Umbach, Simon Lineu ORCID: 0000-0002-2410-9022 (2020). Macroeconomic Forecasting and Evaluation with Supervised and Neural Network Reinforced Factor Models. PhD thesis, Universität zu Köln.

[img]
Preview
PDF
Macroeconomic Forecasting and Evaluation with Supervised and Neural Network Reinforced Factor Models.pdf

Download (1MB) | Preview

Abstract

This thesis comprises three self-contained essays on macroeconomic forecasting with factor models, and on forecast evaluation tests. First, it is analyzed how factor estimates can be tailored to forecasting applications by incorporating the forecasting target directly in the factor estimation process. For this purpose, the Principal Covariate Regression technique is refined and it is analyzed under which circumstances gains in forecast accuracy can be achieved by integrating this form of supervision in the factor estimation. Second, the statistical factor model is aligned with the variational autoencoder framework in the context of macroeconomic forecasting. It is studied whether factor models enriched by neural networks can provide superior forecasting power for macroeconomic time series. In contrast to the original factor model, the resulting neural network reinforced factor model is not subject to the linearity restriction anymore, and can capture nonlinear common dynamics in the set of candidate predictors as well. Furthermore, it is proposed to incorporate the aforesaid supervision aspect within these models. The extended factor models are applied to forecast key monthly macroeconomic variables such as industrial production, inflation, and employment. The findings suggest that their forecasting capability can be significantly improved by the analyzed and refined extensions. Third, an adjustment of the Diebold and Mariano test is proposed. A comparison of two competing forecasts of the same economic quantity requires a formal statistical procedure to distinguish between a better predictive accuracy by coincidence and a fundamental advantage of one over the other. To this end, one of the most popular statistics is the Diebold and Mariano test. This thesis contributes to the literature by showing how the power of the Diebold and Mariano test can be improved when the forecasts are rational, i.e., unbiased and efficient.

Item Type: Thesis (PhD thesis)
Creators:
CreatorsEmailORCIDORCID Put Code
Umbach, Simon Lineusimon.umbach@gmx.deorcid.org/0000-0002-2410-9022UNSPECIFIED
URN: urn:nbn:de:hbz:38-515410
Date: October 2020
Language: English
Faculty: Faculty of Management, Economy and Social Sciences
Divisions: Faculty of Management, Economics and Social Sciences > Economics > Econometrics and Statistics > Professorship for Statistics and Econometrics
Subjects: General statistics
Economics
Uncontrolled Keywords:
KeywordsLanguage
Macroeconomic Forecasting, Factor Models, Neural Networks, Forecast Evaluation TestsEnglish
Date of oral exam: 18 February 2021
Referee:
NameAcademic Title
Breitung, JörgProf. Dr.
Kruse-Becher, RobinsonProf. Dr.
Refereed: Yes
URI: http://kups.ub.uni-koeln.de/id/eprint/51541

Downloads

Downloads per month over past year

Export

Actions (login required)

View Item View Item