FOMC Communication (Trillion Dollar Words)

Labels encode monetary-policy stance — dovish, hawkish, neutral — over sentences drawn from three FOMC communication types the authors describe as the largest tokenized and annotated collection of their kind. The split is 1,984 train and 496 test rows for 2,480 total; the Hugging Face viewer page is commonly read as 1,980 train rows, but the datasets-server count is 1,984. The CSVs load directly from Hugging Face, which also serves the two source annotation workbooks; the GitHub repository adds the training data, grid-search results and look-ahead-bias checks.

What separates this from generic financial sentiment data is the second half of the accompanying work: the paper aggregates the labels into a monetary-policy-stance measure and tests it against treasury yields, stock returns and macro indicators, and the repository ships that analysis as notebooks for treasury, buy-and-hold abnormal returns, and CPI/PPI. A companion classifier, FOMC-RoBERTa, is released under the same organization with a short how_to_use.py that runs it over arbitrary sentences through a transformers text-classification pipeline. The paper is Trillion Dollar Words: A New Financial Dataset, Task & Market Analysis, ACL 2023, pages 6664–6679 (2023.acl-long.368).

The license is the binding constraint here: CC BY-NC 4.0 requires attribution and prohibits commercial use, so the annotations cannot lawfully back a commercial trading signal even though the underlying FOMC documents are US Government works. It was read from the repository’s LICENSE.md, which contains the full CC BY-NC 4.0 text; GitHub’s license detector reports “Other” only because of the .md filename. The project is dormant — last commit 2024-12-17, Hugging Face dataset last modified 2024-12-16 — so the corpus ends well before the present and is not being extended. At 2,480 sentences it is a benchmark and a seed for signal construction, not a corpus for training a model from scratch.