A daily diffusion index of FOMC and Federal Reserve Board communication. Every governor speech, FOMC statement, minutes release, press conference, and Chair testimony is scored by Claude across five policy dimensions (rate path, balance sheet, growth, inflation, financial stability). Higher = hawkish (favors tighter policy); lower = dovish. documents scored to date. Updated .
Two specs of the same underlying data. Flat 30-day (solid) is the pre-reg-locked headline — every doc in the trailing 30 days gets equal temporal weight, then drops off cleanly at day 31. This is the spec that passed the SHIP_PUBLIC backtest gates and drives the headline number above. Persistence-weighted (dashed, half-life 14d) is shown alongside as a sharpening candidate — it weights recent docs more heavily and decays older ones smoothly rather than cutting them off, so single hawkish or dovish events fade gradually instead of vanishing on day 31. The overlay does not replace the headline; if it proves consistently better at flagging regime shifts, a future pre-registration + re-backtest can promote it. Until then, the locked spec stays the official line. Speaker × confidence × source-type weighting (Chair 4×, Vice Chair 3×, Governor 2×, FOMC statements/minutes 5×, others 1×) is identical across both.
Each panel is the same rolling-weighted diffusion as the headline above, but computed per-dimension: rate path (forward policy guidance), balance sheet (QT pace, runoff signals), growth (real-activity reads), inflation (persistence + expectations), fin stab (financial stability + credit). The headline aggregate hides which dimension is doing the work — when one panel runs hot while others sit at zero, that's the regime telling you what specifically the FOMC is positioning around. A +2.0 overall driven entirely by inflation reads differently from a +2.0 driven by rate-path forward guidance.
| Speaker |
n docs |
Weighted avg |
Band |
Most recent |
Confidence-weighted average overall score per speaker over the last 90 days. The scoreboard is observational, not a forecast — speakers shift positions across cycles. Scores are auditable per-doc; click a recent-docs row to inspect.
| Date |
Speaker |
Type |
Score |
Band |
Conf |
Source |
Per-doc scoring contract. Each document is scored on five policy dimensions (rate path, balance sheet, growth, inflation, financial stability) plus an overall composite, with a confidence (0–1) and two pull quotes (hawkish-leaning + dovish-leaning) extracted as evidence. Pull quotes are flagged as verbatim when they are direct, copy-paste quotes from the source — non-verbatim quotes are paraphrased extractions. Every scored document has a full audit trail: model version, prompt version, scoring timestamp, and input/output token counts.
Diffusion construction. Daily reading = weighted rolling average of scored documents, weighted by Claude's confidence × document recency. Single-document spikes don't dominate; the index is designed to capture aggregated regime tilt.
Validation. Backtest gates were locked in pre-registration before the first scoring run — face-validity test required positive slope (hawkish-shift days → next-day implied terminal rate rises) and R² ≥ 0.05. Both passed; ship-public verdict on docs/fedspeak_v0_backtest.md.
What this is not. Not a sentiment analysis demo, not a portfolio sleeve, not a real-time tape reader. No sizing or trade direction. The index is a regime classifier for Fed policy communication — not a forecast of the next FOMC decision.
Sources.