Artificial markets and intelligent agents

In many studies of market microstructure, theoretical analysis quickly becomes intractable for all but the simplest stylized models. This thesis considers two alternative approaches, namely, the use of experiments with human subjects and simulations with intelligent agents, to address some of the limitations of theoretical modeling. The thesis aims to study the design, development and characterization of artificial markets as well as the behaviors and strategies of intelligent trading and market-making agents. Simulations and experiments are conducted to study information aggregation and dissemination in a market. A number of features of the market dynamics are examined: the price efficiency of the market, the speed at which prices converge to the rational expectations equilibrium price, and the learning dynamics of traders who possess diverse information or preferences. By constructing simple intelligent agents, not only am I able to replicate several findings of human-based experiments, but I also find intriguing differences between agent-based and human-based experiments. The importance of liquidity in securities markets motivates considerable interests in studying the behaviors of market-makers. A rule-based market-maker, built in with multiple objectives, including maintaining a fair and orderly market, maximizing profit and minimizing inventory risk, is constructed and tested on historical transaction data. Following the same design, an adaptive market-maker is modeled in the framework of reinforcement learning. The agent is shown to be able to adapt its strategies to different noisy market environments. Thesis Supervisor: Tomaso Poggio Title: Uncas and Helen Whitaker Professor of Brain and Cognitive Sciences Thesis Supervisor: Andrew W. Lo Title: Harris and Harris Group Professor