A Multiresolution Stochastic Process Model for Predicting Basketball Possession Outcomes

Basketball games evolve continuously in space and time as players constantly\ninteract with their teammates, the opposing team, and the ball. However,\ncurrent analyses of basketball outcomes rely on discretized summaries of the\ngame that reduce such interactions to tallies of points, assists, and similar\nevents. In this paper, we propose a framework for using optical player tracking\ndata to estimate, in real time, the expected number of points obtained by the\nend of a possession. This quantity, called \\textit{expected possession value}\n(EPV), derives from a stochastic process model for the evolution of a\nbasketball possession; we model this process at multiple levels of resolution,\ndifferentiating between continuous, infinitesimal movements of players, and\ndiscrete events such as shot attempts and turnovers. Transition kernels are\nestimated using hierarchical spatiotemporal models that share information\nacross players while remaining computationally tractable on very large data\nsets. In addition to estimating EPV, these models reveal novel insights on\nplayers' decision-making tendencies as a function of their spatial strategy.\n