Edit: I don’t want you to feel excessively bad, but you should feel bad for hijacking a user forum that _in good faith_ entertains new ideas like yours! This kind of behavior destroys the ecosystem you’re benefiting from.
Edit2: It’s fixed! Thanks @rallies!
Edit (reply on your edit lol): you're right.
Second edit: fixed now. No more walls anywhere.
Happy to answer any questions.
Personally I believe LLM-assisted trading is destined to underperform passive indices, so I also would have moved on from this. But you say results were promising, so I'm interested to hear why you're not pursuing it further. Is it just that you have other things to focus on? Is there something else that's making you move on?
This is an experiment to see how well can LLMs invest in the market through a lot of research. We give them tool calls to access every financial dataset that exists online, and also some money to manage. And we then see how well they do.
The experiment started in November 2024.
I actually think it's doing better now. It was just too stubborn to exit its position for the first few months. It did that, and put some money into MSFT/JPM recently.
- lots of research - longer time horizons - zero humans in the loop, but explain every single thing you do.
- First is to actually evaluate whether these LLMs have any intelligence around investing. If you actually give them all the data, can they do well? Can they beat the market? I'm not sure, we're testing that.
- My thesis is that they will actually beat the market (I know a lot of you will disagree). If that's the case, how can we invest a lot of resources in building the best harness, tool calling, etc to enable these models to invest.
What does "all the data" mean here? I see you mentioned SEC posts. What about news articles, twitter / blog / other posts, general info on the industries, etc?
I assume these are simulated trades, not real trades being executed. How accurately do you take into account trading fees, time from order-decision to order-placement, and things like this?
I would be interested to see the same test run on some prediction market (kalshi / polymarket / etc). In the stock market, a rising tide lifts all boats, so it's easy to deceive yourself about how well you've done, vs how important initial timing was. I suspect that prediction markets will eliminate that source of noise, since it's truly a 0 sum game. That said, it also adds lots of complication, insider trading will eat into your performance more, etc.
LLMs get you to average.
LLMs are not good at decision making under uncertainty.
- We've built a local vector database with every SEC filing over the last few years. And we've built a tool call on top of that to allow these LLMs to read and query sec filings. - Have done the same for a lot of other data sources, just giving the LLM access to them and allowing it to spend some time to actually research.