Multi-level deep Q-networks for Bitcoin trading strategies

Area of Science:

• * Financial Technology
• * Computational Finance
• * Machine Learning

Background:

• * The Bitcoin market's rapid growth presents opportunities and challenges for traders.
• * Traditional technical analysis and machine learning methods struggle to optimize Bitcoin trading strategies.
• * Existing strategies often fail to balance profit, risk, and trade frequency effectively.

Purpose of the Study:

• * To develop an advanced trading strategy for the Bitcoin market.
• * To address the limitations of current methods by simultaneously optimizing profit, risk, and trade activity.
• * To leverage deep reinforcement learning for enhanced cryptocurrency trading decisions.

Main Methods:

• * Proposed a multi-level deep Q-network (M-DQN) model.
• * Integrated historical Bitcoin price data with Twitter sentiment analysis.
• * Developed an innovative data preprocessing pipeline and a novel reward function for the M-DQN.

Main Results:

• * Achieved a 29.93% increase in investment value.
• * Attained a Sharpe Ratio exceeding 2.7, indicating superior risk-adjusted returns.
• * Demonstrated performance significantly outperforming state-of-the-art Bitcoin trading strategies.

Conclusions:

• * The proposed M-DQN model effectively optimizes Bitcoin trading decisions.
• * The integration of data preprocessing and a novel reward function addresses critical trading factors.
• * This research offers a significant advancement in automated cryptocurrency trading and financial technology.