Pulling the Starting Pitcher: Determining the Optimal Decision in MLB Games

Abstract

In Major League Baseball, teams are increasingly using data-driven methods to optimize all aspects of the game, from roster construction and player development to making in-game strategic decisions. In this paper, we focus on one of the most important in-game decisions for a manager: the optimal time to remove the starting pitcher from the game. I explain how I built upon prior work to construct a model to determine whether a starting pitcher should be removed, based on newly available baseball statistics and machine learning techniques. Specifically, I assembled and analyzed a dataset from MLB’s Statcast consisting of over 90 pieces of raw data about over 500,000 pitches thrown in the 2021 MLB season; I describe how I leverage this new dataset to engineer relevant features and develop a well-performing model. Our best performing model, XGBoost, improved upon previous binary classification and continuous regression models that aimed to solve the same problem.