Investor risk profiles of large language models

Battle Count: 69

The paper is primarily relevant to retail investment advising and robo-advisory systems rather than quantitative trading strategies. However, understanding LLM risk profiles is important for any system that uses LLMs for portfolio allocation decisions, risk-based asset selection, or personalized investment recommendations. The findings about model-specific biases could affect automated advisory systems that inform trading decisions.

The methodology is straightforward: API calls to three LLMs with structured prompts, recording multiple-choice responses, computing scores, and running standard statistical tests (ANOVA, Kruskal-Wallis). No model training or complex architecture is involved. The main effort is in the experimental design (100 iterations per condition) and statistical analysis.

The questionnaire is publicly available (Charles Schwab 2024 version), prompts are fully specified in Table 2, and model versions are identified (gpt-4o-2024-08-06, gemini-1.5-pro, llama3.1-70b). However, no code repository is provided, and the exact scoring methodology for the questionnaire is referenced but not fully reproduced in the paper. The 100-iteration protocol is clearly described.