LLM-assisted strategy research: a workflow that helps vs hype to ignore

Using LLMs for quant research is mostly about speed, not intelligence. If you treat them like a senior analyst, you will get hallucinations. If you treat them like a junior coder who has read every paper but has zero common sense, they become useful.

The real value lies in the boring parts of the workflow. They excel at boiling down a dense 50-page paper into a summary of the core logic, or translating a specific math formula into a Python snippet. I often use them to generate variants of a signal logic I have already built. For example, if I have a momentum filter, I ask for five ways to normalize the input data or handle outliers. It saves me from writing boilerplate code from scratch.

Here is where you need to be careful:

• Never trust them with backtest logic or performance numbers. They will invent plausible-sounding results because they are trained to predict the next word, not the next tick.
• Avoid asking them to generate a full trading strategy from scratch. They tend to overfit the narrative to whatever trend is currently popular.
• Always verify the math. They often swap variables or misapply standard deviation formulas in code.

Tools like Claude or ChatGPT are great for scaffolding, but they cannot replace the intuition needed to judge if a signal is actually capturing a structural market bias. I usually treat their output as a rough draft. If the logic doesn't make sense on a whiteboard, it won't make money in a backtest. How do you integrate these tools into your own pipeline without falling into the trap of over-reliance?