Central Bank Sentiment Index
Unveiling central bank narratives by using a transformer-based deep learning model
Unveiling central bank narratives by using a transformer-based deep learning model
We propose a new framework for analysing the narratives of central banks. We fine-tune BERT and FinBERT models, which relies on Google’s BERT architecture, on a dataset of manually annotated sentences on monetary policy stance. We derive a deep learning domain-specific model (BERT-CBSI and FinBERT-CBSI) ready for sentiment predictions. Based on our framework, we compute sentiment indices for central banks in Czechia, Hungary, Poland, and Romania (negative values are associated with dovish sentiments, while positive values are associated with hawkish sentiments). We compare our sentiment indices to similar macroeconomic indicators and other textual-based sentiment indices. Also, we test the predictive power of our sentiment indices.