Abstract
dc:description.abstractI consider how causal and predictive economic inference problems can be solved using textual data. In doing so, I particularly investigate opportunities created by recent advances in natural language modelling, which appear to remain underexplored as an econometric methodology—a gap this thesis aims to help fill. Through this lens, I revisit three economic inference problems: estimation of the dynamic causal effects of monetary policy, forecasting of macroeconomic aggregates, and predicting financial risk premia. Re-examining the dynamic causal effects of UK monetary policy, I find that orthogonalising conventional instrumental variables with respect to large language model-extracted information appears to help resolve longstanding puzzles within empirical monetary economics. Examining the incremental value of the text of the Federal Reserve System’s ‘Beige Book’ for forecasting US macroeconomic aggregates, I find apparent improvements in forecast accuracy, but also show how the inclusion of language model-generated predictors can make bias-free forecast evaluation challenging. Considering an investor’s inference problem, I make a formal connection between microeconomic theory regarding rankings of information structures and the practical choice of which words or tokens should be prioritised in financial language modelling. Taken together, these findings underscore that language models are a powerful addition to the econometric instrumentarium available to researchers and practitioners. Whilst there are pitfalls to avoid, there can be little doubt that language models will become essential tools for drawing economic inferences.
Degree
thesis:*- Level dc:type.qualificationlevel
- PhD
- Grantor dc:publisher.institution
- University of Glasgow
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Mittendorf, Daniel
Subjects
dc:subject × 1Rights
- Language dc:language
- en
Identifiers
dc:identifier.*- DOI dc:identifier.doi
- https://doi.org/10.5525/gla.thesis.85360