Amazon Forecast: A Guide for Existing Users and Migration Options

What Is Amazon Forecast?

Every business runs on predictions: how many units will sell next quarter, how many staff to schedule next week, how much server capacity to provision tomorrow. Yet many teams still forecast with spreadsheets and basic statistics that struggle with seasonality, irregular trends and the links between price, weather, promotions and the economy. Amazon Forecast uses machine learning to find those patterns and relationships automatically.

Amazon Forecast is a fully managed AWS service for time-series forecasting with machine learning. It is based on forecasting technology Amazon.com has used for demand planning, inventory and capacity for many years. Forecast analyzes your historical time-series data together with related variables, and AWS says it can produce forecasts up to 50% more accurate than traditional non-ML tools.

Amazon Forecast needs no machine learning expertise. You supply historical time-series data (sales, traffic or resource use) plus optional related data (prices, promotions, weather and holidays). The service then handles data cleaning, feature engineering, algorithm selection, tuning and training.

AutoPredictor at a Glance

The AutoPredictor feature picks the best combination of algorithms for each item in your dataset, so you get an ensemble tuned to each time series instead of one model for everything. That item-level tuning sets Forecast apart from tools that apply one algorithm to all data. For example, a product with strong seasonality might suit Prophet, while one with complex cross-item links might suit DeepAR+, and AutoPredictor makes that choice for every item.

Accuracy and Scale

Most forecasting tools give a single point forecast. Amazon Forecast, by contrast, gives probabilistic forecasts at several quantiles: P10, P50 and P90 by default, and any quantile from P1 to P99 on request. So you can pick the forecast that matches your risk tolerance. Use P90 when stockouts cost more than excess inventory, or P10 when tying up capital is the bigger worry. Different teams can also use different quantiles from the same forecast: safety stock from P90, replenishment from P50 and downside financial scenarios from P10.

Amazon Forecast also fits into the wider AWS stack: S3 for data storage, CloudFormation and Step Functions for automated workflows, QuickSight for visualization and Lambda for event-driven use of forecasts. So you can build complete forecasting pipelines on AWS, from raw data to automated delivery into ERP, inventory, workforce scheduling and financial planning systems.

How Amazon Forecast Works

Amazon Forecast follows a four-stage workflow: import data into datasets, train a predictor (the model), generate forecasts and use them through the API or export them for analysis. The service hides the ML complexity, so business analysts and developers can use it without data science skills.

Data Architecture

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

• Target time series (required): 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): 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): 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.

Amazon Forecast supports data frequencies from one minute to one year. It also handles missing values, gaps, irregular timestamps and time zones during preprocessing, which saves building your own cleaning pipelines. Data is imported from S3, and the service checks schema, data quality and time consistency before training starts.

AutoPredictor and Algorithm Selection

AutoPredictor is the default and recommended way to train Amazon Forecast models. Legacy predictors needed you to choose an algorithm, whereas AutoPredictor evaluates several algorithms and selects the best ensemble for your data:

• CNN-QR (Convolutional Neural Network, Quantile Regression): A deep learning algorithm that learns complex patterns across many related time series. It suits datasets with hundreds or thousands of items that share patterns.
• DeepAR+ (Deep Autoregressive): A recurrent neural network that trains one model across all time series, learning shared patterns while adapting to each item. It works well when items share seasonal and trend dynamics.
• Prophet: An additive regression model from Meta that handles seasonality, holidays and trend changes well. It suits business time series with strong seasonal patterns, holiday effects and occasional missing data.
• NPTS (Non-Parametric Time Series): A scalable baseline that forecasts by sampling past observations. It is useful for sparse, intermittent or lumpy demand, which is common in B2B and specialty retail.
• ARIMA and ETS: Traditional statistical methods (AutoRegressive Integrated Moving Average and Exponential Smoothing) that give strong baselines for simple, well-behaved series with clear trends and seasonality.

AutoPredictor does not just pick one winning algorithm. Instead, it builds an ensemble that weights each algorithm by how well it does for each item. That usually beats any single algorithm and removes weeks of manual experiments and tuning from each forecasting project.

Core Amazon Forecast Features

Beyond AutoPredictor, several features make Amazon Forecast strong in production. They address what makes enterprise forecasting hard: weather effects on demand, promotion effects on sales, model drift as conditions change, and the need to explain ML predictions to business people who must trust them enough to commit inventory, staff and budget:

Monitoring and Cold Start

Amazon Forecast tracks predictor accuracy over time as new data arrives. You can measure drift, meaning how far performance has moved from its starting quality, and decide whether to keep the current model, retrain it or rebuild it. That matters because customer behavior, competition and the economy keep changing.

The monitoring view shows accuracy metrics over time, including weighted quantile loss, WAPE (Weighted Absolute Percentage Error), MAPE (Mean Absolute Percentage Error) and RMSE (Root Mean Squared Error). So you can spot decline early and retrain before forecast quality hurts decisions. Automated retraining triggers based on accuracy thresholds keep predictions reliable without constant manual checks.

Amazon Forecast also handles cold start for new items with no history, using item metadata and patterns from similar items. When you add a product, the model uses its category, price range and other attributes to forecast from comparable items, so you get a reasonable estimate from day one. That helps retailers with frequent seasonal launches, media companies publishing daily and any business with a fast-changing catalog.

Talk to an ExpertRunning Amazon Forecast today? Plan your move to SageMaker Canvas with us.

Amazon Forecast Pricing Model

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 the Costs

• Data ingestion: Charged per GB of data imported for training and forecasting, including target and related time series uploaded from S3.
• Predictor training: Charged per training hour based on the instances used. Training covers data cleaning, parallel algorithm evaluation, ensemble optimization, accuracy metrics and explainability. Because Forecast runs several instances in parallel, billed hours can exceed clock time.

Forecast Data Points and Explainability Costs

• Generated forecast data points: Charged per forecast data point. A data point combines one time series (item and dimensions), one quantile and one time step in the forecast horizon. This is often the largest cost for high-frequency, large-catalog forecasting: a retailer forecasting daily demand for 10,000 SKUs across 500 stores at 3 quantiles generates a very large number of data points per run.
• Forecast explanations: Charged per data point explained, based on the number of attributes analyzed. This cost applies only when you use explainability.

Amazon Forecast Security and Compliance

Forecast processes business-critical data, such as sales figures, financial metrics, operational data and demand patterns, so security matters in every enterprise deployment.

Data stored and processed by Amazon Forecast is encrypted at rest with AWS KMS, including customer-managed keys, so you control the encryption keys. API traffic is encrypted in transit with TLS 1.2+. IAM policies also give fine-grained control over who can create dataset groups, train predictors, generate forecasts and read results.

Amazon Forecast runs within your AWS account and chosen region, so your data stays in that region, which helps with residency rules. CloudTrail logs every API call, recording who trained which predictor, when forecasts were generated and which data was accessed. CloudWatch also monitors training jobs, data imports and forecast pipelines. As a result, Forecast suits regulated sectors such as retail, financial services, healthcare and energy, where forecasting data can be commercially sensitive.

What’s New in Amazon Forecast

Amazon Forecast has matured since launch. The biggest updates focused on automation, accuracy and, more recently, the move to newer forecasting approaches:

So Amazon Forecast still works for existing customers, but new AWS forecasting work has moved elsewhere. Amazon’s Chronos foundation models, pre-trained time-series models available through SageMaker, are one direction, because they can forecast from raw data without training. For existing Forecast customers with working pipelines, validated accuracy and trained business teams, however, there is no need to rush. Plan the migration: start new use cases on SageMaker Canvas or SageMaker, and keep existing Forecast pipelines running while you test the alternatives.

Real-World Amazon Forecast Use Cases

Because it combines several data sources, probabilistic forecasts at any quantile and explainability, Amazon Forecast serves industries where accurate predictions drive profit, customer satisfaction and efficiency. Successful deployments usually start from spreadsheet forecasting that was costing money through over-forecasting (excess stock, overstaffing) or under-forecasting (stockouts, understaffing, capacity gaps). Below are the use cases we implement most often:

Most Common Implementations

Amazon Forecast vs Azure Time Series Forecasting

If you are comparing time-series forecasting services across clouds, Amazon Forecast and Azure’s forecasting take different approaches. Amazon Forecast is a standalone service built for analysts with no ML expertise. Azure ML AutoML forecasting, by contrast, is a feature of the wider Azure Machine Learning platform, with more flexibility but more platform knowledge needed. Here is how they compare:

Choosing the Right Forecasting Alternative

For new forecasting projects, the choice depends on your cloud. Amazon Forecast no longer takes new customers, so AWS-based teams should look at Amazon SageMaker Canvas, which AWS recommends, or at AutoGluon-TimeSeries on SageMaker and Chronos foundation models for forecasting without training. Azure-based teams can use Azure ML AutoML forecasting, which offers similar algorithm selection and ensembles.

For existing Amazon Forecast customers, there is no urgent reason to migrate unless you need features Forecast lacks. The service still runs, is still supported and still delivers proven accuracy for established workloads. When you do move, SageMaker Canvas is AWS’s recommended path, and AWS provides a workshop and a notebook that convert Forecast datasets into the single merged dataset Canvas expects.

Organizations that need multi-cloud or cloud-neutral forecasting can also look at open-source options such as AutoGluon-TimeSeries (which runs locally or on any cloud), Prophet (Meta’s open-source library) or NeuralProphet (a neural network extension of Prophet). These avoid vendor lock-in, but they take much more engineering effort to deploy, maintain, scale and monitor than a fully managed service.

Getting Started with Amazon Forecast

For existing customers, Amazon Forecast is simple to set up. You create a dataset group, import data from S3, train an AutoPredictor and generate forecasts, with no ML expertise or infrastructure to manage.

Preparing Your Data

Before you build your first predictor, check that your target time series meets basic quality rules. Each record needs an item identifier, a timestamp and the target value, at consistent intervals with as few gaps as possible. Include at least a year of history to capture seasonality, and more for items with annual cycles. Clean obvious errors too: negative values where only positive ones make sense, outliers from typing mistakes, duplicate records and missing timestamps. Input data quality is the biggest factor in forecast accuracy, and no algorithm can make up for flawed or incomplete history.

Then prepare related time series that move with your target. For retail demand, that means price history, promotion calendars and holiday flags. For workforce planning, add event calendars and weather. The Weather Index brings in weather data automatically for supported locations, but you must prepare every other related variable yourself, as a CSV with matching item identifiers and timestamps.

Building Your First 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}")

For production, automate the pipeline with CloudFormation and Step Functions: schedule daily data imports, weekly predictor retraining and automated forecast generation with delivery to your business systems. For full setup guidance and code samples, see the Amazon Forecast documentation.

Amazon Forecast Best Practices and Pitfalls

From our experience deploying Forecast in enterprises, the patterns below decide whether a forecasting project produces trusted, usable predictions or forecasts that business teams ignore in favor of their spreadsheets and instinct.

Deployment Recommendations

• Always use AutoPredictor: The legacy CreatePredictor API needs manual algorithm selection and usually gives weaker results. AutoPredictor evaluates all available algorithms and builds ensembles automatically, so manual selection rarely makes sense for production workloads with enough data.
• Invest in related time series data: Accuracy improves a lot when the model can learn from related variables. So include price history, promotion calendars, holiday flags, marketing spend and any other time-varying factor that drives your target. Turn on the Weather Index for any location-dependent forecast.
• Choose the right forecast frequency: Match forecast granularity to your planning cycle: daily for retail inventory, hourly for workforce scheduling and monthly for financial planning. Higher frequencies create many more data points and raise costs. So forecast only at the level your teams will actually use; hourly forecasts for a team that plans weekly waste compute and budget.

Improving Accuracy

• Use probabilistic forecasts deliberately: Pick the quantile that fits your risk tolerance. Use P90 for safety stock where stockouts are expensive, P50 for balanced planning and P10 where excess inventory is the bigger worry. Different decisions in the same organization can also use different quantiles from one forecast: supply chain uses P90 for safety stock, finance uses P50 for revenue and procurement uses P10 for minimum orders.
• Monitor predictor drift continuously: Business conditions change as preferences shift, competitors arrive and the economy moves. So use the built-in monitoring to track accuracy over time, and retrain when metrics fall below your thresholds, instead of finding poor forecasts after they have hurt decisions.

Planning Your Migration

Because Amazon Forecast no longer takes new customers, existing users should plan a medium-term migration. AWS recommends Amazon SageMaker Canvas, which offers no-code time-series forecasting, and it provides a transition workshop and notebook. AutoGluon-TimeSeries on SageMaker is another option for teams that want automated algorithm selection and ensembles with the full SageMaker toolset for custom work.

Amazon’s Chronos foundation models take a different approach. They are pre-trained on large amounts of time-series data and can forecast from raw history without training, which gives instant predictions for new use cases. That zero-shot ability suits prototyping, exploration and cases with little training data. For production workloads with domain-specific data and strict accuracy needs, however, trained models (in Forecast, Canvas or SageMaker) often still perform better.

Our ServiceData Analytics, Business Intelligence and Artificial Intelligence

Frequently Asked Questions About Amazon Forecast

References

1. Closed to New Customers July 29, 2024; Move to SageMaker Canvas: AWS Machine Learning Blog — Transition Your Amazon Forecast Usage to Amazon SageMaker Canvas
2. Service Features and Workflow: AWS — Amazon Forecast Developer Guide
3. Pricing Components: AWS — Amazon Forecast Pricing