Election Forecasts Using Spatiotemporal Models

Abstract

There exists a large literature on the problem of forecasting election results. But none of the published methods take spatial information into account, although there is clear evidence of geographic trends. To fill this gap, we use geostatistical procedures to build a spatial model of voting patterns. We test the model in three close elections and find that it outperforms rival methods in current use. We apply kriging (a spatial model) and cokriging (in a spatiotemporal model version) to improve the accuracy of election night forecasts. We compare the results with actual outcomes and also to predictions made using models that use only historical data from polling stations in previous elections. Despite the apparent volatility leading up to the three elections in our study, the use of spatial information strongly improves the accuracy of the prediction. Compared with forecasts using historical data alone, the spatiotemporal models are better whenever the proportion of counted votes in the election night tally exceeds 5%%.