Forecasting the party support in Estonia: comparison of machine learning regression algorithms

Abstract

Forecasting political behavior using economic indicators is not a very new phenomenon with the earliest literature going back as far as the 1930s. In the present day, there exists a lot of research on the topic, but the majority of these studies have been conducted in the context of a very limited number of countries such as the United States or the Western European ones. By comparison, the research on forecasting the political behavior using economic voting in Estonia is almost non-existent. This thesis will be the first in-depth study conducted at that level and forecasts the party support of the Estonian Reform Party and the Estonian Center Party using economic indicators as the predictor variables. Based on the previous economic voting theory, it has been argued that the theoretically correct model to forecast using these variables is the linear regression due to the expected associations between the economic variables and party support. However, this thesis contests this claim and argues that when analyzing the phenomena of forecasting party support using economic indicators, certain modern machine learning algorithms could be considered as legitimate alternatives to the linear regression, as each of them addresses the different shortcomings of the model. For this reason, this thesis compares the methods of linear regression, regularized linear models, autoregressive integrated moving average, and the decision-tree models to see whether the more modern approaches are able to improve upon the default linear regression model.