New Player Evaluation Technique for Players of the NBA

Abstract: I present a combination of player evaluation techniques that significantly outperforms regularized adjusted +/- in prediction accuracy. This approach models a basketball game as a finite state machine, which is a behavior model composed of a finite number of states, transitions between those states, and actions. This state machine consists of four states, a ’normal’ and an ’extra possession’-state for each of the two teams. The ’extra possession’-state can be reached through forcing opponent turnovers or offensive rebounds. Players get positive and negative credit assigned to them for executing certain actions or failure to do so. Variables that determine how credit gets split between players get optimized by minimizing the sum of squared residuals in the training data. NBA games from October through February were used to train each model. All models were evaluated in terms of their ability to predict the outcome of the remaining regular season games, which are not included in each models’ training set. Squared error of predicted vs. actual margin of victory was used as a measure of accuracy

The full paper can be found here

The conference poster can be found here


    Back to Videos