Neural Network Model for Predicting NBA Shot Outcome

To fully exploit the information in the SportVU dataset, the analytic team at the Raptors wanted to move beyond summary statistics and utilize state-of-the-art machine learning algorithms for dealing with dataset of this magnitude. Specifically, one task of interest is “shot outcome prediction”. Given all the information up to the release of shot, can we give a calibrated probability of whether or not the player will make that shot? To achieve this, in the previous internship, we have used neural network models that optimized the data likelihood. Compared to the algorithm that the Raptors analytics team was using before the internship, this new model did better in both interpolation within seen data and extrapolation to unseen data. Towards the end of the previous internship, we started to develop models that could modulate the prediction by the player identity. In this proposed internship, we aim to polish that approach and develop new models that can incorporate and exploit complex interactions of the shooter with the other players and the game situation. Another objective of this research is to develop a time-series model that can extract interesting features, possibly not recognized as significant by the human experts, from the raw tracking data.

Faculty Supervisor:

Richard Zemel

Student:

Partner:

Raptors

Discipline:

Computer science

Sector:

Arts, entertainment and recreation

University:

University of Toronto

Program: