I will be presenting joint, preliminary work of two research projects on natural language processing (NLP) for monetary policy analysis.
First, I will introduce a machine learning benchmark project for monetary policy analysis. Perhaps, it might be time to say goodbye once and for all to the Loughran-McDonald dictionary method, which is still widely used in text analysis across economic and finance.
I will then present work in which we create a novel dataset of US Federal Reserve (Fed) speeches and use supervised multimodal NLP methods to identify monetary policy news effects across three key dimensions – GDP growth, CPI, and unemployment. We analyse how well news shocks in speeches can explain changes in market expectations. Furthermore, we assess whether those news shocks can be associated with changes in intraday financial volatility and tail risk around those speeches. We find that central bankers’ news signals markedly correlate with equity and bond volatility and tail risk. Markets also attend to the content of these signals more closely during ‘extreme’ regimes of GDP and inflation, compared to normal times. These initial results challenge the conventional view that central bank communication primarily resolves uncertainty.