Time-driven feature-aware jointly deep reinforcement learning for financial signal representation and algorithmic trading

摘要

Algorithmic trading is a continuous perception and decision making problem, where environment perception requires to learn feature representation from highly nonstationary and noisy financial time series, and decision making requires the algorithm to explore the environment and simultaneously make correct decisions in an online manner without any supervised information. To address these two problems, we propose a time-driven feature-aware jointly deep reinforcement learning model (TFJ-DRL) that integrates deep learning model and reinforcement learning model to improve the financial signal representation learning and action decision making in algorithmic trading. Concretely, we learn the environmental representation by adaptively selecting and reweighting various features of financial signals and summarize the attention values between historical information and changing trend depending on the current state. Besides, the supervised deep learning and reinforcement learning are jointly and iteratively trained to make full use of the supervised signals in the training data, and obtain more update information and stricter loss function constraints, thereby increasing investment returns. TFJ-DRL is evaluated on real-world financial data with different price trends (rising, falling and no obvious direction). A series of analysis show the robust superiority and the extensive applicability of the proposed method. (C) 2019 Elsevier Ltd. All rights reserved.