Predicting Presidential Election Outcomes from What People Watch

In a recent article by Barfar and Padmanabhan (2015), we demonstrated how television viewership data could predict presidential election outcomes in the United States. In this article, we examine predictive models using a snapshot of Nielsen's national data on television viewership. The study is conducted with high-dimensional low sample size (HDLSS) data, whereby we conduct a comparative analysis with and without feature reduction on the data from the 2012 elections. We find that simple "single-show models" often provided more insights and predictive accuracies than models from feature reduction. Second, beyond the state and county levels of analysis, we show that the results continue to hold at the designated market area (DMA) level, crucial for television broadcasting because programs are often targeted at the DMA level. Finally, we examine the performance of the single-show models in the 2016 election season by applying them to the viewership information during the U.S. presidential primaries. We discuss implications of our findings for research and practice.