Built a system that evolves trading strategies using LLMs as mutation operators.
Inspired by Karpathy's autoresearch — start with a dumb 20-line seed strategy, run three parallel mutations per generation. exploit (conservative tweaks), explore (wild swings), and genesis (completely fresh strategy every 5 gens).
176 generations later it evolved a 200-line monster with dual RSI filters, ATR volatility scaling, MACD confirmation, Bollinger squeeze detection, conviction-based position sizing. no human wrote any of it.
The thing i didn't expect: you stop writing code and start writing markdown files that tell agents what to optimize. the agents write the code. you write the intent.
Coaching loop was the biggest surprise — without it the LLM just keeps recycling the same failed ideas. and diversity matters way more than exploitation. the genesis mutations (fresh starts from scratch) produced some of the biggest fitness jumps.
Full writeup: https://www.jovweb.dev/blog/evolving-trader-autonomous-strategy-evolution
Top comments (0)