FAQ

Is this endorsed by, or does it speak for, any central bank? No. It is an independent open-source project built only from public documents. See the disclaimer. The author works at a central bank; the project uses no non-public information and does not comment on any bank's policy choices.

Is it investment advice? No. Tone scores are noisy research signals with documented limitations.

Why is the Bank of Thailand excluded from the headline results? Its adapter is opt-in (include_bot: false in config/paper.yaml). It is implemented like every other adapter and can be included for appendix analyses.

Why does fetch report MAS as blocked? MAS blocks automated access, and the library does not try to get around a block. Save the pages listed in NEEDED.md by hand and the adapter parses them.

Can I redistribute the corpus? Only where the source's terms allow it. Fed text is public domain, RBA text is CC BY 4.0, and ECB statements may be reproduced with acknowledgement. For the other banks the repository holds URLs and hashes only. Details: LICENSES.md.

Can I use FOMC-RoBERTa commercially? No: it is CC-BY-NC-4.0 and gated on Hugging Face. Access was never granted for this study, so the paper uses a replication trained on the same CC-BY-NC data (experiments/train_fomc_roberta.py, decision D21), published at thanatchon36/cbtone-fomc-roberta-repl under the same non-commercial licence. Neither model may be used commercially.

Do I need a GPU? No. The target machine is a laptop without CUDA. The LLM scorer runs a small model (4B parameters or fewer) through Ollama.

How do I know an LLM score is not just memory? Every LLM result comes with a masked-text variant and a post-cutoff split, and predictive claims are made only where both survive. See Scorers.

Where are the results? In the paper and the one-page summary. Every testable hypothesis (H1–H4) was refuted at its pre-registered threshold. H5 was untestable because the BI and BNM English silver subsets are below 50 sentences, and on MAS the dictionary does not lose accuracy, so it could not have been supported.

What happens when a number looks too good? A headline metric at exactly 0, 1 or 100 %, a correlation above 0.95, or a value more than 2x outside the literature range fails the build until explained in ANOMALIES.md.

How do I add a bank, scorer, language or market series? See CONTRIBUTING.md.