Trying to trade with generalist LLMs is just an exercise in futility, because none of these models have ever seen the inside of a real trading firm. None of that knowledge is in their training sets.
> Plug your agent into the sources where information breaks first. Twitter, Telegram, Discord, on-chain activity. Your agent acts before the market does.
In a world where people are fighting with each other to see who can get closer to the trading systems in order to shave off milliseconds, this seems glacial.
The belief that there is some kind of market-impacting underground "wisdom of the crowd" to be found on all these public social platforms is an artifact of the GameStop craze that never went away.
It isn't so much that. Imagine if you were able to abrogate a pattern out of general communication networks that predicted a meme stock rise or a shitcoin ahead of a pump and dump. That would be extremely lucrative. That being said I don't think LLM is any tool for the job. You'd be better off working with the underlying datasets yourself using some graph based analysis.
>In a world where people are fighting with each other to see who can get closer to the trading systems in order to shave off milliseconds, this seems glacial.
the race to zero is over, nobody is doing that any more.
While most folks are quick to write off agentic trading attempts for various reasons, one strength that I see in using LLMs is its capability to bring together disparate information sources and chart a path forward.I do agree it is a very advanced word calculator but that seems to work most of the time unless there is a massive information gap in the original data sources.
I remember a friend telling me the state highway department used to call them guardrails, but after some lawsuits, they started calling them guide rails. This legal distinction was because they didn't want to imply they could prevent accidents.
I know it's bad manners to point out when "repositories" of "code" are just markdown files, but these are even more pointless markdown than most. They're just telling agents how to use their public, documented API. What is the possible scenario where an agent working with Shuriken would not load "Integrating and building with Shuriken" into context? (Where are the "safe guardrails" in the hallucinated title?)
18 comments
[ 3.9 ms ] story [ 40.9 ms ] threadhttps://en.wikipedia.org/wiki/DeepSeek
In a world where people are fighting with each other to see who can get closer to the trading systems in order to shave off milliseconds, this seems glacial.
the race to zero is over, nobody is doing that any more.
It works
AMA
If it worked, you would not have published a paper or a OSS project -- you would have leveraged it to do actual trading.
Anything in these domains that you find on the internet for free, does not work.