Steps overview:
We manually collected a data corpus consisting of 591 monetary policy minutes, for central banks in Czechia, Hungary, Poland, and Romania, from the central banks websites.
We randomly selected a sample of sentences from the corpus of central banks minutes. We manually labeled each sentence by considering the hawkishness, dovishness, and neutral stance of monetary policy stance.
We used the BERT model, which is pre-trained on BookCorpus and English Wikipedia corpus, and the FinBERT model, further re-pre-trained on Thomson Reuters financial corpus.
We fine-tuned the BERT and the FinBERT models on our dataset of 1,998 labeled sentences on central bank monetary policy stance and we obtained a deep learning domain-specific model, ready for sentiment predictions.Â
We calculated the tone of each minute and we estimated and compared the predictive power of our central bank sentiment index.
All the computations were run in Python. The complete dataset is available by accessing the Harvard Database.
The scripts user are available, upon request, from the authors.