G10 central banks
Hawk/Dove scoring methodology
How RoboMacro scores G10 central-bank speeches from −1.0 (dovish) to +1.0 (hawkish): BIS scrape, LLM scoring, seniority weights, and four named windows (90-day dashboard, 180-day live lean, 12-point EMA, 730-day member history). Recipe and CSV. Free.
What this is
A daily composite of G10 (plus SARB and BCB) monetary-policy stance from public speeches, on a scale from −1.0 (very dovish) to +1.0 (very hawkish). Hawkish = favoring higher rates / tighter policy. Dovish = favoring lower rates / looser policy.
As of 2026-09-29, the GDP-weighted mean of the banks we score is +0.35 (−1 dove / +1 hawk; 17 of 19 banks with a live composite). Cite with that date.
Recipe
- Scrape. Speeches from the BIS Central Bank Speeches database and bank websites, daily.
- Score. Each speech is scored by an LLM with a calibrated prompt: stance, confidence, and excerpts. Not keyword counts.
- Weight. Recency (exponential decay, τ=30 days), speaker seniority (Chair 3×, Deputy 2×, Member 1.5×, External 1×), confidence, and monetary-policy relevance. The dashboard is a 90-day window; member cards are 180 days.
- Composite. Weighted scores → one number per bank, updated daily. Live at /central-banks. This week’s table: /central-banks/weekly.
Four windows, four questions
The same score is not shown four ways. Collapsing them would silently reprice the homepage against the scorecard.
- Live lean (180 days, τ=30). Current member stance on cards and in previews.
- Dashboard (90 days, τ=30). Bank-level composite on the hawk/dove monitor.
- Scorecard recent. 12-point EMA of votes and speeches.
- Member history (730 days, τ=270). The Stance line on a member chart. Votes blend 60/40 where the bank names voters.
GDP-weighted index of the banks we score
A fifth surface, not a fifth window. It reads the dashboard 90-day composites and does not retune live 180-day leans. Three layers on one chart: /stealth/hawk-dove-gdp.
- Grey envelope. Unweighted min/max. Hover names the dove and hawk leaders that month. A 0.4%-of-GDP bank may own an edge.
- Gold line. Weighted mean of those same composites. Default weights are IMF WEO NGDPD (nominal USD), annual, lagged. Euro area once — never DE+FR+IT. A quiet bank keeps its last score; silence is not a change of stance. Toggle to M4/broad money (World Bank FM.LBL.BMNY.GD.ZS × NGDPD; euro-area M3 from the ECB).
- Fan. GDP-weighted quantiles at 10/20/30/40/60/70/80/90 of covered GDP, plus a dashed GDP-weighted median. 50 is not drawn as a second mean.
Copy says “the banks we score” plus the covered share of world GDP. China is a published preview-exclusion and is not in the index; do not call the series global. IMF WEO NGDPD (market rates) is the baseline — about 17% of world GDP for China in 2025. The ~20% figure is PPP (WEO PPPSH). Seniority weights already sit inside each bank composite and are not applied again across banks. GDP-weighted policy-rate trackers (Dallas Fed IMST, BIS) and per-bank communication hawk/dove (Hawkometer) already exist; this page is not marketed as a world first.
What this is not
It is not a vote forecast and not a substitute for the policy rate. Voting records (where published) sit beside the NLP score. Academic lineage: Hansen & McMahon 2016, Shapiro & Wilson 2022.
Download this week’s scores as CSV from the weekly table. Cite as: RoboMacro Hawk/Dove Monitor, robomacro.com/central-banks, date.
License for the public table: quote with attribution. Speeches remain copyright of the issuing banks.
