Three Essays on Big Data in International Finance
- Author(s): Zang, Ziqi
- Advisor(s): Tornell, Aaron
- et al.
This dissertation presents an introduction to big data that can potentially be used in nowcasting key macroeconomic variables for advanced economies. It also explores the forecastability of big data in short-term exchange rate forecasting. Finally, it draws on evidence from a sentiment analysis of Article IV Consultations over the period of 2012 to 2018 and examines the development of member countries' perceptions of IMF policy advice.
Chapter 1 uses big data from Google search data to form better nowcasts of macroeconomic variables. My empirical strategy contributes to the macroeconomic nowcasting literature on three fronts. First, I take a number of steps to identify the most comprehensive set of relevant search queries that capture people's search behavior in relation to each monetary policy variable, such as the unemployment rate and inflation. Second, I consider regularization and dimension reduction methods to handle the underlying high-dimensional regressor space with highly correlated covariates. Third, I evaluate both average point forecasts and conditional point forecasts against benchmark models with DMW test and CSPA test, respectively. According to the test statistics, I find that Google search data offer significant improvements in nowcasting macroeconomic variables both unconditionally and conditionally.
Chapter 2 examines the short-term forecastability of exchange rates using machine learning models in a rich data environment. I investigate the performance of different
machine learning models, such as variable selection models, dynamic factor model, and decision regression trees in obtaining accurate forecasts of three currency pairs (U.S./U.K., Japan/U.S. and U.S./Australia). I consider three types of forecasts: point forecasts, unconditional weighted directional forecasts and conditional weighted directional forecasts. According to the DMW test, out-of-sample forecasts of every currency rejects the null hypothesis of equal forecasting errors with the random walk with at least one machine learning model. Furthermore, the conditional weighted directional forecasts allow us to know when exactly our models are more profitable than the random walk with zero profit. And it turns out that our weighted directional forecasts are significantly positive especially on the tails of the conditioning variable distribution.
Chapter 3 constructs multi-aspect policy sentiment measurements to interpret authorities' tones in response to specific policy advice in IMF Article IV Consultations. Specifically, we use a topic-based sentiment analysis approach that entails the application of a latent Dirichlet allocation (LDA) model as well as sentiment prediction machine learning models. Therefore, we are able to provide the stylized facts that provide useful input for assessing the impact of Fund advice on macroeconomic development of member countries.