Econometric forecasting using ubiquitous news text: Text-enhanced factor model

Abstract

News text is gaining increasing attention as a novel source for econometric forecasting. This paper revisits how narrative information is incorporated into econometric forecasting by effectively quantifying sector-specific textual information without requiring training data. We propose Theme Frequency Indices (TFIs), which utilize domain-specific subject-predicate patterns to measure public perception about the economy. TFIs for 15 sectors, including production, inflation, employment, capital investment, stock and house prices, and others, were examined and integrated into the Text-enhanced Factor Model (TFM), using latent factor structures. Empirical analysis based on over 18 million news articles from Korea reveals that TFM improves the accuracy of near-term GDP forecasts, demonstrating that simple text-mining techniques combined with domain knowledge can effectively leverage qualitative information in the news without costly training. The proposed method is applicable to a wide range of subjects for utilizing narrative information on the economy, offering a rapid and cost-effective approach.