Reproduce “GIFT: LLM-Guided State-Reward Interface for Financial Reinforcement Learning” on OpenResearch

About this paper

Yanyan Wu, Boyi Zhang, Yanlin Liu, Xinyu Fang, Jining Luan, Meiqi Zhang, Jiacheng Liu, Hao Zeng and 5 more

This research introduces GIFT, a framework that employs Large Language Models (LLMs) to dynamically design the state and reward functions for Deep Reinforcement Learning (DRL) agents in financial portfolio trading. The framework significantly enhances out-of-sample risk-adjusted portfolio performance by providing adaptable and informative learning signals.

Key insights

• LLMs can serve as effective, constrained 'financial-knowledge-guided interface designers' for DRL, structuring the learning environment without directly generating trading actions, thereby mitigating risks.
• Dynamically adaptable state and reward signals, designed with LLM guidance and informed by market diagnostics, are crucial for achieving robust and stable financial DRL performance across non-stationary market regimes.
• The iterative diagnostic-guided refinement process is critical for selecting effective state-reward interfaces, enabling the framework to generate semantically coherent and interpretable designs that adapt to specific market conditions.

Summary by alphaXiv.