Comparison of Machine Learning Algorithms to Predict Football Match Outcomes

Abstract Football, being one of the most popular sports in the world, has been open to the betting and gambling world to try and predict the outcomes of its matches. Yet football is a very complex sport and is notoriously hard to predict. In this study, we attempted to predict the football match outcomes of the Indian Super League team, Kerala Blasters, while comparing three algorithms, Random Forest, Logistic Regression and k-Nearest Neighbors (KNN) to understand which model and features work best across datasets. These algorithms were applied to different sets of features, pre-game and in-game features, to determine which led to the most accurate and precise results. We also applied artificial sampling to some models to test if a more balanced dataset found better results. Results showed that Logistic Regression without using SMOTE achieved the highest accuracy (75%) and precision (42%) compared to the other models. Thus, our findings suggest that Logistic Regression is the best model for this dataset. Future work might extend the comparison to more models and apply oversampling techniques to models that don't work as well with unbalanced datasets. Keywords Football, Machine Learning, Random Forest, Linear Regression, KNN