2025/11/14 by David A. Byrd · 1 voice
Decision Sciences · Economics, Econometrics and Finance · #Stock Market Forecasting Methods #Financial Markets and Investment Strategies #Complex Systems and Time Series Analysis
paper · doi:10.1145/3768292.3770424
openalex created_date 2025/11/14 · openalex publication_date 2025/11/14 · openalex updated_date 2026/07/29
Through every economic sector, but especially among financial firms and enthusiasts, agentic AI systems are being tool-enabled, giving them control over large language models (LLM), reinforcement learning (RL) models, and more. We present a timely paper to explore the potential consequences with a novel combined system: a deep RL-based autonomous trading agent which also controls an LLM capable of posting to a simulated social media feed observed by other traders. As the agent trades, it also supplies order flow information to the LLM, which produces and posts natural language market analysis at the agent’s direction. We empirically investigate the performance and impact of such an agent using two DeepRL algorithms, finding that it learns to augment profit by manipulating sentiment in a sort of accidental pump and dump scheme. Along the way, we present confidence-building baseline results and specific insights from our investigation, before concluding with a discussion of results, limitations, and suggestions for future work.