From Narrative to Numbers: Large Language Models and Economic Policy Uncertainty

This study examines the impact of adopting machine-generated data in economic analysis, taking the Economic Policy Uncertainty index as the research object. It finds that large language models fine-tuned on human-audited labels identify nearly twice as many audited marked articles as keyword-based alternative methods, and are robust to noisy text and language drift. The choice of classifier, threshold, and aggregation method significantly affects index performance and interpretation, thereby altering the relationship between policy uncertainty and returns.