A Comparative Study of an LLM-Based Agent in Synthetic Markets with Memory

This paper presents a comparative analysis of four trading strategies---termed chartist, fundamentalist, noise, and a Large Language Model (LLM)-based agent---within a synthetic market environment characterized by different memory regimes, modeled by the Hurst exponent ($H$). Price series were generated exogenously using Fractional Gaussian Noise. The LLM agent received textual descriptions of price trends and responded with trading actions. Results indicate that the LLM agent achieved superior performance compared to other heuristic strategies across all regimes, particularly in persistent markets ($H = 0.7$). This study is not intended as an agent-based model (ABM) with endogenous interactions, but rather as a comparative backtest of strategies on exogenous price series. The objective is to explore the potential of LLMs as tools for modeling financial decision-making, serving as a baseline for future research on bounded rationality and adaptive behavior.