Natural language processing for economic and financial modelling

Thesis

Over the past years, researchers have introduced various natural language processing (NLP) methods to the fields of economics and finance. It is time to take stock and systematically evaluate which NLP methods work best for the most common domain specific tasks, under the most common domain specific data characteristics, and to let those findings guide future research. We underline the importance of evaluating multimodal (text and numeric) data tasks, since many text datasets in economics and finance come accompanied by potentially relevant numeric metadata. Particularly in the field of modern monetary economics, multimodal data analysis is of vital importance. Ever since the global financial crisis, central bank communication has become a key monetary policy tool. The development of multimodal data analysis frameworks is pivotal for modern day monetary policy research, yet the integration of NLP methods into economic and financial modelling processes is far from being conclusively answered.

In this DPhil thesis, we introduce work towards establishing an NLP benchmark foundation for economics and finance that includes multimodal dataset evaluation tasks. Next, we introduce Bayesian Topic Regression (BTR). BTR is a multimodal NLP algorithm based on a supervised topic model that can jointly model text and numeric data. With BTR, we provide the research community with a method for more reliable causal inference with text data, in an identification framework commonly used in the economics and finance literature. BTR also demonstrates competitiveness in prediction tasks that rely on text and numeric data. Finally, we introduce a multimodal NLP framework for the application space of monetary policy analysis. We establish a new monetary policy shock series based on central bank speeches along three key macroeconomic measures: GDP growth, inflation, and unemployment. Based on empirical estimates on our newly constructed central bank communication dataset, our monetary news signals imply that news on macroeconomic outlooks in central bank communication can help explain equity and bond market volatility and tail risk. The news signals carry relevant information to which markets attend and react. We also derive a monetary news dispersion index, measuring the degree of alignment amongst different central bankers in their policy communication with the markets. Our findings suggests that more misaligned policy communication is associated with stronger market surprises at policy announcement time.

Authors: Ahrens, M

• DOI: 10.5287/ora-7e1qdbj0k
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: LLM; LLMs; NLP for finance; topic models; NLP for economics; monetary econometrics; natural language processing; NLP benchmarks for finance; macroeconometrics; NLP; large language models