Words or numbers? Macroeconomic nowcasting with textual and macroeconomic data

Abstract

This paper performs the nowcasting of GDP growth rate and inflation expectation in China with traditional macroeconomic and novel textual data estimated by the latent Dirichlet allocation (LDA) model. We combine the MIDAS model with various machine learning techniques to handle the mixed-frequency and high-dimensional problems. Our empirical findings are threefold. First, we collected 866234 articles published over 20 years of Chinese economic newspapers. We systemically decomposed the textual data into news attention time series, which provide narrative descriptions of the economic and social conditions. Second, news attention data can provide similar or even better precision for nowcast, especially for inflation expectation compared with traditional macroeconomic data. Random forest delivers the most accurate forecast among the three machine learning methods, even for longer horizons. Thirdly, the most informative predictors for the nowcast align with existing literature, and news attention variables provide narrative realism for the forecast targets.