Time-Varying Dictionary and the Predictive Power of FED Minutes

This paper proposes a method to extract the most predictive information from FED minutes that is specifically adapted to the problem of forecasting. Instead of considering a dictionary (set of words) with a fixed content, we construct a dictionary whose content is allowed to change over time. Specifically, we utilize machine learning to identify the most predictive words (the most predictive content) of a given minute and use them to derive new predictors. We show that the new predictors improve real time forecasts of output growth by a statistically significant margin, suggesting that the combination of supervised machine learning and text regression can be interpreted as a powerful device for out-of-sample macroeconomic forecasting.