Aplicação de técnicas de Machine Learning na combinação de pesquisas eleitorais

Abstract

The Brazilian electoral polls are made, mostly, from a non-probabilistic sample plan by quotas. Thus, there is an accumulation of uncertainties regarding the estimates obtained, since there is no way to guarantee that the margin of error is being respected and there is no way to calculate the confidence intervals. As a way to reduce the error in the estimation of the intention of votes, this work proposes the use of techniques of Machine Learning to combine electoral polls. We used the electoral polls for the first round of the presidential election of 2014 in Brazil, published by the institutes Datafolha, Ibope, MDA, Sensus and Vox Populi. The combinations were made using Local Regression, Random Forest and Support Vector Machine (SVM). With the exception of Random Forest and SVM with linear kernel, the other techniques presented as good options for combining the polls.