Predicting the U.S. Unemployment Rate Using Natural Language Processing Models

open_access
Open Access
Author:
Bhusari, Sanchita
Area of Honors:
Finance
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
Brian Spangler Davis, Thesis Supervisor
Christoph Hinkelmann, Thesis Honors Advisor
Keywords:
Sentiment Analysis
Federal Reserve
Unemployment Rate
Economic Forecasting
Abstract:
This research project investigates the predictability of the unemployment rate during economically uncertain periods through the analysis of Federal Open Market Committee (FOMC) meeting minutes, using Natural Language Processing (NLP) techniques such as FinBERT and Principal Component Regression (PCR) modeling. Traditionally, the relationship between inflation and unemployment follows an inverse pattern, where shifts in employment levels influence consumer strength and pricing power. However, recent economic conditions—shaped by post-pandemic inflationary pressures, geopolitical events, and monetary policy interventions—have complicated this relationship, challenging existing predictive models. This study aims to bridge the predictive gap by developing a model that translates the sentiment and tone of FOMC communications into unemployment rate forecasts. The methodology proposed involves text scraping of Federal Reserve meeting minutes from key periods of economic uncertainty (2003-2022), applying sentiment analysis to assign sentiment scores, and conducting regression analysis to assess the sentiment-unemployment relationship. By training a PCR model on historical unemployment data, this study seeks to create a nowcasting tool to forecast short-term unemployment trends. This project ultimately aims to identify how Federal Reserve communications reflect and potentially forecast labor market responses, offering a vital resource for anticipating economic shifts in real time.