Predicting Pitch Outcomes in Major League Baseball (Student Presentation, Group 11)

This is a STAT 451 class project presentation

by Micah Jona, Nathan Kolbow, and Roshan Poduval

This presentation is shared with the students' permission.

Abstract:

In this study, we evaluate the abilities of numerous ma- chine learning algorithms to predict the outcome of a pitch in Major League Baseball (MLB) using real time data. This is a good task for testing the efficacy of various ma- chine learning methods because it is a highly random and complex task that is accompanied by a plethora of data. This study focuses chiefly on the relative abilities of each model in performing the classification task, but as a consequence of tackling a novel problem we also offer performance baselines for any future endeavors. We compare the performances of four gradient boosting machines, k- Nearest Neighbors (kNN) clustering, as well as neural net- works with and without convolutional layers preceding fully connected multiple perceptron layers. To this degree, when comparing learning algorithms we varied hyperparameters and attempted to achieve the best possible local maxima for each method. Extensive optimization of each algorithm revealed that XGBoost performed best on this dataset.