Does generative AI facilitate investor trading? Evidence from ChatGPT outages
Researchers are using outages of tools like ChatGPT to test whether generative AI is influencing stock trading behavior, prompting speculation about how LLMs might be used for tasks such as news and sentiment analysis in markets. Commenters note that while AI can help parse information and lower the social cost of asking “basic” finance questions, current models suffer from issues like hallucinations, sycophantic answer changes, and weak reasoning. The thread also veers into debate over Bitcoin’s legitimacy and macroeconomic properties, highlighting broader skepticism about both AI-driven and crypto-driven investing narratives.
Uses of Generative AI in Trading
- Proposed uses: parsing news, earnings-call transcripts, and other corporate text/audio into structured sentiment or signals for trading models.
- Some argue prompt quality and limiting hallucinations are key; others say “right” in trading just means doing what others do, slightly faster.
- Alternative view: skip LLMs and use simple, transparent sentiment methods (e.g., word lists and pattern matching), which may be more reliable.
- Recognition that many such data sources are trailing indicators, more explanatory than predictive.
LLM Behavior, Sycophancy, and Reliability
- Multiple comments note ChatGPT often concedes when challenged, even when correct, attributed to “sycophancy” and subservient tuning.
- Some suggest custom/system prompts and chain-of-thought prompting can improve behavior, but this doesn’t change base error rates.
- Example given where an EV-efficiency question and its inverse both get confidently but mutually contradictory answers, illustrating lack of true reasoning.
- One view: LLMs don’t “see” mistakes; they generate likely responses from patterns of conversation about being wrong.
- Idea of external intervention layers to correct or constrain LLM answers is floated, but others warn such non-differentiable hacks are technical debt.
Investor Behavior, Crypto, and LLM Influence
- Anecdote: a long-time finance professional became pro-Bitcoin after private Q&A with ChatGPT, highlighting LLMs as low-embarrassment learning tools.
- Others caution that LLMs “repeat what everyone is saying” and may confidently provide wrong or biased crypto takes.
- Heated debate on crypto/Bitcoin:
- Pro side: Bitcoin as “strong money” with fixed supply, censorship resistance, and large aggregate market value.
- Skeptical side: pervasive scams, volatility, limited real-world delivery, risks of liquidity crises, miner concentration, and fixed supply being macroeconomically harmful.
- Technical back-and-forth on miners vs verification nodes, 51% attacks, censorship limits, mempools, and incentives.
Privacy and Risk Perception
- Some users avoid asking LLMs sensitive or “stupid” questions due to fear of data logging and potential leaks; others value the lack of social embarrassment more.
- Example of a third-party app leak is cited; distinction drawn between core providers and wrappers.
Market-Timing vs Long-Term Investing
- Discussion revisits “time in the market vs timing the market.”
- Viewpoints:
- With small capital, aggressive timing may feel like the only path to meaningful gains (lottery-ticket analogy).
- With larger capital or fiduciary duties, steady returns and lower risk become preferable.
- High-frequency trading cited as proof that timing can work at scale, though individual competition is difficult.