Modelos dinâmicos hierárquicos generalizados: uma aplicação com dados eleitorais brasileiros

Abstract

In the context of the modern world, in which globalization is already a reality, politics has a fundamental role in relations and dialogues between people and also between countries. Brazil has a multiparty system, which today is the fourth largest in terms of electorate, just behind India, the United States and Indonesia. Brazilian parties are divided into labor, socialist, communist, ecological, democratic, liberal, among others. The interest of this dissertation is to use one of the classes of dynamic models, more specifically, generalized hierarchical dynamic linear models. The model that we used was proposed by a group of researchers from Germany in 2017, in order to predict the result of the presidential election of the same year. The model is specified to predict election results in multiparty scenarios such as the Brazilian case. It consists of a combination of two other models, thus being done in two steps. The first stage is responsible for predicting the shared votes of each party based on the results of past elections and explanatory variables. The second stage is responsible for including information to predict the intention of votes for each party, candidate or group (which depends on the specified modeling), based on electoral surveys that are released during the electoral period. The proposed model was applied to the dataset of the 2018 Brazilian presidential elections, specifically for data referring to the first round. A reorganization of this model was proposed in this dissertation for this dataset. After adjustments, the results are presented and it is possible to verify that the applied model was efficient in predicting the final result of the 2018 election according to the presented assumptions.