Do Language Models Need a Markowitz-Style Pivot?

Using Harry Markowitz’s transformation of intuitive diversification into modern portfolio theory as a reference, this article explores the application direction of large language models in finance. Current LLMs can already effectively perform text interpretation, classification, and prediction for individual securities. Related research shows that their assessments of financial news can capture immediate market reactions and predict some subsequent drift, although this advantage diminishes as LLM adoption increases. The article argues that the next core question is how to use information from language models for portfolio construction and risk management, and to uncover the implicit economic relationships between companies embedded in text.