Towards robust deep reinforcement learning-based quantitative trading with neuro-symbolic trend analysis

Area of Science:

• Artificial Intelligence
• Quantitative Finance

Background:

• Deep reinforcement learning (DRL) has advanced quantitative trading (Q-trading).
• Current DRL models struggle with market trend identification, leading to missed opportunities and significant losses during market volatility.
• Incorporating abstract human expertise is challenging but crucial for improving DRL in trading.

Purpose of the Study:

• To develop a novel framework that leverages abstract human expert knowledge for Q-trading.
• To address the limitations of existing DRL models in identifying market trends and managing risk.
• To introduce a universal logic-guided deep reinforcement learning approach for enhanced trading performance.

Main Methods:

• Proposed Logic-Q, a universal logic-guided deep reinforcement learning framework for Q-trading.
• Implemented a neuro-symbolic trend analysis mechanism (NeSy-TA) for dynamic market trend determination.
• Utilized program synthesis by sketching to integrate symbolic reasoning with neural networks.

Main Results:

• Logic-Q demonstrated superior performance compared to state-of-the-art baselines in quantitative trading tasks.
• The framework effectively identified market trends and adapted tuning parameters across different modalities.
• Outperformed recent multimodal LLM-based trading strategies, indicating a significant advancement.

Conclusions:

• Logic-Q successfully integrates human expert knowledge into DRL for Q-trading.
• The neuro-symbolic approach offers a robust method for market trend analysis and adaptive trading.
• This framework represents a significant improvement over existing DRL methods in quantitative trading.