A dynamic feature selection based LDA approach to baseball pitch prediction

Abstract

Baseball, which is one of the most popular sports in the world, has a uniquely discrete gameplay structure. This stop-and-go style of play creates a natural ability for fans and observers to record information about the game in progress, resulting in a wealth of data that is available for analysis. Major League Baseball (MLB), the professional baseball league in the US and Canada, uses a system known as PITCHf/x to record information about every individual pitch that is thrown in league play. We extend the classification to pitch prediction (fastball or nonfastball) by restricting our analysis to pre-pitch features. By performing significant feature analysis and introducing a novel approach for feature selection, moderate improvement over published results is achieved.