- The paper demonstrates that LLM agents replicate human behavioral biases such as the disposition effect and extrapolative expectations, shaping bubble dynamics.
- It employs an equilibrium asset market simulation with explicit chain-of-thought reasoning and structured forecasts to quantify mispricing and market fluctuations.
- The study shows that prompt-level cognitive interventions can modulate trading behaviors and bubble magnitudes, offering novel strategies for AI-based market regulation.
Expert Analysis of "Dissecting AI Trading: Behavioral Finance and Market Bubbles" (2604.18373)
Experimental Framework and Design
The study establishes a rigorous, equilibrium-based asset market simulation leveraging autonomous LLM agents. The market architecture replicates the classic Smith et al. (1988) open-call auction paradigm, featuring two assets: risk-free cash and a risky stock with stochastic dividends. Notably, a terminal buyout ensures a constant fundamental value (FV=14) throughout the 20 trading periods, facilitating unambiguous measurement of mispricing and bubble dynamics. Fourteen state-of-the-art LLMs populate both homogeneous (single-model) and heterogeneous (mixed-model, "battle royale") environments. The agent design incorporates structured chain-of-thought memory, explicit forecasting incentives, and JSON-logged reasoning, yielding a multimodal dataset linking textual cognition, forecasts, trading actions, and market outcomes.
Micro-Level Behavioral Patterns in LLM Agents
The investigation rigorously documents that LLM agents manifest canonical behavioral finance patterns:
- Disposition Effect: Agents exhibit statistically significant propensity to sell assets realizing gains while holding onto losers, even when controlling for forward expectations. This parallels persistent disposition effects observed in human investors, but emerges despite explicit knowledge of fundamental value and the absence of real-world frictions.
- Extrapolative Expectations: Agents' price forecasts are disproportionately influenced by recent market returns, with regression coefficients demonstrating heavy recency bias and horizon amplification. Longer-term forecasts extrapolate short-term trends with increasing magnitude, a behavior closely linked to bubble emergence in experimental and real-world markets.
- Tight Belief-Action Coupling: Contrary to empirical results in humans (e.g., Giglio et al. 2021), LLM agents' trading decisions show strong sensitivity to stated expectations. The autoregressive architecture ensures frictionless translation of beliefs into portfolio allocations, amplifying the impact of cognitive biases.
Aggregation into Equilibrium Market Dynamics
Individual-level heuristics aggregate into market phenomena that replicate classical experimental findings:
- Bubble-and-Crash Trajectories: Average prices rise above FV in early periods, tracing hump-shaped bubbles, with trading volume peaking synchronously. The severity of mispricing exhibits substantial heterogeneity across models (e.g., MSE(FV) from near zero in Deepseek V3 to >100 in Meta Llama 3.1.70B), highlighting architecture-dependent behavioral profiles.
- Adaptive Price Adjustment: Bid-offer gaps (excess demand) robustly predict future price changes, confirming non-equilibrium, feedback-driven market dynamics. This result directly replicates Smith et al. (1988) and supports adaptive, expectation-induced price trajectories.
- Disagreement-Volume Link: Cross-sectional dispersion in agents' price forecasts tightly correlates with trading volume. Regression results indicate strong, statistically significant relationships, consistent with heterogeneous belief models (Hong & Stein 2007).
Diagnostic of Cognitive Mechanisms: Textual Reasoning Analysis
A distinct methodological advance is the audit of agents' reasoning using a twenty-mechanism framework ("LLM-as-a-judge" taxonomy). The direct extraction and scoring of reasoning features reveal:
- Bubble Episodes Characterization: During bubbles, agent texts manifest elevated scores in momentum-chasing, rational speculative bubble logic, and synchronization risk. Agents articulate explicit awareness of overvaluation, herding behavior, and strategic timing problems, echoing human narratives during historic bubbles.
- Loss of Fundamental Anchoring: Bubble periods exhibit significant abandonment of value anchoring, as textual evidence demonstrates pure extrapolative logic and narrative-driven trading.
These results provide direct, causal evidence linking aggregate price dynamics to micro-level behavioral heuristics.
Prompt-Based Policy Interventions
A critical contribution is the demonstration of market-level effects from prompt-level cognitive interventions:
- Amplification and Suppression Treatments: Market-wide prompt shocks amplifying or suppressing bubble-prone heuristics causally alter bubble magnitudes. Amplification prompts intensify speculative logic and momentum; suppression prompts elevate fundamental anchoring and risk discipline, reducing mispricing.
- Programmability and Policy Implications: LLM-driven markets are fundamentally programmable at the cognitive level. Regulators can deploy behavioral-finance-informed prompts as "cognitive guardrails" to mitigate bubble risk and stabilize equilibrium, offering an actionable lever absent in human-dominated markets.
Practical and Theoretical Implications
The findings have significant implications:
- Credibility of LLMs as Experimental Economic Agents: LLM-populated markets endogenously reproduce both individual cognitive biases and equilibrium-level market regularities observed in human studies. This validates their utility as laboratories for studying asset price formation and systemic risk. The methodology enables direct observation and manipulation of internal cognition—an advance over opaque human subject experiments.
- Market Stability and AI Regulation: The programmability of AI agents creates novel regulatory paradigms. Prompt-level "debiasing" can preempt speculative crises, shifting focus from traditional capital or circuit-breaker mechanisms to cognitive design. Financial institutions deploying LLM-based trading systems must consider behavioral vulnerability and guardrails at deployment.
- Long-Term Systemic Risks: The architecture- and prompt-dependence of bubble susceptibility suggests that rapid evolution in LLMs may lead to shifts in aggregate market behavior. Further, mixed populations of rational and bubble-prone agents exhibit nontrivial interactions, implying that the market composition is a key parameter for stability.
- Open Questions:
- The generalizability of behavioral control in more complex market structures (e.g., continuous trading, derivatives).
- Interactions between human and AI agents in hybrid markets.
- Mechanistic interpretability: whether observed cognitive biases constitute emergent agent behavior or are merely pattern reproduction from training corpora.
- Dynamic evolution: whether inter-agent learning can modify or entrench behavioral profiles over time.
Conclusion
This study rigorously establishes that autonomous LLM agents, when deployed in a classical equilibrium asset market, inherit and amplify behavioral finance patterns from their training data, manifesting in both micro-level trading heuristics and macro-level bubble dynamics. The frictionless belief-action coupling characteristic of LLMs renders their cognitive biases rapidly transmissible to market outcomes. Crucially, the behavioral profiles of AI agents are amenable to explicit prompt-level intervention, enabling direct market stabilization via cognitive guardrails—a regulatory innovation not possible in conventional human-dominated financial environments. These findings underscore both the power and responsibility accompanying the adoption of AI-driven trading systems, and set the stage for continued research on the interplay between AI cognition and systemic stability in financial markets.
