Forecasting Stochastic Time Series using Reinforcement Learning

Abstract

This study aims to analyse the effectiveness of machine learning algorithms, specifically deep reinforcement learning (DRL), in predicting financial time series. This research focuses on identifying the actions that must be taken at specific market states to maximize profit. Analysing these algorithms in the context of financial time series prediction is a crucial step toward improving investment strategies and maximizing returns. The outcomes of this research will provide insights into the potential of DRL for financial forecasting and inform the development of more accurate predictive models. In this work, authors applied deep reinforcement learning, specifically the Proximal Policy Optimization algorithm, to predict financial time series. The authors analysed the structure of the time series, performed data pre-processing, and developed a custom environment that mimics the logic of a stock exchange. The authors evaluated the performance of trained reinforcement learning agents using four basic models based on two strategies: buy and hold and random actions. The authors` agent outperformed the benchmark models significantly, and the authors assessed its risk by analysing the number of negative and positive returns for each day of testing. The results showed that the risk of its use is minimal but present. This study highlights the importance of a detailed analysis of the subject area, pre-processing of data, and development of a custom environment for financial time series prediction. Reinforcement learning shows promise in addressing the challenges of financial time series forecasting and should be further explored with larger data sets, multiple cryptocurrencies, and real-time deployment.