I built this dataset and I run the dashboard linked at the bottom.<\/p>\n
I have been crawling monetary policy communications from 26 central banks and labelling them at the sentence level. The whole thing is CC-BY-4.0.<\/p>\n
What is in it:<\/p>\n
- \n
- 225,101 annotated sentences across 15,055 documents, Feb 1995 to Aug 2026<\/li>\n
- Policy statements, rate decisions, meeting minutes, press conference transcripts<\/li>\n
- 12 sentiment labels: rate_hike, rate_cut, rate_hold, guidance_hawkish, guidance_dovish, dissent_hawkish, dissent_dovish, liquidity_ease, liquidity_tight, reserve_ease, reserve_tight, neutral<\/li>\n
- 9 topic labels: inflation, interest_rate, economic_activity, labor_market, exchange_rate, credit, financial_stability, fiscal_policy, governance<\/li>\n
- 21 source languages, with an English translation on every non-English sentence in text_en<\/li>\n
- 19,387 economic indicator rows (policy rates, FX, CPI) so you can join labels against outcomes<\/li>\n
- Parquet, loads with datasets.load_dataset<\/li>\n<\/ul>\n
Sources are the central banks’ own sites (federalreserve.gov, ecb.europa.eu, boj.or.jp and so on). Every document keeps its source URL.<\/p>\n
The taxonomy follows IMF Working Paper WP\/25\/109, “From Text to Quantified Insights”. Labels are model-generated with gpt-4o-mini rather than hand-annotated, so spot-check them for your bank and period if you are using this for anything that matters.<\/p>\n
Dataset: https:\/\/huggingface.co\/datasets\/aufklarer\/central-bank-communications<\/a> Dashboard built on it: https:\/\/monetary.live<\/a><\/p>\n
Happy to take criticism of the taxonomy, especially the dissent and guidanthe hardest to pin down.<\/p>\n<\/div>\n
submitted by \/u\/ivan_digital <\/a>
[link]<\/a><\/span> [comments]<\/a><\/span><\/p>
