A sober look at ChatGPT for stock trading
A sober look at ChatGPT for stock trading
If it's too good to be true, it usually is.
Welcome to Market Sentiment. We curate the best ideas from thousands of research sources and distill them into weekly actionable insights. Based on the poll in last week’s ideastorm, 45% of you voted for a deep dive on using chatGPT for investments.
If you are new here, join 35,000+ others who receive curated financial research.
Actionable Insights
For stock picks using historical stock price data, ChatGPT underperforms even simple methods such as regression and random forests.1
But, ChatGPT outperformed traditional sentiment analysis tools in evaluating financial news and its implications2. GPT-4 was advanced enough to be on par with human reasoning in decoding Fedspeak.3
Long-short strategies derived from sentiment analysis on GPT 3.5 considerably outperformed the market — with a lot of caveats.
On May 11, 1997, an event took place that forever changed the field of artificial intelligence. A classic man-versus-machine story for the ages. The reigning world chess champion, Garry Kasparov was defeated by Deep Blue, a supercomputer developed by IBM. It was the first time a computer definitively beat the world’s best chess player under tournament conditions.
This milestone was significant because chess is a game that requires strategy, foresight, and logic — qualities that make up human intelligence. Finding the best move is not just a brute-force problem that computers can solve4 as there are more possible move variations in chess than there are atoms in the observable universe.
Scientists had been working on a chess-playing computer since the late 1940s, and it took more than 50 years of development to beat the world's best chess player. In 2016, AlphaGo developed by Google went one step further by beating eighteen-time world title winner Lee Sedol in the game of Go — A game with more than 10170 moves making it considerably more complicated than chess (at least computationally).
What was stunning was that AlphaGo invented new moves and unconventional strategies that never existed before.
AlphaGo played several inventive winning moves, several of which - including move 37 in game two - were so surprising that they upended hundreds of years of wisdom.
Using Machine Learning and Artificial Intelligence is nothing new for the stock market. Jim Simon’s Medallion Fund has been consistently beating the market over the last 30 years by using quantitative models and advanced trading algorithms.
With the release of ChatGPT, cutting-edge Natural Language Processing (NLP) is now accessible to almost everyone. With the accessibility comes the allure of using it for stock picking — while there are a lot of dubious sources claiming to find alpha using ChatGPT, let’s take a look into the latest research:
No free lunches
First, it’s important to get this out of the way — ChatGPT standalone cannot pick stocks. As you might have guessed by now, contrary to popular belief, you cannot feed historical data into ChatGPT and expect it to pick the right stocks.
Research [1] indicates that ChatGPT underperforms even traditional methods like linear regression for predicting stock price movement using historical data. This is expected as ChatGPT is just an NLP program developed by training on a massive corpus of text data. While this can give state-of-the-art performance in text completion, summarization, etc. it cannot easily translate to deep financial know-how.
Expecting ChatGPT to pick stocks would be akin to expecting Buffett bot to run Berkshire.
What is ChatGPT really good at?
ChatGPT is excellent at understanding context and nuances. GPT-4 was advanced enough to be on par with human reasoning in decoding Federal Reserve communications.
Leveraging this, researchers at the University of Florida were able to develop a long-short strategy using ChatGPT that generated a 550% return (Oct '21 to Dec '22 - during which, the S&P 500 was down ~10%). Instead of asking ChatGPT to pick stocks directly, they used it to evaluate the sentiment of news surrounding companies.
ChatGPT showed remarkable accuracy in identifying the context that traditional sentiment analysis tools lacked.
Compared to the above response, researchers found that one of the most developed sentiment analysis tools gave a sentiment score of -0.52, indicating the news was unfavorable for Oracle. The ability of ChatGPT to understand complex financial nuances might enhance the performance of quantitive trading strategies.
Let's dig in: