Previsão de resultados de jogos do campeonato brasileiro de futebol utilizando aprendizagem de máquina

Abstract

Over the past few years, the demand for sports betting has grown, and several people started living off this market. Using Machine Learning with the goal of making game analysis for betting more comfortable as well as measuring profits, a database for the Brazilian Soccer Championship Série A 2017 was created, encompassing features from two opposing teams and the match. For testing in an evaluation model, the Naive Bayes technique was chosen, simulating a Brazilian Championship round, which contains ten matches. The evaluation model was executed both with and without correct probability restrictions, with the goal of reducing errors. Three scenarios were used on the evaluation model, the first having two classes that consider goals by both teams, the second regarding the total match goals (over or below 2.5 goals) and the third considering all three possible match results (a victory for one side, a tie, or a victory for the other side). The results without probability restriction reached its higher value at 5.51% of average and total profit. However, regarding the results with probability restriction, the best average gain was 36.05%, and the best total profit was 39.13%, both using 99% correct probability restriction.