Counterfactual Optimization of Baseball Pitch Sequences and Estimation of Its Impact on Season-Level Statistics

This study introduces a Transformer-based machine learning model that analyzes pitch sequencing in baseball, utilizing MLB Statcast data to predict in-play outcomes based on pitch types and locations. By conducting counterfactual analyses, the research demonstrates that optimizing both final and setup pitches can significantly enhance season-level statistics, with potential improvements exceeding 1.0 in K/9. These insights underscore the critical role of pitch sequencing strategies, emphasizing the importance of pitch command and the effective use of middle-velocity pitches.