Hacker News Reader

Comment by ta12653421 | Hacker News Reader

Thanks for the question:

Not really in the field of trading, rather prediction-based approaches etc.; Im not sure if LLM could be of any use here? The approaches based on statistical arbitrage are purely math/number models. from my own experience, LLM are absolutely useless when it comes to "trading ideas" (I use them for code generation, instead), this is because they are dicing together values in their output which are not really related, because of their hallucis.

Also, for fundamental analysis they are too often incorrect - so running an auto-approach based on LLM-fundamental output would be an "interesting" idea :))

View on HN

Interesting, so you're doing more fundamental quantitative analysis and prediction. Yes, the issue with using an LLM is that they are too often incorrect, however, a human in the loop could solve this at the expense of automation.

Your project sounds ready cool though. If you ever feel like collaborating, give me a shout at drknyt05@gmail.com.

Thats one of the primary edges: That there is _no_ human in the loop in my approach; actually, the system is build around a simple idea (regardless which approach taken): Out of 700+ stocks globally (blue chips only), there is always something in a "robust trend situation", you just have to scan the market automated and pick the ones with the highest probability. To do so, you need an application (otherwise you would die of boringness in front of the screen), and the app is submitting all trades automaticly (sure, you can cancel one, but a manual action is usually not required)

What most people dont get: This is not about "predicting" where the price will be, this is about taking momentum and weeding out the bad options - and applied on a large set of stocks, there is always something you can ride with for a couple of days (with applied leverage, there is usually a substantial profit)

you could certainly encode data into an transformer using custom tokens and fine-tune but that's not trivial.

This is what i referred to as "too high costs" :)

Im sure, there are ton of other options by applying whatever "mega-tech", but if the result is only slightly better while having much more costs (and complexity!), for me its not worth (as i'm not a company, but an individual invstor)

I mean the crazy thing is that I have implemented transformers in another language so I know exactly what needs to be done and I know why it's such a pain in the ass in python, but I just don't have the discipline to sit down and do it with a risky reward (I spend my risk tokens on other things). Now, if I had a patron, I could probably figure it out.

Reply on news.ycombinator.com