加载中...
This paper proposes a benchmark framework that generates 9,500 mathematically well-defined portfolio-optimization multiple-choice problems to evaluate LLMs' quantitative decision-making in asset allocation, comparing GPT-4, Gemini 1.5 Pro, and Llama 3.1-70B across objectives (variance, return, Sharpe, MDD, CVaR), constraint types, and distractor-generation methods, and finds model-specific strengths (GPT: risk objectives; Gemini: return-focused) and shared limitations on multi-criteria optimization such as Sharpe and CVaR .
| Investment objective | GPT (%) | Gemini (%) | Llama (%) |
|---|---|---|---|
| Min variance | 91.73 | 67.89 | 46.79 |
| Max return | 68.42 | 79.26 | 28.47 |
| Max Sharpe ratio | 10.47 | 9.53 | 1.21 |
| Min MDD | 73.68 | 56.21 | 40.42 |
| Min CVaR | 42.74 | 39.67 | 16.58 |
元数据与概览(引言与报告概览)
逐节深度解读
图表与图片的逐一深度解读(关键环节)
估值/评分与数值实验细节(第 4 节实证)
风险因素评估(报告识别的风险与潜在影响)
批判性视角与细微差别(审慎评价)
结论性综合(总结与对图表的关键见解)
补充建议(基于报告内容的后续改进方向)
(注:文中所有引用的结论或直接推断均已在句末标注原文页码以便溯源。)
2603.09301.pdf