Amazon Forecast Cheat Sheet

A fully-managed time-series forecasting service that utilizes machine learning to generate highly accurate predictions without requiring deep ML expertise.

It helps businesses predict future outcomes such as demand, sales, inventory, staffing, or resource usage by analyzing historical time-series data and related variables.

Features

• Fully managed service that handles data preprocessing, model training, tuning, and hosting for you.
• Uses machine learning or ML models that are purpose-built for time-series forecasting.
• Automatically detects seasonality, trends, and holidays.
• Supports probabilistic forecasts (P10, P50, P90) instead of single-point predictions.
• Scales to millions of time series with no infrastructure to manage.
• Supports automatic model selection and hyperparameter tuning via AutoML.
• Integrates seamlessly with S3, IAM, CloudTrail, and AWS KMS.
• Supports custom predictors for advanced use cases.

Key Concepts

Dataset
A container for your time-series data. Amazon Forecast supports:

• Target Time Series: The main variable you want to predict (e.g., product demand).
• Related Time Series: Additional time-dependent data that influences predictions (e.g., price, promotions).
• Item Metadata: Static attributes for items (e.g., category, brand, region).

Dataset Group
A logical grouping of related datasets (target, related, and metadata) used together for training.

Predictor
A trained forecasting model created from a dataset group. It handles training, tuning, and evaluation.

AutoPredictor (AutoML)
Automatically selects the best algorithm and hyperparameters based on your data.

Forecast
The generated prediction output for future time periods, including multiple quantities.

Forecast Export
Exports forecast results to S3 for downstream consumption.

High-Level Architecture Diagram

Amazon Forecast uses historical time-series data stored in Amazon S3 to train fully managed predictors that generate probabilistic forecasts, which can be exported and consumed by business applications

Common Use Cases

You can use Amazon Forecast for:

• Retail demands and sales forecasting
• Inventory and supply chain planning
• Workforce and staffing forecasts
• Energy consumption and load prediction
• Financial and revenue forecasting
• Capacity planning for infrastructure or services

Best Practices

• You should provide at least one year of historical data to capture seasonality.
• Use related time series to improve forecast accuracy.
• Start with AutoML before creating custom predictors.
• Retrain predictors periodically as new data becomes available.
• Validate accuracy using backtesting metrics before production use.
• Export forecasts to S3 for integration with dashboards and applications.

Pricing

You are charged based on dataset storage, model training, and forecast generation.

Depending on dataset size and predictor complexity, pricing may vary.

There are no upfront costs or infrastructure charges. You only pay for what you use.

Note that exported data stored in S3 incurs standard S3 charges.

Security

• Use IAM roles to control access to datasets, predictors, and forecasts.
• Supports fine-grained permissions for Amazon Forecast resources.
• All API activity is logged via AWS CloudTrail.
• Data is encrypted at rest and in transit.
• Supports AWS KMS for encryption key management.
• Integrates with S3 bucket policies for secure data storage.

References

• https://aws.amazon.com/forecast/
• https://docs.aws.amazon.com/forecast/
• https://docs.aws.amazon.com/forecast/latest/dg/how-it-works.html