Exploring the Value of Prediction in an Artificial Stock Market

An action selection architecture is described which uses two learning modules: one to predict future events (the PPM) and one to decide upon appropriate actions (the IALM). These are only connected by the fact that they have access to each other’s past results and the IALM can use the predictions of the PPM as inputs to its action selection. This is instantiated in a model which uses GP-based learning for the two modules and is tested in an artificial stock market. This used to start exploring the conditions under which prediction might be helpful to successful action selection. Results indicate that prediction is not always an advantage. I speculate that prediction might have similar role to learning in the ”Baldwin Effect” and that the ”momentum” of the system might be a significant factor.