Using large language models to generate baseball spray charts in the absence of numerical data

Abstract

Although baseball has been revolutionized by analytics, not all teams have access to high quality data. While many high school, collegiate, and club teams do not have high speed cameras and radars, they often do record a text-based play-by-play account of the game. The purpose of this study is to demonstrate how to use large language models to convert play-by-play information into quantitative data. We walk through the specific example of spray charts, which depict where on the baseball diamond a hitter tends to put the ball in play. Spray charts are a particularly relevant example because of their use in informing in-game strategy decisions (e.g., the infield shift). This study successfully generates spray charts for collegiate baseball players with 95% accuracy.