Predicting Counter-Strike Matches using Machine Learning Models

Abstract: Sports betting is a widespread industry where predictive modeling play a big role. The goal of this thesis is to explore the possibilities of machine learning within the realm of e-sport prediction. The data used for this thesis is publicly available data was recorded over a three year period. The chosen variables are defined as the difference in player performance between two teams in order to create conditional probabilities. The paper focuses on two machine learning models for evaluating predictability within the data, Logistic regression with parameter regularization and Random Forest. Both models were optimised with cross-validation and their effectiveness is compared to a benchmark which in this case is the betting odds of multiple bookmakers. Measures such as accuracy, Log-Loss and the κ-parameter are our main points of comparison. The findings suggest that our models were capable of achieving an accuracy exceeding 50/50 on the test data, implying a certain level of predictability. The models were also applied to a professional tournament, and although a small sample size with large standard errors had an influence, we conclude that the evaluated models did not surpass the performance of the benchmarks.