Using cost-sensitive learning to forecast football matches

Abstract

Forecasting football match outcomes have been investigated previously, with the primary goal of these studies being to accurately predict the outcome for the highest number of matches. This thesis takes a different approach, comparing different methods to investigate which would result in the highest profit, rather than focusing on predictive accuracy. Data of eleven seasons of the Dutch football league, the Eredivisie, were used to train six cost-sensitive decision trees. Two of our cost-sensitive models use the odds of a single bookmaker. Since multiple bookmakers publish their odds for the same match, four cost-combining methods are compared as well. The thesis shows what the strengths and weaknesses of each method are and how they behave both when betting on all matches, and when the strategy of value betting is employed. Though unfortunately this study could not provide any robust conclusions for which cost source leads to the best result in terms of profit, our results show potential for future research.