Three cooperating agents
The pipeline is built from three cooperating agents rather than a single prompt. The Monitor Agent continuously tracks and stores Fed documents; the Analyst Agent reads each release and extracts its tone, section by section; the Strategist Agent turns the smoothed tone, together with macro data, into a rate-move forecast. A read-only layer then exposes the finished results to this dashboard.
Monitor · Analyst · StrategistRuns on a fixed schedule
Sectioned tone scoring
Rather than scoring a release as a whole, the pipeline first splits it into its functional sections — forward guidance, inflation, labor market, and the general assessment — and asks a large language model for a tone from −1 (dovish) to +1 (hawkish) on each. Sections carrying no signal are excluded so that short documents are not pulled toward neutral, and every model response is cached, so re-running the analysis costs nothing.
Scale · −1 dovish · +1 hawkishResponses cached
Weights live in the database
The document's composite tone is a weighted average of its section tones, and those weights are stored as data rather than hard-coded into the model. Because they sit outside the language model, they can be recalibrated against historical FOMC outcomes at any time — without re-running any of the text analysis. Forward guidance carries the most weight; the general section is retained for completeness but contributes almost no signal.
Forward guidance45%
Inflation25%
Labor market15%
General15%
The finance model
Tone is smoothed with a half-life EWMA (21-day half-life over the calendar gap between releases), then fed to an ordered-probit nowcast whose latent index blends tone with the core-CPI gap and the unemployment gap. The bucket probabilities give a tone-implied next rate; its divergence from the 2-year Treasury proxy is the headline signal.
S_t = α_t·tone_t + (1 − α_t)·S_{t−1}, α_t = 1 − exp(−ln2/21 · Δt)
η_t = β_S·S_t + β_π·(CPI_yoy − 2) + β_u·(U − U_baseline)
P(Y_t = j_k) = Φ(c_k − η_t) − Φ(c_{k−1} − η_t), j ∈ {−50, −25, 0, +25, +50}
Data & infrastructure
Macroeconomic inputs come from FRED — core CPI, unemployment, the 2-year Treasury yield, and the policy rate — joined on observation month. Everything the pipeline produces is persisted in a single database spanning documents, tone scores, weights, and calibration history. A synthetic test environment mirrors the Fed's public structure, so the monitoring can be validated end to end before it ever touches live data.
Inputs · FRED macro seriesSingle store · fully reproducible