UC Davis

This paper uses modern data analysis tools to measure sentiment overtime of pub- lished Central Bank communications. We employ natural language processing and FinBERT techniques (machine learning models adapted to analyze financial news) to construct sentiment based on FOMC minutes, and ECB press conferences. We also construct sentiment based on speeches, statements and CB overviews of economic con- ditions, and find high correlation across communications. Fed and ECB sentiment tends to move together, but there is little evidence of one leading the other. Using local projection analysis, we find that CB sentiment leads policy rates and the Taylor rule. In turn, stock market returns tend to lead Central Bank sentiment. Our find- ings imply that there is important information in CB communication sentiment, which could be used to identify monetary policy shocks.