Nowcasting German GDP with Text Data

This paper investigates the impact of news media information on improving short-term GDP growth forecasts by analyzing a large and unique corpus of 12.4 million news articles spanning from 1991 to 2018. We extract business cycle-related sentiment from each article using an annotated dataset from Media Tenor International and a Long Short-Term Memory neural network. This sentiment is then applied to adjust the sign of daily topic distributions estimated through the Latent Dirichlet Allocation algorithm. For the forecasting experiment, we select 10 sign-adjusted topics that show strong correlations with GDP growth, are highly interpretable, and economically relevant. An encompassing test reveals that these topics provide valuable information beyond professional forecasts. In an out-of-sample forecasting experiment, we also find that combining Dynamic Factor Model (DFM) forecasts—derived separately from hard data and text information—consistently outperforms the DFM model relying solely on hard data across all forecasting horizons, with the greatest improvements seen in nowcasts. These results underscore the effectiveness of integrating news media information into economic forecasting, in line with existing literature.