This paper presents the method we used in the prediction challenge organised by the Sports Analytics Lab of the KU Leuven for the European football(soccer) championship. We built a generalised linear model to predict the score of a match. This score was modelled as the joint probability of a Poisson distribution, representing the total number of goals, and a binomial distribution, representing the goals of one team given that total number of goals. This model was trained on the matches of the past year using gradient descent to maximise the log-likelihood with l2 regularisation. Special care was taken to construct a model that is symmetrical and does not involve any home advantage, with the exception of the host team. The features considered were both team-based and player-based, using a randomised approach to select the players based on their past selections. A simulation of the tournament was then built on this match model to predict how far each team would go in the tournament.
© Copyright 2016 Machine Learning and Data Mining for Sports Analytics ECML/PKDD 2016 workshop. Published by Department of Computer Science, KU Leuven. All rights reserved.