A data mining approach to predict probabilities of football matches

Abstract

With the increasing growth of the amount of money invested in sports betting markets it is important to verify how far the machine learning techniques can bring value to this area. A performance evaluation of the state-of-art algorithms is performed and evaluated according to several metrics, incorporated in the CRISP-DM methodology that goes from web-scraping through to generation and selection of features. It is also explored the universe of ensemble techniques in an attempt to improve the models from the point of view of bias-variance trade-off, with a special focus on neural network ensembles.