I wrote in "Brave New Backtest" about how LLMs have architectural limitations that make them bad at trading research. Proactive interference, mode collapse, serious training data limitations.

I thought the article would very effectively warn people off of using AI for early-stage trading research. After all, it provides the theoretical basis for what any experienced practitioner immediately observes from the output of the vibe quanting crowd - slop that can't possibly work.

But what you put out into the world doesn't always have the effect you assumed it would.

The number of comments and DMs I received suggesting ideas for "fixing" the AI to make it more effective at the task was not what I expected. It's like every problem resembles a nail and AI has become the only hammer in town.

Vibe quanting can't possibly work. For all the reasons I laid out previously. And if you need more convincing, read any of the "AI in trading" slop on Substack and ask yourself "why would I get paid for doing that?" We're all grown ups. Deep down you know it's wishful thinking.

I kind of get the appeal around "fixing" the AI to do better at trading problems... After all, what AI can do in the programming domain is extraordinary. I can see the appeal in having a tool like that on your trading desk. But it can't achieve that same level of performance.

Instead, we need a bit of curiosity, discipline of thought, and time. This is what's needed to develop a theory of edge: Who's paying you on this trade, and why will they keep doing it? Everything flows from that. But the good news is that anyone can learn this stuff. It's not rocket science. But sometimes smart people want to solve hard problems even more than they want to make money...

Anyway, I'll start laying out this "theory of edge" for you in this week's article.