Accuracy, bias and precision: Comparing forecasting methodologies along three dimensions.

Forecasting open-door and closed-door event outcomes is a popular activity, and with reason: large event outcomes can have significant economic impacts. As such, it is economically important, as well as of academic interest, to determine the forecasting methods that have performed best historically. In this paper we compare forecasts of election outcomes from a number of prediction markets and surveys of opinion over recent years. Opinion polls are surveys of the voting intentions of a sample of voters, while prediction markets allow participants to trade contracts whose value is contingent on some outcome occurring. We consider all opinion polls from two common aggregators of polling information – Pollster and Real Clear Politics - and look at a number of well-known prediction markets. While polls in their present form can be dated back only about 80 years, prediction markets have a longer history, offering a body of evidence dating back more than 500 years. This paper contributes to this growing body of evidence. Even so, it is their internet-based electronic variants that have been attracting the most attention in recent years, and these are our focus. In particular, we examine the hypothesis that decentralized prediction markets provide better forecasts of outcomes than more centralized surveys of opinion. Nonetheless, there is published evidence that corrected polls – allowing for historical incumbency advantage and number of days to the election - can perform comparably well. In line with this, we find that our corrected polls do indeed exhibit the least bias amongst our forecasts, yet they seem generally less accurate and less precise. In deriving these results, we employ a Mincer-Zarnowitz regression methodology. We also note that the larger prediction markets perform better than smaller markets and this appears to be borne out on each dimension. A number of previous comparison studies between prediction markets and polls rely on data from a single prediction market rather than from the wider selection that we are able to consider. We believe that this may be driving the divergence from previous results within the literature.