The tone of the Fed, 1936-2026

Net tone by year, 1936–2026, against the policy rate

Each point pools all documents of a year: net tone = (hawkish matches - dovish matches) / (all matches), so +1 is all-hawkish vocabulary and -1 all-dovish. Years with fewer than 20 pooled matches are left blank rather than plotted (1936, 1938, 1939, 1940, 1941): the early corpus is thin and the index does not pretend otherwise. The policy rate is the BIS series for the United States, which begins in 1954; no policy-rate series in the estate reaches back further.

The index reads the eras the way the record says it should: pooled tone is +0.38 across 1979-1981 (the Volcker disinflation), -0.34 across 2008-2015 (crisis and zero lower bound), and +0.56 across 2022-2023 (the fastest tightening cycle since the early 1980s). Those three eras are the sanity gate this index had to pass before shipping; the full validation panel is below.

Source: FinObservatory fed_tone build over the fed_docs corpus | BIS central bank policy rates (USA) | NBER recessions via FRED USREC Methodology

Era validation: does the index read known eras correctly?

Pooled tone across well-understood policy episodes, for the shipped lexicon and for the published benchmark it replaced (Apel and Blix Grimaldi 2012, implemented verbatim first and rejected; the methodology section documents exactly why). Mixed episodes read near zero by design: in 2004-2006 and 2017-2018 the Fed was raising rates while still calling its stance accommodative, and the vocabulary balance reflects both.

Source: FinObservatory fed_tone_annual, pooled hawkish and dovish matches per episode Methodology

What the Fed talks about, by decade

Five keyword families (patterns published below), each decade’s bar showing every family’s share of all family matches, in percent. Attention, not sentiment: a financial stability mention is a mention, whether reassuring or alarmed. Shares are of matched terms only, so they always sum to one hundred.

Source: FinObservatory fed_tone_annual, topic family matches pooled by decade Methodology

What the index is made of, era by era

The corpus changes character across 90 years, and any 90-year index inherits that. Before the 1970s the index reads one document type only: the minutes and their predecessor records (historical minutes, executive committee minutes, memoranda of discussion, records of policy actions). Verbatim transcripts enter in the 1970s and stop at the FOMC's roughly five-year release embargo; post-meeting statements begin in the 1990s (the FOMC only started issuing them in 1994), and Board speeches and the Beige Book enter in the 1990s with the Fed's web archive. Speeches then dominate the modern corpus by document count. This table is the exhibit to read before trusting any single year.

Source: Board of Governors of the Federal Reserve System, federalreserve.gov US government work, public domain. Corpus coverage details on the FOMC materials page. Methodology

Method, verbatim

• Corpus. The fed_docs full-text corpus: 4,498 documents extracted from the estate's Federal Reserve text acquisition (the FOMC materials page documents the file-level catalog; this corpus drops PDF copies of documents also held as HTML). It is what the estate holds, not the complete universe of Fed communication: no testimony, no press conferences, no staff documents (Greenbook/Tealbook), and nothing a release embargo has not yet published.
• Counting. A hit is one case-insensitive regex match of a pattern over a document's extracted text; hawk and dove are the total hits of the eight patterns on each side. Patterns are counted independently, so a phrase can legitimately match twice ("rising inflationary pressures" counts under both the inflation-direction and the inflationary pattern). No negation handling, no weighting, no model. Net tone = (H - D)/(H + D), computed per document only when H + D >= 5, and per year, pooling all documents' counts, only when the pooled total reaches 20.
• Known impurities, disclosed. “Tighten” also matches bank-lending-standards language (“tightening of credit conditions”, common in 2008-2009 minutes), “easing” is ambiguous symmetrically, and “restraint” includes fiscal restraint, a minority of its uses. These are left in deliberately: the index is a vocabulary balance, the ambiguity runs in both directions, and the era panel above shows the balance still signs every well-understood episode correctly.
• The published lexicon that was tested and rejected. The Apel and Blix Grimaldi (2012) dictionary, direction adjectives (high/higher, strong/stronger, fast/faster, increasing, increased against low/lower, weak/weaker, slow/slower, decreasing, decreased) immediately preceding seven noun stems (inflation*, price*, wage*, oil price*, cyclical position*, growth*, development*), was implemented verbatim first. On this corpus it reads the Volcker disinflation as neutral (the ABG benchmark column above): stagflation-era descriptions of weak activity (“slow growth”, “lower growth”) dominate its dovish side exactly when policy was at its most hawkish, and 1970s Fed English says “rate of inflation” and “inflationary pressures”, which form almost no adjective-noun bigrams. A hawk-dove index that misses Volcker is broken, so the FinObservatory lexicon below drops direction language about real activity entirely and instead counts inflation-as-threat language plus policy-restraint vocabulary against disinflation/deflation language plus policy-accommodation vocabulary. The ABG idea of direction vocabulary is kept; the credit is theirs, the word lists are ours.
• Policy rate and recessions. The overlay is the BIS central bank policy rate series for the United States (monthly, end of period), annual averages, available from 1954; recession shading is derived from the NBER chronology as published in the FRED USREC series, never hardcoded.

The FinObservatory Fed tone lexicon, in full

Topic families

Citations

• Apel, M. and M. Blix Grimaldi (2012), “The Information Content of Central Bank Minutes”, Sveriges Riksbank Working Paper Series No. 261. archive.riksbank.se (the tested-and-rejected benchmark; source of the direction-vocabulary idea).
• Hansen, S., M. McMahon and A. Prat (2018), “Transparency and Deliberation Within the FOMC: A Computational Linguistics Approach”, Quarterly Journal of Economics 133(2), 801-870. academic.oup.com (the reference point for dictionary-plus-direction text analysis of FOMC records).
• Aruoba, S. B. and T. Drechsel (2024), “Identifying Monetary Policy Shocks: A Natural Language Approach”, NBER Working Paper 32417. nber.org (the NLP-shocks frontier; what this page deliberately is not).