Amazon Forecast

Core components

Amazon Forecast is built on four main elements:

• Datasets: These are the data collections we supply to Amazon Forecast for training and making predictions. There are three dataset types: historical data, which includes the time-series information we wish to predict; related data, which encompasses additional details that might influence the predictions, like product characteristics, holidays, or sales events; and item data, which holds metadata about the items we’re forecasting, such as product codes, categories, or prices.

• Predictors: These are the machine learning algorithms that Amazon Forecast uses to learn from our data. We have the option to select from a variety of algorithms or let Amazon Forecast pick the most suitable one for our data. We can tailor the algorithm settings, including the number of tests, the prediction period, and the performance metrics.

• Forecasts: These are the projections made by Amazon Forecast using the trained predictors. We can create predictions for individual items or groups of items. We can also define the confidence level and the range of outcomes we want to see in the predictions. These forecasts can be sent to Amazon S3 for storage or viewed directly in the Amazon Forecast console.

• Forecast explanations: These insights provided by Amazon Forecast help us understand the impact of different factors in our datasets on the predicted values. We can see how significant each factor is and whether its effect is positive or negative for each item and time point. We can also compare these insights across different items, moments, or predictors.

How Amazon Forecast works

To use Amazon Forecast, we need to follow these steps:

• Prepare data: Gather historical time series and related data that could influence the predictions, like product details, holiday dates, or promotional activities. Decide on the frequency of our forecasts (hourly, daily, weekly, etc.).

• Import data: Upload our data to Amazon S3 and then import it into Amazon Forecast using the console, APIs, or SDKs. We can enhance our model with built-in datasets from Amazon Forecast.

• Train predictor: Choose a forecasting model or let Amazon Forecast automatically find the best fit for our data. Adjust the model settings, such as the number of tests, the prediction period, and the performance metrics.

• Generate forecasts: Use the trained model to predict an item or a group of items. We can store these forecasts in Amazon S3 or view them in the Amazon Forecast console.

• Evaluate forecasts: Compare the predictions with actual results and use performance metrics to judge our model’s performance. Consider the range of potential errors and confidence levels to gauge the reliability of our forecasts. We can assess our predictions through the AWS Management Console, download them as a CSV file, or use the service’s API endpoint.

• Explain forecasts: Utilize the forecast explanation feature to see how different dataset attributes influence our predictions. We can also compare these explanations across various items, moments, or predictors.

Use cases

Amazon Forecast can be used for various use cases, such as:

• Retail and inventory forecasting: Amazon Forecast can help minimize excess stock, enhance inventory turnover, and ensure product availability by predicting demand at specific probability levels. It also aids in refining inventory restocking and distribution strategies by anticipating demand across various locations, sales channels, and product categories.

• Workforce planning: The service can project staffing needs in 15-minute intervals to better align with fluctuating demand periods. This helps enhance customer satisfaction and reduce labor costs by predicting the ideal number of staff needed for each shift, location, or skill set.

• Travel demand forecasting: Amazon Forecast can predict the number of visitors and demand across different channels, aiding in efficiently managing operational expenses. It also supports revenue growth and maximizes occupancy by forecasting the most effective pricing strategies and promotions for each destination, season, or customer segment.

Pricing

Amazon Forecast’s usage-based pricing model has no mandatory fees or initial commitments. The costs are categorized into four types:

• Imported Data: Charges are based on the volume of data imported into Amazon Forecast for training and prediction purposes.

• Training a Predictor: Costs are incurred for the time spent using infrastructure to develop a custom predictor or to monitor its performance.

• Generated Forecast Data Points: Fees are associated with the number of unique forecast values produced across all combinations of time series, items, and dimensions.

• Forecast Explanations: Charges apply to analyzing how different attributes or related data influence the forecasts for each item and time point.

We can use the AWS Pricing Calculator to estimate the cost of our architecture solution using Amazon Forecast. We can also use the AWS Free Tier to get started with Amazon Forecast for free for the first two months.