The World Central Banks (WCB) paper introduces the most comprehensive dataset of central bank communications to date, spanning 380k+ sentences from 25 central banks across 28 years. It includes expert-annotated labels for stance detection (hawkish/dovish/neutral), temporal classification, and uncertainty estimation, designed to evaluate both pretrained and large language models in high-stakes financial policy contexts.
Key highlights:
- Benchmarked 15k+ experiments across state-of-the-art NLP models, including large language models.
- Demonstrated that cross-bank training yields significantly better performance than single-bank training.
- Released the WCB dataset, along with scripts and evaluation pipelines, to support further research on monetary policy and financial NLP.
† Indicates core contributors.