Predicting football match result using fusion-based classification model

Abstract

Over the last decades, various football experts and researchers around the world seek to forecast football match result. Although there are multiple types of prediction models constructed for this purpose, most of them still depends on integrating in-game statistical numbers, such as number of successful passes in one game. This kind of information has huge positive impact on predicting outcome but is not desired as it requires the match to finish first. Thus, this thesis aims to propose more accurate models, which are not relied on in-game numbers. Two forms of fusion-based classification models are proposed in this study, including hierarchical model and ensemble model. Player and team ratings from video games are incorporated to assist in match prediction. Experiments are designed to compare the proposed models with six traditional classification models, by measuring accuracy with data sets collected from English Premiere League (EPL) competition, from season 2010/2011 to 2015/2016. Accurate rates achieved by both models are at 56.533% and 56.800% by using a total of 36 features, which are all superior to comparative models with similar environment.