The purpose of central bank minutes is to give an account of monetary policy meeting discussions to outside observers, thereby enabling them to draw informed conclusions about future policy. However, minutes are by necessity a shortened and edited representation of a broader discussion. Consequently, they may omit information that is predictive of future policy decisions. To investigate this, we compare the information content of the FOMC's minutes and transcripts, focusing on three dimensions which are likely to be excluded from the minutes: 1) the committee's degree of hawkishness; 2) the chairperson's degree of hawkishness; and 3) the level of agreement between committee members. We measure committee and chairperson hawkishness with a novel dictionary that is constructed using the FOMC's minutes and transcripts. We measure agreement by performing deep transfer learning, a technique that involves training a deep learning model on one set of documents - U.S. congressional debates - and then making predictions on another: FOMC transcripts. Our findings suggest that transcripts are more informative than minutes and heightened committee agreement typically precedes policy rate increases.
Central Bank Communication Monetary Policy Machine Learning