AI summary by TwinMind
Using ChatGPT to Trade Stocks - Let's Talk
The Plain Bagel
12m 24s
Introduction - AI Trading Hype in 2023
- Host: Richard Coffin, The Plain Bagel
- AI dominated 2023 investing discussions
- ChatGPT leading charge with impressive Q&A capabilities
- People asking "how to make money" with ChatGPT
- Rise of ChatGPT stock trading: strategies, trading bots, stock picks
Research Claims Supporting ChatGPT Trading
- Finder.com article: ChatGPT vs 10 UK popular funds - ChatGPT outperformed
- University of Florida research: ChatGPT earned 500% return in 1+ year using sentiment analysis
- CNN article: "ChatGPT can pick stocks better than your fund manager"
- Thousands of TikToks claiming large returns with ChatGPT trading bots
- GPT Portfolio Twitter account:
- Started with $50,000 using ChatGPT
- Outperformed S&P 500 in first week (May)
- Now has $27 million of other people's money trading alongside
How ChatGPT Actually Works
- Large language model (LLM)
- Processes vast text data: billions of words from online sources, web pages, Wikipedia
- Develops neural network for understanding natural language structure/patterns
- Fine-tuned using human feedback
- Excellent at language tasks: summarizing, translation, answering questions
- NOT a stock analyzing system
- Pre-trained on data only up to 2021 (stale company data)
- Stock picks based entirely on web pages from pre-training period
- Internet search feature just summarizes web pages
- Doesn't calculate economic forecasts, revenue growth, or financial variables
- "Regurgitating what it reads online" - not necessarily best stock pick source
Contradictory Research Findings
- Different 2023 research: ChatGPT underperformed simple models like linear regression for numerical data stock price prediction
University of Florida Study Deep Dive
Methodology
- Researchers gave ChatGPT headlines to analyze
- Asked to determine if headline was positive/negative/neutral for company
- Scoring system: -1 (bad), +1 (good), 0 (neutral)
- Multiple headlines averaged per stock
- Buy/short based on positive/negative sentiment
- Exit position at next market close
- ChatGPT wasn't actually picking stocks - just indicating if headlines were good/bad news
ChatGPT's Language Understanding Advantages
- Better at nuanced headline interpretation
- Could determine if competitor's bad news = good news for stock in question
- Previous research showed ChatGPT could interpret "fedspeak" (Federal Reserve announcements)
- Could pick up nuances to determine Fed's rate intentions (hiking/lowering/stable)
Major Study Limitations
- 500% return figure ignores transaction costs
- With 25 basis point transaction cost: return drops to ~50% over 1+ year
- No mention of taxes (would reduce returns significantly due to trading frequency)
- Strategy worked best for micro-cap companies (bottom 10th percentile market caps)
- Liquidity issues: might not find buyers/sellers for hypothetical trades
- Report lacks peer review, missing calculations, basic details like trade sizing
- No outlined assumptions
Version Performance Discrepancy
- 500% return came from ChatGPT 3.5 (older, less extensive model)
- ChatGPT-4 (current, better system): nearly half the return of 3.5
- Better language understanding = worse returns
- Suggests performance due to chance, not superior sentiment analysis
Accuracy and Reliability Issues
- ChatGPT sentiment analysis: only 51% accuracy likelihood
- "Little better than tossing a coin" for individual headlines
- Works through aggregation across multiple companies/days
- Examples of ChatGPT inaccuracy: lawyers facing fines for using fabricated case citations
Alternative Explanations - Role of Chance
- Michael Reeves: fish outperformed NASDAQ
- Random portfolios often outperform market over given timeframes
- ChatGPT success likely "dumb luck"
- Short timeframe doesn't prove sustainable strategy
Real-World Performance Updates
- Twitter ChatGPT fund: lagging S&P 500 since May
- UK funds comparison: ChatGPT still underperforming S&P 500
- AI ETFs intended for stock selection: many still lag S&P 500
Why Automated Systems Struggle
- Opportunities are marginal and fleeting
- Very small returns requiring tens of thousands of trades
- Opportunities disappear over time
- Majority of market trading already algorithmic/AI-driven
- Existing systems compete over smaller arbitrage opportunities
- Quantitative/hedge funds already use proprietary AI for stock analysis
- Widely available AI (ChatGPT) unlikely to provide trading edge
Scam Alert - Social Media Claims
- TikTok videos/YouTube ads showing "crazy returns" mostly scams
- Not all fraudulent - some representation bias (seeing one success among thousands of failures)
- People exploiting public misunderstanding of how ChatGPT works
- Reference to "How Money Works" video on ChatGPT scams
Legitimate Applications and Limitations
- ChatGPT applications in investing/trading are supplemental only
- Dangerous to rely on ChatGPT for actual trades/stock picks
- "Might as well be throwing dart at board"
- ChatGPT's coding ability for trading bots:
- Requires user to specify strategy
- Often contains bugs preventing proper function
- Technological advancement ≠ investment opportunity
- Warning against get-rich-quick scheme mentality
