Using the Random Forest to Estimate In-Game Win Probability

In-game win probability is the estimated probability of victory for each competitor in the midst of a competition. We discuss the common methods of estimating in-game win probability values and present an approach using random forests that is uniformly applicable to all head-to-head competitions. The random forest is a non-parametric machine learning methodology common in big data regression and classification problems. Applying our method to the NHL, NBA and NFL demonstrates the performance and consistency of the approach.  In addition to being intrinsically interesting our win probability values can have several useful applications, from enhancing the fan experience to evaluating coaching or player decisions. We also discuss utilizing win probability values to evaluate the overall contributions of NBA players.