Exploration Techniques for a Deep Reinforcement Learning Trading Agent

Nowadays nearly all financial orders across the globe are performed by automated trading systems. In particular, machine learning and deep learning have been utilised to try and predict the trend of different financial securities. Financial market analysis has always been critical and, in recent years, more and more opportunities are available to extract new information from all kinds of financial data. In the past, reinforcement learning has been used to solve various financial problems, including portfolio optimisation, securities trading and risk management. Among those, stock trading is considered the most complex application of reinforcement learning in quantitative finance, primarily due to the inherent stochastic behaviour of financial markets. In this thesis, we argue in favour of deep reinforcement learning as a practical approach to time series forecasting and automated stock trading. The objective is to automatically generate a profitable and reasonably safe trading strategy in any financial market. Furthermore, we propose a deep reinforcement learning algorithm to automatically trade financial securities, which can consistently outperform market indexes and generate positive returns while exposing itself to minimal risk. We also discuss how a reinforcement learning agent can learn to trade in highly volatile markets, plus methods to build a robust and stable algorithm across different markets.