Adrian Reid argues that large language models can't build reliable trading systems because they're trained on mixed data with no way to separate good advice from bad. LLMs regurgitate patterns from training data rather than rigorously backtest strategies, so any trading algorithm they generate needs human validation and quantitative testing before you'd risk real money.
AI does not know how to trade, okay? So I'm just going to say that right up front. It is not a miraculous super intelligence that can look at a chart and figure out exactly how to trade profitably, right? That's just not what it is, not right now anyway. Right now we have large language models which are trained on huge amounts of data, some of which talk about trading, as in the data that it was trained on talks about trading. And so it has learned some stuff about trading from the material it was trained got. Now, some of that stuff might be good and profitable and sensible trading wisdom. Some of it might be complete horseshit. And the trouble is the AI was trained on all of that, doesn't necessarily know the difference. So when you ask AI for a profitable trading system, it's going to look in its training data essentially and give you what it thinks is a profitable trading system based on how it's been trained. It's not going to go and do some analysis on 30 years of history and run backtests and come back to you with a profitable trading system that it just tested and created using a sensible methodology. So it has a lot of information, it can do a lot of stuff, but it is not doing what most people probably think it's doing in terms of coming up with a profitable trading system. So we're just going to burst that bubble right there. I hope that's not a surprise to most of you, but it's, it's not as intelligent as we think it is, but it's really good at doing a bunch of things that are useful.
