How the number is made

Inputs

Three kinds of input feed each member's reading. Votes: a dissent for a hike reads hawkish and a dissent for a cut dovish; a dissent for a hold reads against whatever the Committee did; a vote with the majority reads balanced. Record: every dissent for tighter or easier policy in a member's whole FOMC record, from the St. Louis Fed's history of dissents since 1936, leaves a permanent mark, so a member's past votes keep saying what kind of policymaker they are. Dissents on other grounds leave none.

Words: a language model reads the full published text of each appearance and returns the speaker's own stance on the path of policy, from strongly dovish to strongly hawkish, with a confidence for how much of the text is about policy and one verbatim sentence as evidence. A speech about payments or supervision barely moves a reading. A reviewer can replace any score. Of 322 texts, 114 are scored this way and 0 of those have been reviewed; the rest are scored against a fixed list of hawkish and dovish phrases.

How the parts rank

A dissent outweighs any single speech, and a vote with the majority counts for little. A press conference counts more than a speech, testimony or statement, and an interview less. Recent events count most: words fade faster than votes, both fade over months, and nothing from more than half a year ago counts, except a member's dissent record, which never fades. A member's reading is the weighted average of what remains, and is shown only when there is enough recent evidence to average.

Composite

Member readings are averaged with the Chair weighted most, this year's voters next and non-voters least. Bands: strongly dovish below −50, dovish to −15, balanced through +14, hawkish to +49, strongly hawkish from +50.

Why votes are weighted this way

The record since 1994 covers 270 meetings, 2872 votes and 134 dissents, a dissent rate of 5 percent per vote. Of the 97 dissents for tighter or easier policy that had another meeting the same year, 47 were followed by the same dissent at that next meeting and 49 by a vote with the majority. Only 6 members have dissented in both directions. Members persist rather than swing back, and half of them rejoin the majority without changing their view, so a vote with the majority is weak evidence of balance and past dissents count for good. In the same record a dissent for easier policy at a meeting that held rates was followed by a cut more often than other holds were; a dissent for tighter policy at a hold did not foretell a hike.

Limits

A model reading a speech can still misjudge a hedge or a quoted view, which is what the reviewer's check is for, and a score is one reading of one text. Single readings are noisy. The record marks assume a member's type is fixed; the few who have changed sides carry their old dissents the wrong way, and a member with no recent text reads from their votes alone. Interviews and Q&A without a published transcript are not scored at all, and Kansas City, St. Louis and Atlanta's interim president are not yet covered. Treat the trend as the signal.

Did the words lead the votes?

For every dissent since 2025, the dissenter's words-only reading on the day before the meeting, next to the words-only average of everyone else. 8 of the 10 dissents had enough recent text to score. The dissenter's words leaned the way they then voted in 8 of those, and leaned further that way than the rest of the Committee in 6.