[2Yin-A] Poster 3

[2Yin-A-45] A Comparative Study of Predictive Models for Baseball Pitch Calling as Joint Classification of Pitch Type and Zone

〇Yuto Kanoh¹, Takuichi Nishimura¹ (1. Japan Advanced Institute of Science and Technology)

Keywords:

Deep Learning,Pitching predict,baseball,Machine Learning,Artificial Intelligence

Pitch calling, jointly decided by a catcher and a pitcher, is a core baseball decision that directly affects outcomes. Because pitch selection depends on complex contextual factors, consistently optimal choices are difficult for humans. We formulate pitch calling as selecting a (pitch type, pitch location zone) pair at time using information available up to time , and treat each pair as a single pitching label. Using MLB Statcast data from 2017–2024 and time-blocked 5-fold cross-validation, we compare diverse machine learning and deep learning models and evaluate pitch type, zone, and pitching-label prediction with Accuracy, Precision, Recall, and F1-score. Transformers achieve the best overall performance, particularly for pitching labels, whereas tree-based ensembles degrade under sparse joint classes due to majority-class bias. Future work will improve interpretability via attention and feature-attribution analyses.