Prediction markets are used in real life to predict outcomes of interest such as presidential elections. In this work we introduce a mathematical theory for Artificial Prediction Markets for the purpose of supervised classifier aggregation and probability estimation. In this direction, we bring the following contributions. First, we derive the market equations starting from the total budget conservation condition, show the market price uniqueness and give efficient algorithms for computing it. Second, we present a method for classifier aggregation using the market price as the estimated conditional probability given the evidence presented to the market participants though a feature vector x. We also show how to train the market in a supervised manner, updating the participants budgets based on the market price and the amount they bet. Third, we introduce classifier specialization as a new type of differentiating characteristic of classifiers. Finally, we present an application to multi-class classification using random decision rules as specialized classifiers and show that the prediction market consistently outperforms Random Forest on an array of datasets with Bayes errors ranging from 0 (very easy) to 0.5 (impossible). UCLA Department of Statistics 2009-09-14T10:16:17-07:00 2009-11-10 Statistics true 4:00 PM 417 4660 Geology Bldg. Florida State University false true UCLA Department of Statistics Seminar Adrian Barbu 3:00 PM 2009-11-04T07:05:01-08:00