Browsing by Author "Coelho, Carlos Manuel Maia"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- OddAssist - Um sistema de recomendação de apostas em eSportsPublication . Coelho, Carlos Manuel Maia; Pereira, Isabel Cecília Correia da Silva Praça GomesIt is globally accepted that sports betting has been around for as long as the sport itself. Back in the 1st century, circuses hosted chariot races and fans would bet on who they thought would emerge victorious. With the evolution of technology, sports evolved and, mainly, the bookmakers evolved. Due to the mass digitization, these houses are now available online, from anywhere, which makes this market inherently more tempting. In fact, this transition has propelled the sports betting industry into a multi-billion-dollar industry that can rival the sports industry. Similarly, younger generations are increasingly attached to the digital world, including electronic sports – eSports. In fact, young men are more likely to follow eSports than traditional sports. Counter-Strike: Global Offensive, the videogame on which this dissertation focuses, is one of the pillars of this industry and during 2022, 15 million dollars were distributed in tournament prizes and there was a peak of 2 million concurrent viewers. This factor, combined with the digitization of bookmakers, make the eSports betting market extremely appealing for exploring machine learning techniques, since young people who follow this type of sports also find it easy to bet online. In this dissertation, a betting recommendation system is proposed, implemented, tested, and validated, which considers the match history of each team, the odds of several bookmakers and the general feeling of fans in a discussion forum. The individual machine learning models achieved great results by themselves. More specifically, the match history model managed an accuracy of 66.66% with an expected calibration error of 2.10% and the bookmaker odds model, with an accuracy of 65.05% and a calibration error of 2.53%. Combining the models through stacking increased the accuracy to 67.62% but worsened the expected calibration error to 5.19%. On the other hand, merging the datasets and training a new, stronger model on that data improved the accuracy to 66.81% and had an expected calibration error of 2.67%. The solution is thoroughly tested in a betting simulation encapsulating 2500 matches. The system’s final odd is compared with the odds of the bookmakers and the expected long-term return is computed. A bet is made depending on whether it is above a certain threshold. This strategy called positive expected value betting was used at multiple thresholds and the results were compared. While the stacking solution did not perform in a betting environment, the match history model prevailed with profits form 8% to 90%; the odds model had profits ranging from 13% to 211%; and the dataset merging solution profited from 11% to 77%, all depending on the minimum expected value thresholds. Therefore, from this work resulted several machine learning approaches capable of profiting from Counter Strike: Global Offensive bets long-term.