SOCCER ANALYTICS: AN EXAMINATION OF CURRENT APPROACHES FOR GAME OUTCOME PREDICTION AND THE POTENTIAL OF DEEP LEARNING WITH ARTIFICIAL NEURAL NETWORKS

Open Access
- Author:
- Gomez Lorente, Juan
- Area of Honors:
- Industrial Engineering
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Hui Yang, Thesis Supervisor
Dr. Catherine Mary Harmonosky, Thesis Honors Advisor - Keywords:
- Machine Learning
Deep Learning
Neural Networks
Artificial Neural Networks
Poisson Regression
Soccer
Sports Analytics
Betting
Predictive Modelling
Moneyball Analytics
SoftMax Classifier
Regression
effects encoding
feed-forward networks
backpropagation
gradient descent
resilient backpropagation
Newton's Method
Taylor Series expansion
Manhattan Rule
Broyden–Fletcher–Goldfarb–Shanno algorithm
BFGS
Quasi-Newton Methods - Abstract:
- As the world has fully embraced analytics and data across a great variety of fields, the growth sports analytics have generally not lagged behind. With the economic and social power behind many professional sporting events, the use of data analytics has become mainstream for games like football, basketball, or baseball. However, the most popular sport in the globe has been hesitant to take a similar path. Soccer is widely known as a highly random, complex game. Its unpredictable nature is often pointed out as its most attractive feature. Professionals of the sport, pundits, and fans often trust their gut feeling, experience and qualitative perceptions over statistics and data. The applied research detailed in this thesis focuses on predictive modeling of soccer game outcomes in the 2015-16 Spanish LaLiga championship. Initially, extensive analysis and review are performed of generalized linear Poisson regression models, the most common approaches encountered in the literature. This allows for a discussion regarding the validity and appropriateness of these methods as well as the establishment of a baseline for comparison. Then, machine learning techniques under the scope of supervised learning are explored as possible alternative approaches. In particular, artificial neural network models are used to create predictive analysis tools with both regression and classification approaches to output prediction. Different input estimators are explored, including a detailed review about the utilization of betting odds as game outcome predictors. The results obtained are compared to the baseline set with the Poisson regression models to demonstrate the potential of machine learning in this growing field of applied data analytics.