Prediction and Error

In this dissertation project, I make three separate contributions on how we can improve aggregate election forecasting models with respect to modeling choices, interpretability, and performance. Two of the three papers are applications to particular cases, the U.S. and France specifically, while the third points to a cross-national pattern in polling errors. The first paper addresses how we can make more reasonable prior choices for key parameters – such as the variability of non-sampling error – by using past pre-election polls. I showcase this approach on U.S. presidential elections. The second paper shows how to create and aggregate predictions in a multi-party contest while keeping the individual forecasts intact. This is useful to see convergences or divergences in the forecasts which might affect our confidence in the aggregate prediction. I develop a new aggregate forecasting model for French presidential elections to demonstrate this idea. The last paper shows and investigates a pattern in polling errors. We see that across multiple countries and electoral systems, polling errors favor the lesser party in two-party contests, i.e. polling errors favor Democratic candidates in Republican states and vice versa. We demonstrate a simple adjustment procedure based on this pattern to reduce the mean absolute polling error. We achieve a 16% reduction in the 2016 U.S. presidential election.