All Men Count with You, but None Too Much: Information Aggregation and Learning in Prediction Markets.

Abstract

Prediction markets are markets that are set up to aggregate information from a population of traders in order to predict the outcome of an event. In this thesis, we consider the problem of designing prediction markets with discernible semantics of aggregation whose syntax is amenable to analysis. For this, we will use tools from computer science (in particular, machine learning), statistics and economics. First, we construct generalized log scoring rules for outcomes drawn from high-dimensional spaces. Next, based on this class of scoring rules, we design the class of exponential family prediction markets. We show that this market mechanism performs an aggregation of private beliefs of traders under various agent models. Finally, we present preliminary results extending this work to understand the dynamics of related markets using probabilistic graphical model techniques.

We also consider the problem in reverse: using prediction markets to design machine learning algorithms. In particular, we use the idea of sequential aggregation from prediction markets to design machine learning algorithms that are suited to situations where data arrives sequentially. We focus on the design of algorithms for recommender systems that are robust against cloning attacks and that are guaranteed to perform well even when data is only partially available.