Essays in macroeconomics and machine learning

The unifying theme of this thesis is the use of techniques from machine learning and data science to address questions in macroeconomics. It makes both theoretical contributions by applying neural networks as a learning algorithm in models that are indeterminate under rational expectations, and empirical contributions by developing and applying natural language processing methods to datasets including news media articles and central bank communication.

The first Chapter, Resolving Indeterminacy with Neural Network Learning: Sinks become Sources, aims to make a theoretical contribution to the literature on indeterminacy in rational expectations models. Indeterminacy (i.e. non-uniqueness of equilibrium under rational expectations) is a pervasive and often neglected challenge for macroeconomists. Previous literature on learning in macroeconomics has shown that the equilibria in linear indeterminate models are almost always not learnable. This Chapter examines the equilibrium sophisticated learning agents converge to in models where there is indeterminacy, but it is bounded. This neural network learning converges to a stable equilibrium, in which indeterminate regions are sources and determinate regions are sinks. Furthermore, this equilibrium is consistent with Rational Expectations. There are multiple steady states and agents have correct beliefs about transitions between them, which means that transitory shocks can have permanent effects. This is demonstrated with an application to a well-known example of indeterminacy: a New Keynesian model with a Zero Lower Bound in interest rates.

The resolution to the challenge of indeterminacy presented here is plausible, as it’s based on learnability, and also appealing because the identified equilibrium is globally stable, can have multiple steady states with well-defined transitions between them, and passes tests for rationality. It also demonstrates the value of using machine learning, as the neural network acts as a very flexible function approximator that is fast to train, allowing the use of learnability as an equilibrium selection device in non-linear models. Alternative learning algorithms like Recursive Least Squares yield qualitatively similar results, but are not sufficiently flexible to pass tests of rationality.

Chapter 2, Bayesian Topic Regression for Causal Inference has a more method- ological focus, developing Bayesian Topic Regression, a model for causal inference with text data. This methodology is then applied in Chapter 3 to help identify a potentially causal effect of media coverage on stock price volatility. The Bayesian Topic Regression model jointly estimates topics in text documents and a regression using these topics and associated numerical data to predict a response variable. As well as showing that per- forming text feature extraction and prediction in separate stages can lead to incorrect inference, we benchmark our model on two real-world customer review datasets and show markedly improved out-of-sample prediction in comparison to competing approaches.

Chapters 3 and 4 use text data to address to empirical questions relating to how agents in the economy acquire information and on what they focus their attention. In Chapter 3 Financial news media and volatility: is there more to newspapers that news? identifies a co-movement media coverage in the Financial Times newspaper and a firm’s intra-day stock price volatility is identified. I argue that part of this co-movement is causal, relying on an identification strategy based on the publication time of the newspa- per, controlling for persistence and anticipation effects as well as the content of articles using the Bayesian Topic Regression introduced in Chapter 2. These results are consis- tent with a salience-based view of the media’s role in financial markets: media coverage does not (only) provide information, but influences where investors choose to direct their focus. This identified effect also has interesting spillovers to firms in sectors that are linked by the production network.

In Chapter 4 The Shifting focus of Central Bankers. I use an unsupervised topic model to quantify the focus of central bank communication and economic news media. I offer an explanation for the variation of this focus over time, and identify a robust co-movement between central bank and media focus. A model of multidimensional uncertainty and limited attention is proposed to explain the shifting focus of central bank communication. Evidence from the Survey of Professional Forecasters is used to support this explanation, showing that focus shifts to cover variables about which there is greater uncertainty. An event study approach is used to show a potentially causal influence of Federal Reserve communication on the focus of US news media and on the communication of other central banks.