Utilizing Markov Models to Simulate MLB Games

Abstract

This paper utilizes a Markov model to simulate Major League Baseball (MLB) games. The simulation uses Retrosheet data and translates game outcomes into state transition matrices. This methodology allows for predicting multiple aspects in a game, compared to traditional Machine Learning approaches, which generally allow for only a single prediction. The model compares the results to publicly available betting data to predict which team will win, whether the favored team will win by enough runs to exceed the expected run difference, whether the run total will be above or below the betting line, and whether either team will score in the first inning. The model slightly outperformed the baseline accuracy when predicting the winner at around 54.2%. It was more successful in predicting which team will cover the spread at 57.2%. The run total predictions were inaccurate, since the model consistently overestimated the number of runs scored in a game. The predictions for whether either team will score in the first inning were on par with the baseline values.