Learning is neither sufficient nor necessary

Financial time series data exhibits long memory phenomena, where certain behaviours in the market have a persistent influence on the market over time. It has been suggested that imitation of successful trader strategies by other less successful traders is an important factor in contributing to this persistence. We test this explanation by using an existing adaptive agent-based model and we find that the robustness of the model is directly related to the dynamics of learning; models in which learning converges to a stationary steady state fail to produce realistic time series data. In contrast, models in which learning leads to dynamic strategy switching behaviour in the steady state are able to reproduce the long memory phenomena. We demonstrate that a model which incorporates contrarian trading strategies results in more dynamic behaviour in steady state, and hence is able to produce more realistic results. We also demonstrate that a non-learning contrarian model that performs dynamic strategy switching produces long memory phenomena and therefore that learning is not necessary. Models that can be validated against properties of empirical high frequency financial data should allow exploration of the robustness and reliability qualities of market mechanism modifications.