Amazon
catalog-scale probabilistic demand forecasting
Amazon's SCOT team forecasts demand for more than 400 million products across 185 countries using probabilistic deep learning models (DeepAR, MQ-Transformer) and achieved a 15x improvement in accuracy after moving to a unified deep learning model.
Key points
- Catalog-scale probabilistic demand forecasting (SCOT team).
- In-house deep learning models (DeepAR, MQ-Transformer) on Apache Spark.
- Forecast accuracy improved 15-fold, across more than 400 million products.
- Evidence level B, confirmed active status.
Objective
Predict demand for every product in the catalog to decide how much to order and where to place stock, so that availability holds without tying up capital in overstock.
The deployment
Amazon sells more than 400 million products across over 185 countries. Its SCOT team (Supply Chain Optimization Technologies) forecasts demand for each item to drive replenishment, stock placement, and logistics capacity decisions. The system has evolved since 2008: from classic time series methods, to a sparse quantile random forest, and finally to a unified deep learning model. Published work includes DeepAR (autoregressive recurrent networks producing probabilistic forecasts) and MQ-Transformer (multi-horizon forecasts with attention). The forecast is probabilistic: rather than a single number, it produces a distribution, which allows the trade-off between stockout risk and overstock cost.
Results Proof B
Technical publications from Amazon Science (Amazon's scientific organization) that name Amazon's forecasting system and its quantified accuracy gain. Official interested source, not independent, hence B.
How it works
Documented architectureThe stack in detail
- llm DeepAR Autoregressive recurrent networks for probabilistic forecasting, published by Amazon and available open source via GluonTS.
- llm MQ-Transformer Multi-horizon quantile model with attention, published by Amazon Science, core of the unified deep learning model.
- infra Plateforme ML maison sur Apache Spark Training and scoring pipeline at the scale of 400 million items.
- outil Systemes SCOT de reapprovisionnement Internal systems that consume the demand distributions to decide order quantities and stock placement.
How it runs, concretely
For ops teams-
1Data preparation ML platform / data team
Sales, price, availability, seasonality, and product attributes are consolidated into time series at catalog scale.
-
2Probabilistic forecasting AI model
The deep learning model (DeepAR / MQ-Transformer type) produces, for each item, a demand distribution over several horizons.
-
3Stock decision SCOT systems
The replenishment systems consume these distributions to decide order quantity and placement location.
-
4Learning loop ML platform
Actual sales and observed stockouts flow back as training data and refine subsequent forecasts.
The predicted demand distribution per product. If the sales history is short (new product) or disrupted (stockout, exceptional promo), the forecast becomes noisy and the system relies on comparable products.
How your customers perceive this type of use
Sourced studiesC'est la famille la moins acceptee : 68% des Americains jugent inacceptable un score financier personnel calcule par algorithme et 67% l'analyse video automatisee d'entretiens d'embauche (Pew Research, 2018). La demande d'explication et de recours est massive : 83% veulent savoir quelles donnees l'IA utilise et 91% veulent pouvoir corriger des donnees erronees (Consumer Reports, 2024). A l'echelle mondiale, seuls 46% se disent prets a faire confiance aux systemes d'IA et 70% jugent une regulation necessaire (KPMG / Universite de Melbourne, 2025).
Acceptance conditions
- Transparence sur les donnees utilisees : 83% des Americains la reclament (Consumer Reports 2024)
- Droit de correction des donnees erronees : 91% le demandent (Consumer Reports 2024)
- Explication de la logique de decision : 44% des consommateurs sont plus enclins a utiliser un agent IA si sa logique est clairement expliquee (Salesforce 2024)
- L'acceptabilite depend du contexte de la decision : 50% des Americains jugent equitable un score de risque criminel pour la liberation conditionnelle, contre 32% pour un score financier applique aux consommateurs (Pew Research 2018)
Red lines
- La decision opaque et sans recours sur l'emploi, le credit ou le logement : 45% tres mal a l'aise pour l'embauche, 39% pour le pret, 39% pour le logement (Consumer Reports 2024)
- Le scoring des personnes a partir de donnees comportementales : 68% le jugent inacceptable pour les offres financieres (Pew Research 2018)
Sources: Pew Research Center 2018 · Consumer Reports 2024 · KPMG / Universite de Melbourne 2025 · Salesforce 2024
How to replicate
Inference, not sourcedData prerequisites
- multi-year sales history per SKU
- availability and price signals
- product attributes to link new items to comparables
Org prerequisites
- supply chain data science team
- replenishment system able to consume probabilistic forecasts
Possible stack
- in-house forecasting (DeepAR via GluonTS, MQ-Transformer)
- Amazon Forecast / SageMaker
- Google Vertex AI Forecast
- demand planning platforms
The plan, step by step
- Step 1Consolidate the time series per SKU: sales, price, availability, product attributes over several years.Deliverable: Multi-year dataset ready for training
- Step 2Replace the point forecast with a probabilistic forecast (quantiles, DeepAR type) on a high-variability category.Deliverable: Quantile model evaluated in backtest against the incumbent
- Step 3Connect the demand distributions to the stock decision: order quantity and placement.Deliverable: Ordering rule that consumes the quantiles
- Step 4Run the pilot in parallel with the existing system and measure stockouts and overstock.Deliverable: Quantified comparison on the pilot category
- Step 5Extend to the rest of the catalog and industrialize the retraining loop.Deliverable: Production forecasting pipeline with accuracy monitoring
First step: Replace a point forecast (mean) with a probabilistic forecast (quantiles) on a high-variability category, and measure the gain on stockouts and overstock.
Sources
- S1 The history of Amazon's forecasting algorithm Interested party archive pending
- S2 Probabilistic demand forecasting at scale Interested party archive pending
An error, newer info, a source?
This page lives on its accuracy. If a figure has moved, if the deployment has changed, or if you have a higher-quality source, tell us. Every sourced correction is verified before publication.