Real Bubbles, Synthetic Traders: AI-Agent-Based Simulations of the Speculative Market

Abstract

With the rise of agentic AI comes a powerful opportunity to rethink how we model economic complexity and nonlinear systems. I create an extensible, AI-agent-based market simulation sandbox for exploring behavioral finance questions and speculative dynamics. This paper presents the first AI-agent-based simulation of a speculative stock market built on a general-purpose agentic framework—a high-quality, consumerfacing software development kit (SDK) rather than a bespoke, special-purpose scaffold around a LLM using chat completions API features. The framework is fully customizable, enabling researchers to design and run experiments by defining trader archetypes, injecting news shocks, testing different market microstructures, and observing interactions—including those via a simulated social platform, enabling rich experimentation with speculation, herding, pricing dynamics, and financial stability. Empirical simulation results echo findings from prior literature while introducing new nuance. Across markets with varying share of trader archetypes, results show that trading volume negatively predicts future returns when irrational agents dominate, suggesting that high-volume events can signal upcoming price reversals. In contrast, even though results show that rationalists reduce momentum overall, when they do trade in high volume, it’s often with conviction, which can support trend continuation (i.e., real momentum). These findings point to a possibly more nuanced relationship between volume and returns — one distinctly modulated by the rationality of market participants. Notably, this does not imply that rational markets exhibit more volume overall (simply because prices may appear more supported); rather, it reflects a shift in the informational character of volume—from noise to signal—as market composition increases in rational agents. Lastly, volume and rationality in my simulations are uncorrelated, suggesting their separate influences on price dynamics are not confounded — but rather, arise endogenously from the system’s own emergent complexity. This study demonstrates the viability of AI-agent-based simulations and their potential to generate nuanced insight into enduring financial economics questions.