SUERF Policy Brief No. 528 on Fed Speeches Meet Transformer Models

This short policy article discusses the use of state-of-the-art methods from natural language processing to examine the evolution of the Fed’s interpretation of its mandate over time, using the largest corpus of Fed speeches assembled to-date:

It is based on our recently published working paper: Bertsch, Christoph, Hull, Isaiah, Lumsdaine, Robin L. and Xin Zhang (2022). “Central Bank Mandates and Monetary Policy Stances: through the Lens of Federal Reserve Speeches,” Sveriges Riksbank Working Paper Series No. 417, 2022:

Riksbank WP No. 417 on Central Bank Mandates

Central Bank Mandates and Monetary Policy Stances: through the Lens of Federal Reserve Speeches” with Isaiah Hull, Robin Lumsdaine and Xin Zhang

  • Abstract: When does the Federal Reserve deviate from its dual mandate of pursuing the economic goals of maximum employment and price stability and what are the consequences? We assemble the most comprehensive collection of Federal Reserve speeches to-date and apply state-of- the-art natural language processing methods to extract a variety of textual features from each paragraph of each speech. We find that the periodic emergence of non-dual mandate related discussions is an important determinant of time-variations in the historical conduct of monetary policy with implications for asset returns. The period from mid-1996 to late-2010 stands out as the time with the narrowest focus on balancing the dual mandate. Prior to the 1980s there was a outsized attention to employment and output growth considerations, while non dual-mandate discussions centered around financial stability considerations emerged after the Great Financial Crisis. Forward-looking financial stability concerns are a particularly important driver of a less accommodative monetary policy stance when Fed officials link these concerns to monetary policy, rather than changes in banking regulation. Conversely, discussions about current financial crises and monetary policy in the context of inflation-employment themes are associated with a more accommodative policy stance. (C63, D84, E32, E7)
The figure above shows a word cloud of concerning terms that appear in statements with low dual mandate content scores during the period 1984-2017. Such statements are identifed using extractive question answering with the RoBERTa model.
  • Keywords: Natural Language Processing, Machine Learning, Central Bank Communication, Financial Stability, Zero Shot Classification, Extractive Question Answering, Semantic Textual Similarity.

Forthcoming in the Journal of International Money and Finance: “Spread the Word: International Spillovers from Central Bank Communication”

“Spread the Word: International Spillovers from Central Bank Communication” with Hanna Armelius, Isaiah Hull and Xin Zhang, Journal of International Money and Finance, Volume 103, Pages 1-32, May 2020. – Lead article –

    • Abstract: We construct a novel text dataset to measure the sentiment component of communications for 23 central banks over the 2002-2017 period. Our analysis yields three results. First, comovement in sentiment across central banks is not reducible to trade or financial flow exposures. Second, sentiment shocks generate cross-country spillovers in sentiment, policy rates, and macroeconomic variables; and the Fed appears to be a uniquely influential generator of such spillovers, even among prominent central banks. And third, geographic distance is a robust and economically significant determinant of comovement in central bank sentiment, while shared language and colonial ties have weaker predictive power. (JEL E52, E58, F42)

The figure show the normalized rolling net sentiment scores associated with ECB speeches. Sentiment scores are computed using a dictionary-based approach documented in Loughran and McDonald (2011).