Lusha Xu

University of California, Riverside · Economics · 2026–27 job market

Fields

Macroeconomics, Econometrics, Finance

Job market paper

From Financial Reports to Macroeconomic Forecasts: An LLM-Based Approach

Do qualitative narratives in corporate financial reports carry forward-looking information about the aggregate economy? Using a large language model and a structured questionnaire, I extract 38 filing-based measures from 60,236 Form 10-K and 10-Q filings submitted by S&P 500 firms between 1996 and 2026. Each measure is aligned by its public EDGAR filing date and distinguishes between explicitly stated disclosures and contextual model inferences. This corporate soft information is economically meaningful at both the micro and macro levels: the extracted measures predict subsequent firm-level outcomes, and their aggregates closely track target macroeconomic series. Using variables selected through 2012, I construct weekly leading and coincident indices to evaluate real-time recursive forecasts. Over 2013–2019, incorporating these indices reduces mean squared forecast error relative to an autoregressive benchmark by 17 percent for the current quarter and 18 percent one quarter ahead, averaged over weeks 5–13. The indices also improve GDP quantile forecasts, with larger gains in lower-tail accuracy. Beyond GDP, filing-based measures improve industrial production point forecasts and year-ahead core inflation quantile forecasts. Labor-market gains are less consistent, and production and inflation point-forecast gains weaken after 2020. These findings show that corporate soft information complements aggregate statistics by capturing firms' assessments of business conditions and anticipated adjustments in economic activity.