MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models
Researchers have developed MadEvolve, a system that combines large language models and evolutionary algorithms to enhance Bitcoin trading strategies. The framework boosts out-of-sample Sharpe ratios by up to 1.8 points and increases impact-adjusted profits up to 25 times over traditional methods. The most notable gains come from evolving order placement strategies. The study shows that AI-driven evolutionary search can find real trading improvements, though results are based on simulations and may not fully translate to live markets.
What it examines
This paper studies how large language models (LLMs) and evolutionary algorithms can automatically improve trading strategies in finance, especially for Bitcoin. Using the MadEvolve framework, the authors test if AI-driven optimization can find real trading improvements without just overfitting to noisy data.
What it concludes
The results show that LLM-driven evolutionary search can create better trading strategies, improving out-of-sample performance and risk-adjusted returns. These methods could help automate quantitative research and trading. However, real-world testing is needed, and future work should address live trading and more realistic market conditions.
Evidence objects
MadEvolve, a new framework combining large language models and evolutionary algorithms, dramatically improves Bitcoin trading strategies, boosting out-of-sample Sharpe ratios by up to 1.8 and profits by 25 times over hand-crafted baselines.
key_findings bullet 1 · key_findings · validation V0
Surprisingly, the greatest performance gains come from evolving order placement strategies, with joint optimization of trading features and execution logic yielding the highest risk-adjusted returnseven when tested on noisy, realistic financial data.
key_findings bullet 2 · key_findings · validation V0
The study introduces terms like 'agentic loop' and shows that LLM-driven evolutionary search finds robust trading improvements, not just overfitting, but notes results rely on simulated data and may not fully transfer to live markets.
key_findings bullet 3 · key_findings · validation V0
This paper uniquely adapts the agentic loop framework, inspired by AlphaEvolve, to algorithmic trading, leveraging LLM-driven evolutionary optimization for feature engineering and execution strategies. Its rigorous approach to overfitting, comparative analysis of agentic search methods, and insights into LLM ensemble contributions make it a compelling, innovative advance in autonomous quantitative finance.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … We explore the application of LLM-driven algorithm … MadEvolve to optimize algorithmic trading strategies and … of AI-driven agentic and evolutionary algorithms for …
Source row: 1275 · abstract type: snippet