Heterogeneous Beliefs, Intelligent Agents, and Allocative Efficiency in an Artificial Stock Market

Various studies of asset markets have shown that traders are capable of learning. In this paper we replace human traders with artificial-intelligent software agents in a simulated stock market. They make predictions about the future, randomly submit their quotes, and transact at certain price. A simplified double auction market mechanism is employed. Three types of agents are included, value traders, momentum traders, and noise traders. Value traders form future expectation with an artificial neural network (ANN). They use ANN to predict dividend growth rate. Three computational experiments in terms of market participants are conducted. Time series from this market are analyzed from the standpoint of well-known empirical features in real markets, including GARCH, CAPM, and EMH tests. I extend earlier research in three ways. First, I employ a double auction market mechanism. I use this mechanism for two reasons. First, major real financial markets are organized as double auctions. Second, laboratory double auctions with human traders are known to yield data that approximate the equilibrium predictions of economic theory in a variety of environments. My second contribution is that I do not focus solely on equilibrium selection and convergence. I emphasize the behavior of the learning and market dynamics themselves. I analyze the portfolio returns and stock returns from this market to see whether the market exhibits characteristics cited in the empirical literature, including volatility persistence, GARCH, portfolio performance evaluation and allocative efficiency. The third extension is that I introduce the noise trader. The noise trader does not act strategically, but rather randomly posts a market order, and the resulting trade quantity is a randomly distributed variable.