How Accurate are Prediction Markets?
In class recently we have been talking about prediction markets, which use the aggregate information provided by many betters and their beliefs to predict the outcome of events. A very simple example of such a market is a horse race. Betters come to the race with some belief about who will win, bet some proportion of their available wealth on who they believe will win, and either win some payout or nothing depending on the true outcome of the race. As more betters participate, the odds that any given horse will win are updated according to the portion of total bets placed on them. A larger example of a prediction market is the Iowa Electronic Market which, among other things, runs an online futures market where participants can bet on the outcome of presidential elections. Prediction markets are often used to, well, predict things, but how accurate are they really?
In 2012, the Iowa Electronic Market ran a futures market to predict vote shares for the Republican and Democratic presidential nominees. In the days before the popular vote came in, the market predicted a roughly 51% / 49% split of the votes, with Barack Obama winning. The real results gave Obama the win with 52.1%, Mitt Romney for the G.O.P came in at 46.5% of the vote, with the remaining votes split among the other candidates. The Iowa Electronic Markets have successfully predicted outcomes on many other occasions.
In a paper published in the International Journal of Forecasting, professors Berg, Nelson and Rietz of the Tippie College of Business, University of Iowa argue that prediction markets are notably more accurate than polls at predicting presidential election outcomes. The authors compared the Iowa Electronic Market predictions to 964 polls over five presidential elections. They had four major findings. First, the prediction market gives a noticeably different picture of the race as a whole, and is often significantly different from polls for long periods of time. During these periods, the market tends to be much closer to the final outcome than polls. Second, the prediction market does not experience “convention bounce,” where one party rises in the poles during its convention and falls after it ends. Third, the market tends to predict outcomes better than polls. Finally, polls are significantly more volatile than the prediction market, and much more subject to day-to-day shifts in voter sentiment.
The third point above is perhaps the most significant. Analyzing this data, the authors found that the prediction market is closer to the actual outcome than are polls 74% of the time. Additionally, they found that prediction market “significantly outperforms the polls in every election when forecasting more than 100 days in advance.” Here we have our answer then. In almost three out of every four cases, prediction markets are more accurate than polls, the next-best prediction method we have. And in long term forecasts, prediction markets tend to be significantly more accurate than polls. Of course, these findings only answer the question of relative accuracy of prediction markets rather than absolute accuracy. Perhaps another post will answer the question of absolute accuracy, but for now, we have an answer. How accurate are prediction markets? Better than polls.
Sources:
Iowa 2012 Presidential Election
http://www.politico.com/2012-election/results/president/iowa/
https://iemweb.biz.uiowa.edu/graphs/graph_Pres12_VS.cfm
Berg, Nelson and Rietz:
http://www.sciencedirect.com/science/article/pii/S0169207008000320