Extent of LLM agents’ inherent tendencies on trading recommendation reliability

Determine, across diverse market scenarios, the extent to which the inherent tendencies and learning capabilities of large language model-based trading agents affect the reliability of stock recommendations and the effectiveness of quantitative trading strategies.

Background

The paper introduces StockAgent, a LLM-based multi-agent framework for simulating stock trading and investor behavior. A central motivation is understanding how external factors and model-specific behaviors influence trading outcomes and recommendations.

The authors note uncertainty about how the inherent tendencies and learning capabilities of LLM-driven agents (e.g., GPT-3.5-Turbo and Gemini-Pro) may impact the reliability and effectiveness of stock recommendations and quantitative strategies. Their experiments suggest different LLMs exhibit distinct trading behaviors, underscoring the need to quantify and generalize these effects across scenarios.