Deep Q Network Applied in Trading Portfolio of Virtual Currencies

Abstract

Virtual currency is the basic form of currency in the coming WEB 3.0 era. It has the characteristics of high volatility, high liquidity, decentralized and no government supervision. These properties are suitable for high-frequency trading, but make it difficult to obtain higher profits for investors. Although researchers have introduced Deep Reinforcement Learning (DRL) into the financial market, many strategies are only related to one stock. This paper will study the daily trading strategy of the portfolio of two virtual currencies (Bitcoin and Ethereum). A quantitative analysis of bitcoin shows that the price increasing rate is a Leptokurtic, which means the tail of the distribution is higher than that of the normal distribution. Consider of this situation, we add random noise on the currency price with setting threshold to simulate some sharp price changes. In our paper, Deep Q-Network (DQN) algorithm is applied to solve this financial auto-trading problem. We use a small trick to speed up the convergence, which gradually reduces the exploration rate ϵ. The experimental results show that the final profit return is as high as 51.1%.