Do you see any edge using LLMs that more traditional ML doesn't have?
In my limited experience, small models trade off noise, whereas frontier models are too slow to be usable for something like HFT. I find you really want something trained on market data, not generalist next-token predictors.