Predicting play types in American football with machine learning.

My bachelor thesis dives into using machine learning to predict the types of plays in American football, specifically whether a play will involve passing or running the ball. It goes deeper, providing details about where passes are thrown and where runs are directed on the field. The main focus is on leveraging specialized knowledge about football strategies.

To gather this knowledge, I interviewed head coach of the national team of Austria Max Sommer, who shared crucial insights into how plays are decided. Using this information, different machine learning models were tested: logistic regression, a support vector machine, and a neural network. The support vector machine performed the best, achieving an impressive maximum accuracy score of 87.9%.

These predictions can be valuable for defensive coordinators, helping them anticipate the opposing team's moves. Moreover, the models also offer insights beyond just pass or run, aiding in more detailed play-calling decisions. After testing various features, a set of five key factors proved most effective for prediction. These include the time left in the game, the distance to the goal line, the distance needed for a new first down, the score difference, and the offensive team's past tendency to pass on each down. The data was taken from the 10 previous seasons of the NFL.

Importantly, the models were trained separately for each down and team to capture the unique strategies employed by teams in different situations during a game.