NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 25·8 min read
MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models
TL;DR
Designing effective algorithmic trading strategies requires manually tuning many interdependent components, a time-consuming process prone to overfitting. MadEvolve, an LLM-driven evolutionary optimization framework, automatically generates and refines trading algorithms by iteratively improving feature sets and execution logic.
✦ Why It Matters
Engineers can apply LLM-driven evolutionary search to automate strategy discovery and reduce manual tuning overhead in quantitative finance.
Key Takeaways
- MadEvolve automates trading strategy design by iteratively proposing and testing algorithmic improvements using LLMs.
- Framework optimizes both feature pipelines (signal generation) and execution logic independently or jointly.
- Explicitly measures p-hacking probability to validate that improvements reflect real patterns, not statistical artifacts.
- Outperforms alternative agentic search methods like Claude Code on Bitcoin trading backtests.
Full Summary
Quantitative trading relies on algorithmic strategies that combine feature engineering (extracting predictive signals from market data) with execution logic (deciding when and how to trade). Manually optimizing these components is labor-intensive and often produces strategies that overfit to historical data.
MadEvolve adapts DeepMind's Alpha-Evolve framework—which uses large language models to propose and test algorithmic improvements—for financial applications. The system iteratively generates candidate trading strategies, evaluates them against historical Bitcoin price data, and uses feedback to guide the next generation of proposals.
Results demonstrated improvements across feature generation, individual strategy components, and joint optimization of the full pipeline. The authors compared MadEvolve against Claude Code (an AI coding agent) and explicitly measured p-hacking risk—the probability of finding false patterns by chance—to ensure findings were statistically valid rather than artifacts of excessive testing.