Adaptive Stock Trading Algorithms Jointly Driven by Multi-Agent Collaborative Decision Making and Large Language Models in Complex Financial Market Environments

Abstract

To address the challenges of heterogeneous information sources, rapidly changing market conditions, and insufficient utilization of semantic information in traditional trading methods within complex financial markets, this paper proposes a multi-agent collaborative decision-making approach for automated stock trading, jointly driven by a large language model. This method, centered on the unified modeling requirements of structured market data and unstructured financial text, constructs an overall framework comprised of unified state encoding, agent interaction and collaboration, adaptive coordination and fusion, and action selection. This enables the effective fusion of multi-source content, including market prices, transaction characteristics, volatility information, and news and announcement semantics, within the same decision-making chain. By introducing a multi-agent collaborative mechanism, different functional agents can participate in trading decisions from the perspectives of trend perception, event judgment, risk identification, and execution control, thereby improving the completeness and stability of trading judgments under complex market conditions. Simultaneously, a large language model is used to extract key semantic clues from financial texts, enhancing the model's understanding of potential market disturbances and implicit trading signals. At the trading execution level, this paper further integrates explicit buy and sell conditions, as well as profit-taking and stop-loss constraints, to construct an adaptive trading mechanism for practical application scenarios. This ensures that the strategy output not only has strong information integration capabilities but also clear execution logic. The results show that the proposed method performs well in terms of profit generation, drawdown control, and overall transaction quality, verifying the effectiveness and application value of the constructed framework for automated stock trading in complex financial market environments.