Optimization of dynamic crypto portfolios via deep reinforcement learning against supervised learning under transaction costs

The management of a cryptocurrency portfolio is difficult since it involves the high volatility and non-stationarity of the cryptocurrency market. Therefore, financial management strategies cannot be directly applied. In this paper, a deep reinforcement learning optimization technique (DRL-PPO) is discussed and compared with supervised predictive models like LSTM, Random Forest, and XGBoost in the task of optimization of a portfolio with ten different cryptocurrencies, taking into account the transaction fee of 0.1%. The results have shown that DRL-PPO outperforms other methods in this case, providing an annualized return of 37.22% and Sharpe ratio of 0.87. It is worth mentioning that the DRL-PPO technique is market neutral (beta = − 0.08) and alpha equals 46.41%. Meanwhile, the disadvantages of supervised predictive models are revealed in periods of reversal of the trend; LSTM provides 6.75%, while Random Forest provides 23.54%, being subject to fluctuations in the market (beta approximately equal to 1). Finally, XGBoost provides a negative number (-1.48%) which does not even cover the transaction fees. Summarizing, the success of DRL-PPO is not based on directional accuracy but rather on the ability to optimize trading efficiency and manage risks.