Well, a simple average is a pretty reasonable ways to make predictions when you have no better information... like a regression model with 0 predictors.

But really the idea is more about a feedback mechanism in which successful predictors those who are successful at making predictions end up with more money to invest back into the market making more predictions. So its a weighted average of the masses, in which those who are successful predictors get weighted more heavily in virtue of their having more financial capital with which to invest. For that to work well, you really need a reasonable level of participation. Also, if the inputs from new participants continually outweighs those from experienced quality participants, you are not going to get good predictions. Also, you have to have a rock-solid mechanism to decide the market, or it will be gamed.

You should visit the good judgment project website. They've studied this for a while and have documentation outlining and explaining their claims.

The idea is not that an untrained mass of people are necessarily better than experts. Instead, it's that some people, with training, can learn how to make good predictions about many different fields. These predictions exceed the predictions of domain experts.

Sorry, no citation handy, and this might be semi-tangential (more "wisdom of he crowds" per se), but I recall a study that reviewed those "guess how many jellybeans are in this jar" -type contests, in which the guesses' median value converged with the correct answer...