Simulating prediction markets that include human and automated agents

Abstract

In this work I study the interaction of sophisticated trading agents with simpler agents in a prediction market. The goal is to simulate markets with both human and computer agents, and investigate ways to maximize the performance of these markets. I start with the neural net-based agent that is currently used in CCI's collective prediction experiments on football plays. By tuning their training and risk affinity, I configure a "smart" agent to represent the sophisticated computer traders. I implement three types of simple agents to approximate human traders - two are rule based, and one uses aggregate human data from lab experiments. By exploring different combinations of smart versus simple agents, I showed that it is possible for mixes of agents to outperform either types alone. This result is consistent with the larger goal of the collective prediction project, which is to show that humans and computer agents combined in a prediction market can do better than either alone.