The advent of reinforcement learning (RL) in financial markets is driven by several advantages inherent to this field of artificial intelligence. In particular, RL allows to combine the "prediction" and the "portfolio construction" task in one integrated step, thereby closely aligning the machine learning problem with the objectives of the investor. At the same time, important constraints, such as transaction costs, market liquidity, and the investor's degree of risk-aversion, can be conveniently taken into account. Over the past two decades, and albeit most attention still being devoted to supervised learning methods, the RL research community has made considerable advances in the finance domain. The present paper draws insights from almost 50 publications, and categorizes them into three main approaches, i.e., critic-only approach, actor-only approach, and actor-critic approach. Within each of these categories, the respective contributions are summarized and reviewed along the representation of the state, the applied reward function, and the action space of the agent. This cross-sectional perspective allows us to identify recurring design decisions as well as potential levers to improve the agent's performance. Finally, the individual strengths and weaknesses of each approach are discussed, and directions for future research are pointed out.
financial markets reinforcement learning survey trading systems machine learning