Please use:
- Amazon Forecast, for a low-code/no-code fully managed time series AI/ML forecasting service.
- AutoGluon, if you prefer more control over the forecasting model exploration, training, and evaluation processes.
Amazon Forecast Accelerator (AFA) is an open-source application that:
- Enables users to run, test, and validate forecast accuracy in minutes rather than weeks,
- Performs model selection across 75+ statistical forecasting and machine learning techniques, and
- Exports forecasts and accuracy results as CSV files for benchmarking against existing forecast solutions.
see cdk/README.md
- Create/Login to AWS Account (a new AWS account is recommended for simplicity and testing purposes)
- Note: For AWS Employees using internal AWS accounts - a new internal account is required.
- Click on a "Launch Stack" button corresponding to your nearest AWS Region below:
| Region name | Region code | Launch |
|---|---|---|
| US East (Ohio) | us-east-2 |  |
| US West (Oregon) | us-west-2 | |
| Asia Pacific (Tokyo) | ap-northeast-1 | |
| Asia Pacific (Sydney) | ap-southeast-2 | |
| Asia Pacific (Singapore) | ap-southeast-1 | |
| EU (Frankfurt) | eu-central-1 | |
| EU (Ireland) | eu-west-1 | |
OR
Use the supplied CFN Template from see cfn/cfntemplate.yaml
-
Enter your e-mail address in the "Parameters" section of the form, as shown below:
-
Acknowledge and Accept the cloudformation deployment, and click "Create stack" (which will begin deployment) as shown below:
-
During the deployment, you will recieve an e-mail:
- a subscription confirmation email with the subject heading "AWS Notification - Subscription Confirmation" from
AWS Notifications <no-reply@sns.amazonaws.com>, by clicking "Confirm subscription" in the message body, you will recieve e-mails notifying when the AFA dashboard is deployed and when ML forecasting jobs are complete:
This must be accepted prior to the deployment completing, therefore we advise that you click "Confirm subscription" as soon as you receive the e-mail.
Otherwise, if the deployment completes before confirming the subscription, you will not receive notifications and will need to monitor the deployment progress and access the application via the AWS Console, as follows:- Enter "Cloudformation" in the AWS Console search bar and select "CloudFormation" from the results list:
- Deployment is complete when the four stacks below reach a "CREATE_COMPLETE" status:
- Once the deployment is complete, navigate to the Amazon SageMaker console via the AWS Console search bar:
- Select "Open JupyterLab" in the list of Notebook instances:
- Open "Landing_Page.ipynb" in the file list on the left and if prompted with a "Select Kernel" window, click the "Select" button. This will bring you to the AFA landing page, which contains instructions on getting started with your forecasting.
- a subscription confirmation email with the subject heading "AWS Notification - Subscription Confirmation" from
-
The deployment will complete in 15-20mins, click the URL in the notification e-mail, which will bring you to the AFA landing page (if you see a number of tiled icons, Open "Landing_Page.ipynb" in the file list on the left and if prompted with a "Select Kernel" window, click the "Select" button).
-
The Landing Page contains instructions on how to use the Amazon Forecast Accelerator application to generate forecasts and validate their performance.
By default, Amazon Forecast Accelerator can process datasets of up to 5,000 timeseries (1 timeseries = unique SKU x unique Channel) and uses default AWS service limits for EC2 and Lambda. Refer to the table below for resource requirements based on # time-series in your dataset. A limit increase will be required for larger data sets.
| # Timeseries | SageMaker Notebbok Instance Type | # Concurrent Lambdas | Est. Run-time | Est. Cost per Forecast ($USD) w/ AWS Free-Tier | Est. Cost per Forecast ($USD) w/o AWS Free-Tier |
|---|---|---|---|---|---|
| 1–5,000 | ml.t2.medium (default) | 1,000 (default) | 1–5 mins (default) | <$0.10 | <$0.30 |
| 10,000* | 10s–1 min | <$0.10 | <$0.30 | ||
| 5,000–10,000 | ml.t3.xlarge* | 1,000 (default) | 5–15min (default) | <$0.10 | $0.30–$1.70 |
| 10,000* | 30s–1.5 min | <$0.10 | $0.30–$1.70 | ||
| 10,000–50,000 | ml.t3.2xlarge* | 1,000 (default) | 15–45min (default) | <$0.10–$2.00 | $1.70–$9.00 |
| 10,000* | 30s–1.5 min | <$0.10–$2.00 | $1.70–$9.00 | ||
| 50,000–100,000 | ml.m4.4xlarge* | 1,000 (default) | 45+ min (default) | $2.00–$10.00+ | $9.00–$16.80+ |
| 10,000* | 5+ min | $2.00–$10.00+ | $9.00–$16.80+ |
A limit increase request is required to process larger datasets, which can be made in one of two ways:
- Self-Service (~24-48hr):
- Contact your AWS Account Manager (instant approval for SageMaker Notebook Instance type limit increass only)
These estimates are for the statistical forecasting models only and were based on datasets with three years of historical (weekly) demand for each time-series. The machine learning model run-time and costs are defined by the Amazon Forecast service and take longer to train (typically hours). Please refer to the Amazon Forecast pricing example for expected costs.
The frequency of the data (e.g. daily, weekly, monthly) significantly impacts the run-time. Datasets containing monthly demand will yield the fastest run-times and can typically run using smaller SageMaker Notebook Instance types when compared to weekly or daily demand data with the same number of time-series.