Evolving Portfolio Heuristics: A Self-Correcting LLM Framework for Portfolio Optimization

The application of advanced artificial intelligence to portfolio management presents a critical dilemma. While methods such as deep reinforcement learning offer powerful capabilities for navigating market dynamics, their inherent 'black-box' nature fundamentally limits their utility in high-stakes financial environments where transparency, expert validation, and trust are paramount. Addressing this challenge, we introduce EPH (Evolving Portfolio Heuristics), a framework that leverages Large Language Models not for direct optimization, but for the automated discovery of interpretable investment strategies. EPH operates on a population of heuristics, each represented dually by a natural language 'thought' and its corresponding executable code, making the evolutionary design process itself transparent. To bridge the gap between heuristic discovery and practical application, the framework integrates a self-correction mechanism that guides the evolutionary search. This mechanism strongly favors the selection and development of heuristics that yield portfolios compliant with predefined constraints on cardinality, asset allocation bounds, and overall risk. Through extensive back-testing on historical market data, we demonstrate that EPH discovers novel and effective heuristics that yield superior risk-adjusted returns compared to established financial and AI baselines, while demonstrating high fidelity to the specified risk budget.