Election Forecasting in a Multiparty System

Abstract

This bachelor thesis in statistics covers the subject of election forecasting in a multiparty system, using polling data, that is data collected to measure party support, and dynamic linear models (DLMs) with Kalman filtering. In terms of decision-making the outcome of an election can be thought of as an uncertainty. Forecasts of election results can reduce risks for decision-makers and thereby facilitate decision-making. To be able to foresee the outcome of an event can be of use for experts in several different fields, for instance political strategists, nancial investors and policy makers. A DLM considers an observable time series to be a linear function of a latent, unobservable series and random disturbance. In the case of election forecasting we can think of the observable series as being polling data, and the underlying series to be true measures of party support. The purpose of using the Kalman filter is then to retrieve the latent series representing true party support. Altogether three different models are explored in the thesis; a Gamma- Normal, a time-invariant and a multivariate time-invariant model. The main difference between the frameworks concerns the variance term in the distribution of the noise terms in the DLM. The models are applied to the Swedish election of 2018, using polling data for the period stretching from September 2014 to September 2018. The polling data is then disregarded for three di erent time periods; the last month, the last six months andthe last twelve months before the election. For those periods, we instead use simulated data which together with the polling data is the basis of our forecasts. We find that the Gamma-Normal model performs slightly better than the two other models, when forecasting the election result one month ahead, while the multivariate time-invariant model is slightly better for the two other time frames. For the one year forecast this model predicts the election result with an average absolute prediction error of 1.28 percentage points for each party. Finally, the forecasting capability of the models are discussed and evaluated in the analysis section of this thesis.