By election day, the markets with an average absolute error of around 1.5 percentage points, were considerably more accurate than the Gallup poll projections, which erred by 2.1 percent.
Prediction markets -- also known as information markets or events futures -- first drew widespread attention in July 2003 when it was revealed that the Pentagon's Defense Advanced Research Projects Agency (DARPA) was establishing a Policy Analysis Market to allow trading in various forms of geopolitical risk, including economic and military scenarios. The objective was to discover whether trading in such contracts could help predict future events. Bowing to a storm of criticism that it was proposing "terrorism futures," DARPA dropped the program. But other prediction markets, dealing with everything from sports and entertainment to elections and finances, have emerged and gained growing interest and participation.
In Prediction Markets (NBER Working Paper No. 10504), authors Justin Wolfers and Eric Zitzewitz describe the types of contracts that might be traded in prediction markets and then survey several applications, with special attention to market design issues. Finally, they assess the predictive value of such markets.
Wolfers and Zitzewitz begin by noting that much of the enthusiasm for prediction markets derives from the efficient markets hypothesis. In a truly efficient prediction market, the market price will be the best predictor of the event, and no combination of polls or other information can be used to improve on the market-generated forecasts. Wolfers and Zitzewitz do not insist that prediction markets are literally perfectly (or fully) efficient; however, they acknowledge that a number of successes in these markets, both within firms and with regard to public events such as presidential elections, have generated substantial interest among both political and financial economists.
In a prediction market, the researchers note, payoffs are tied to unknown future events, and how the design of how the payoff is linked to those events can elicit the market's expectations of many things. A in a "winner-takes-all" contract, for example, the contract costs a specific amount and pays off a specific amount, and only pays if a specific event occurs, such as a particular candidate winning an election. The price on a winner-takes-all market represents the market's expectation of the probability that an event will occur. By contrast, for in an "index" contract, the amount that the contract pays varies in a continuous way based on a number that rises or falls, like the percentage of the vote received by the candidate. Finally, in "spread" betting, traders bid on the cutoff that determines whether an event occurs, like whether a candidate wins more than a certain percentage of the popular vote.
The various types of contracts may reveal the market's expectation of a specific parameter: a probability, a mean, or median, respectively. But prediction markets also can be used to evaluate uncertainty about these expectations, for example a family of winner-takes-all contracts that pays off only if the candidate earns 48 percent of the vote, 49 percent, 50 percent and so on. This family of winner-takes-all contracts then will reveal almost the entire probability distribution of the market's expectations. A family of spread-betting contracts can yield similar insights.
With these factors in mind, Wolfers and Zitzewitz examine the data compiled from analyses of the University of Iowa's Iowa Electronic Market, which has offered trade on presidential election contracts since 1988. Charting the price bids for the past four presidential elections, the data show that as election day drew nearer, the prediction markets' projected candidate vote shares grows more accurate. Prediction markets also beat appeared better calibrated than independent analysts on the probability of the ouster of Saddam Hussein. The Hollywood Stock Exchange likewise has proved highly accurate in predicting opening weekend box office success and Oscar winners.
Even some prediction markets with very small participation have shown striking results. An internal market at Hewlett-Packard produced more accurate forecasts of printer sales than did the firm's internal processes, and at Siemens an internal market predicted the firm would definitely fail to deliver a software project on time, even as traditional planning tools said the deadline could be met. In each firm, the traders numbered only between 20 and 60 employees.
Wolfers and Zitzewitz maintain that the success of prediction markets, like all markets, depends on their design and implementation. Some key design issues include: how buyers are matched to sell traders; the specification of the contract; whether real money is used (some prediction markets operate for entertainment purposes and use make-believe currency); and the kind of information available to provide a basis for trading. When such factors are weighed judiciously, Prediction markets are better at pricing some events than others. The markets, like many individuals, are not always well calibrated on small probability events. In addition, markets on complex events, or events where there is likely to be inside information, often fall to attract sufficient liquidity.
Wolfers and Zitzewitz conclude with cautious optimism. They believe that prediction markets are extremely useful for estimating the market's expectation of certain events. Simple market designs can elicit expected means or probabilities, while more complex markets can elicit variances, and contingency markets can be used to elicit the markets' expectations of covariances and correlations.
Prediction markets have their limitation, the researchers caution, but they may be useful as a supplement to more traditional means of prediction, such as opinion surveys, expert panels, consultants, and committees.
-- Matt Nesvisky