AlphaSharpe: LLM-Driven Discovery of Robust Risk-Adjusted Metrics (2502.00029v2)

Abstract: Financial metrics like the Sharpe ratio are pivotal in evaluating investment performance by balancing risk and return. However, traditional metrics often struggle with robustness and generalization, particularly in dynamic and volatile market conditions. This paper introduces AlphaSharpe, a novel framework leveraging LLMs to iteratively evolve and optimize financial metrics to discover enhanced risk-return metrics that outperform traditional approaches in robustness and correlation with future performance metrics by employing iterative crossover, mutation, and evaluation. Key contributions of this work include: (1) a novel use of LLMs to generate and refine financial metrics with implicit domain-specific knowledge, (2) a scoring mechanism to ensure that evolved metrics generalize effectively to unseen data, and (3) an empirical demonstration of 3x predictive power for future risk-returns, and 2x portfolio performance. Experimental results in a real-world dataset highlight the superiority of discovered metrics, making them highly relevant to portfolio managers and financial decision-makers. This framework not only addresses the limitations of existing metrics but also showcases the potential of LLMs in advancing financial analytics, paving the way for informed and robust investment strategies.

• The paper introduces AlphaSharpe, a novel LLM-driven framework that iteratively evolves risk-adjusted metrics for superior predictive performance.
• The methodology applies iterative crossover, mutation, and ranking to refine metrics, enhancing correlations and portfolio Sharpe ratios compared to traditional measures.
• Experiments on 15 years of US market data, including the 2020 crash, show that AlphaSharpe metrics achieve over 3x higher ranking correlations and improved portfolio optimization.

This paper introduces AlphaSharpe, a novel framework that utilizes LLMs to discover and refine financial metrics, specifically risk-adjusted return metrics like the Sharpe ratio. The core idea is to address the limitations of traditional financial metrics, such as their sensitivity to outliers, reliance on normality assumptions, backward-looking nature, and limited generalization across different market conditions. AlphaSharpe leverages the generative capabilities of LLMs to propose innovative financial metrics, iteratively optimizing them for robustness, generalization, and predictive performance through crossover, mutation, scoring, and ranking.

The methodology involves an iterative four-step process. First, the LLM generates variations of existing metrics, drawing inspiration from financial literature and mathematical principles. Second, these metrics are refined via crossover (combining top-performing metrics) and mutation (small, deliberate modifications). Third, the mutated metrics are evaluated using scoring functions that consider robustness, generalization, and predictive power. Finally, metrics are ranked based on their quality and diversity, with only the top candidates retained for further iterations. LLMs are prompted to integrate domain-specific insights, encourage innovation, and use efficient tensor operations.

The paper details the scoring functions used to evaluate the evolved metrics, focusing on Spearman's Rho, Kendall's Tau, and Normalized Discounted Cumulative Gain (NDCG) to assess the correlation between metric scores on historical data and future Sharpe ratios. These functions are designed to reduce sensitivity to outliers and non-linearities.

Experiments were conducted using 15 years of historical data from 3,246 US stocks and ETFs, split into overlapping folds for time-series cross-validation. The evolved metrics were tested during the 2020 COVID-19 market crash to assess their robustness. The results demonstrated that AlphaSharpe metrics, specifically αS1 to αS4, achieved significantly higher ranking correlations than traditional metrics, leading to improved portfolio performance. For example, αS4 achieved over 3x higher Spearman and Kendall correlations compared to the Sharpe Ratio and Probabilistic Sharpe Ratio (PSR). Moreover, portfolios constructed using AlphaSharpe metrics showed significant improvements in Sharpe Ratios compared to portfolios based on traditional metrics.

The paper concludes by discussing the implications of AlphaSharpe for portfolio management, risk assessment, and financial decision-making, emphasizing its ability to handle small datasets, integrate distributional nuances, and adapt to changing market regimes. The AlphaSharpe Portfolio optimization method, which emerged from the LLM-driven discovery process, is also highlighted for its superior performance compared to traditional portfolio allocation techniques like Risk Parity and Equal Risk Contribution. The discovered AlphaSharpe Portfolio integrates inverse covariance risk-adjusted returns, stability-adjusted weighting, entropy regularization, and volatility normalization.