AlphaPortfolio: Discovery of Portfolio Optimization and Allocation Methods Using LLMs

Abstract

Traditional long-only portfolio allocation strategies, including equal-weighted, risk-parity, and equal-risk contribution approaches; often struggle with rigid assumptions, excessive concentration in low-volatility assets, and sensitivity to changing market conditions. These methods frequently fail to balance risk and return optimally in dynamic financial environments. This paper introduces AlphaPortfolio, a novel framework that leverages Large Language Models (LLMs) to iteratively generate, refine, and validate portfolio optimization methods. By integrating inverse covariance risk-adjusted returns, entropy-based diversification, and volatility normalization, AlphaPortfolio significantly outperforms classical allocation techniques in both risk-adjusted performance and drawdown resilience. Our experimental findings, cross-validated across 15 years of historical financial data from 3,246 US stocks and ETFs, demonstrate that AlphaPortfolio achieves a 71.04% increase in the Sharpe Ratio, a 73.54% improvement in the Sortino Ratio, and a 116.31% boost in the Calmar Ratio. Additionally, AlphaPortfolio reduces maximum drawdowns by 53.77%, ensuring greater stability in turbulent market conditions. These results highlight the transformative potential of LLM-driven discovery in redefining portfolio optimization methodologies, providing institutional investors and portfolio managers with more resilient allocation strategies.