evolution-engine

Evolution Engine

Overview

The Evolution Engine autonomously discovers trading strategies from raw price data. It uses LLM-powered pattern generation combined with vectorized backtesting to evolve, test, and graduate viable trading rules — without manual rule writing.

This is not parameter optimization on a known strategy. It's open-ended strategy discovery: the LLM proposes novel entry/exit logic, the engine validates it against real data, and natural selection eliminates the losers.

How It Works

The Evolution Loop

OHLCV Data → LLM Generation → Vectorized Backtest → Selection → Mutation → Repeat
                                                          ↓
                                                    Out-of-Sample Validation
                                                          ↓
                                                    Graduated Strategies

Step-by-Step

1. Data Fetch: Pull OHLCV candles from Binance public API (no key needed)
2. Generate: LLM analyzes price patterns and proposes N candidate strategies (entry/exit rules, position sizing, stop loss)
3. Backtest: Each candidate is backtested vectorized (numpy, no loop-per-candle) for speed
4. Score: Candidates scored by Sharpe ratio, win rate, max drawdown, total return
5. Select: Top K candidates survive. Bottom candidates are eliminated (graveyard).
6. Mutate: LLM takes survivors and generates variations (parameter tweaks, rule modifications)
7. Repeat: Steps 3-6 for N generations
8. Validate: Final survivors are tested on held-out out-of-sample data
9. Graduate: Strategies that pass OOS validation are marked as graduated

Key Design Decisions

• LLM generates rules, not parameters. The engine doesn't optimize MACD(12,26,9) → MACD(14,28,10). It discovers entirely new rule combinations.
• Vectorized backtesting. No candle-by-candle loops. Numpy vectorized operations make backtests 100x faster than event-driven simulators.
• OOS validation is mandatory. In-sample performance means nothing. Only OOS-validated strategies graduate.
• Graveyard is data. Failed strategies are logged with failure reasons. This prevents re-discovering the same dead ends.