Future Work
Possible directions for further exploration
There are several obvious future steps that could be taken with this project. For example, we could include more features in the model to attempt to improve the accuracy of our classifiers. One option would be to include data about the quality of the players in key positions, such as running back, quarterback, and wide receiver. This could potentially be useful because a team with a top ten running back may be more likely to run the football, especially in short yardage situations, than teams with a bottom ten running back. The same reasoning applies for quarterback and wide receiver. To some extent, the effect of this additional feature is probably already being captured, since we would expect a team with a better quarterback to pass more, which would increase the prior passing percentage of the team, a feature of our current models.
Another potentially useful feature is the in-game and season performance of certain key players, or of the team overall. For example, it’s possible that a team has called a low percentage of running plays over the previous year, but that the rushing game has been highly successful over the past few games or even in the current game. We would expect a team who has had in-game success running the ball to continue doing so, and our current model is not able to capture the effect of this in-game performance.
A final feature to consider including in our models is the scheme which is used by an offense, since some offensive schemes are more run-oriented and others are more pass-oriented. Including information about an offense’s scheme might be able to improve the accuracy of our classifiers. Somewhat tangentially, analyzing offensive schemes outside the context of play prediction may be an interesting direction for future work, because it’s possible that some offensive schemes are better overall than others, and it would be helpful for teams to be aware of this.
There is even more potential outside the context of play prediction. For example, with either our current classifier or one which has been improved by the features discussed above, we could build an interactive tool for users to predict plays. While this is not likely to provide additional insight into playcalling, it might be fun to allow website users to test their own play-predicting ability against our model or our model could be adapted into a live predictor which could be used for future NFL games.