Dissecting AI Trading: Behavioral Finance and Market Bubbles

Battle Count: 72

Highly relevant to quantitative trading. The paper directly addresses how LLM-based trading agents behave in equilibrium markets, documents specific behavioral biases (disposition effect, extrapolative expectations) that affect trading decisions, shows how these aggregate into market-level phenomena (bubbles, volume dynamics), and provides actionable intervention mechanisms (prompt-level cognitive guardrails) for stabilizing AI-driven markets. The bid-offer gap as a leading indicator and disagreement-volume relationship are directly applicable to market microstructure analysis and trading strategy design.

The experimental framework requires: (1) multi-agent LLM orchestration with persistent memory (PLANS.txt, INSIGHTS.txt), (2) a custom market-clearing algorithm for open-call auctions, (3) structured JSON parsing of agent outputs, (4) incentivized forecasting mechanisms, (5) NLP-based reasoning extraction and 20-mechanism LLM-as-judge scoring, (6) prompt engineering for targeted interventions, and (7) panel data regression with multiple fixed effects. The simulation pipeline is complex but well-documented. Running 14 different LLMs with 20-24 agents per market across multiple sessions adds computational and API cost complexity.

The paper provides extremely detailed experimental design including full prompt architecture (Appendix A), JSON schema for agent outputs, market-clearing algorithm, 20-mechanism audit prompt (Appendix B), and complete regression specifications. However, no code or data repository is explicitly linked. The use of 14 specific LLMs with version identifiers aids reproducibility, but API-based LLM outputs may vary over time.