Reference project for building a DeepAR forecasting model to predict store sales. This will build a Step Function that will launch a training job on SageMaker then update a prediction endpoint which can be used to forecast future store sales.
This library is licensed under the Apache 2.0 License.
After the model has been built and deployed you can test it with a UI tool included in this project.
History (blue color) is loaded from S3
Forecasts (red color) are loaded from SageMaker endpoint
Note: Supported regions are Oregon, Virginia, Dublin.
- Download this CloudFormation template: https://raw.githubusercontent.com/aws-samples/amazon-sagemaker-deepar-retail/master/initialization-template.json
- Execute the CloudFormation template, provide a Stack Name of retail-deepar - be sure to use lowercase letters.
Once the CloudFormation template has finished it will kick off a CodeBuild project that will clone this repository into a private CodeCommit repository in your account. A CodePipeline will also be created which will be triggered when the clone happens. This will build the source code and deploy it with 2 more CloudFormation templates:
- retail-deepar-deployment - this includes all the code that was built, i.e. Lambdas and Step Function.
- retail-deepar-base - this includes the base resources for the project.
After all 3 CloudFormation templates have run you can kick off the Step Function named Retrain. Note: This will look for a file named SageMaker/train.csv inside the bucket that was created from the retail-deepar-base CloudFormation template. Besure to place that file at that location in that folder. You can download that file here: https://raw.githubusercontent.com/aws-samples/amazon-sagemaker-deepar-retail/master/train.csv
After the Retrain Step Function has completed you can test the forecast model you built with this test tool: https://raw.githubusercontent.com/aws-samples/amazon-sagemaker-deepar-retail/master/public/index.html
- Download that file to your desktop, then open it in Chrome browser. (only Chrome is tested, so use that).
- Set the IdentityPool. You can find that in the Resources tab of the retail-deepar-base CloudFormation template in the AWS console.
- Pick a store and item then load the last 90 days of data by clicking the Load History button.
- Click the Load Forecast button to load the forecast for the next 90 days. It will be appended to the end of the table so you scrole to the right to view.
If you want to retrain on data as it's dropped into S3, you can build a trigger that kicks off the Retrain Step Function every time new data is placed in S3.
