Nowcasting German GDP with Text Data

Research context

Mariia Okuneva, Philipp Hauber, Kai Carstensen and Jasper Bär analyse 12.4 million news articles from 1991–2018 in CESifo Working Paper 11587 (2024). They use Media Tenor’s annotated data in a sentiment model. Their out-of-sample tests find that combining text and hard-data forecasts improves on the hard-data model alone, with the largest gains in nowcasting.

English editorial overview of the original source. Historical findings relate to the period described, not necessarily the source archive’s listing date.