Three essays in macroeconomic forecasting using dimensionality reduction methods

Guo, Yu (2023) Three essays in macroeconomic forecasting using dimensionality reduction methods. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2022GuoPhD.pdf] PDF
Download (3MB)


This thesis consists of three studies that concentrate on the dimensionality reduction methods used in macroeconomic forecasting.

Chapter 2 (the first study) aims to investigates the predictive ability of several indicators of consumer sentiment and perceptions about the economy. Based on seven key qualitative questions in the University of Michigan survey of consumers, I employ various quantification approaches to construct six indexes namely sentiment, disagreement, pessimism, uncertainty, price pressure, and interest rate pressure. I establish that these six indexes convey predictability for key macroeconomic indicators beyond and above the information found in existing, popular macroeconomic and financial indicators. I also provide a deep explanation of consumer indexes by monitoring their response to supply, demand, monetary policy and financial shocks using a VAR model with sign restrictions. The results indicate that price pressure and interest rate pressure are mainly correlated with financial and uncertainty shocks, while the other indicators reflect the formation of opinions that are sensitive to shocks related to supply, demand, and monetary policy.

Chapter 3 (the second study) explores the dimensionality reduction algorithm by extracting factors from a large number of predictors that take into account correlation with the predicted (target) variable, using a novel time-varying parameter three pass-regression-filter algorithm (TVP-3PRF). The benchmark 3PRF algorithm (Kelly and Pruitt, 2015) assumes that a predictor is relevant for forecasting over the whole sample and can be represented using a series of OLS regressions. I extend this approach using time-varying parameter regressions that are conveniently represented as a series of high-dimensional time-invariant regressions which can be solved using penalized likelihood estimators. TVP-3PRF algorithm allows for a subset of variables to be relevant for extracting factors at each point in time, accounting for recent evidence that economic predictors are short-lived. An empirical exercise confirms that this novel feature of TVP-3PRF algorithm is highly relevant for forecasting macroeconomic time series.

Chapter 4 (the third study) determines which of the two main types of algorithms in the field of dimensionality reduction truely reflect the true way variables enter the model. It is know that in the area of modelling and forecasting highdimensional macroeconomic and financial time series, two main methods, sparse modelling and dense modelling, are both popular. However, instead of simply viewing each a method for avoiding overfitting, a question that is worth exploring is which of these models can represent the real structure of the data. Another question that arises is whether the uncertainty of variable selection will affect the prediction. In line with Giannone et al. (2021), I used their spike and slab prior to explore the scenarios for six economies when forecasting production growth. The results indicate that the way macroeconomic data are employed in the model of all the economies have an obvious sparse structure albeit with different degrees. However, the pervasiveness of uncertainty causes the sparse model to fail and the model averaging technique to become the preferred method. Moreover, what is surprising is that the dense model(ridge regression) dominated after the pandemic began.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Subjects: H Social Sciences > HB Economic Theory
H Social Sciences > HG Finance
Colleges/Schools: College of Social Sciences > Adam Smith Business School
Supervisor's Name: Korobilis, Professor Dimitris and Tsoukalas, Professor John
Date of Award: 2023
Depositing User: Theses Team
Unique ID: glathesis:2023-83690
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 29 Jun 2023 11:51
Last Modified: 29 Jun 2023 11:51
Thesis DOI: 10.5525/gla.thesis.83690

Actions (login required)

View Item View Item


Downloads per month over past year