| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 6.94 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Com o crescimento do mercado das apostas desportivas a nível mundial [1] e o facto de um ténis ser um dos desportos mais populares para os apostadores [2], cresce a necessidade da existência de plataformas que ajudem os apostadores na tomada de decisão. O principal objetivo deste projeto passa pela criação de um modelo de previsão baseado em machine learning que consiga prever resultados de jogos profissionais de ténis. Uma plataforma para apostadores que disponibiliza previsões de forma automática, e tendo como base a análise de dados das últimas dezoito épocas desportivas, irá permitir aos apostadores pouparem tempo nas suas análises sem comprometerem os seus ganhos. Existem alguns trabalhados desenvolvidos relacionados com a previsão de resultados de jogos de ténis, alguns destes utilizam modelos de machine learning e outros utilizam apenas técnicas de análise de dados históricos, os resultados obtidos nestes trabalhos variam entre os 62.6% e os 69.9% de taxa de acerto a prever o vencedor de um jogo de ténis. A solução proposta é constituída por três componentes, o componente chamado Deuce Brain que é responsável pelo treino e teste do modelo de previsão, o componente chamado Deuce Services que é responsável por disponibilizar previsões através de um API, e por fim, o componente chamado Deuce Application que é uma aplicação web para disponibilização de previsões a apostadores. Durante o projeto foram feitas algumas experiências, onde se testaram modelos treinados com diferentes conjuntos de variáveis e diferentes abordagens. Foram desenvolvidas cinco experiência com conjuntos de variáveis diferentes, e para cada uma dessas experiências foram testados um modelo de regressão logística, uma rede neuronal artificial e um modelo SVM (Support-vector machine). O modelo que no final obteve maior taxa de acerto foi um modelo de regressão logística, com os rankings dos jogadores, a categoria do torneio e a superfície do court. Este modelo conseguiu uma taxa de acerto de 68%, e um retorno do investimento de 4.32% nos jogos do US Open de 2019. No geral, os modelos de regressão logística foram os mais precisos, seguidos das redes neuronais artificiais com taxas de acerto muito semelhantes, e por último os modelos SVM com uma diferença significativa.
With the growth of the sports betting market worldwide [1] and the fact that tennis is one of the most popular sports for gamblers [2], there is a growing need for platforms that help gamblers in their decision making process. The main goal of this project is the development of a prediction model based in machine learning that can predict the result for professional tennis matches. A platform for gamblers that provides automatic predictions based on data analysis from the last eighteen sports seasons, will allow the gambler to save time in their analysis without compromising their profits. There are some academic works related to the prediction of tennis match results, some of them use machine learning models and the others only use historical data analysis techniques, the results for these studies range from 62.6% to 69.9% of accuracy predicting the winner of a tennis match. The proposed solution is made up of three components, a component called Deuce Brain which is responsible for training and testing the prediction model, a component called Deuce Services which is responsible for providing predictions through an API, and finally a component called Deuce Application which is a web applications for making predictions available to gamblers. During the project were developed some experiments that tested models with different sets of variables and different approaches. It was tested five experiments with different sets of variables, and for each of these experiments it was produced a logistic regression model, an artificial neural network and a SVM (Support-vector machine) model. The model with highest accuracy was a logistic regression model with the players rankings, the tournament category and the court surface. This model achieved 68% of accuracy, and a return of investment of 4.32% during the US Open 2019. Overall, the logistic regression models were the most accurate, followed by the artificial neural networks with very similar accuracy, and lastly the SVM models with a significant difference.
With the growth of the sports betting market worldwide [1] and the fact that tennis is one of the most popular sports for gamblers [2], there is a growing need for platforms that help gamblers in their decision making process. The main goal of this project is the development of a prediction model based in machine learning that can predict the result for professional tennis matches. A platform for gamblers that provides automatic predictions based on data analysis from the last eighteen sports seasons, will allow the gambler to save time in their analysis without compromising their profits. There are some academic works related to the prediction of tennis match results, some of them use machine learning models and the others only use historical data analysis techniques, the results for these studies range from 62.6% to 69.9% of accuracy predicting the winner of a tennis match. The proposed solution is made up of three components, a component called Deuce Brain which is responsible for training and testing the prediction model, a component called Deuce Services which is responsible for providing predictions through an API, and finally a component called Deuce Application which is a web applications for making predictions available to gamblers. During the project were developed some experiments that tested models with different sets of variables and different approaches. It was tested five experiments with different sets of variables, and for each of these experiments it was produced a logistic regression model, an artificial neural network and a SVM (Support-vector machine) model. The model with highest accuracy was a logistic regression model with the players rankings, the tournament category and the court surface. This model achieved 68% of accuracy, and a return of investment of 4.32% during the US Open 2019. Overall, the logistic regression models were the most accurate, followed by the artificial neural networks with very similar accuracy, and lastly the SVM models with a significant difference.
Description
Keywords
Machine Learning Classificação Aprendizagem Supervisionada Ténis Apostas Regressão logística ANN SVM Classification Supervised Learning Tennis Betting Logistic Regression