Using play-by-play data from all 2014-15 regular season NBA games, we build a generative model that accounts for substitutions of one lineup by another together with the plus/minus rate of each lineup. The substitution model consists of a continuous-time Markov chain with transition rates inferred from data. We compare different linear and nonlinear regression techniques for constructing the lineup plus/minus rate model. We use our model to simulate the NBA playoffs; the test error rate computed in this way is 20%, meaning that we correctly predict the winners of 12 of the 15 playoff series. Finally, we outline several ways in which the model can be improved.
© Copyright 2015 Machine Learning and Data Mining for Sports Analytics ECML/PKDD 2015 workshop. Published by Department of Computer Science, KU Leuven. All rights reserved.