Do party leadership contests forecast British general elections?

Abstract

When assessing election forecasts, two important criteria emerge: their accuracy (precision) and lead time (distance to event). Curiously, in both 2010 and 2015 the most accurate forecasts came from models having the longest lead time—albeit at most 12 months. Can we increase the lead time further, supposing we tolerate a small decrease in accuracy? Here, we develop a model with a lead time of more than 3 years. Our Party Leadership Model relies on the votes of MPs when selecting their party leader. We assess the forecasting quality of our model with both leave-one-out cross-validation and a before-the-fact forecast of the 2019 general election. Compared to both simple forecasting methods and other scientific forecasts, our model emerges as a leading contender. This result suggests that election forecasting may benefit from developing models with longer lead times, and that party leaders may influence election outcomes more than is usually thought.