ADS Capstone Chronicles Revised

15

influence from the game flow. Future research should explore methods for integrating these data sources effectively into predictive modeling frameworks. ACKNOWLEDGMENTS This capstone article would not have been possible without the guidance and support of several individuals such as previous professors and peers. A special thanks goes out to our University of San Diego Academic Director Ebrahim Tarshizi, PhD. References Gifford, M., & Bayrak, T. (2023). A predictive analytics model for forecasting outcomes in National Football League games using decision tree and logistic regression. Decision Analytics Journal , 8 , 100296. https://doi.org/10.1016/j.dajour.2023.100 296 Joash Fernandes, C., Yakubov, R., Li, Y., Prasad, A. K., & Chan, T. C. Y. (2020). Predicting plays in the National Football League. Journal of Sports Analytics , 6 (1), 35–43. https://doi.org/10.3233/JSA-190348 Kaggle. (n.d.). NFL team stats 2002 - Feb. 2024 (ESPN) . https://www.kaggle.com/datasets/cviaxmi wnptr/nfl-team-stats-20022019-espn/code Pantle, I. (2022). Analyzing the predictive power of NFL statistics . Dartmouth Sports Analytics. https://sites.dartmouth.edu/sportsanalytics /2022/02/01/analyzing-the-predictive power-of-nfl-statistics/ Pelechrinis, K., & Papalexakis, E. (2016, December 22). The anatomy of American football: Evidence from 7 years of NFL

Moreover, our incorporation of feature engineering, particularly through the creation of one-hot encoded features for home and away teams, significantly bolsters the predictive power of our model. By accounting for variables such as team strength and home advantage, we endeavor to refine the accuracy of game outcome predictions. Nevertheless, further fine-tuning in feature selection and correlation management is imperative to ensure the comprehensive consideration of relevant features without introducing biases or exacerbating overfitting. Despite the efficacy of logistic regression, challenges pertaining to overfitting persist in other models like Random Forest and Gradient Boosting. Our endeavors in hyperparameter tuning have yielded modest improvements, signaling the necessity for more advanced techniques or alternative algorithms to effectively tackle these issues. 6.2 Recommend Next Steps/Future Studies Future studies should focus on refining model techniques to address overfitting, on the explored but unused models, as well as enhance generalization capability. Although this article attempted regularization techniques, they did not yield significant improvements in performance. Therefore, subsequent efforts may involve exploring ensemble methods or advanced feature selection algorithms to improve the robustness of predictive models. Incorporating external data sources, such as comprehensive player information or expert predictions, may provide valuable additional information to enhance model accuracy and robustness. Additionally, first-half only data or data from the first three quarters would allow for better scoring predictions that utilize passing and rushing attempts as descriptive features. This would incorporate scoring without as harsh of

42