FOMC Statement Sentiment vs. 2Y Yield Reaction

Overview

Implementation

Core libraries

transformers and torch for FinBERT, statsmodels for HAC OLS, BeautifulSoup and lxml for scraping, pydantic for schemas, pandas, pyarrow, Matplotlib, Plotly

Text corpus

138 cleaned FOMC statements, 2010 to 2026, committed as processed text

Scoring

Sentence-level FinBERT with a pinned model name and revision, a hawkish/neutral/dovish label map, and a recorded scorer version

Market leg

A Barchart ZTY00 nearby-futures intraday export, ingested from 15 numbered raw chunks; daily FRED DGS2 as reference only

Tests

9 modules covering schemas, event windows, modelling and a pipeline smoke path

Components

src/fomc_scraper.py

Statement acquisition and release-date extraction

src/text_cleaning.py, src/sentence_splitter.py

Boilerplate removal and the sentence segmentation FinBERT consumes

src/sentiment.py

The FinBERT scorer wrapper, pinning model name, revision, label mapping and scorer version into the artifact

src/yield_data.py

Intraday market-leg ingestion and the daily reference series

src/event_windows.py

Release-minute alignment and the baseline and alternate reaction windows

src/event_dataset.py

The joined event panel every regression reads

src/cycle_labels.py

Tightening, easing and hold labels for the per-cycle breakdown

src/modeling.py

HAC OLS plus window, cycle and outlier sensitivity and the headline metric

src/validation.py, src/schemas.py

Schema contracts and the sentiment validation pass against manual labels

scripts/preflight_checks.py

The gate run before a report is rendered

Pipeline

1. ingest-statements Scrape, clean and store statement text with release timestamps.
2. ingest-yields Ingest the intraday market leg from the numbered raw chunks.
3. score Score every sentence with FinBERT and aggregate to a per-statement signal.
4. validate Check schemas and compare scored sentiment against the manual label set.
5. event-study Build the event panel and fit the baseline and sensitivity regressions.
6. report Render figures and the brief, behind preflight checks.

Reproducibility and validation

• The scorer records its model name, revision and version into every artifact, so a silent upstream model change cannot be mistaken for a change in the data.
• Sentiment scores are validated against a manually labelled subset following a documented labelling guide, rather than being trusted because the model is well known.
• Forty-eight releases lack verified minute-level timestamps and are excluded from the baseline regression instead of being assigned an approximate time.
• The intraday leg is a nearby-futures proxy rather than a cash two-year yield feed; legacy yield_* field names are kept for schema compatibility, but the rendered report interprets them as proxy moves, and the finished run is frozen as a dated bundle under reports/published/.

Availability