AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


Machine Learning in Football Betting: Testing Profitability on the Betfair Exchange

Authors

;

Term

4. term

Education

Publication year

2025

Abstract

This thesis examines whether machine learning can build profitable football betting strategies on the Betfair Exchange across the top five European leagues (2017–April 2025). We extend Hubáček et al. (2019) by training an XGBoost model with a custom loss designed to decorrelate predictions from market odds, focusing on the 1X2 market. Strategies are constructed using the Kelly Criterion and Modern Portfolio Theory (MPT) and evaluated with rigorous walk-forward cross-validation plus separate validation and out-of-sample sets to mitigate backtest overfitting. Unfiltered strategies produce negative returns, but simple rule-based filters improve performance: only bet when model probabilities exceed market-implied probabilities by 25%, and cap odds below 2.0. The best-performing approach—MPT combined with the probability filter—achieves a 3.71% annual growth rate over the full backtesting period. Although modest, the returns are lowly correlated with traditional assets and may support diversification and enhance the value of bookmakers’ freebets. The thesis also discusses market efficiency, participant behavior, and the Covid-19 regime shift affecting betting fundamentals.

Denne afhandling undersøger, om maskinlæring kan bruges til at konstruere profitable fodboldstrategier på Betfair-børsen i de fem største europæiske ligaer (2017–april 2025). Vi bygger videre på Hubáček et al. (2019) ved at træne en XGBoost-model med en specialdesignet tabsfunktion, der har til formål at afkorrelere modelens forudsigelser fra markedsodds, og vi fokuserer på 1X2-markedet. Strategierne konstrueres med Kelly-kriteriet og Modern Portfolio Theory (MPT) og evalueres med streng walk-forward cross-validation, samt separate validerings- og out-of-sample-sæt for at begrænse backtest-overfitting. Ufiltrerede strategier giver negative afkast, men simple regelbaserede filtre forbedrer resultaterne: der satses kun, når modellens sandsynligheder overstiger markedsimlicerede sandsynligheder med 25%, og odds begrænses til under 2,0. Den bedst performende tilgang—MPT kombineret med sandsynlighedsfiltret—opnår en årlig vækstrate på 3,71% over hele backtestperioden. Selvom afkastet er moderat, er det svagt korreleret med traditionelle aktiver og kan bidrage til diversifikation og til at maksimere værdien af bookmakeres freebets. Afhandlingen diskuterer også markedseffektivitet, deltageradfærd og Covid-19 som et regimeskifte, der påvirker fundamentale forhold i betting.

[This apstract has been generated with the help of AI directly from the project full text]