Predicting Passing-Play Outcomes in American Football: A Comparative Study of Traditional Machine Learning Algorithms and Fine-Tuned Large Language Models

Abstract

This study examines American football passing play strategies based on an analysis of 100 NFL games, addressing the notable lack of research on offensive passing strategies. The yardage prediction task is formulated as a regression task, and the touchdown prediction task is formulated as a classification task. A regression analysis was conducted using four traditional machine learning models, including Support Vector Machines (SVMs), K-Nearest Neighbors, XGBoost, and LightGBM, and a LoRA-fine-tuned Llama 3 (8B) model. The SVM model showed optimal performance with a Mean Absolute Error of 6.996 yards, marginally outperforming the Llama 3 model (7.059 yards). In contrast, the fine-tuned Llama 3 significantly outperformed all traditional algorithms for the classification task, with a Macro F1 of 0.712 compared to the highest Macro F1 of 0.614 for traditional algorithms. Therefore, I advocate large language models (LLMs) for classification tasks and traditional machine learning algorithms for regression tasks. Sports analytics models can be selected based on the findings of these studies.