I have to admit, as an economist I should love the idea of prediction markets: a free market in which an actual monetary stake is taken in the outcome should tap the 'wisdom of crowds' and result in startlingly accurate predications. But to be honest, I believe that markets are more often than not driven by fads, speculation and, especially in the elections markets, very incomplete information leading to predications that aren't much better than any other prediction. David Leonhardt of the New York Times has an excellent article that suggests I may be right in my pessimistic attitude.
The basic idea is this: we all have biases and make mistakes, but gather enough of us together and on average we will converge to the true mean (think of is as a 'law of large numbers of people'). But if we all make the same mistakes (too influenced by, say, an article in the NYT) or are influenced by each others behavior, we can go badly wrong. Take the last two bubbles, the dot com bubble and the housing bubble, if we are all so smart why is there a bubble?
Prediction markets may also have investors that may have ulterior motives - they support a candidate and want to improve their showing in the prediction markets so that others will think they are successful and be less reticent to support them (creating a self-fulfilling prophecy).
Still, in the end, large samples do have a tendency to minimize the individual biases if they are not all in the same direction, so these prediction markets are almost always better than polls.