When AI Meets Finance (StockAgent): Large Language Model-based Stock Trading in Simulated Real-world Environments

A research team has introduced StockAgent, a multi-agent system powered by large language models that simulates investor behavior in realistic stock-market conditions to study how external factors influence trading, according to a paper posted on arXiv [1]. The system, detailed in a submission last revised in June 2026, is designed to let users evaluate the effect of macroeconomic shifts, policy changes, company fundamentals, and global events on trading activity and profitability [1][2]. Unlike earlier AI-based trading simulators, StockAgent is built to prevent the model from drawing on prior knowledge of the test data, a flaw known as test-set leakage [1][2]. The paper’s authors, including Chong Zhang, tested several large language models inside the StockAgent framework and report that the simulations capture how external factors shape trading behavior and stock-price fluctuation patterns [1][2]. Large language models are a class of machine-learning systems with many parameters, trained on vast text corpora for tasks such as language generation [6]. StockAgent deploys them as autonomous agents inside a market environment, an approach that draws on the broader tradition of agent-based modeling. Agent-based models simulate the actions and interactions of individual agents — each following simple decision rules — to understand how macro-scale phenomena emerge from micro-level behavior [8]. Such models have been applied across biology, ecology, and the social sciences [8]. The StockAgent paper argues that the patterns identified through its simulations can inform LLM-based investment advice and stock recommendations [1][2]. The system’s code has been made publicly available on GitHub [1][2]. The research arrives as artificial intelligence continues to expand into finance, where machine-learning techniques are already used for credit scoring, decision-making, and e-commerce applications [3]. The initial version of the paper was submitted in July 2024, with the most recent update posted nearly two years later, suggesting ongoing refinement of the framework [1].