Adaptive Multi-Agent LLM Systems for Financial Trading: A Framework for Realistic Simulation and Dynamic Prompt Optimization

Abstract

Large Language Models (LLMs) have demonstrated strong potential in complex decision‑making tasks, showing promise for financial trading applications that demand the integration of diverse information sources, reasoning under uncertainty, and adaptation to rapidly changing market conditions. However, progress in this direction is limited by the absence of realistic market environments tailored to the assessment of trading agents, as well as a shortage of well‑optimized frameworks capable of fully leveraging LLM capabilities in such challenging domains.

This thesis addresses these gaps through a two‑stage contribution. First, we present StockSim, an open‑source simulation platform that realistically models the behavior of financial markets and supports the development of LLM‑based trading agents. StockSim extends beyond simplified historical backtesting by emulating the dynamics of live trading, where agents place fully specified orders - including type, price and quantity - that execute in alignment with actual price movements. The platform also captures order timing, execution delays, and the effects of market activity on price formation. Its flexible configurations, diverse data sources, and multi‑agent support enable the creation and evaluation of advanced trading strategies under conditions that closely mirror real‑world scenarios.

Building on this foundation, we introduce ATLAS (Adaptive Trading with LLM Agent Systems), a coordinated multi‑agent trading framework that integrates specialized analysts for market trends, financial news, and company fundamentals, synthesizing these perspectives into coherent strategies. At its core, ATLAS incorporates Adaptive‑OPRO, an enhanced Optimization by PROmpting method that iteratively refines decision‑making based on trading outcomes, yielding progressively better performance. Extensive experiments show that Adaptive‑OPRO consistently outperforms both traditional quantitative strategies and existing LLM‑based approaches, while ablation studies confirm the complementary contributions of each component in the framework. ATLAS also improves transparency in the decision‑making process, fostering trustworthy collaboration and enabling more reliable deployment alongside financial experts.

Our results reveal distinctive behavioral patterns in LLMs, providing new insights into their capabilities and limitations in high‑stakes financial contexts. These findings are grounded in rigorous multi‑run evaluation protocols that expose severe reliability issues in the single‑run assessments common in prior literature, underscoring the need for robust evaluation in complex sequential decision‑making settings.