Image Representation of Time Series for Reinforcement Learning Trading Agent

Guinther K. da Costa; Leandro dos S. Coelho; Roberto Z. Freire1

doi:10.48011/asba.v2i1.1108

Guinther K. da Costa Pontifical Catholic University of Parana
Leandro dos S. Coelho Pontifical Catholic University of Parana, Federal University of Parana
Roberto Z. Freire1 Pontifical Catholic University of Parana

DOI: https://doi.org/10.48011/asba.v2i1.1108

Keywords: Stock trading, Algorithmic trading, Financial market, Machine learning, Reinforcement learning, Time series forecasting

Abstract

The availability of diverse data has increased the demand for expertise in algorithmic trading strategies. Reinforcement learning has shown interesting applicability in a wide range of tasks, especially in some challenging problems as trading, where slow model convergence, inference speed, and reduced model accuracy appear as barriers in this type of application. In this paper, we propose the transformation of time series into images considering a transfer learning based on a semi-supervised model with deep Q learning agents, where labels were generated by an evolutionary algorithm to improve both training speed and performance measures.