Essays in macroeconomic and financial forecasting using big data econometric methods

Student thesis: Doctoral Thesis


This thesis explores several aspects of econometric methods in time series forecasting of both macroeconomic and financial variables. The contribution is provided in three essays. The first essay (Chapter 2) contributes to the econometric literature and develops models for regional nowcasting. We use Bayesian mixed frequency methods estimated at the common lower frequency. Moreover, we propose a procedure which allows model estimation with stochastic volatility and large datasets. We produce high frequency state-level GDP nowcasts that will assist policymakers in understanding the impact of greater regionalisation on economic growth in the U.S., and evaluate its impact on present and future economic conditions in a more timely fashion. We evaluate the accuracy of point and density forecasts, by making comparisons across models with constant and stochastic volatility. We provide results on the accuracy of nowcasts of real-time economic growth in the U.S. from 2006 to 2018. Empirical results suggest that models with stochastic volatility outperform models with constant volatility at nowcasting. The second essay (Chapter 3) develops a Mixed Frequency Vector Autoregressive model (MF-VAR) for producing timely monthly nowcasts and historical estimates of GDP growth at the state level in the U.S. economy. The variables in the MF-VAR include GDP growth at the state and country level, as well as additional monthly variables at the state and country level. The variables are observed at different frequencies, leading to a complicated high-dimensional MF-VAR. A computationally-fast approximate Bayesian Markov Chain Monte Carlo (MCMC) algorithm is proposed for estimating the MF-VAR coefficients and nowcasting. Empirical results explore the nature and magnitude of spillover effects among the U.S. states. Further, the proposed model produces historical estimates at monthly frequency for both the U.S. economy and U.S. states. The third essay (Chapter 4) proposes a deep quantile estimator, using neural networks and their universal approximation property to examine a non-linear association between the conditional quantiles of a dependent variable and predictors. The proposed methodology is versatile and allows both the use of different penalty functions, as well as high dimensional covariates. We present a Monte Carlo exercise where we examine the finite sample properties of the proposed estimator and show that our approach delivers good finite sample performance. We use the deep quantile estimator to forecast Value-at-Risk and find significant gains over linear quantile regression alternatives, supported by various testing schemes. The Chapter also contributes to the interpretability of neural networks output by making comparisons between the commonly used SHAP values and an alternative method based on partial derivatives.
Date of Award4 Oct 2022
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SupervisorGary Koop (Supervisor) & Stuart McIntyre (Supervisor)

Cite this