Real-Time Reinforcement Learning with Data Expansion using GAN in the Forex Market

Abstract

There are two major problems in applying machine learning to financial markets: the number of data per unit time is small, and the data is non-stationary. In general, these problems can be a bottleneck for machine learning methods that require a large amount of data as training data and handle stationary data. In this paper, we propose a "real-time learning framework for reinforcement learning models using GAN data expansion" to address the above problems. There are two major novelties in the proposed method. The first is that the proposed method can generate alternative data to the market data by using the price forecasting model as a generative model, thus solving the data shortage. The second is that the proposed method can easily cope with non-stationary movements of data by incorporating a real-time learning framework. In this study, we show that the proposed method has an improved rate of return compared to conventional deep reinforcement learning methods using foreign exchange market data. In addition, in order to examine the ability of the proposed method to cope with non-stationarity, we confirmed that the model copes well with exceptional events such as the new coronavirus and the Lehman shock.