不断进化的交易者：自主策略演化

A system has been built that evolves trading strategies using large language models (LLMs) as mutation operators. Inspired by Karpathy's autoresearch, it starts with a simple 20-line seed strategy and runs three parallel mutations per generation: conservative tweaks, wild swings, and a completely fresh strategy every five generations. After 176 generations, it evolved into a complex 200-line strategy featuring dual RSI filters, ATR volatility scaling, MACD confirmation, Bollinger squeeze detection, and conviction-based position sizing, all without human-written code. One unexpected outcome was the shift from writing code to creating markdown files that instruct agents on what to optimize, with agents generating the code. The coaching loop emerged as a significant surprise—without it, the LLM would merely recycle existing content.