Order this book
International Business School, Brandeis University
As one of the earliest agent-based approaches to financial models it has served as a kind of benchmark model inspiring many others. It is in the class of agent-based models which are highly computational, using many tools from machine learning to model the activities of learning agents. It set these learning agents into a relatively simple economic environment and explored the dynamics of prices, trading volume, and their responses to certain key parameters. From the early stages of model building the research team realized that they would never be able to completely explore all parameter settings, or model specifications, in our limited runs done for publication. The hope was to lay out a kind of computational platform for others to explore. History has shown that for most people this was an overly optimistic dream. The general complexity of the model, and daunting amounts of computer code that one needed to understand the model kept most researchers away. Norman Ehrentreich was one of the daring few to take on the model, and he has summarized his work and findings in this excellent book.
This book is much more than just an analysis of the SFI market. It is a useful primer for anyone interested in getting started in the area of agent-based finance. The early chapters give the reader a background on agent-based modeling, and learning models in economics. There is also an excellent chapter on some of the important empirical puzzles that are well known from the financial econometrics world. Many of these empirical features have influenced research in agent-based financial modeling, so this chapter is a much needed part of the book.
In the second part of the book Ehrentriech goes on to carefully test, and critique 3 parts of the SFI market. In the process of doing this he gets his own version of the software up and running, and replicates most of the features of the original market. He explores three different areas which include sensitivity to the design of the mutation operator, genetic drift, and finally the impact of wealth in evolutionary systems.
In an important set of experiments he implements modifications to the mutation operator in the SFI market. This changes the way in which new and novel rules are presented to agents as they are considering how to modify their behavior. He presents a strong case for why this learning mechanism might make more intuitive sense than the one used in the original market. The results with the new operator show a dramatic change in the learning outcomes. The original SFI market experiments showed a strong bias toward agents using technical trading strategies. His new results show that this feature of the original model was sensitive to the type of mutation operator considered, and there may be a broad set of parameters for which convergence to the rational expectations equilibrium is possible.
In a later chapter (chapter 9) the attraction of rules to the no technical trading situation is further explored. It is shown that much of this pull may be due to genetic drift as opposed to a clear definitive rejection of technical trading. There are nice experiments that greatly improve on tests done in the early SFI paper. The most important of these shows that the forecast errors for the technical trading rules do appear to be lower than for other benchmark rules, indicating that something interesting is probably going on. Ehrenttriech's conclusions to this chapter are very important, and should be noted by anyone doing research in the field. From his runs he's worried that fitness signals are weak in the world of noisy financial data, and that changes in strategies may be more dominated by drift than purposeful selection. I think this is a very important point for agent-based modelers to remember. Agents will continue exploring all strategies over time unless there is a significant fitness penalty attached to technical trading. In the end, noisy financial data may not put much fitness pressure on any particular set of strategies to survive or die off, leaving a diverse pool of trader activity. For me this conclusion seems to resonate well with both simulation and real world experience.
A little known problem with the SFI market, and many other agent-based approaches is that they ignore long range wealth dynamics. Traders' fitness is often tied to short range forecast accuracy, or utility measures, and not long range wealth growth. Several nice experiments show that the growth and accumulation of wealth may be related to many things in the model that are difficult to interpret. This is because the objective ignores long range wealth, and one can't really make statements about the rationality or lack of rationality and how it is connected to eventual wealth accumulation. Finally, the constant absolute risk aversion preference structure demands that agents position be independent of wealth. Several others (including myself) have pointed out these problems in the SFI structure, but this chapter is a very complete summary of these problems along with some extensive experiments.
I often think of agent-based market simulations as consisting of three levels. At the lowest level are detailed questions about software design and computer language platforms. At the highest level are general questions about economic structure of the model, equilibrium, and generic dynamics. In between there is a middle realm where questions pertain to the details of the learning dynamics. This area has often been under explored. Ehrentreich's work is a good example of this area. It is a complete analysis of some of the mechanisms behind the SFI market. It is much more detailed than any of our earlier papers were. It is essential reading for anyone interested in the dynamics of the SFI market in particular, but I also recommend it for others as a useful resource on agent-based financial market design as well. My only criticism is not of his work, but of the SFI platform itself. Reading the book reminds one just how complex this market is. My general hope (and research goal at the moment) is that we can eventually settle on simpler platforms which still offer the rich evolutionary dynamics of the SFI market with fewer moving parts.
Return to Contents of this issue
© Copyright Journal of Artificial Societies and Social Simulation, 2009