2025-12-15 · 1 min read
Can LLMs Manage a Stock Portfolio? My URECA Research
Co-authored URECA paper on Perplexity-built S&P 500 portfolios. The high-risk book hit a 7.545 Sharpe, but the returns trace to sector tilts, not stock-picking.
The question
Can a prompt-engineered large language model construct equity portfolios that reflect different investor risk profiles and produce measurable differences in risk and return? Our URECA paper (Jain, Shah, Lee, Ha; supervisor Dr. Danling Jiang) tested this on S&P 500 equities via WRDS Compustat data.
Three risk profiles
We prompted Perplexity API to generate 25-stock portfolios under three specifications:
- High Risk - aggressive growth, cyclical/tech exposure
- Low Risk - defensive sectors, capital preservation
- No Risk Specification - outperform SPY without explicit risk guidance
Verified results (Apr 30 - Aug 6, 2025)
| Strategy | Total Return | Sharpe | Beta vs SPY |
|---|---|---|---|
| High Risk | 31.53% | 7.545 | -0.267 |
| Low Risk | 3.87% | 4.795 | -0.138 |
| No Risk Spec | 20.88% | 5.540 | 0.817 |
Key finding
Returns traced to systematic sector and style exposures. High-risk portfolios concentrated in Information Technology and Consumer Discretionary. LLMs encode risk appetite through sector tilts, not persistent stock-selection alpha.
Takeaway
Generative AI works as a decision-support tool for retail portfolio construction inside a disciplined validation framework. It is not an autonomous investment system.
Related project
Generative AI in Stock Portfolio Management
URECA paper testing Perplexity-built S&P 500 portfolios. The high-risk book returned 31.5% over three months, but the edge came from sector tilts, not stock-picking.