Forecasting elections based on the aggregation of election polls

Abstract

Elections polls have been known to exist since 1824 [14], to fulfill the objective of what is happening and may happen. In this thesis it is investigated what the performance of election polls is and if the aggregation of polls give a better forecast than the polls themselves. The data is used of Spain during years 2012 until 2017 and The Netherlands during years 2010 until 2021. Furthermore, in this thesis the results of Spain and The Netherlands are compared. The aggregation of polls is done by using The Classical Model of Roger Cooke [2] and by using two types of Equal weighting. When using the Classical Model two different methods are used, namely a window-shift way and a cumulative way. The comparison between the aggregation of polls and the polls themselves is done by looking at the unnormalized weights and the total absolute differences. The results were quite different for the two countries. However, for both countries it appeared that method 1 was better when looking at the unnormalized weights, while method 2 was better when considering the total absolute differences. When looking at the unnormalized weights for the Spanish data the forecast when aggregating the polls, using the Item weight decision maker, was better than most of the polls’ forecasts. In addition when looking at the absolute total differences of the Spanish data, the forecast when aggregating the polls, also using the Item weight decision maker, was better than most of the polls’ forecasts. Similarly for the Dutch data when looking at the unnormalized weights, the forecast of the election when aggregating the polls, using the Performance-based decision maker, was better than most of the polls’ forecasts. In addition, when looking at the total absolute differences, the aggregation using the Equal weight decision maker based on distribution was better than most of the polls’ forecasts.