Autor: Aromí J. Daniel*, Heymann Daniel**

Institución: *IIEP UBA-Conicet, FCE UCA, **IIEP UBA-Conicet

Año: 2023

JEL: E5, E3


We apply natural language processing techniques to infer sentiment expressed in FOMC meetings. The sample period covers the Great Recession and its aftermath (2003-2012). We infer meetings’ tone implementing large language models (BERT, NLI, ChatGPT) and traditional dictionary methods (Loughran & MacDonald 2011, Aromí 2020). Suggesting policymakers are advantageously informed, we find that tone in FOMC meetings anticipates media sentiment, consumers’ confidence, and financial market dynamics. Furthermore, meetings’ tone also anticipates growth forecast errors from Fed staff and private sector analysts. The findings are robust to changes in text processing methodologies and show a persistent anticipatory ability that extends over multiple quarters. We observe that, despite some discrepancies and evidence of underreaction, the tone of FOMC is closely replicated in meetings’ minutes. Our analysis shows that, despite its availability, analysts fail to incorporate the information on policymakers’ deliberations in an adequate manner.