| Popis: |
Prediction of elections is a subject that excites the population, especially in the last few months before an election. In Brazil, there is a wide availability of political, economic and social data, in institutions such as TSE, IBGE and opinion research institutes that can be used as sources to create prediction models. Therefore, this work aims to build multivariate linear regression and regression tree models to predict the percentage of votes received by the situational candidate for the presidency of Brazil. The multivariate linear regression model had the smallest prediction errors, with MAE of 1.45 in the first round and 1.48 in the second, with margins smaller than 1\% in 2002, 2006 and 2018. The proposed models seemed to be more accurate than other models found in the literature. As main contributions, it was possible to observe that the sampling of data by state and the use of the illiteracy rate and the popular vote intention contributed directly to the performance of the models. |
