Text Representations for Central Bank Communication Changepoint Detection

This is my Dissertation codebase exploring how different text representations (scalar sentiment scores, TF-IDF, and transformer embeddings) affect changepoint detection on Bank of England Monetary Policy Committee (MPC) minutes, using both real BoE data and synthetic LLM-generated corpora.

Contents

• text_mining_data_processing.ipynb — scrapes and cleans BoE MPC minutes (1997–2026) from bankofengland.co.uk, extracts features (bank rate, decision direction), and produces the cleaned dataset used downstream.
• synthetic_text_generation.ipynb — generates synthetic BoE-style documents via a locally-served Ollama model, including a graded hawkish-dovish sentiment ladder, topic-shift corpora, and extreme positive/negative tone corpora, for controlled changepoint simulations.
• simulation_study.ipynb — Monte Carlo simulation study comparing scalar/TF-IDF/embedding representations and kernel choices for changepoint detection accuracy.
• CPD_real_BoE_data.ipynb — applies changepoint detection to the real scraped BoE minutes corpus.
• embedding_exploration.ipynb — exploratory analysis of transformer embedding representations.
• topic_prevalence_experiment.ipynb — topic prevalence analysis over time.
• utils/ — shared helper modules (sentiment.py, text_mining.py, kernel_mmd.py, ruptures_wrappers.py).
• data/ — scraped BoE minutes and synthetic corpora (CSV).

Setup

python -m venv .venv
.venv\Scripts\activate   
pip install -r requirements.txt

Some notebooks (synthetic_text_generation.ipynb) require a locally running Ollama server with a pulled model (e.g. ollama pull mistral) reachable at http://localhost:11434.

Suggested run order

1. text_mining_data_processing.ipynb
2. synthetic_text_generation.ipynb
3. simulation_study.ipynb
4. CPD_real_BoE_data.ipynb

Notes

This is a research/dissertation repo, not a packaged library — code prioritizes clarity and reproducibility of experiments over production-readiness.