Persona-Driven LLM Interaction in Stock Market Simulations

ACL ARR 2025 May Submission4710 Authors

Abstract:

This paper investigates the behavior of large language models (LLMs) in stock trading by assigning each model a distinct trading persona: Competitive, Adaptive, or Strategic. These personas represent different risk tolerances and decision-making styles, inspired by real-world trading psychology. The study is conducted in three stages. First, each LLM is tested individually using hypothetical trading scenarios to evaluate alignment with its assigned persona. Next, the models participate in an interactive stock market simulation where their
decision-making behaviors are observed in response to real-world market data. Finally, we enable direct interaction among the LLMs within the simulation to study how their trading strategies adapt through collaboration and mutual influence. Our results highlight the effectiveness of persona-driven prompting in guiding LLM decision-making and introduce a novel framework for examining agent interactions in simulated economic environments.

Paper Type: Short

Research Area: Human-Centered NLP

Research Area Keywords: human-centered evaluation, human factors in NLP

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 4710