mboaglio/forecasting-workshop

A workshop on timeseries forecasting with Amazon SageMaker and Amazon Forecast

Timeseries Forecasting on AWS

This demo repository walks through an example time-series forecasting problem, comparing pre-built timeseries forecasting engine Amazon Forecast with other models trained on data science platform Amazon SageMaker. We pay particular attention to modern forecasting procedures DeepAR and Prophet.

The demo is presented as a series of Python notebooks designed to be run in Amazon SageMaker, but with some parts requiring activity in the AWS Console.

Getting Started

Setting Up Your Notebook Instance

First you will need to create a new Notebook Instance, to do that begin by logging into the AWS Console.

Next, ensure you are in a supported region for Amazon Forecast. We test primarily in Asia Pacific (Singapore).

Under Find services in the main body of the page, enter SageMaker, then select it from the drop-down.

To the left, will see a category titled Notebook inside that, click Notebook instances.

Click the orange Create notebook instance button.

Give the instance a name unique in the account you are using. If a shared account, place your name first like FirstNameLastNameForecastDemo. The default Instance type is fine.

The Next component to change is the IAM role. Under the drop-down click Create a new role. Then for S3, select Any S3 Bucket, finally click Create role.

Note that the role itself has become a link. Open that link in a new tab.

Here you will update the policies of your instance to allow it to work with Forecast. Click the Attach policies button. Search and check the box next to the following policies: -IAMFullAccess -AmazonForecastFullAccess

Finally click the Attach policy button on the bottom right corner.

Now click on Trust relationship tab > click on Edit trust relationships button > update the json file with the following: "Service": [ "forecast.amazonaws.com", "sagemaker.amazonaws.com" ]

Next click the Create policy button at the top. In the new page, click the JSON tab.

Erase all of the content that is in the editor and paste the content in IAM_Policy.json.

After pasting, click the Review policy button. Give the policy again a personalize name like FirstNameLastNameForecastIAMPolicy.

For the description, enter in something about it being used to demo Forecast. Finally click Create policy. Close this tab or window.

Once closed you should see the tab for adding permissions to your SageMaker role. Click the Filter Policies link, then select Customer managed. After that, you should see the policy you just created, if the list is long, just paste the name in the search bar to reduce the number of items. If you do not see it still, click the refresh icon in the top right of the page.

After clicking the checkbox next to the policy, click Attach policy at the bottom of the page. Then close this window.

Back at the SageMaker Notebook Instance creation page, expand out the Git repositories section and select to "Clone a public Git repository into this notebook instance only": Provide the URL of this repository.

Now you're ready to click Create notebook instance at the bottom of the page. This process will take 5-10 minutes to complete. Once the status says InService you are ready to continue.

Working Through The Notebooks

To begin with your new notebook instance, click Open Jupyter or Open JupyterLab, this will take you to the default interface for the Notebook Instance.

The forecastworkshop folder should already be created for you: Enter it and start by opening Notebook 1.

If prompted for a kernel, select conda_python3.

From here you will follow the instructions outlined in the notebook.

Read every cell FULLY before executing it!