The stakes are too low. People betting on prediction markets are typically playing with small amounts of money (compared with big traders in capital markets) or with fake money. Therefore, one theory holds, they lack sufficient incentive to ensure that their calls are rigorous—they are lazy bettors.
This will ring true with anyone who's ever played online poker for fake money. The lack of any stings attached to losing leads players to stay in hands holding terrible cards; it brings down the quality of play and rewards reckless bets. A similar dynamic is at work in prediction markets, which is why conventional wisdom has long held that real-money prediction markets are more reliable than fake-money ones.
Yet that conventional wisdom has been challenged this year. InTrade is a real-money site, and it failed to accurately predict the California primary. Neither the play-money Hollywood Stock Exchange nor InTrade correctly predicted that Tilda Swinton would win the Academy Award for best supporting actress, but, somewhat surprisingly, the play-money site at least had Swinton in second place. And the incentives argument certainly can't explain recent failures in prediction markets, because nothing has changed: They offer the same minor incentives as they always have.
They're too slow to react to events. David Leonhardt made this argument in the New York Times. Citing Barack Obama's ups and downs in the Intrade market after the contests in Iowa (he won and went up), New Hampshire (he lost and went down), and South Carolina (he won and went up again), Leonhardt asserts that "the impact of each contest took surprisingly long to sink in."
Which only raises the question: compared with what? Leonhardt's yardstick is the stock price of drug maker Schering-Plough, which appears to react to market news almost instantaneously. That's true, but it's an individual equity, not a futures contract scheduled to expire at a prearranged date. As such, few would argue that its current price is intended to predict anything in particular about the company eight months down the road. One might just as usefully compare the political-prediction market with the prices of yachts or Beanie Babies. Moreover, there are counter-examples in which political futures seem to react pretty rapidly: On the day of the New Hampshire primary, a contract that Hillary Clinton would be the Democratic nominee traded as low as 18.1 cents and as high as 58.7 cents on the Iowa presidential market.
None of this is to suggest that all of the above theories are wrong, but they are each incomplete and unproven. I suspect that a large part of the problem has to do with expectations: Because prediction markets have been more accurate than other prognostications in the past, we crave them to be perfect. They aren't, and so long as they are built on bets in which no one has perfect information—like Oscar winners and political candidates—they never will be. Every bettor's perception of who will win is filtered through a myriad of inputs: a sense of who is the leader and who is the underdog; the influence of other people's bets; even something as mundanely human as whom we want to win. To get the maximum use out of them, we must—as with political polls—learn to read them in a discriminating, critical fashion. This year, that process seems to have begun.