal with knowledge formatting, covariate administration, and fine-tuning with LoRA in a couple of strains of code:

# pip set up autogluon.timeseries
from autogluon.timeseries import TimeSeriesDataFrame, TimeSeriesPredictor

tsdf = TimeSeriesDataFrame.from_data_frame(df, id_column="item_id", timestamp_column="timestamp")
future_tsdf = TimeSeriesDataFrame.from_data_frame(future_df, id_column="item_id", timestamp_column="timestamp")

predictor = TimeSeriesPredictor(
    prediction_length=52,  # forecast horizon
    goal="gross sales",  # column to forecast
    known_covariates_names=["price", "store_count"],  # options identified sooner or later
)

predictor.match(
    tsdf,
    hyperparameters={
        "Chronos2": {
            "model_path": "amazon/chronos-2",
            "fine_tune": True,  # Activate fine-tuning; if False the mannequin is utilized in zero-shot mode
        }
    },
)

predictor.predict(tsdf, known_covariates=future_tsdf)

Outcomes and enterprise influence

Decathlon measured the influence of the Chronos-2 deployment alongside three dimensions: forecast accuracy, enterprise outcomes, and operational effectivity.

Forecast accuracy enhancements

Decathlon’s deployment of Chronos-2 fine-tuned has delivered important accuracy beneficial properties throughout each areas and horizons, compared with the beforehand used legacy forecasting device:

pp = proportion factors

Enterprise influence

Every proportion level of WAPE enchancment on the 12-week horizon interprets to:

• 0.3 days of stock financial savings, decreasing working capital and warehousing prices.
• 0.3 factors of product availability, fewer stockouts on cabinets.
• 0.4 factors of gross sales improve for every level of availability gained, which is roughly 0.12 factors of gross sales per WAPE level (common throughout all zones).

With 11–15 factors of WAPE enchancment on the replenishment horizon, the compounding impact on stock effectivity, availability, and income is substantial.

Operational effectivity

Past accuracy, the migration to Chronos-2 decreased operational overhead throughout a number of dimensions:

The discount in deployment time from 6 months to 2–3 months is especially impactful for Decathlon’s growth into new markets. With Chronos-2, the workforce can deploy forecasting capabilities to a brand new area by working fine-tuning on native historic knowledge with no structure redesign required.

Manufacturing timeline

Chronos-2 is already in manufacturing for the South East Asia and Latin America provide zones, that are the areas the previous outcomes are drawn from. Decathlon is now extending the answer throughout its remaining provide zones, with full multi-zone manufacturing rollout focused for 2026 and the Center East and Africa areas as the subsequent growth targets.

Conclusion and classes realized

Decathlon’s journey demonstrates that point collection basis fashions, and Chronos-2 particularly, are prepared for production-grade retail demand forecasting. By combining a rigorous mannequin analysis course of with environment friendly fine-tuning with LoRA, Decathlon achieved 11–15 factors of WAPE enchancment on the 12-week horizon. On the similar time, the workforce decreased deployment time from 6 months to 2–3 months per area. For retailers evaluating basis fashions for forecasting, Decathlon’s expertise presents a sensible blueprint.

Classes realized

1. Benchmark by yourself knowledge: World leaderboards are helpful for shortlisting candidates, however mannequin rankings can differ considerably on domain-specific knowledge. In Decathlon’s case, some fashions that ranked larger globally carried out worse on retail distribution datasets. Testing throughout 25,000 merchandise and 101 cutoffs additionally revealed that Chronos-2’s simple fine-tuning with AutoGluon was a decisive benefit, turning a robust zero-shot mannequin into the clear chief.
2. Superb-tuning unlocks the complete potential: Even rare fine-tuning (each 6 months) with Low-Rank Adaptation (LoRA) measurably improves accuracy over zero-shot inference. The mix of a robust pre-trained basis with area adaptation is the profitable formulation.
3. Begin with a targeted strategy, iterate: Decathlon started with a single fine-tuned mannequin in manufacturing and is progressively including covariates. This pragmatic strategy reduces threat whereas delivering worth early.
4. Basis fashions democratize forecasting: Working Chronos-2 on a single m6i.8xlarge CPU occasion with inference instances of 40–75 seconds for 7,000–15,000 time collection makes forecasting at this scale accessible with out GPU infrastructure. At roughly $0.03 per weekly inference run, the compute price is negligible.

Subsequent steps

Trying forward, Decathlon plans to deploy a Combination of Specialists (MoE) ensembling strategy in manufacturing. Their benchmarks present that combining a number of TSFMs constantly outperforms the only finest knowledgeable, with different fashions nonetheless profitable on roughly 40 p.c of merchandise. The workforce can also be increasing to Center East and Africa areas and integrating exterior knowledge comparable to worth or climate as covariates. As well as, they’re exploring Chronos-2’s cross-learning capabilities for cold-start merchandise.

To discover Chronos-2 on your personal forecasting use instances or to study extra about Chronos-2, go to the Amazon Science weblog put up, the analysis paper, or strive the quickstart pocket book. For fine-tuning with AutoGluon, see the AutoGluon Chronos tutorial.

Concerning the authors