optimizing dynamic portfolio management in the cryptocurrency market using multi-agent deep reinforcement learning and the fear and greed index

Selecting an appropriate investment portfolio is the foundation of modern financial theories and is particularly crucial in the highly diverse, complex and risky cryptocurrency market. traditional models, such as the markowitz model, are inadequate for addressing the dynamics of these financial environments, particularly in dealing with high computational complexity. enhanced investment performance, and the fulfillment of diverse investor objectives hinge on selecting the right model and strategy. this study introduces a deep reinforcement learning environment capable of adapting to a dynamically changing state space with various assets. the proposed model adopts a multi-agent approach, with each agent assigned to a specific asset. it utilizes two dqn neural network models for action selection and lstm neural network for predicting trend. for selecting appropriate actions, nine different indicators were utilized, including the fear and greed index to identify market sentiment and other technical indicators. the dataset comprises 11 non-stable cryptocurrencies and one stable coin to preserve capital value. two different strategies were used to test the model. the cumulative profit and the sharpe ratio were employed as evaluation metrics. the results indicate that the average profitability of the proposed model is 2.32 times higher, and the sharpe ratio is 1.45 times greater than the buy-and-hold strategy. additionally, the use of lstm alongside dqn leads to more appropriate action selection, ultimately optimizing and enhancing the profitability of the investment portfolio.