ishtiaqrahman

thanks for the suggestion. I am still new to the Discord game. TBH, if it continues to be this successful I'll start a hedge fund or something.

I appreciate your encouragement. I'll continue the effort!

Nope that's not me. I think they abandoned their project.

You misunderstood my research, I think my wording wasn't clear enough. Let me clarify.

I am comparing SPY's return with the GPT return for the exact timeframe during which a particular position was recommended by the GPT Investor. For example, some positions were held for only 7 days, while others were held for 3 months, etc. I generated 19 different sets of recommendations with varied timeframes and at different points in the last year. This link, where I track everything, should clarify any confusion: https://www.gptinvestor.co/the-gpt-investor/

I am comparing SPY's return with the GPT return for the exact timeframe during which a particular position was recommended by the GPT Investor. For example, some positions were held for only 7 days, while others were held for 3 months, etc. I generated 19 different sets of recommendations with varied timeframes and at different points in the last year. This link, where I track everything, should clarify any confusion: https://www.gptinvestor.co/the-gpt-investor/

Yes. GPT-3.5 made some terrible calls and didn't really beat $SPY. GPT-4 turned it around for team LLM.

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%)

I am starting to use longer timeframes, such as one year. I consider this a long-term experiment and will continue my research. Based on the improvement in returns I have seen from GPT-3.5 to GPT-4, my hypothesis is that it will continue to outperform $SPY as the models become smarter. For reference, GPT-4's return was approximately 156.86% better than GPT-3.5's return. GPT-3.5 didn't actually beat $SPY.

I think you're confusing me with someone else. I do not buy any ads on Twitter. Maybe I should! What do you think?

I do charge a tiny subscription fee for full access to all my research to fund further research on this.

GREAT QUESTIN!

My theory is that since LLMs (Large Language Models) are the ultimate generalists when it comes to knowledge, they "know" a decent amount about every single topic and concept known to the entire human race. I do not believe there is a single human alive today who possesses knowledge about so many different subjects. For example, a physicist may know 100 times more about physics and mathematics than an LLM, but the LLM probably knows a decent amount about 10,000 more disciplines (like plant biology) that the physicist has little to no understanding of.

I believe this multidisciplinary ability makes LLMs uniquely qualified to pick stocks based on the millions of variables that may impact stock prices.

For example, one of the most successful strategies I deployed was when I asked GPT-4 to come up with the attributes of a hypothetical "most investable stock on the NYSE and NASDAQ based on current market conditions." Once it generated the attributes, I used another instance of GPT-4 to find me a stock that matched these attributes. It came up with NVIDIA. The actual prompt I used was more detailed, but you get the idea.

I also noticed that. It recommended NVIDIA 5 times.

I am definitely going to continue the experiments. The ROI improvement from GPT-3.5 to GOT-4 is very encouraging. I also got better at prompting.

haha the thing is that I got better at prompting as I ran more and more experiments.

For example, in the beginning I was asking GPT to recommend me 3 stocks at a time, which was not a good idea. I revised the strategy to ask for only 1 stock recommendation at a time.

Also, GPT-4 was MUCH better at it than GPT-3.5 which makes sense and is also very exciting.

I published the all the methodologies and data, do take a look if you're curious.

Thanks! I know a lot of people wondered if it was possible.

thank you!