Key Takeaways
- •ChatGPT overvalued Chinese stocks by 12.5% average
- •Forecast errors 13% larger than DeepSeek
- •More frequent “buy” signals from ChatGPT
- •Study reveals AI bias on foreign equities
- •Investors should scrutinize LLM stock recommendations
Summary
Harvard Business School researcher Charles C.Y. Wang compared ChatGPT and DeepSeek on stock analysis for roughly 5,000 publicly traded Chinese firms. ChatGPT’s price forecasts were on average 12.5% higher and it issued more "buy" recommendations, yet its prediction errors were 13% larger than DeepSeek’s. The study labels this tendency a "foreign bias," suggesting the model overestimates the value of non‑U.S. equities. Findings raise questions about the reliability of large language models for cross‑border investment advice.
Pulse Analysis
The Harvard Business School investigation adds a new layer to the growing conversation about artificial intelligence in finance. While large language models like ChatGPT are praised for their ability to synthesize massive data sets, the study demonstrates that their training on predominantly English‑language, U.S.-centric sources can skew outputs when applied to foreign markets. By tasking both ChatGPT and the Chinese‑focused model DeepSeek with evaluating 5,000 Chinese companies, researchers uncovered a systematic overvaluation trend, highlighting the importance of domain‑specific data in AI model performance.
Results show ChatGPT projected stock prices 12.5% higher and issued more "buy" calls, yet its forecast error margin was 13% larger than DeepSeek’s. This discrepancy suggests that the model’s confidence does not translate into accuracy, especially when dealing with markets that differ culturally, regulatory-wise, and in reporting standards. Investors using LLM‑generated recommendations must therefore treat such signals as preliminary insights rather than definitive advice, supplementing them with traditional analysis and localized expertise.
The broader implication is a call for greater transparency and calibration of AI tools used in investment decision‑making. Financial firms deploying LLMs should incorporate bias‑detection frameworks, diversify training corpora, and continuously back‑test model outputs against real‑world performance. Regulators may also consider guidelines to ensure that AI‑driven advice meets fiduciary standards, protecting investors from inadvertent overexposure to mispriced assets. As AI integration deepens, rigorous validation will be essential to harness its potential without compromising market integrity.
Comments
Want to join the conversation?