Non-parametric Bayesian Forecasts

Although there is much literature that covers election outcome forecasting, few to no methods of prediction have been able to consistently deliver accurate results. This problem essentially stems from the fact that election results are greatly influenced by idiosyncratic factors. This makes model selection difficult as, at the time of election, it is not clear which (type of) model will perform best. In this research, the problem of election forecasting is approached with a non-parametric Bayesian individual-level model using voting intentions and sociodemographic variables on Dutch elections in 2010 and 2012. Making use of a Dirichlet Process mixture (DPM) model, a flexible model specification is proposed. This specification is useful as the model’s flexible nature allows it to be able to adapt to the characteristics of a new election. Furthermore, results of previous elections can be incorporated by adapting the prior specification. The results show that the DPM model improves on the forecast of the benchmark election models. Using the outcome of the DPM model applied to earlier years, forecasts of present elections can be further improved.