Machine Learning in Football Betting: Prediction of Match Results Based on Player Characteristics

Abstract

In recent times, football (soccer) has aroused an increasing amount of attention across continents and entered unexpected dimensions. In this course, the number of bookmakers, who offer the opportunity to bet on the outcome of football games, expanded enormously, which was further strengthened by the development of the world wide web. In this context, one could generate positive returns over time by betting based on a strategy which successfully identifies overvalued betting odds. Due to the large number of matches around the globe, football matches in particular have great potential for such a betting strategy. This paper utilizes machine learning to forecast the outcome of football games based on match and player attributes. A simulation study which includes all matches of the five greatest European football leagues and the corresponding second leagues between 2006 and 2018 revealed that an ensemble strategy achieves statistically and economically significant returns of 1.58% per match. Furthermore, the combination of different machine learning algorithms could neither be outperformed by the individual machine learning approaches nor by a linear regression model or naive betting strategies, such as always betting on the victory of the home team.

Investigating inefficiencies of bookmaker odds in football using machine learning

“…On the other hand, when applying betting strategies that use only limited amount of information (compared to the odds), one can expect systematic losses over time. This is in line with the results of Stübinger et al (2020), where strategies that use no or little information (random-betting or favoring the home team) never gain any positive returns. However, using todays computation power and data, one can generate an historical information advantage in times where the odds did not reflect that knowledge which results in positive returns for earlier periods.…”

“…However, the analysis is based on econometrical models, only, whereas Stübinger et al (2020) cover the application of state-of-the-art machine learning technique with a strong emphasis on the prediction of the correct outcome of a game rather than modeling the mechanics between the influencing factors. Stübinger et al (2020) point out diminishing profitability of Benedikt Mangold, Johannes Stübinger systematic betting approaches over time, which can be explained by improved modeling techniques and the improved availability of processing power and data. In this section, we link the aforementioned results of Stübinger et al (2020) to the efficiency hypothesis.…”

“…Stübinger et al (2020) point out diminishing profitability of Benedikt Mangold, Johannes Stübinger systematic betting approaches over time, which can be explained by improved modeling techniques and the improved availability of processing power and data. In this section, we link the aforementioned results of Stübinger et al (2020) to the efficiency hypothesis.…”

“…We use the predictions to (virtually) place a bet on historic odds to analyze if we could systematically outperform the believes of the market, measured by the return of the placed bet. In the following, we revisit the results of the paper by Stübinger et al (2020) with respect to the success of the betting strategy, the implications for the efficiency of the offered odds and some detailed analyses from the perspective of the bookmakers during the covered period.…”

“…The first section revisits the simulation design and important results of Stübinger et al (2020), the second section discusses the findings with respect to the well-known market efficiency hypothesis. The third section concludes this manuscript.…”