Amazon Forecast: The Complete Guide to AWS Time-Series Forecasting

What Is Amazon Forecast?

Undeniably, every business depends on predictions — how many units will sell next quarter, how many staff to schedule next week, how much server capacity to provision tomorrow. However, traditional forecasting relies on spreadsheets and basic statistical models that struggle with seasonality, irregular trends, and the complex relationships between variables like price, weather, promotions, and economic conditions. Amazon Forecast replaces this guesswork with machine learning that automatically discovers the hidden patterns, complex correlations, and non-linear relationships that traditional statistical methods simply cannot capture.

Amazon Forecast is a fully managed time-series forecasting service from Amazon Web Services powered by machine learning. Essentially, built on the same forecasting technology that Amazon.com has developed over two decades for demand planning, inventory management, and capacity forecasting, Amazon Forecast analyzes your historical time-series data and combines it with related variables to generate predictions that are up to 50% more accurate than traditional non-ML forecasting tools.

Importantly, Amazon Forecast requires no machine learning expertise. Simply provide historical time-series data (sales figures, traffic counts, resource utilization) along with optional related data (prices, promotions, weather, holidays), and the service automatically handles data cleaning, feature engineering, algorithm selection, hyperparameter tuning, and model training.

Amazon Forecast AutoPredictor

The AutoPredictor feature selects the best combination of algorithms for each individual item in your dataset — not just one model for all items, but an optimized ensemble tailored to each time series. This item-level optimization is what sets Forecast apart from generic forecasting tools that apply a single algorithm uniformly across all data — a product with strong seasonal patterns might be best served by Prophet, while a product with complex cross-item dependencies might benefit from DeepAR+, and AutoPredictor automatically makes this selection for every item independently.

Amazon Forecast Accuracy and Scale

Moreover, unlike most forecasting tools that generate only point forecasts (a single predicted value), Amazon Forecast generates probabilistic forecasts at multiple quantiles — P10, P50, and P90 by default, with support for any quantile between P1 and P99. This allows you to choose the forecast that matches your business risk tolerance: use P90 when the cost of under-forecasting (stockouts) exceeds the cost of over-forecasting (excess inventory), or P10 when minimizing capital commitment and avoiding excess inventory is the priority. Different quantiles from the same forecast can drive different business decisions simultaneously — safety stock calculations use P90, replenishment orders use P50, and financial projections use P10 to model downside scenarios.

Additionally, Amazon Forecast integrates natively with the broader AWS ecosystem — S3 for data storage, CloudFormation and Step Functions for automated end-to-end workflows, QuickSight for forecast visualization, and Lambda for event-driven forecast consumption. Consequently, you can build complete forecasting pipelines entirely within AWS, from raw data ingestion through model training to automated forecast delivery to downstream business systems — ERP platforms, inventory management tools, workforce scheduling applications, and financial planning dashboards.

How Amazon Forecast Works

Fundamentally, Essentially, Amazon Forecast operates through a four-stage workflow: import your data into datasets, train a predictor (model), generate forecasts, and consume those forecasts through the API or export them for downstream analysis. The service handles all ML complexity behind the scenes, making the process accessible to business analysts and developers without data science expertise.

Data Architecture for Amazon Forecast

Every Amazon Forecast deployment starts with a dataset group containing up to three types of datasets:

• Target time series (required): Essentially, your primary historical data — the values you want to forecast. Each record includes an item identifier, a timestamp, and the target value. For retail demand forecasting, this would typically be daily sales units per individual product per store location. For workforce planning, this would be hourly inbound call volume per contact center location.
• Related time series (optional): Additionally, variables that change over time and correlate with your target. Examples include product prices, promotional flags, marketing spend, web traffic, and economic indicators. Related time series can include both historical values and known future values (such as planned promotions, scheduled price changes, and confirmed marketing campaigns), allowing the model to incorporate forward-looking information that dramatically improves forecast accuracy for periods affected by planned business actions.
• Item metadata (optional): Furthermore, static attributes that describe each item — product categories, brands, store locations, warehouse regions, seasonal classifications, or any other categorical attributes that help the model identify shared behavioral patterns and demand characteristics across groups of similar items.

Importantly, Importantly, Amazon Forecast supports multiple data frequencies — from one minute to one year — and automatically handles missing values, data gaps, irregular timestamps, and timezone normalization during the comprehensive data preprocessing phase — saving significant engineering effort compared to building custom data cleaning pipelines. Specifically, data is imported from S3, and the service validates schema compliance, data quality, and temporal consistency before training begins.

AutoPredictor and Algorithm Selection

Currently, the AutoPredictor is the default and recommended approach for training Amazon Forecast models. Unlike legacy predictors that required manual algorithm selection, AutoPredictor automatically evaluates multiple algorithms and selects the best ensemble for your specific data:

• CNN-QR (Convolutional Neural Network – Quantile Regression): Specifically, a deep learning algorithm that excels at learning complex patterns across large numbers of related time series. Best for datasets with hundreds or thousands of items where cross-item patterns exist.
• DeepAR+ (Deep Autoregressive): Similarly, a recurrent neural network that trains a single model across all time series in your dataset, learning global patterns while adapting to individual item behavior. Performs particularly well when items share common seasonal dynamics and trend patterns.
• Prophet: Additionally, an additive regression model developed by Meta that handles seasonality, holidays, and trend changes effectively. Well-suited for business time series with strong seasonal patterns, holiday effects, and occasional missing data points.
• NPTS (Non-Parametric Time Series): Furthermore, a scalable baseline algorithm that generates forecasts by sampling from past observations. Particularly useful for sparse, intermittent, or lumpy demand patterns common in B2B and specialty retail scenarios.
• ARIMA and ETS: Finally, traditional statistical methods (AutoRegressive Integrated Moving Average and Exponential Smoothing) that provide strong baselines and work well for simple, well-behaved time series with clear trends and seasonality.

Consequently, Importantly, AutoPredictor does not simply pick the single best algorithm — it builds an optimized ensemble that combines multiple algorithms, weighting each based on its performance for specific items in your dataset. This ensemble approach consistently outperforms any single algorithm in production benchmarks and completely eliminates the need for costly and time-consuming manual algorithm experimentation and hyperparameter tuning — saving weeks of data science effort per forecasting project.

Core Amazon Forecast Features

Beyond the AutoPredictor engine, several capabilities make Amazon Forecast particularly powerful for production forecasting deployments. These features address the real-world challenges that make forecasting genuinely difficult in enterprise environments — weather impacts on demand, promotional effects on sales patterns, model drift as business conditions change, and the critical need to explain ML-generated predictions to business stakeholders who must trust these numbers enough to make inventory, staffing, and financial commitments based on them:

Monitoring and Cold Start in Amazon Forecast

Amazon Forecast automatically tracks predictor accuracy over time as new data is imported. You can systematically quantify model drift — how much the predictor’s performance has deviated from its initial quality metrics — and make informed decisions about whether to continue using the current model, retrain with updated data, or rebuild from scratch. This monitoring capability is essential for maintaining forecast accuracy in dynamic business environments where consumer behavior, competitive landscapes, and economic conditions change continuously and unpredictably.

Specifically, the monitoring dashboard shows key accuracy metrics over time — weighted quantile loss, WAPE (Weighted Absolute Percentage Error), MAPE (Mean Absolute Percentage Error), and RMSE (Root Mean Squared Error) — allowing you to detect degradation early and trigger retraining before forecast quality impacts business decisions. Setting up automated retraining triggers based on accuracy thresholds is a best practice that ensures your predictions remain reliable without requiring constant manual oversight.

Furthermore, Amazon Forecast handles cold start scenarios for new items with no historical data by leveraging item metadata and patterns from similar items in the dataset. When you add a new product to your catalog, the model uses its category, price range, and other attributes to generate initial forecasts based on the behavior of comparable items — providing reasonable demand estimates from day one rather than waiting months for enough historical data to accumulate. This cold start capability is particularly valuable for retailers who frequently introduce new seasonal products, media companies publishing new content daily, and any business with a rapidly changing catalog where waiting for historical data would make forecasting irrelevant by the time sufficient history exists.

Amazon Forecast Pricing Model

Fundamentally, Amazon Forecast uses pay-as-you-go pricing with four cost components. Rather than listing specific dollar amounts that change over time, here is how the cost structure works:

Understanding Amazon Forecast Costs

• Data ingestion: Essentially, charged per GB of data imported into Amazon Forecast for training and forecasting. This includes both historical target time series and related time series data uploaded from S3.
• Predictor training: Specifically, charged per training hour based on the compute instances used. Training time includes data cleaning, parallel algorithm evaluation, ensemble optimization, accuracy metric calculation, and explainability insight generation. Because Amazon Forecast deploys multiple instances in parallel, billed training hours may exceed actual clock time.

Forecast Data Points and Explainability Costs

• Generated forecast data points: Similarly, charged per forecast data point generated. A data point is the combination of a unique time series (item × dimensions), a quantile, and a time step within the forecast horizon. This is typically the largest cost component for high-frequency, large-catalog forecasting deployments — a retailer forecasting daily demand across 10,000 SKUs in 500 stores at 3 quantiles generates a substantial number of data points per forecast run.
• Forecast explanations: Finally, charged per data point explained, based on the number of attributes being analyzed for impact. This cost applies only when you use the explainability feature.

Amazon Forecast Security and Compliance

Since Forecast processes business-critical data — sales figures, financial metrics, operational data, and customer demand patterns — security is essential for enterprise deployments.

Specifically, all data stored and processed by Amazon Forecast is encrypted at rest using AWS KMS with support for customer-managed CMKs, giving you full control over encryption keys. Furthermore, all API communications are encrypted in transit using TLS 1.2+. Moreover, IAM policies provide fine-grained access control over which users, roles, and applications can create dataset groups, train predictors, generate forecasts, and access forecast results.

Additionally, Amazon Forecast operates within your AWS account and region — your data never leaves the region you select for processing, ensuring compliance with data residency requirements. Furthermore, CloudTrail logs all API calls for audit compliance, providing a complete record of who trained which predictors, when forecasts were generated, and which data was accessed. CloudWatch provides operational monitoring of training jobs, data import status, and forecast generation pipelines. These security features make Forecast suitable for organizations in regulated industries including retail, financial services, healthcare, and energy where forecasting data may contain commercially sensitive intelligence requiring strict access controls, encryption, and comprehensive audit trails.

What’s New in Amazon Forecast

Indeed, Amazon Forecast has matured significantly since its launch, with the most impactful updates focusing on automation, accuracy, and the transition to next-generation forecasting approaches:

Consequently, while Amazon Forecast remains fully operational for existing customers, the strategic direction for new AWS forecasting workloads has shifted. Amazon’s Chronos foundation models — pre-trained time-series models available on SageMaker and Bedrock — represent the next generation of forecasting technology. Notably, Chronos models can generate forecasts from raw time-series data without training, making them faster to deploy for new use cases. However, for existing Forecast customers with established pipelines, validated accuracy metrics, and trained business teams, there is no immediate need to migrate — the service continues to function with full API access, training capabilities, and full AWS technical support. Migration should be planned strategically rather than rushed, prioritizing new use cases for SageMaker while maintaining existing Forecast pipelines until a formal transition timeline becomes clear.

Real-World Amazon Forecast Use Cases

Given its ability to incorporate multiple data sources, generate probabilistic forecasts at any quantile, and provide forecast explainability, Amazon Forecast serves organizations across industries where accurate predictions directly impact profitability, customer satisfaction, and operational efficiency. The common thread across all successful deployments is that traditional spreadsheet-based forecasting was producing unacceptable error rates that cost the business money through either over-forecasting (excess inventory, overstaffing) or under-forecasting (stockouts, understaffing, capacity shortfalls). Below are the use cases we implement most frequently:

Most Common Amazon Forecast Implementations

Amazon Forecast vs Azure Time Series Forecasting

If you are evaluating time-series forecasting services across cloud providers, the comparison between Amazon Forecast and Azure’s forecasting capabilities reveals two fundamentally different architectural approaches. Amazon Forecast is a purpose-built, standalone forecasting service designed for business analysts with no ML expertise. Azure ML AutoML forecasting is a feature within the broader Azure Machine Learning platform that offers more flexibility but requires more platform knowledge. Here is how they compare:

Choosing the Right Forecasting Alternative

Clearly, for new forecasting projects, the choice depends on your cloud ecosystem. Since Amazon Forecast no longer accepts new customers, AWS-native organizations should evaluate Amazon SageMaker with AutoGluon-TimeSeries (which provides similar automated forecasting capabilities within the SageMaker environment) or Chronos foundation models for zero-shot forecasting without training. Azure organizations can leverage Azure ML AutoML forecasting, which provides comparable algorithm selection and ensemble capabilities within the Azure ecosystem.

Furthermore, for existing Amazon Forecast customers, there is currently no compelling reason to migrate unless you need capabilities beyond what Forecast provides. The service remains fully operational, continues to receive support, and delivers proven accuracy for established forecasting workloads. When AWS eventually announces an end-of-life timeline, SageMaker with AutoGluon-TimeSeries will be the natural migration path — it supports the same data formats and similar workflow patterns, ensuring a relatively smooth and low-risk transition for teams that have invested in Forecast-based forecasting pipelines.

Moreover, organizations that require multi-cloud or cloud-agnostic forecasting capabilities should evaluate open-source alternatives like AutoGluon-TimeSeries (which can run locally or on any cloud), Prophet (Meta’s open-source forecasting library), or NeuralProphet (a neural network extension of Prophet). These tools provide production-quality forecasting without vendor lock-in, though they require significantly more engineering effort to deploy, maintain, scale, and monitor in production compared to the fully managed Forecast service where AWS handles all infrastructure, scaling, and operational concerns automatically.

Getting Started with Amazon Forecast

For existing customers, Amazon Forecast provides a straightforward setup experience. You create a dataset group, import data from S3, train an AutoPredictor, and generate forecasts — no ML expertise or infrastructure management required. The free trial provides two months of generous usage to build, validate, and refine your forecasting pipeline before committing to production costs.

Preparing Data for Amazon Forecast

Before building your first predictor, ensure your target time series data meets quality requirements. Each record needs an item identifier, a timestamp, and the target value — with consistent time intervals (no gaps in the series where possible). Include at least one year of historical data to capture seasonal patterns, though longer history generally produces better forecasts for items with annual seasonality. Clean any obvious data errors — negative values where only positives are valid, extreme outliers from data entry mistakes, duplicate records that could skew training, and missing timestamps that create gaps in the series. The quality of your input data is the single most important factor determining forecast accuracy — no algorithm — no matter how sophisticated — can compensate for fundamentally flawed, incomplete, or inaccurate historical data.

Additionally, prepare related time series data that correlates with your target metric. For retail demand, this includes pricing history, promotional calendars, and holiday flags. For workforce planning, include event calendars and weather data. Notably, the Weather Index feature automatically handles weather data ingestion for supported locations, but all other related variables must be prepared and uploaded separately in a correctly structured CSV with matching item identifiers and timestamps.

Building Your First Amazon Forecast Predictor

Below is a minimal Python example that creates a forecast from an existing predictor:

import boto3

# Initialize the Forecast client
client = boto3.client('forecast', region_name='us-east-1')

# Create a forecast from a trained predictor
response = client.create_forecast(
    ForecastName='my-demand-forecast',
    PredictorArn='arn:aws:forecast:us-east-1:123456:predictor/my-predictor',
    ForecastTypes=['0.10', '0.50', '0.90']
)

# Query the forecast for a specific item
query_client = boto3.client('forecastquery', region_name='us-east-1')
result = query_client.query_forecast(
    ForecastArn=response['ForecastArn'],
    Filters={'item_id': 'product-123'}
)

# Print probabilistic forecast values
for quantile, predictions in result['Forecast']['Predictions'].items():
    print(f"Quantile {quantile}:")
    for point in predictions[:5]:
        print(f"  {point['Timestamp']}: {point['Value']:.1f}")

Subsequently, for production deployments, automate the entire pipeline using CloudFormation and Step Functions — schedule daily data imports, weekly predictor retraining, and automated forecast generation with downstream delivery to your business systems. For complete setup guidance and code samples, see the Amazon Forecast documentation.

Amazon Forecast Best Practices and Pitfalls

Based on our experience deploying Forecast across enterprise environments, the following patterns consistently determine whether a forecasting implementation delivers actionable, trustworthy predictions or produces forecasts that business teams ultimately ignore and continue relying on their existing spreadsheet-based approaches and institutional intuition.

Recommendations for Amazon Forecast Deployment

• First, always use AutoPredictor: Specifically, the legacy CreatePredictor API requires manual algorithm selection and produces inferior results. AutoPredictor automatically evaluates all available algorithms and builds optimized ensembles — there is no documented scenario where manual algorithm selection outperforms the automated ensemble approach for production forecasting workloads with sufficient data.
• Additionally, invest in related time series data: Importantly, forecast accuracy improves dramatically when the model can learn from correlated variables. Include price history, promotional calendars, holiday flags, marketing spend, and any other time-varying factors that influence your target metric. The Weather Index should be enabled for any location-dependent forecast.
• Furthermore, choose the right forecast frequency: Specifically, match the forecast granularity to your actual planning cadence. Daily forecasts for retail inventory management, hourly for workforce scheduling, monthly for financial planning. Higher-frequency forecasts generate exponentially more data points and significantly increase costs — forecast only at the granularity your business teams will actually consume and act upon. Generating hourly forecasts when your planning team reviews weekly is a waste of both compute and budget.

Optimizing Amazon Forecast Accuracy

• Moreover, use probabilistic forecasts strategically: Select the quantile that matches your business risk tolerance. Use P90 for safety stock calculations where stockouts are expensive. Select P50 for balanced planning. Choose P10 for scenarios where minimizing excess inventory matters most. Different business decisions within the same organization may warrant different quantiles extracted from the same forecast — supply chain uses P90 for safety stock, finance uses P50 for revenue projections, and procurement uses P10 for minimum order commitments.
• Finally, monitor predictor drift continuously: Inevitably, business environments change — consumer preferences shift, competitors enter markets, economic conditions fluctuate. Consequently, use the built-in monitoring to track prediction accuracy over time and retrain proactively when accuracy metrics degrade beyond your defined acceptable thresholds rather than discovering poor forecasts after they have already impacted business decisions.

Planning for the Future Beyond Amazon Forecast

Since Amazon Forecast is no longer accepting new customers, existing users should develop a migration strategy for the medium term. Currently, the most natural migration path is Amazon SageMaker with AutoGluon-TimeSeries, which provides automated time-series forecasting within the broader SageMaker ML platform. Importantly, AutoGluon-TimeSeries supports similar data formats, automated algorithm selection, and ensemble optimization — with the added benefit of access to the full SageMaker feature set for custom model development when needed.

Additionally, Amazon’s Chronos foundation models represent a fundamentally new approach to time-series forecasting. Essentially, pre-trained on billions of time-series data points, Chronos models can generate forecasts from raw historical data without any training — providing instant predictions for new use cases. Consequently, this zero-shot capability makes Chronos ideal for rapid prototyping, exploratory analysis, and scenarios where training data is limited. However, for production workloads with domain-specific data and high accuracy requirements, trained models (whether via Forecast or SageMaker) typically still outperform foundation model approaches.

Frequently Asked Questions About Amazon Forecast