This is a selection of readings from a conference at New York University. Since I'm reviewing this for JASS rather than a finance journal, my objective is not to summarize its wide-ranging content, but rather to focus on the part most relevant to simulation and artificial societies.
The book is divided into seven parts, one of which is called "Heterogeneous Agents." In this section, there are three very interesting papers that simulate various aspects of agents and markets. The first paper concerns a project which is close to my heart, the Santa Fe Artificial Stock Market. In their paper, "Technical Trading Creates a Prisoner's Dilemma: Results from an Agent-Based Model," Joshi, Parker, and Bedau describe simulation model originally proposed by Palmer, Arthur, Holland, LeBaron, and Taylor. They explore the question of whether agents will gravitate to the use of investment strategies which rely on trends rather than fundamentals. The agents are "bit forecasters," which means they take information about the state of the world as a string of 0's and 1's, then they select the bits they are interested in and formulate a price prediction. These agents are supposed to be boundedly-rational, able to comprehend only a fragment of the data that exists out in the "real world." The simulations show that agent rules tend to put more emphasis on bits that reflect trends and noise, rather than fundamental stock values. Agents trade on trends. That makes all investors worse off, as the observed price is noisy and less meaningfully reflective of the underlying value of the stock. This is a very interesting paper, and if you are interested in digging into the ASM code base on your own, I have been updating and cleaning it up to run with the current version of Swarm, an agent-based modeling library. For more information, consult http://ArtStkMkt.sourceforge.net. That site has information about the model, the current revision of it, and access to code snapshots and a cvs repository.
Goldbaum's paper, "Cycles of Market Stability and Instability Due to Endogenous Use of Techical Trading Rules," has themes in common with the previous paper. The implementation is different, however. This is not an agent-based simulation in the same sense. The approach is rather more reminiscent of evolutionary game theory. There are not individual agents making decisions. There is rather a pool of agents whose actions follow a pattern that reflects the fitness of competing strategies. There are 3 kinds of information signals, two of which reflect the underlying value of a stock and a third which reflects a technical trading rule (a forecast based on a moving average of stock prices). The proportions of agents who buy/sell according to the signals are assumed, in the aggregate, to reflect the probabilities implied by a categorical choice model (multinomial logit). So, for example, if the model says the probabilities of selecting the information sources are given by (0.3, 0.4, 0.3), then the proportion of agents using the first rule is 0.3, and so forth. Demand is inferred from this division of strategies, the fortunes of the agents who follow each signal are adjusted, and then the process is repeated. Under some parameter settings, this model produces a pattern which I find to be most interesting. The market trends upwards as trend-based investing grows in popularity. The bubble eventually bursts and the market plummets, and trend-based strategies fare poorly and they fall into disfavor. Then, as the price grows again, the mechanics prosper again, and the process repeats itself.
The paper by Taghu, Rao, and Sen, "Relative Performance of Incentive Mechanisms in Delegated Investments: A Computational Study," investigates the famous principal-agent problem that arises in delegated investment management. Several investment agents work for a single principal, and the principal's problem is to design a contract which entices the agents to work hard to improve the principal's return. Although this paper is included in the section on heterogeneous agents, the investment managers are all assumed to labor under identical contracts and utility functions. The simulation model allows the agents to accumulate information with which to estimate the underlying value of stocks and (possibly) to invest wisely on behalf of the principal. The model yields conclusions that will be comforting as well as disconcerting to the people who manage my investment portfolio. On the comforting side, the model shows that if the contract limits the agent's liability on the down-side (meaning that they share in gains, but bear none of the losses), then the performance of the portfolio is enhanced. Managers like that, of course. On the disconcerting side, principals who give their agents a smaller share of the profit are tend to observe better in stock performance. The obvious recommendation: people who hire portfolio managers don't need to offer them a very big share of the pie, but they have to make up for that by letting the agents lose their money without sharing any of the risk.
The rest of the papers in this book are quite interesting, but they are not exactly about simulation of artificial societies. Rather, they are mostly written from the more standard applied finance perspective, looking for patterns in stock market returns , improving methods to calculate various theoretically-important quantities, or devising new methods to trade and evaluating them against historical data (in time series or case study). Most of this work will be recognizable to social scientists, who often posit theories which imply patterns in data, patterns (or parameters) which are to be discerned through statistical methods.
The importance of these other papers came clear to me while I listened to presentations at a recent colloquium that was hosted by the U.S. National Academy of Sciences ( http://national-academies.org/nas/colloquia ; look for the October 4-6 meeting entitled "Adaptive Agents, Intelligence and Emergent Human Organization: Capturing Complexity through Agent-Based Modeling"). One of the presentations, by Charles Plott, presented experimental evidence that human traders who are told the value of a stock behave in a way that is consistent with the theory of demand: they try to buy and sell low, but the price adjusts toward at a market clearing price. Another presentation, by Jeffrey Kephart, showed that, in the labs at IBM, they have developed software for agents that buy and sell stock in an experimental market, and that these artificial agents typically generate higher investment profits than humans who are competing with them. The suggestion was that human traders are likely to be replaced by software in the near future.
I don't really have much doubt that a finely honed computer algorithm can earn higher profits than a human who is told to buy low/sell high. ( I can't beat artificial agents in chess or bridge, I hasten to admit.) The experimental studies described in the previous paragraph begin at a late stage in the investment scenario: they assume that each agent has decided that an asset is worth X and the only decision is how much to pay or what to charge. Without trivializing that part of the problem, one should note that our understanding of this brave new stock market can be complete only if we consider where agent expectations and valuations come from. There are many papers in this volume that propose alternative methods of making these calculations against "real world" data. Rather than viewing these papers as depictions of optimal rational action (representative of real people) or information gathering by humans, we can take them as proposed strategies of artificial agents in a stock market of the future. Some interesting twists of emphasis are likely to emerge. For example, in existing artificial stock markets, there has been an emphasis on developing models of "realistic," "boundedly rational" investors. If we expect the real market to be driven by software in the future, that point of emphasis will probably receive less weight. The literature on empirical tools to analyze markets will be changed as well, because the behavioral foundations on which these models lie might be uprooted. Models of human markets typically assume stable underlying patterns of behavior. We expect that changes of fashion in stock picking can be relatively slow, so observations might shed light on the underlying reality. If software agents start using (very expensive, proprietary) software to develop valuations and buy and sell, we won't necessarily have that element of comfort.
Paul E. Johnson
Dept. of Political Science
University of Kansas