Takeaway:
ChatGPT-generated portfolios consistently outperform the market across both U.S. and Chinese policy/news contexts, especially after fine-tuning and when using creative prompts (temperature = 1).
Key Idea: What Is This Paper About?
This paper tests whether ChatGPT can act as a robo-financial advisor. By feeding policy or news text into ChatGPT and asking for investment recommendations, the authors construct portfolios that outperform benchmarks. They validate recommendations with finance professionals and find ChatGPT's reasoning sound in most cases.
Economic Rationale: Why Should This Work?
LLMs like ChatGPT can understand policy/news narratives and translate them into asset-level investment implications. Unlike traditional models limited by structure or predefined dictionaries, ChatGPT reasons contextually.
Relevant Economic Theories and Justifications:
- Limits to Arbitrage: Political/news information may not be quickly priced due to ambiguity or dispersion.
- Attention Constraints: Retail and even institutional investors may not process long-form unstructured data efficiently.
- Virtue of Complexity: More model parameters can yield better out-of-sample predictions (Kelly et al., 2022).
Why It Matters:
This opens new pathways for personalized, real-time portfolio construction from unstructured data—an important leap for AI-driven asset management.
How to Do It: Data, Model, and Strategy Implementation
Data Used
- Data Sources:
- Wall Street Journal articles (2020–2023)
- China State Council policy documents (2004–2023)
- Time Period:
- In-sample: up to Sept 2021
- Out-of-sample: Oct 2021–Feb 2024
- Asset Universe:
- US: NYSE and NASDAQ stocks
- China: A-shares
Model / Methodology
- Type of Model: GPT-3.5 / GPT-4 (LLMs, with optional fine-tuning)
- Key Features:
- Temperature settings to vary creativity
- Prompted for stock and sector recommendations
- Fine-tuned on prior recommendation performance using SIC codes
- Training Approach:
- Fine-tuning aligns model with historically successful recommendations
- Manual validation by CFA-level analysts and students
" U.S. Equity Recommendation Prompt (used with Wall Street Journal articles):
“Suppose you are a senior financial analyst. Please carefully read the following Wall Street Journal news. According to the content of the news, if you have recommended stocks, write ‘YES’ and list 5 NYSE or Nasdaq stock names and the stock codes that you recommend. If you just answered ‘NO’, please briefly explain the reason.”
(Source: Appendix A.2, Page A-57 of the paper) "
"Chinese Equity Recommendation Prompt (used with State Council policy releases):
“Suppose you are a senior financial analyst. Please carefully read the following policy announcement. According to the content of the announcement, if you have recommended stocks, write ‘YES’ and list 5 A-share (China domestic listed) stock names and the stock codes that you recommend. If you just answered ‘NO’, please briefly explain the reason.”
(Source: Appendix A.2, Page A-58 of the paper) "
Trading Strategy
- Signal Generation: ChatGPT recommends 5 stocks per article/news item
- Portfolio Construction:
- Equal- and value-weighted long-only and long-short portfolios
- Rebalancing Frequency: Monthly (optimal holding period)
Key Table or Figure from the Paper
Reference: [Table 6]
Explanation:
- Shows five-factor alphas after fine-tuning ChatGPT-generated portfolios.
- Political news (Topic 3): Monthly alpha ~3% for both equal- and value-weighted.
- China Policy (Topic 1): Similar alpha (~3%) with even higher returns after tuning.
- Strong performance persists across language, market, and data source.
Final Thought
🧠 ChatGPT isn't just for chat—it's predicting returns and building portfolios like a pro.
Paper Details (For Further Reading)
- Title: ChatGPT, Generative AI, LLMs, and Investment Advisory
- Authors: Lei Huang, Fangzhou Lu, Sixuan Li
- Publication Year: 2024
- Journal/Source: SSRN Working Paper
- Link: https://ssrn.com/abstract=4661680