Reading the Fed when the wording is the signal

Start with what already exists

The most common mistake in building a research workflow is bolting on a tool before checking what the source already gives you for free. Central banking is where that mistake is most expensive, because the Federal Reserve publishes an unusually complete record and publishes it openly — no paywall, no registration, HTML and PDF.

Here is what arrives for a single FOMC meeting, and when. The order matters more than the list: each layer describes the same meeting in more detail than the one before it, and the gaps between them are the whole reason the earlier layers get read so closely.

1. Decision day The statement, and the implementation note

   A few hundred words, released the moment the meeting ends. Short enough to read in full, formulaic enough that any change in it is deliberate. This is the document people diff.

2. Decision day · four times a year The Summary of Economic Projections

   Published at four of the year's meetings alongside the press conference — participants' individual projections, compiled. Numbers rather than words, and the one part of this list that is not a language problem.

3. Decision day · same afternoon The press conference, and its transcript

   The Chair takes questions. The Fed posts a full transcript as a free PDF: the opening statement first, then question-and-answer, each utterance attributed by speaker, with a status marker in the page header so you can tell a settled FINAL text from an earlier one. A recent example ran 28 pages.

4. Three weeks later The minutes

   The Fed states that minutes of regularly scheduled meetings are released three weeks after the date of the policy decision. These summarise the discussion — who argued what, and roughly how many held each view. A summary, not a verbatim record.

5. About five years later The full meeting transcript

   The Fed publishes transcripts, background materials and meeting agendas with roughly a five-year lag. This is the verbatim record of the room. Nothing released earlier contains it, and no amount of tooling can make it arrive sooner.

The one technique: diff the statement

The statement is short, highly formulaic, and reissued every six weeks or so. Those three properties together make it almost uniquely suited to comparison — most of it is identical to last time by design, so the parts that are not identical stand out completely once you stop reading and start diffing.

The mechanics take about two minutes. Save each statement as a plain text file named by date. When the new one lands, run a diff against the previous one — diff -u fomc-2026-03-18.txt fomc-2026-04-29.txt in a terminal, or paste both into any text-comparison tool. What comes back is a handful of changed phrases rather than a document.

This is not a clever technique and it is not a secret one — every desk covering this does some version of it. It is here because it is the concrete thing this article is asking you to actually do, and because it demonstrates the general principle the rest of the workflow rests on: with formulaic recurring documents, comparison is cheap and reading is expensive. The same move works on any body that publishes on a fixed cadence in a fixed house style.

Phrases that carry weight because they recur

The reason a diff is legible at all is that this material runs on standing phrases — constructions that appear meeting after meeting and mean something precise inside the document's own vocabulary. Learning the recurring set is what lets you see a substitution rather than just new words. These all appear in published Fed material:

Build your own list this way rather than borrowing one: take the last four statements, find every phrase that appears in all four, and that set is the boilerplate. Anything outside it is either new or gone, and both are worth a look.

The press conference is a different document

The statement is negotiated by a committee. The press conference is one person answering questions they did not write, which means it is the only same-day material where the phrasing is not fully controlled in advance — and the transcript's structure lets you use that.

Because the opening statement is separated from the Q&A, you can read the two halves as what they are: the prepared half restates and expands the committee's language, and the unprepared half is where a question can push at a phrase until it either holds or gets rephrased. When a journalist asks the same question three different ways, the useful thing is not the answer — it is whether the wording of the answer stayed constant.

Two habits make the transcript worth more than the broadcast. First, read it against the statement, not on its own: a phrase from the statement restated with an extra qualifier in the Q&A is a specific observation you can only make with both texts in front of you. Second, keep the PDFs. They are stable URLs on federalreserve.gov, but a local folder of them going back two years is a corpus you can search in one command, and that is a capability you cannot assemble retroactively in a hurry.

Where the words exist only as audio

Everything above is a text problem with a text solution, which is why we have not mentioned transcription once. Now for the part of this beat where text does not exist.

The Fed archives speeches and testimony by year — 2006 onward on the main page, with an older archive beyond that — and posts the prepared text. But the prepared text is frequently not the whole event. A speech delivered at a university or a conference is often followed by a moderated conversation or audience questions, and that portion commonly has no published transcript even when the prepared remarks do. The same is true of panel appearances, fireside-format interviews at industry conferences, and broadcast interviews. Officials in other jurisdictions vary widely in what they publish; a video with no accompanying full text is a normal outcome.

That material is where a transcript is worth something, for the plain reason that you cannot search audio. A ninety-minute panel is unsearchable, unquotable and uncomparable until it is text — and its interesting sixty seconds are somewhere in the middle of it with no way to find them but listening to all ninety.

Two practical notes on capture, both of which are about not wasting your time:

• If it is streaming in a browser tab and there is no file to save, record the tab. The Whipscribe Chrome extension captures tab audio and returns a transcript, which is the case a URL cannot cover.
• If the source is DRM-protected, no tool can fetch it. Spotify is the routine example and we refuse those links up front rather than letting you wait for a failure. The answer is to find the same remarks where they are published openly — for central bank material they very often are — not to look for a way around the protection.

What the finished workflow looks like

Per meeting, this is maybe thirty minutes and it does not grow over time:

1. Save the statement as dated plain text. Diff against the previous one. Note the changed phrases.
2. Download the press conference transcript PDF. Skim the opening statement for those same phrases; read the Q&A properly.
3. Three weeks on, search the minutes for the phrases your diff surfaced, to see how the discussion behind them was characterised.
4. Between meetings, keep speech texts in the same folder. Transcribe the events that have no text, so the folder stays complete.
5. Search the whole folder, not individual documents. One grep across two years of statements, transcripts and minutes answers questions like when did this construction first appear? — which is unanswerable by memory and trivial with a corpus.

The point of the folder is that it makes the wrong kind of question cheap enough to ask. Most searches will return nothing interesting. That is fine: the cost of checking has to be near zero, or you will stop checking and start relying on your impression of what was said, which is precisely the failure mode this whole exercise exists to prevent.