Building the Aqueduct March Madness Workflow

Last week, we shared our Aqueduct pipeline to predict March Madness. You can see our live bracket here. So far, 30% of our predictions have been correct. 😳

In this post, we’ll walk through how we set up the Aqueduct pipeline to make these predictions — you can play around with the Aqueduct workflow in a GitHub Codespace here.

---

As we mentioned last week, Kaggle provides an excellent dataset with a large history of NCAA tournament games.

First things first, we needed a data source. We set up a Postgres server and loaded the Kaggle datasets into it. The Kaggle data is broken down into a number of smaller datasets, including regular season data, tournament performance data, and tournament seeding, and so on.

With the data in Postgres, we broke down our pipeline into a few typical steps: data cleaning, train-test split, model training, model validation, and ultimately inference. The pipeline publishes predictions to an S3 bucket. The final workflow in Aqueduct looks like this:

As you can see, we load data from a few different tables, rank the teams, compile the relevant regular season data, and train a random forest model. The model we used was adapted from the winning submission from last year’s Kaggle competition, which you can find here. Once the model was built, putting together the Aqueduct workflow took us less than an hour.

With the data and pipeline ready to go, we were able to build a simple bracket visualization using react-tournament-brackets and hook up our predictions. You can see the live predictions on our website.

Of course, as we said above, the predictions we’ve generated have been ~30% accurate so far this year. In other words, a coin flip would be more accurate than our model! 😬 This is the peril of relying on historical data going back 40 years. As the saying goes, past performance is not an indicator of future returns. For example, Alabama is a #1 seed in this year’s tournament but hasn’t been a historically strong basketball school and was actually predicted to lose their first round matchup. (Of course, there’s also Purdue…)

---

Once we had a workflow running, we wanted to make it reusable to let you all play around with it. GitHub Codespaces makes it easy to create sharable environments. We created a GitHub Codespace that you can clone here.

This Codespace runs a Dockerfile with an Aqueduct environment pre-configured and with the Aqueduct March Madness profile deployed. The Aqueduct server running in this Codespace has a connection to the Postgres server with the Kaggle data described above and the Aqueduct workflow pre-described.

When you create the Codespace, you should see the notebook that we used to create the Aqueduct workflow. You will also see a pop-up in the bottom right that there’s a service running on port 8080, which will have the Aqueduct UI and the deployed workflow.

Let us know what you think! If you’re interested in learning more, check out the open-source project or join our Slack community.