MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models

Battle Count: 78

Directly addresses core quantitative trading problems: alpha generation, execution optimization, feature engineering, and joint strategy design. Provides rigorous statistical framework for evaluating whether evolutionary search produces genuine alpha vs. p-hacking. Demonstrates practical improvements (up to 25x PnL, +1.83 Sharpe OOS) on a realistic trading simulation with market impact modeling. The methodology is directly applicable to trading desk research workflows.

Requires setting up a minute-bar backtesting engine with market impact modeling, configuring an LLM ensemble with API access to multiple providers, implementing MAP-Elites and island model population management, defining EVOLVE-BLOCK markers in strategy code, and managing the evolutionary loop infrastructure. The framework is publicly available but integrating it with a custom trading strategy and simulation requires significant engineering effort. Hyperparameter calibration adds another layer of complexity.

Framework is publicly available at https://madevolve.org. Detailed simulation setup, skeleton code for base strategy and forecaster, market impact model parameters, and hyperparameter calibration details are provided in appendices. Data source (Polygon BTCUSD 1-minute bars) is specified. However, exact LLM API calls, random seeds for some runs, and full evolution logs are not fully disclosed. The backtest is acknowledged as not exchange-specific.