AI/ML in Sports

One of the focuses at Stats Perform was to introduce as much automation and - where feasible and valuable - proper machine learning into Stats Perform products. In addition to managing products, I also had responsibility to organize the delivery of machine learning driven predictions across the Stats Perform portfolio to meet fan engagement use cases of clients in the search, tech, media, broadcast, and sports betting/fantasy spaces.

I became experienced in integrating machine learning systems into customer facing products, including the interactions between modeling efforts and the end product such as; scoping human-in-the-loop tooling, defining data feed/API outputs, versioning outputs, and other model specifications to meet business requirements.

This meant building a process whereby I would:

Meet with each of the other product managers at the company on a regular cadence to understanding their roadmaps
Partner with UX Design and Research teams to mock up potential use cases and share them with clients and users
Share the vision and desired outcomes with the various data science and data engineering teams
Monitor progress as data engineering created data pipelines and new features for model training and data scientists trained models
Work with data delivery teams to prepare delivery via API so that individual engineering teams could consume and build features using the outputs

Several projects, experiments, and outcomes are described below:

Guided development of a new B2B service providing machine learning driven sports predictions and probability distributions, delivered via API, to sportsbook operators, enabling wider coverage of player specific betting markets for multiple US Pro and College sports
- As part of the above delivery, predictions of the number of minutes each NBA player might play in a given game were created as a way to automate further the prediction of each individual statistic. When a player was ruled out for a given game, human-in-the-loop users would indicate via front end tooling that the player was not available, and the minutes projection would scale and adjust based on the updated list of available/unavailable players, and the individual statistic predictions and probabilities would be reproduced accordingly.
US Sport pre-game and in-game win probability modeling for Basketball, Football, Hockey, and Baseball
Season Prediction modeling for Soccer, Football, and Basketball, predicting team ratings/strength and simulating the season to determine where teams would be likely to finish in the standings of their league
Predictions - ongoing throughout a match - of expected stoppage time in Soccer, to aid sportsbook operators in keeping betting windows open for as long as possible
Value/interest ranking of automatically generated text insights, based on novelty/rarity and usefulness to clients