Amazon Forecast is a fully managed service that automatically uses machine learning to deliver accurate forecast for any time series data sets. A time series data set is a set of data points that are ordered by a unit of time. For example, monthly sales of a product, daily inventory in a warehouse, hourly Internet of Things, sensor readings, or even weekly website traffic. Time series forecasting is a method to predict the future data points in the series based on its historical trend.

Forecasting is important for business results because under-forecasting errors result in missed opportunities. And over-forecasting errors result in wasted resources for your business. This is applicable to multiple use cases like Amazon EC2 instance capacity planning where over-forecasting results in unused infrastructure and under-forecasting results in unmet demand. For the business of demand and inventory planning, over-forecasting results in excess inventory and under-forecasting results in loss sales.

Finally, for workforce planning, over-forecasting results in unused labor and under-forecasting results in overtime costs. Each of these use cases represent a data set domain. A data set domain indicates a predefined data set schema for a common use case. Amazon Forecast includes predefined data set domains for retail, inventory planning, easy to capacity planning, workforce prediction, web traffic forecast, metrics like sales and revenue and a custom category for all other types of time series forecasting.

Amazon forecast delivers the accuracy of machine learning forecasting technology powering amazon.com yet requires no machine learning experience. It draws from 20 years of forecasting at Amazon and packages a suite that include deep learning algorithms and statistical methods. The deep learning algorithms improve accuracy by up to 50% compared to traditional models. You can receive completion time estimates for ongoing data set imports, predicted training jobs and forecast jobs.

Amazon Forecast data import can use three types of datasets. They are the first, the Target Time Series. Which represents the historic time series data of items to forecast. The second, the Related Time Series, which represents related time series data such as price or web hits. Finally, the Item Metadata. Which represents attributes of the items such as category, genre or brand. The time series is the only data set that is required to generate forecast. As simple time series is a collection of item IDs, timestamps and values. Where each item ID and timestamp payer has an associated value. Time series can have multiple timestamps and values for each item ID.

A more complex time series may have additional dimensions where you just don't have a unique value for each timestamp X item ID combination. But rather a value for each timestamp X item and perhaps an additional dimension combination. In this example of a retail store, a dimension could be the location of a store. While the item IDs remain unique, the timestamps and the values. Amazon Forecast automatically sets up a data pipeline, training and prediction. Making it simple to use following a general three-step sequence. First, you import your data because data sets are required to train predictors, which are used to generate forecasts. Then you train the predictor, which is a custom model with underlying infrastructure that Amazon forecast trains with your datasets. Finally, you generate forecasts by deploying your train predictors and exploring forecast results.