The Use of Features to Enhance the Capability of Deep Reinforcement Learning for Investment Portfolio Management

Cryptocurrencies trading algorithms: A review

“…With evolution in cryptocurrency and advances in the creation of centralized and decentralized exchanges, accurate information on prices has become accessible and therefore studies are emerging in this line of research using Neural Networks and Deep Learning to analyze market volatility (Bu & Cho, 2018; Miura et al, 2019), forecast future prices (Betancourt & Chen, 2021b; Bu & Cho, 2018; Ji et al, 2019; Lahmiri & Bekiros, 2019, 2021; Lee, 2020; Li et al, 2020; Livieris et al, 2021; Loh & Ismail, 2020; Lucarelli & Borrotti, 2019; Miura et al, 2019; Nithyakani et al, 2021; Sattarov et al, 2020; Sun et al, 2021; Zanc et al, 2019), and managing portfolios with Bitcoin in an automated way (Betancourt & Chen, 2021a; Jiang & Liang, 2016; Ren et al, 2021; Shi et al, 2019; Sun et al, 2021).…”

“…Ren et al (2021) used a neural network with reinforcement learning and a set of characteristics such as traded volume, moving average, Elliot oscillator, and stochastic oscillator in 11 cryptocurrencies with the highest market volume of past 30 days before the backtesting strategy and compared them with traditional strategies such as OLMAR and CWMR, achieving better performance than all the strategies compared.…”

“…As can be seen in Table 2, most authors prefer to work with deep learning and machine-learning algorithms to forecast the prices of cryptocurrencies and manage smart portfolios. with Bitcoin in an automated way (Betancourt & Chen, 2021a;Jiang & Liang, 2016;Ren et al, 2021;Shi et al, 2019;Sun et al, 2021).…”

“…Betancourt and Chen (2021a) proposed a dynamic portfolio management strategy using deep learning and learning by reinforcement on data from August 2017 to November 2019 from the Binance API with 85 cryptocurrencies and used buy and sell data of 30 min, 6 h, and 1 d. Compared with other strategies, this is the only one that presents a more dynamic approach to the market. Ren et al (2021) used a neural network with reinforcement learning and a set of characteristics such as traded volume, moving average, Elliot oscillator, and stochastic oscillator in 11 cryptocurrencies with the highest market volume of past 30 days before the backtesting strategy and compared them with traditional strategies such as OLMAR and CWMR, achieving better performance than all the strategies compared. Shi et al (2019) proposed in their paper a deep learning framework with a convolutional network with inception, a method already used in image analysis and applied to portfolio management.…”

Quantitative Trading of Stocks Based on TD3 Algorithm

“…However, in different cycles, investors' actions are also crucial. Therefore, some scholars combine deep learning and reinforcement learning methods, using CNN or DRSN for feature extraction, and then using DQN or DDPG algorithms for decision-making, making the action space of the agent continuous [2][3][4][5][6][7]. Some scholars also use Q-learning and Actor-Critic algorithms for decision-making, and use GAN to generate noise and enhance generalization ability [8][9].…”