The Temptation of Emergence or: Don't Rush into Economic(al) Explanations

One of the pioneering stock market simulations, the Santa Fe Institute Artificial Stock Market (SFI-ASM), showed an influence of learning speed on the aggregate outcome. For slow learning rates, the neoclassical properties of a homogeneous rational expectations equilibrium (hree) could be confirmed. However, a complex regime with higher price volatility, GARCH-behavior, cross-correlation between price and trading volume, and significant levels of technical trading emerged for faster learning speeds. In Ehrentreich (200x) it was shown that these results are based on the specific design of the mutation operator which introduced an upward bias in the level of set trading bits in the classifier system. Since agents in a corrected version endogenously gave up the use of their classifier system, it was implied that it did not provide any profitable trading information. This is in contrast to the original SFI-ASM. Joshi, Parker, and Bedau (1998) found that agents with access to technical trading bits did significantly better with respect to wealth levels than those without. In their 2002 study they found that faster learning agents outperformed slow learning agents. A sensitivity analysis of wealth levels with both the original and the corrected mutation operator was able to replicate the Joshi, Parker, and Bedau results, however only for specific parameter settings. It was revealed that differences in wealth levels are a result of the size of the active rule set that an agent possesses. It was a peculiarity of design in the SFI-ASM that trading rules were not always logical. The more trading bits were set, the higher the fraction of illogical trading rules. For faster learning rates, the number of logical and thus potentially activated trading rules decreased significantly. This had an effect on the efficiency of the selection mechanisms (Select Best and Roulette Wheel) that were used to determine the rules on which agents acted upon. Contour plots show that the differences in wealth levels between two agent types, who differ in the size of the rule set and learning speeds, can be positive or negative, depending on the specific parameter combination. Any economic interpretation of differences in wealth levels are thus too hasty, since these differences are again a result of technicalities in the model, unless one is willing to interpret cognitive or economic substance into the working of the selection mechanisms. Unexpected results that are hard to explain on the basis of the constituting elements in a simulation model pose a general temptation to explain them as emergent behavior, yet they can still be a result of unsuspicious programming details. Since the two selection mechanisms are widely used in the agent-based community, the necessity of having a constant number of active strategies over the course of the simulation is demonstrated. Because of this, the above analysis could be of general interest to the agent-based simulation community.