RD-AGENT Lecture 3: What do LLM-written factors look like: The Research Taste of AI
第三讲 · LLM 写的因子长什么样：AI 的研究品味 第一章第三讲（本章收尾） 学习目标：通过真实案例建立对 LLM 因子假设质量的判断力；识别"伪逻辑假设"的特征；理解代码质量与逻辑质量的不对称。 1. 开场 前两讲我们隔着玻璃看工厂：循环怎么转（第一讲）、成绩单怎么读（第二讲）。这一讲开门进去，把 AI 写的因子摆上解剖台。材料全部来自真实战役产出（A 股 290 因子池、期货两轮、crypto 两轮）。 先给本章的总结论，然后逐类验证它： AI 的代码比它的金融直觉诚实得多。 代码 bug 是显性的——跑不通就报错，反馈即时；逻辑空洞是隐性的——假设再空，回测照样吐出一个数字，而这个数字还会进入排行榜参与选拔。 2. 案例一：像样的——但要看清"像样"的来源 期货面板上真正有产业逻辑的因子，集中在人给的两列特有数据上： 基差（期现价差） 和 持仓量…
A study on factors generated by Large Language Models (LLM) reveals that while LLM's code is reliable and honest, its financial intuition can be flawed. The study analyzed 290 factors generated for the A-share market, as well as factors from futures and crypto markets, and found that LLM's research quality is limited by the quality of the data panel it is given.
The study identified three types of flawed factors: those that appear good but are based on human-prepared data, those with pseudo-logic that can be identified by features such as high adjustability and vague mechanisms, and those that fail due to errors in code or logic. The study concludes that LLM's ability to generate new factors is not the bottleneck, but rather humans' ability to verify the factors.
Written by urgent.news from Dev.to's report — not a translation of it. Machine-written — may contain errors; check the original before relying on it.
