Exactly one year ago, on May 16th, 2023, I started investing in the stock market using LLM-powered autonomous agents, which I named The GPT Investor. These investments were real, made with actual cash in the NYSE and NASDAQ, not just paper trading.
My goal was to determine if autonomous agents powered by Large Language Models, such as GPT-4, could "beat the market."
"Beating the market" refers to achieving investment returns that exceed the performance of a benchmark index, such as the S&P 500. It implies that an investor's portfolio has generated a higher return compared to the average market return over a specified period. This concept is often used to evaluate the success of investment strategies or the skill of portfolio managers.
I wanted the GPT Investor to compete against the SPDR S&P 500 ETF Trust, also known as $SPY, an exchange-traded fund (ETF) that aims to track the performance of the S&P 500 Index. This index includes 500 of the largest publicly traded companies in the U.S., making $SPY a popular investment vehicle for those seeking broad exposure to the U.S. stock market.
$SPY is a formidable opponent because, over the last 10 years, fewer than 10% of active U.S. stock funds have managed to outperform index funds like $SPY.
If The GPT Investor could beat $SPY, it would mean it outperforms 90% of professional fund managers.(e.g. In the last one year, Warren Buffet couldn't beat $SPY)
Methodology:
I experimented with various methods and technology stacks to generate stock recommendations, meticulously documenting and reporting the results to our subscribers. The common thread among these methods was:
-Using OpenAI's LLMs (GPT-3.5 and GPT-4.0)
-Enabling the LLMs to autonomously search the web
-Generating stock recommendations for specific durations, such as three months
I utilized platforms like ChatGPT, Godmode, and BabyAGI UI to generate the stock recommendations. Each of these platforms can perform multi-step reasoning, a crucial attribute of autonomous agents. For example, based on a prompt, the agent can create its own to-do list and independently execute the steps to arrive at a result.
I conducted 19 experiments, investing CAD $300 in each, for a total of CAD $5,700 invested. I used a Wealthsimple brokerage account to execute the trades. Since each stock recommendation had a specific duration, I closed the positions at the end of each duration and compiled the returns as part of The GPT Investor portfolio.
For every experiment I ran, I published the entire methodology (tech stack, prompt, LLM, etc.) and results on this platform—The GPT Investor (www.gptinvestor.co)
Results:
-Total Invested: CAD 5,700
-Number of Experiments: 19
-Shortest Experiment Duration: 7 days
-Longest Experiment Duration: 1 year
-Number of Stocks Recommended by The GPT Investor: 31
Total Return:
-The GPT Investor: 11.54% (CAD $658.20)
-$SPY: 8.89% (CAD $507.10)
Overall, The GPT Investor Portfolio return was approximately 29.78% better than the $SPY's return.
Number of experiments by LLM
-GPT-4: 11 experiments
-GPT-3.5: 8 experiments
The average return for the two LLMs used by the GPT Investor is as follows:
-GPT-4: 15.54% (with a corresponding average $SPY return of 9.74%)
-GPT-3.5: 6.05% (with a corresponding average $SPY return of 7.74%)
*GPT-4's return was approximately 156.86% better than GPT-3.5's return.
The results raise the exciting possibility that as LLMs become more powerful, the returns of The GPT Investor should improve even further.
I publish the real-time status of The GPT Investor here: https://www.gptinvestor.co/the-gpt-investor/
What were max drawdowns on spy and your experiment?
If yours was smaller, then it's interesting, because it means higher returns with lower volatility.
* Also, looking at SPY data it looks like its 1-year return from today was 26.05% rather than the 3x smaller number you gave. Even considering that you operate in CAD that's quite a discrepancy...