Data-driven Agent Design for Artificial Market Simulation

Abstract

This study proposes a new scheme for implementing actual data into artificial market simulations at the level of trader agents. In artificial market simulations, the reliability of the simulations highly depends on the design of the agents. Traditionally, agent design is performed by humans, so the reliability of the trader agents depends on the sense of the model designer. Because humans can introduce bias or overlook the important features of actual traders, we implemented the actual data and automated the strategy learning (imitating) of agents using machine learning. We then ran artificial market simulations in the treader model, which imitates the actual trading behaviors in a machine learning (ML) architecture. The model that successfully predicted the actual traders' behaviors in given actual situations generally failed to replicate the features of those behaviors in the simulation environment. This inverse proportional relationship depended on the number of parameters in the ML model. When the number of parameters was small, the simulation better reproduced the features than the conventional model. Through this study, we demonstrate the potentials and limitations of the proposed scheme. In future work, we will consider the evaluation metrics of the simulation and develop a method that determines appropriate ML architectures.