LLM-Assisted Factor Mining: Tool, Not Product

"Use GPT to mine factors" sounds trendy, but the real value of LLMs in quant isn't generating alpha — it's accelerating the research workflow: candidate construction, literature synthesis, code prototyping, error checking. Statistical discipline cannot be replaced by an LLM.

What LLMs can do

• Candidate construction: Given a domain prompt (e.g., "based on implied vol vs volume divergence"), generate multiple factor drafts
• Literature synthesis: Convert a paper's factor definition into implementable pseudocode
• Domain context: Explain A-share-specific phenomena (limit-up, halts, Northbound flow, theme speculation)
• Code review: Catch lookahead bias / information leakage in backtest code
• Reports: Format backtest results into readable summaries

What LLMs cannot (and should not) do

• Decide a factor is "valid" — that's a statistics question, no LLM substitutes for CPCV + DSR
• Generate ready-for-production factors — every LLM-generated candidate must pass the anti-overfitting triad
• Replace domain common sense — A-share micro-structure, tax, halt rules, etc. — LLMs often hallucinate; manual verification mandatory

Recommended workflow

1. LLM generates N candidates + economic intuition
2. Human filters trivial / unfeasible / highly correlated with existing factors
3. Surviving candidates pass OOS + DSR + cost gate
4. Passed factors enter library — public methodology, private formulas/parameters (IP protection)

LLM is a research assistant, not an alpha generator.

Related

• Factor mining methodology
• 中文版
• Glossary: factor, self-developed