Stock Market Simulation by Micro-Macro GAN

In recent years, research on the fusion of computer simulations with real data, as seen in data assimilation, has been active. In addition, with the development of machine learning, especially deep learning, computer simulations are increasingly using deep learning models in addition to physical models. In the field of finance, research has also been conducted to use machine learning models when simulating markets, for example, to emulate synthetic orders placed by virtual traders. However, it is fundamentally difficult to generate realistic macro dynamics such as the price from micro order dynamics. In this study, we propose a new market simulation method by machine learning model to generate macro dynamics from micro order dynamics. The market simulator built by the machine learning model, which is called Micro-Macro GAN, is trained by coupling two mechanisms. The first mechanism generates micro order dynamics and trains a generator of micro order dynamics to distinguish the real from the synthetic order data using a Wasserstein Generative Adversarial Network (WGAN), which is called the Micro GAN. The second mechanism generates macro dynamics from the micro order dynamics generated by the Micro GAN and trains a generator of macro dynamics to distinguish the real from the synthetic macro data using a Sig-W GAN, which is called the Macro GAN. The macro dynamics data is converted to a signature, and finally, the Sig-W metric is computed as a loss function. As a demonstration, training was performed on Toyota Motor Corporation taken from the Flex Full data provided by the Japan Exchange Group (JPX). The order dynamics data generated from the Micro GAN and the Micro-Macro GAN were compared to the real order dynamics data. The results showed that the Micro-Macro GAN results were more similar to the real dynamics data than the Micro GAN results.