Bayesian Forecasting of Election Results in Multiparty Systems

Abstract

I present a Bayesian forecasting model particularly suited for multiparty systems. The method I develop systematically combines (i) information from a Multinomial logit regression model fitted on historical data and (ii) estimates of current party support produced by a Dynamic Linear Model for multinomial observations. I apply the method to the Norwegian multiparty system, and assess the performance of the model on past elections. As of present, the model is ready to be updated as the Norwegian parliamentary elections of 2013 draws closer. The current forecast for the upcoming election is that the four opposition parties will obtain a majority in parliament with a probability of 0.775.