Skip to main content
Open Access Publications from the University of California

Using Recurrent and Mixture Density Network Architectures to Model National Basketball Association In-Game Win Probabilities

  • Author(s): Poole, Henry
  • Advisor(s): Wu, Ying Nian
  • et al.

There are a number of possible machine learning approaches to modeling the win probability of an NBA game. Previous publically available research suggests that a mixture density network performs best in modeling win probability. In this paper, I explain and reproduce a number of previously shared approaches to modeling win probability. Unlike previous research, I fit each model with an identical set of inputs to fairly evaluate and compare the performance of each model. Furthermore, I create a recurrent mixture density network approach based off the recommendation of previous research. I find that the recurrent mixture density network has the highest measured accuracy in comparison to all other tested models.

Main Content
Current View