As I've noted before, prediction markets are not a panacea. They do some things much better than any existing alternatives. In other cases they simply aggregate stupidity more quickly and efficiently, as Barry Ritholtz points out over at Big Picture with regards to the Michael Jackson verdicts. I'm not surprised - that the markets were volatile and in the end wrong, that is.
Divining the discrete actions of a small, opaque group with no previous track record is one kind of problem for which prediction markets are poorly suited. Trial verdicts, panel-driven game-show outcomes, regulatory rulings and even the election of the pope are good examples. The markets may be right reasonably often on such questions, but they are often spectacularly wrong.
By contrast, problems with highly distributed information and no clear locus of control (e.g., public health, the behavior of customers and the outcome of elections) are of a completely different nature. In those cases, thousands if not millions of people whose observations are not filtered through a handful of final decision-makers have something to contribute. They are first-order rather than second-order actor-observers. Markets are the best way I know of bringing together that hyper-distributed knowledge into a point prediction.