An Ensemble Method of Deep Reinforcement Learning for Automated Cryptocurrency Trading

We propose a novel ensemble method for trading that addresses challenges due to large volatility in cryptocurrency markets and improves the out-of-sample performance of deep reinforcement learning based trading strategies. The ensemble method utilizes a mixture distribution to combine models selected from multiple validation periods. We test our proposed method on a 4 -year historical period that encompasses various cryptocurrency market conditions, including bullish and bearish markets. The results empirically demonstrate the profitability and robustness of the ensemble method compared with a deep reinforcement learning benchmark and a passive investments strategy in annualized returns and risk-adjusted returns.