Testing for Non-Linear Structure in an Artificial Financial Market

We present a stochastic simulation model of a prototype financial market. Our market is populated by both noise traders and fundamentalist speculators. The dynamics covers switches in the prevailing mood among noise traders (optimistic or pessimistic) as well as switches of agents between the noise trader and fundamentalist group in response to observed differences in profits. The particular behavioral variant adopted by an agent also determines his decision to enter on the long or short side of the market. Short-run imbalances between demand and supply lead to price adjustments by a market maker or auctioneer in the usual Walrasian manner. Our interest in this paper is in exploring the behavior of the model when testing for the presence of chaos or nonlinearity in the simulated data. As it turns out, attempts to determine the fractal dimension of the underlying process give unsatisfactory results in that we experience a lack of convergence of the estimate. Explicit tests for non-linearity and dependence (the BDS and Kaplan tests) also give very unstable results in that both acceptance and strong rejection of IIDness can be found in different realizations of our model. All in all, this behavior is very similar to experience collected with empirical data and our results may point towards an explanation of why robustness of inference in this area is low. However, when testing for dependence in second moments and estimating GARCH models, the results appear much more robust and the chosen GARCH specification closely resembles the typical outcome of empirical studies.