Quantitative Tuning of Artificial Market Simulation using Generative Adversarial Network

This study focuses on parameter tuning of artificial market simulations and aims to replace the traditional qualitative evaluation metrics based on stylized facts with the proposed quantitative metrics. Traditionally, for the evaluation of artificial market simulations, the replication of stylized facts, a common phenomenon among financial markets and observed in empirical studies, is verified by humans. However, this prevents large-scale parameter tuning owing to the complexity of automation. Hence, this study utilizes a generative adversarial network (GAN) for this replacement because we assume that the GAN's learning architecture has a good fit for evaluating the distributional features of actual markets and can learn stylized facts implicitly. In the proposed parameter-tuning method, the simulated data are input into the critic of the GAN, and the outputs are employed as the objective value of the tuning. The parameter tuning results show that we successfully tuned the high-dimensional parameters of artificial market simulations and confirmed that the optimized parameter could replicate the stylized facts employed in traditional qualitative evaluation metrics.