1. From Financial Reports to Macroeconomic Forecasts: An LLM-Based Approach
Abstract: Does the granular, qualitative information in corporate financial reports improve our understanding of aggregate economic activity? This paper bridges the gap between micro-level corporate disclosures and macroeconomic forecasting by using a Large Language Model (LLM) to perform structured information extraction from over 20,000 SEC 10-K and 10-Q filings (2001–2025). Unlike traditional dictionary-based methods, I employ a structured questionnaire approach to quantify firm-level operating conditions, labor shortages, and investment plans, deriving these metrics alongside supporting textual evidence and confidence scores. I aggregate these micro-signals into size-weighted quarterly indices to capture economy-wide trends. Using dynamic factor models and partial least squares, I demonstrate that these “textual factors” significantly improve out-of-sample GDP growth forecasts relative to standard benchmarks. The findings suggest that corporate narratives contain timely, forward-looking information that precedes traditional macroeconomic indicators, providing a scalable framework for converting unstructured text into systematic economic data.