LLM-generated stress scenarios depend on prompts, not model, study finds

A new research paper explores the use of Large Language Models (LLMs) for generating macroeconomic stress scenarios in portfolio risk simulation. The study found that LLM-generated scenarios' severity is primarily determined by prompt design rather than the model's inherent capabilities. While GPT-5.6-luna followed prompted severity bands, Llama-3.1-8B exhibited more dispersion and contradictions. The research suggests LLMs are best suited for narrative elicitation around a predefined severity path, rather than autonomous calibration of stress magnitudes. AI

IMPACT LLMs are better suited for narrative generation in financial risk scenarios than for autonomous calibration of stress magnitudes.

RANK_REASON Research paper published on arXiv detailing LLM capabilities for financial risk simulation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• GPT 5.6 Luna
• International Monetary Fund
• Llama-3.1:8b
• Masoud Soleimani
• World Economic Outlook

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →