Hawkish vs dovish central banks: scoring what they say against what they do

Quantamentry — how the communication-stance dimension reads 900+ central bank statements.

TL;DR

• Hawkish vs dovish central banks is a tone question, not a rate question. We score it separately from the rate level — and the two routinely disagree.
• Every English-language central bank statement is scored hawk/dove by FinBERT (CentralBankRoBERTa) and reconciled with Claude Haiku — a 60% model / 40% LLM blend.
• Communication captures direction; behind-the-curve captures the level. A bank can talk tough while its policy rate sits below the rule, or sound cautious while it's already running tight.
• On our June 29, 2026 snapshot, Indonesia talks hawkish (communication 51.3) while its rate is the most behind the rule (behind-the-curve 14.4) — and Colombia does the opposite, with middling tone (41.7) but the tightest rate on the board (79.8). Words and rate point different ways.
• The dimension carries a 10% composite weight — deliberately small, because words move faster than mandates.

It's the cheapest tell in central banking. A bank changes its rate eight times a year; it changes its tone in every speech, minutes release, and press conference in between.

What is a communication stance score?

A communication stance score is a 0–100 measure of how hawkish or dovish a central bank's published language is, independent of where its policy rate actually sits. We take every English-language statement, minutes release, and press-conference transcript a bank puts out, score each one on a hawk–dove axis, and aggregate to a single per-country number that updates as new statements land. High means the bank's words lean hawkish (tightening bias); low means the words lean dovish (easing bias). It is a read on intent and direction of travel — not on the level of rates, and not a forecast of the next decision.

That last distinction is the whole point of keeping it as its own dimension.

How is hawk/dove tone measured?

Two models, on purpose. One is fast and consistent; the other is good at context. We blend them.

1. FinBERT — CentralBankRoBERTa (60% weight). This is a transformer fine-tuned specifically on central bank communication, not general financial news. It classifies each sentence on a hawkish–neutral–dovish axis. Because it's a fixed model, it's consistent: the same sentence scores the same way every time, with no drift. That consistency is exactly what you want as the backbone of a time series — you're measuring change in tone, so the ruler can't move.

2. Claude Haiku reconciliation (40% weight). FinBERT is literal. It can miss conditionals ("we would hike if inflation re-accelerates"), sarcasm, and forward guidance that hedges. We pass the same statement to Claude Haiku for a stance read that accounts for context, negation, and the difference between describing a risk and committing to an action. Where the two disagree, the blend splits the difference, model-weighted.

The 60/40 split is a judgment call: anchor on the consistent model, let the LLM correct for the cases where literal sentence classification gets the meaning wrong. The result is a per-statement hawk/dove score, averaged into the per-country communication dimension you see in the composite.

You can read the full weighting and reconciliation logic on the methodology page.

A coverage note before the numbers

Communication scoring is weekend-only — the LLM reconciliation runs on a batch cadence — and it needs a recent English-language statement to score. That means on any given day only a subset of central banks carry a live communication value; the rest are simply blank until their next English release lands and gets scored. On our June 29, 2026 snapshot, seven banks had a current number:

Read this as a snapshot, not a census: a country missing from the table isn't neutral, it's unscored today. Banks that publish primarily in their domestic language, or release infrequently, show up here less often. (See coverage for which banks have deep statement histories.)

Why direction and level are different things

Here is the insight the dimension exists to capture.

Behind-the-curve asks: is the policy rate where inflation and output say it should be? It's a level — a Taylor-rule deviation. (We wrote that one up in the most behind-the-curve post.)

Communication asks a different question: which way is the bank leaning, in its own words, right now? It's a direction.

These come apart all the time:

A bank in the top-left is doing what it says. A bank in the bottom-left — hawkish language, rate still below the rule — is the interesting one: the words are a promise the rate hasn't kept yet. Markets price that gap.

Talk matches action: Brazil

Brazil is the clean diagonal. The BCB's statements read the most hawkish on the board (communication 55.3) — explicit about keeping real rates high until expectations re-anchor — and the action backs it up: a behind-the-curve score of 60.4, comfortably ahead of the Taylor rule, inside a composite of 59.8. Words and rate point the same direction. When a bank's communication and its rate level agree, the stance is cheap to trust: there's no gap to arbitrage.

When talk and rate diverge

Now the cases where the two dimensions split — and these are the ones worth watching.

Indonesia — talk tighter than the rate. Bank Indonesia's communication reads hawkish (51.3, second only to Brazil), yet on the rate level it scores 14.4 on behind-the-curve — the most behind the rule of any scored bank. The language leans tight; the policy rate sits well below where the model says it should be. Hawkish words, behind-the-curve action: the words are running ahead of the rate.

India — talk and action both loose. The RBI is the mirror corner. Its communication is the most dovish we scored (12.9), and the rate agrees: a behind-the-curve score of 21.5, also below the rule. Here the two dimensions point the same way — dovish-sounding and structurally loose — so there's no contradiction to flag, just a consistently accommodative posture.

Colombia — action far tighter than the talk. The most striking divergence runs the other direction. BanRep's tone is merely middling (communication 41.7, right in the neutral band), but on the rate level it posts 79.8 on behind-the-curve — the tightest reading on the entire board, furthest ahead of the rule. The action is far more hawkish than the words. A reader who judged Colombia off its statements alone would badly underestimate how restrictive its policy actually is.

The lesson: never read tone off the rate, or the rate off the tone. That's why they're two columns in the score, not one.

What the communication score is not

We flag these because they come up in DMs.

1. It is not a forecast of the next decision. A hawkish stance score does not mean a hike is coming next meeting. Banks talk hawkish to anchor expectations precisely so they don't have to hike. The score reads the leaning, not the calendar.
2. It carries an English-statement coverage bias. We score English-language communication, on a weekend batch cadence. Banks that publish primarily in their domestic language with thin or delayed English releases are scored on less text, and the cadence of their statements affects how quickly the number moves — which is why only a handful of banks carry a live value on any given day. Gold and silver tiers with rich English output are scored most reliably; sparse-coverage banks carry more noise. (See coverage for which banks have deep statement histories.)
3. It is not a credibility verdict on its own. Tone is one input. A bank can talk perfectly while missing its target for three years — which is why communication is a 10% weight, and why credibility itself is built from seven dimensions, not one. Words are evidence, not the verdict.
4. It is not sentiment about the economy. We're scoring the bank's policy lean — tighten vs ease — not whether the statement is optimistic or pessimistic about growth. A bank can be gloomy on growth and still hawkish on rates.

Why only 10%?

Because words are the easiest thing for a central bank to change and the cheapest thing for us to mis-score. Tone leads action — that's its value — but tone also bluffs, hedges, and gets lost in translation. Weighting it at 10% lets the communication dimension tilt a country's composite when language and fundamentals diverge, without letting a single dovish press conference swamp a credible rate path. It's a thumb on the scale, not the scale.

What's coming next on Quantamentry

• Next: the geopolitical-pressure dimension — what GDELT event flow does to a sovereign's score, and why the mean (not the sum) of event impacts is the right call.
• Soon: a live hawk–dove leaderboard — communication stance for every gold- and silver-tier central bank, updated as statements land.
• Soon: the Quantamentry API. Country risk and macro intelligence, REST, $49–$499/mo.

If you want the daily snapshot or early access — [subscribe to Quantamentry]. We'll email you the moment the hawk–dove leaderboard goes live.

— Quantamentry, June 29, 2026