The Language of Central Banking: Probing Global Monetary Policy Communications Spillovers and Central Bank Shocks with Natural Language Processing Tools and a Novel Text Database

Abstract

The discipline of macroeconomics relies mainly on structured data for empirical research, despite unstructured text data being vastly more abundant. This text data, particularly central bank communications, holds untapped potential for monetary economics research due to their influence on market expectations and policy outcomes like inflation. To help guide monetary policy researchers in exploring the growing universe of text data, this research lays out a foundational framework, both in terms of coding infrastructure and Natural Language Processing (NLP) methods. The first step in building out this infrastructure is through the creation of a new open-source central bank text database consisting of monetary policy communications from 14 countries consisting of 2,418 monetary policy statements. I leverage this novel database to explore the literature on "information effects," which has mainly relied on structured data for empirical analysis despite the possibility that the phenomenon itself is attributable to the linguistic elements or sentiment expressed via central bank communications. Chapter 1 (The Anatomy of a Central Bank Statement and Information Shocks) details the steps necessary to create a reproducible and scalable database of monetary policy statements from a diverse group of countries using the latest open source technologies and modern data science practices. I find that positive co-movement between policy rates and equities (what the literature defines as an "information shock") is a common event, with almost half of all policy rate increases (decreases) occurring alongside higher (lower) equity prices. With linguistic regressions and part-of-speech annotations, I provide novel linguistic evidence that information shocks are likely related to both the future state of the economy \parencite{nakamura2018high} and inflation expectations \parencite{boehm2021beyond}. Chapter 2 (Sentiment Analysis-From Past to Present) develops a novel approach for extracting sentiment at the sentence level using cutting-edge transformers models, the architecture behind many large language models (LLMs). My research demonstrates that transformer models as well as the traditional lexical methods employed in the economic literature, can produce starkly divergent results when applied to the same monetary policy statement. This highlights the critical need to utilize multiple sentiment measures to ensure the robustness of any findings derived from textual analysis. Reinforcing the linguistic evidence from Chapter 1, I show that positive (negative) sentiment is associated with positive (negative) information shocks providing further evidence the shocks are driven by the language of the statement itself. I also show that positive sentiment is associated with higher GDP growth in the quarters following a monetary policy statement. Chapter 3 (Central Bank Shocks and Global Spillovers), aggregates sentiment measures from the previous chapter to produce what I call the Global Policy Stance (GPS). I find that the GPS, led by the U.S., Japan, and Switzerland, tends to co-move with the global financial cycle (Global Asset Prices Factor from \textcite{miranda2020global}). I also find that domestic sentiment, rather than U.S. or global sentiment, is predictive of future policy rate changes suggesting that markets may be more sensitive to the communications of the home country's central bank. This thesis sets a rigorous standard for database transparency and code reproducibility above and beyond what is standard practice in the economics literature today. I will publicly release the codebase encompassing data retrieval, cleaning processes, figure generation, model development, all of which were produced utilizing the open-source Python programming language. Through this public release, I will provide researchers with valuable coding infrastructure that supports the operationalization of best practices in data management, enabling (1) the creation of open-source databases fostering collaboration and automation, as well as (2) the development of reproducible, scalable algorithms for text classification and text cleaning processes. In the future, I intend to further build out the central bank database to include other types of monetary policy communications (e.g., minutes and speeches) while also separately maintaining a repository of text classification algorithms (e.g. positive and negative sentiment) including lexical dictionaries from the literature as well as fine-tuned transformer models.