Prediction markets, when they work well, solve a fundamental problem: how to aggregate individual beliefs into a meaningful quantitative estimate of the probability that a given event will occur. They also provide incentives for people to disseminate privately-held information. I will describe one way to help these markets work better: incorporating a learning agent who provides liquidity, called a market maker. Along the way, the design of this agent raises and solves some fundamental problems in reinforcement learning and Bayesian reasoning. I will also discuss the deployment of this market-making agent in two different settings with human participants. One of these settings is a novel experiment for comparing market structures. Another one, the RPI Instructor Rating Market, allows students to trade on the ratings their professors will receive, thus providing dynamic feedback to instructors on the progress of their classes; we find that market prices are, in fact, better than past ratings at predicting future ratings.
Joint work with Aseem Brahma, Mithun Chakraborty, Allen Lavoie, Malik Magdon-Ismail, and Yonatan Naamad.

]]>