Glossary

DeepAR

An autoregressive recurrent neural network model that produces probabilistic forecasts by learning a parametric distribution over future time steps, natively handling cold-start items across related time series.

Get in touch Learn more

ML engineer managing model training cluster on laptop, GPU utilization visible, technical deep learning setup.

PROBABILISTIC FORECASTING

What is DeepAR?

DeepAR is an autoregressive recurrent neural network (RNN) model developed by Amazon for producing accurate probabilistic time series forecasts by learning a parametric distribution over future time steps.

DeepAR is a supervised deep learning algorithm that generates probabilistic forecasts by training a global model across hundreds of related time series. Unlike classical methods like ARIMA, it uses an autoregressive recurrent neural network architecture based on Long Short-Term Memory (LSTM) units to capture complex, non-linear seasonal patterns and long-range dependencies. The model outputs the parameters of a user-specified likelihood distribution, such as Gaussian or negative binomial, enabling it to produce full prediction intervals with associated quantiles rather than single point estimates.

A key advantage of DeepAR is its native ability to handle the cold start problem by learning shared patterns across a portfolio of time series and incorporating static item-specific covariates. It also gracefully manages intermittent demand with zero-inflated distributions. During training, the network learns to predict the next time step given the previous ground-truth values; at inference, it samples from the predicted distribution to generate multiple future trajectories, allowing for robust quantification of forecast uncertainty.

Probabilistic Forecasting Architecture

Key Features of DeepAR

DeepAR is an autoregressive recurrent neural network that produces probabilistic forecasts by learning a parametric distribution over future time steps, natively handling cold-start items across related time series.

Probabilistic Output via Parametric Distributions

Unlike point-forecast models, DeepAR learns the parameters of a user-specified likelihood function at each time step. The model outputs mu and sigma for a Gaussian or alpha for a negative binomial, enabling it to predict entire prediction intervals.

Quantifies forecast uncertainty with quantile regression
Trained by maximizing the log-likelihood of observed data
Supports Gaussian, Beta, Negative Binomial, and Student's t distributions

Native Cold-Start Handling

DeepAR uses static categorical features—such as product category, brand, or location—to learn a global model across thousands of related time series. For new items with no history, the model conditions on these static covariates to generate reasonable initial forecasts.

Shares information across similar time series via learned embeddings
Eliminates the need for separate fallback models for new SKUs
Critical for retail catalogs with high product churn

Autoregressive Recurrent Architecture

The model uses an LSTM encoder-decoder structure. The encoder summarizes the historical context into a hidden state, while the decoder generates predictions one step at a time, feeding each sampled output back as input for the next step.

Captures complex, non-linear temporal dependencies
Handles variable-length time series without padding or truncation
Supports multi-horizon forecasting with a single model

Scale Handling via Item-Level Normalization

DeepAR applies per-item normalization by dividing each time series by its historical mean before training. This prevents high-volume items from dominating the loss function and allows the model to learn patterns across items with vastly different sales volumes.

Stabilizes training across heterogeneous product portfolios
Enables a single model to forecast both fast-movers and long-tail items
Predicted samples are re-scaled to the original magnitude post-inference

Sample Path Generation for Decision-Making

During inference, DeepAR generates multiple sample paths from the predicted distribution. These Monte Carlo samples can be used to compute empirical quantiles for safety stock optimization or to simulate complex downstream supply chain scenarios.

Compute any quantile (P10, P50, P90) without retraining
Evaluate risk metrics like expected stockout probability
Integrate directly with stochastic inventory optimization systems

Covariate Integration for Contextual Awareness

DeepAR ingests both time-varying covariates (price, promotions, day-of-week) and static covariates (product category, location) to condition forecasts on known future information and item metadata.

Time-varying known features are fed directly to the decoder
Static features are embedded and concatenated with the LSTM input
Enables causal forecasting by incorporating planned marketing activities

DEEP LEARNING FOR PROBABILISTIC FORECASTING

Frequently Asked Questions

DeepAR is an autoregressive recurrent neural network architecture developed by Amazon Research for producing accurate probabilistic forecasts across related time series. Below are common questions about its mechanisms, advantages, and implementation.

DeepAR is a supervised deep learning model based on an autoregressive recurrent neural network (RNN) that produces probabilistic forecasts by learning a parametric distribution over future time steps. Unlike classical models that forecast a single point, DeepAR outputs the parameters of a user-specified likelihood function—typically a Gaussian for real-valued data or a negative binomial for count data—at each prediction step. The architecture uses an LSTM encoder to summarize the historical context of a time series into a fixed-length hidden state, then an LSTM decoder recursively generates predictions. At each step, the model's own previous sample or the ground-truth value is fed back as input, making it autoregressive. Crucially, DeepAR is trained on a portfolio of related time series simultaneously, learning a global model that shares statistical strength across items. This allows it to generate forecasts for items with little or no history by leveraging patterns from similar series, directly addressing the cold-start problem that plagues traditional univariate methods.

PROBABILISTIC DEEP LEARNING VS. CLASSICAL STATISTICS

DeepAR vs. Traditional Forecasting Models

A feature-level comparison of Amazon's DeepAR algorithm against ARIMA and Exponential Smoothing for demand forecasting at scale.

Feature	DeepAR	ARIMA	Exponential Smoothing
Model Family	Autoregressive RNN (Deep Learning)	Classical Statistical Model	Classical Statistical Model
Output Type	Probabilistic (full distribution)	Point forecast	Point forecast
Handles Cold-Start Items
Learns Across Multiple Time Series
Native Support for External Regressors
Handles Intermittent Demand
Requires Stationarity
Scalability (Thousands of Series)	High (shared parameters)	Low (one model per series)	Low (one model per series)

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

DEEP LEARNING FORECASTING

Related Terms

Explore the core concepts and advanced architectures that define DeepAR's probabilistic approach to time series forecasting.

Probabilistic Forecasting

Unlike point forecasts that output a single number, DeepAR generates a full probability distribution (e.g., a negative binomial or Gaussian) for each future time step. This allows supply chain systems to quantify uncertainty and set dynamic safety stock levels based on specific quantiles (P10, P50, P90) rather than relying on a single, fragile estimate.

P10-P90

Typical Prediction Interval

Autoregressive Recurrent Neural Network

DeepAR is built on an LSTM (Long Short-Term Memory) encoder-decoder architecture. The model is autoregressive, meaning its prediction for time step t is fed as an input for time step t+1. This recurrent loop allows the network to learn complex, non-linear seasonal patterns and long-range dependencies directly from hundreds of related time series without manual feature engineering.

LSTM

Core Architecture

Cold-Start Item Handling

A key advantage of DeepAR is its ability to generate forecasts for items with zero or sparse sales history. By learning a global model from similar items and conditioning on static categorical features (e.g., brand, color, size), the model can infer a reasonable demand distribution for a newly launched SKU. This is critical for retail and spare parts logistics where historical data is absent.

Zero History

Minimum Data Requirement

Negative Binomial Likelihood

For count data like product demand, DeepAR often uses a negative binomial distribution as its output likelihood. This distribution is specifically suited for over-dispersed integer data, where the variance exceeds the mean. It naturally handles intermittent demand patterns with many zero values, avoiding the bias introduced by Gaussian assumptions in classical models like ARIMA.

Count Data

Optimal Target Type

Cross-Series Learning

Instead of fitting an isolated model per SKU, DeepAR trains a single global model across thousands of related time series. This shared learning approach allows the model to borrow statistical strength from high-volume items to improve forecasts for low-volume items. The network learns typical seasonal shapes and trend patterns that generalize across the entire product catalog.

1 Model

For Thousands of Series

Temporal Fusion Transformer (TFT)

An evolution of the DeepAR paradigm, the Temporal Fusion Transformer adds interpretable attention mechanisms to the probabilistic forecasting framework. While DeepAR uses recurrent networks, TFT employs multi-head self-attention to capture long-range dependencies and uses variable selection networks to explicitly identify which external regressors (e.g., price, promotions) are driving the forecast at each time step.

Attention

Interpretability Mechanism

EXPLORE

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.