Forecasting football match results - A study on modeling principles and efficiency of fixed-odds betting markets in football

Abstract

Objectives of the study

This thesis is about the statistical forecasting of (European) football match results. More specifically, the purpose of this thesis is to assess how a statistical forecast model that uses only publicly available information fares against public market odds in forecasting football match outcomes.

Academic background and methodology

The forecasting of sports results has been widely researched because it provides important insight into how betting markets operate. Football and betting associated with it has been the most popular topic because of the global popularity of the sport and because the betting markets associated with it capture large annual turnover.

In spite of research by numerous authors, there is still room for improvement in terms of developing more accurate forecast models. Therefore, we contribute to existing literature by developing a regression model for forecasting football results. We assess the model's performance with forecast accuracy measurements and betting simulations. The principal idea of the model is based on the ELO rating system which assigns relative performance ratings to teams.

Findings and conclusions

In terms of accuracy measurements and betting simulations, the model developed in this thesis is able to match or surpass the results of existing statistical models of similar build. The measurements also indicate that the model can on average match the accuracy of the forecasts implied by the publicly quoted odds. However, the model is unable to generate positive betting returns. Together these results indicate that the publicly quoted odds for extensively betted football matches are slightly inefficient, but that this inefficiency does not make statistical betting algorithms consistently profitable. The results also indicate that historical league match results are the most important components of a statistical football forecast model, and that supplementing these components with other data yields only modest improvements to forecast accuracy.