Let’s start with the very first building block: what makes data a “time series.”

What is a time series?

A time series is simply a sequence of data points recorded over time — like your store’s daily sales for the past three years, or the hourly temperature readings from a weather station. What makes forecasting possible is that many time series contain patterns: things that repeat weekly, monthly, or seasonally, and trends that gradually rise or fall over time.

What was Amazon Forecast?

Amazon Forecast was a fully managed service that used statistical and machine learning algorithms to generate accurate time-series predictions, based on the same forecasting technology originally built for Amazon.com’s own retail demand planning. It required no machine learning expertise — you provided historical data, and Forecast handled selecting and training an appropriate model automatically.

To appreciate what Forecast offered, picture predicting the future without it.

Before managed forecasting services existed, businesses either relied on simple spreadsheet-based averages (which ignore complex seasonal patterns), or hired data scientists to build custom statistical or machine learning models from scratch — a process requiring significant expertise in choosing the right algorithm, tuning it correctly, and validating its accuracy before trusting it with real business decisions.

Forecast addressed this by automatically testing multiple different forecasting algorithms against your specific data and selecting whichever performed best for your particular patterns — removing the need for a business to have deep in-house forecasting expertise to get accurate, tailored predictions.

A small, precise vocabulary makes everything else about time-series forecasting click into place.

Let’s see how raw historical data became an actual future prediction.

flowchart TD
    A[Historical Data in S3 - CSV Files] --> B[Dataset Group]
    B --> C[AutoPredictor Training]
    C --> D[Trained Predictor]
    D --> E[Generate Forecast]
    E --> F[Forecast Results Exported to S3]
    F --> G[Business Application Uses Predictions]
        

You began by uploading historical data — typically as CSV files stored in Amazon S3 — organized into a dataset group containing your target time series and any related time series or item metadata. Forecast’s AutoPredictor feature then automatically trained a model tailored to your data. Once training completed, you generated an actual forecast, which was exported back to S3 or queried directly, ready for your applications or dashboards to use.

The three dataset types

This is the part most tutorials skip. Let’s open the hood.

When you trained an AutoPredictor, Forecast did not simply pick one algorithm and hope for the best. Behind the scenes, it automatically tested several different forecasting algorithms — ranging from classical statistical methods to deep learning-based neural network approaches — against a portion of your historical data set aside specifically for evaluation, then selected whichever approach produced the most accurate results for your particular patterns.

Every forecasting project followed a clear, repeatable sequence.

stateDiagram-v2
    [*] --> DATA_IMPORT
    DATA_IMPORT --> TRAINING
    TRAINING --> BACKTESTING
    BACKTESTING --> FORECAST_GENERATION
    FORECAST_GENERATION --> EXPORT
    EXPORT --> MONITORING
    MONITORING --> DATA_IMPORT
        

Notice this lifecycle loops back to data import. As new actual results came in — this week’s real sales, for instance — they were fed back into the dataset, and predictors were periodically retrained to reflect the most current patterns. A forecast trained on data from two years ago, never refreshed, would gradually become less reliable as real-world conditions changed.

A key idea beginners often miss: good forecasts express uncertainty, not false confidence.

How did Forecast handle businesses needing predictions for thousands of products at once?

Forecast was designed to train and predict across many individual time series simultaneously — for example, generating a separate forecast for every product in a retailer’s entire catalog within a single training job, rather than requiring a separate manual model for each item.

As a fully managed AWS service, reliability was largely handled for you.

Amazon Forecast operated on AWS-managed infrastructure with built-in redundancy, meaning you didn’t need to configure servers or failover mechanisms yourself. Your responsibility instead focused on data quality and process reliability — ensuring your source data pipeline consistently delivered accurate, timely historical data for retraining, since a forecasting service is only as reliable as the data feeding it.

Business forecasting data is often commercially sensitive, so access control mattered throughout.

Trusting a forecast for real business decisions requires understanding how accurate it actually is.

Forecast provided built-in accuracy metrics from its backtesting process, such as weighted quantile loss, letting you quantitatively assess how well a predictor performed before trusting it in production. Amazon CloudWatch could also track operational metrics like dataset import status and training job completion.

Forecast rarely operated alone — it typically fed into broader business systems.

For teams starting new forecasting projects today, AWS points toward Amazon SageMaker Canvas, which carries forward similar automatic algorithm selection concepts within the broader SageMaker ecosystem, alongside SageMaker AI’s AutoML capabilities for more customized pipelines.

Existing customers can continue using it as normal, but it is no longer available to new customers as of July 2024, with AWS recommending Amazon SageMaker Canvas for new projects.

While it could produce some prediction with limited data, having at least several full seasonal cycles of clean historical data generally led to noticeably better accuracy.

It refers to generating reasonable predictions for brand-new items with little or no history, by learning from the patterns of similar items sharing comparable metadata attributes.

A single number hides uncertainty and can lead to overconfident decisions, while a range with associated probabilities better reflects real-world unpredictability and supports more resilient planning.

AWS points new customers toward Amazon SageMaker Canvas for a similar no-code experience, or directly to SageMaker AI’s AutoML capabilities for more customizable pipelines.

Amazon Forecast brought automated, machine learning-powered time-series prediction to businesses without requiring deep in-house forecasting expertise, by testing multiple algorithms against historical data and selecting whichever performed most accurately. While it is now closed to new customers in favor of Amazon SageMaker Canvas, the underlying concepts — target and related time series, item metadata, probabilistic quantile forecasts, and the importance of ongoing accuracy monitoring — remain foundational to understanding time-series forecasting on AWS today, wherever you ultimately build it.