Modeling basketball play-by-play data

• A method for modeling basketball play-by-play data is proposed.
• The model facilitates simulations of a basketball game between two distinct teams.
• We improve on the state-of-the-art in both forecasting accuracy and plausibility.
• Modeling the non-homogeneous parts of game improves the quality of the simulations.

We present a methodology for generating a plausible simulation of a basketball match between two distinct teams as a sequence of team-level play-by-play in-game events. The methodology facilitates simple inclusion into any expert system and decision-making process that requires the performance evaluation of teams under various scenarios. Simulations are generated using a random walk through a state space whose states represent the in-game events of interest. The main idea of our approach is to extend the state description to capture the current context in the progression of a game. Apart from the in-game event label, the extended state description also includes game time, the points difference, and the opposing teams’ characteristics. By doing so, the model’s transition probabilities become conditional on a broader game context (and not solely on the current in-game event), which brings several advantages: it provides a means to infer the teams’ specific behavior in relation to their characteristics, and to mitigate the intrinsic non-homogeneity of the progression of a basketball game (which is especially evident near the end of the game). To simplify the modeling of the transition distribution, we factorize it into terms that can be estimated with separate models. We applied the presented methodology to three seasons of National Basketball Association (NBA) games. Empirical evaluation shows that the proposed model outperforms the state-of-the-art in terms of forecasting accuracy and in terms of the plausibility of the generated simulations.