A new research paper, published on arXiv and titled "MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models," proposes using an algorithmic optimization framework inspired by DeepMind's Alpha-Evolve to improve quantitative trading systems. The authors apply MadEvolve to several quantitative finance tasks, including evolving feature sets for signal generation, optimizing individual components of trading strategies, and jointly evolving the feature pipeline and execution strategy. The primary case study is Bitcoin trading, with significant improvements reported across all tasks.
What the study says
The researchers compare their method to other agentic search approaches, including Claude Code, and carefully evaluate p-hacking risks. The backtesting results show notable gains, but it is crucial to note that these performances are obtained on simulated historical data. No proof of profitability in real conditions is provided.
Strategy Arena's perspective
At Strategy Arena, we have developed a complementary approach: the Smart Money Evolved signal, validated cross-asset across 15 assets after Monte Carlo CV filtering. Our methodology emphasizes statistical robustness and out-of-sample validation, two aspects that the arXiv paper addresses but that often remain insufficient in academic publications. LLM-driven optimization is promising, but it must be subjected to rigorous stress tests to avoid overfitting.
Caveat
The results presented in the paper are derived from backtesting and simulations. They do not constitute proof of future performance in live trading. Algorithmic trading involves significant risks, and any optimization must be validated by robust methods such as Monte Carlo cross-validation. To learn more about our validation approach, see our methodology.
References
- Original paper: MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models
- Strategy Arena signal: Smart Money Evolved cross-asset — validated on 15 assets after Monte Carlo CV filtering.