Data Augmentation of High Frequency Financial Data Using Generative Adversarial Network

Abstract

In recent years, the field of machine learning, as seen in deep learning, has grown tremendously and is being applied to the financial field. For example, in the area of high frequency trading, machine learning techniques are used to predict stock prices. However, the non-stationary nature of financial time series makes it quite difficult to use machine learning to create a good stock price prediction model in the area. A large amount of training data may be required to overcome the problem. As another way to solve the problem, this study proposes a data augmentation method using artificial market simulations based on Generative Adversarial Networks (GANs). We use the Wasserstein GAN (WGAN), which is one of several GANs and is formulated for the optimal transport problem. Our method trains the GAN using a time series of trading data provided by the Tokyo Stock Exchange called FLEX Full. The results confirmed that the probability distribution of synthetic order events generated by the GAN was close to reality, and we also obtained data augmentation of execution prices from the artificial market simulations using the synthetic order events. The results showed that the accuracy of the prediction of increase/decrease in execution prices was better than that of the case without data augmentation.