What's that noise? Analysing sentiment-based variation in central bank communication

To which degree can variation in sentiment-based indicators of central bank communication be attributed to changes in macroeconomic, financial, and monetary variables; idiosyncratic speaker effects; sentiment persistence; and random "noise" ? Using the Loughran and McDonald (2011) dictionary on a text corpus containing more than 10,000 speeches and press statements, we construct sentiment-based indicators for the ECB and the Fed. An analysis of variance (ANOVA) shows that sentiment is strongly persistent and influenced by speaker-specific effects. With about 80% of the variation in sentiment being due to noise, our findings cast doubt on the reliability of conclusions based on variation in dictionary-based indicators.