Using LLMs to Discover Better Portfolio Performance

Can large language models (LLM) help improve portfolio performance metrics, portfolio optimization and strategy feature discovery? In his three January-February 2025 papers entitled “AlphaSharpe: LLM-Driven Discovery of Robust Risk-Adjusted Metrics”, “AlphaPortfolio: Discovery of Portfolio Optimization and Allocation Methods Using LLMs” and “AlphaQuant: LLM-Driven Automated Robust Feature Engineering for Quantitative Finance”, Kamer Yuksel explores use of specially trained LLMs to discover new:

• Enhanced portfolio risk-return metrics that outperform traditional approaches such as Sharpe ratio.
• Better portfolio optimization methods.
• Robust investment strategy features.

The development processes are iterative for continuous improvement. He assesses usefulness of discoveries with 15 years of historical data for 3,246 US stocks and ETFs, of which he uses the last few years for out-of-sample equal-weighted portfolio testing. Using these methods and this dataset, he finds that:

(more…)