How well do markets forecast major event outcomes?

Many different sources of information can be used to forecast major event outcomes, such as national elections. Of these, much attention has been paid in the published academic literature to the relative predictive power of statistical models, opinion polls and expert opinion. More recently, a great deal of attention has been devoted in the literature to the role that prediction markets can play in adding forecasting accuracy to the range of other established methodologies. In particular, there has been an increasing focus on the value of de-biasing polls and prediction markets for established time horizon biases, and comparing predictive power at various stages in the run-up to polling day and on election day itself. This paper seeks to build on the established literature using data collected from US presidential elections since 2004 and also the UK general election of 2015, to arbitrate between the relative accuracy of these forecasting methodologies at different time horizons, as well as to identify any systematic biases. The data is collected from various prediction markets as well as polling archives, and records of the forecasts derived from statistical models and expert judgment. On the basis of established literature, we might expect that prediction markets outperform polls at a distance from the election but that any advantage diminishes or even reverses at the latest stages of the cycle. We seek to test this using our new data sets as well as to assess the value of combining forecasts in maximizing predictive power. Our results will add to the growing debate in the literature about the value of combining alternative methodologies in improving forecasts of major event outcomes.