Sentiment analysis for monetary policy research

The article examines the experience of using sentiment analysis methods in the field of monetary policy for the research of public opinion, central bank communication, macroeconomic analysis and forecasting. Natural language processing (NLP) methods overcome the disadvantages of traditional approaches based on opinion polls in the context of measuring inflation expectations and trust in the central bank. Sentiment analysis is actively used in the study of various aspects of central bank communication and the assessment of media sentiment. As a tool for quantifying qualitative data, sentiment analysis allows to build high-frequency and operational estimates of indicators. It is noted that thematic indicators based on sentiment analysis can in some cases improve the accuracy of macroeconomic forecasting by enriching information extracted from classical economic variables. The development of machine learning and the emergence of new neural network architectures expand the possibilities of using sentiment analysis, but they also raise the issue of choosing tools. The article provides a comparative analysis of various sentiment analysis methods, including classical machine learning models and neural networks, for the task of building an indicator of trust in the Central Bank. The approach based on fine-tuning of pre-trained Bidirectional Encoder Representations from Transformers (BERT) is recognized as the most effective. It was revealed that fine-tuning large language models can significantly improve the quality of solving the problem of sentiment analysis by adapting to the subject area, which is especially important for research in the field of monetary policy.