Financial Stability Implications of Generative AI: Taming the Animal Spirits

Battle Count: 74

The paper is highly relevant to understanding how AI agents behave in trading contexts, particularly regarding herd behavior and information cascades. While not directly proposing trading strategies, it provides micro-foundations for understanding AI-driven market dynamics. The findings about AI agents being more rational than humans but still susceptible to biases (e.g., color associations) are directly relevant to designing AI-assisted trading systems. The optimal AI agent experiments show how fine-tuning for profit maximization can induce herding behavior, which has implications for algorithmic trading design.

Implementing the experimental framework requires: (1) API access to multiple LLMs, (2) careful prompt engineering following the disclosed prompts, (3) Bayesian market maker price updating logic, (4) random seeding for reproducibility, (5) structured output parsing from LLMs, (6) multiple experimental variations. The core logic is straightforward but requires attention to matching the original human experiment protocol. No specialized ML training is needed - it uses pre-trained models via API calls.

The paper provides detailed experimental design, prompts (Prompts 1-4), and methodology descriptions. However, no code repository is provided. The human benchmark data comes from Cipriani and Guarino (2009), a published paper. The LLM experiments use API calls to specific models (Claude 3.5/3.7, Llama 3, Nova Pro) with specified temperature settings. Random seeds are used for comparability. Reproduction would require API access to these specific models and careful implementation of the experimental protocol.