Predicting Win-Loss outcomes in MLB regular season games – A comparative study using data mining methods

Baseball is a statistically filled sport, and predicting the winner of a particular Major League Baseball (MLB) game is an interesting and challenging task. Up to now, there is no definitive formula for determining what factors will conduct a team to victory, but through the analysis of many years of historical records many trends could emerge. Recent studies concentrated on using and generating new statistics called sabermetrics in order to rank teams and players according to their perceived strengths and consequently applying these rankings to forecast specific games. In this paper, we employ sabermetrics statistics with the purpose of assessing the predictive capabilities of four data mining methods (classification and regression based) for predicting outcomes (win or loss) in MLB regular season games. Our model approach uses only past data when making a prediction, corresponding to ten years of publicly available data. We create a dataset with accumulative sabermetrics statistics for each MLB team during this period for which data contamination is not possible. The inherent difficulties of attempting this specific sports prediction are confirmed using two geometry or topology based measures of data complexity. Results reveal that the classification predictive scheme forecasts game outcomes better than regression scheme, and of the four data mining methods used, SVMs produce the best predictive results with a mean of nearly 60% prediction accuracy for each team. The evaluation of our model is performed using stratified 10-fold cross-validation.