Essays on time series analysis and statistical machine learning

Research output: Thesis › Doctoral thesis

Authors

  • Johanna Meier

Research Organisations

View graph of relations

Details

Original languageEnglish
QualificationDoctor rerum politicarum
Awarding Institution
Supervised by
  • Philipp Sibbertsen, Supervisor
Date of Award5 Oct 2023
Place of PublicationHannover
Publication statusPublished - 2023

Abstract

This thesis encompasses three research articles contributing to the fields of time series analysis and statistical machine learning. Firstly, we develop a peaks-over- threshold approach, which captures both short- and long-term correlations in the underlying time series in order to model the clustering behaviour in high-threshold exceedances. The suggested model is motivated by and applied to oceanographic data. Secondly, we propose an efficient discrepancy-based inference approach for intractable generative models based on quasi-Monte Carlo methods. We demonstrate that this method substantially reduces the computational cost of estimating the model parameters in various applications of academic and practical interest. Thirdly, we suggest training methods for deep sequential models, which improve the forecast precision when facing structural breaks in the in-sample period. These mitigation strategies are examined in an extensive simulation study and utilised to forecast energy data. As the developed theory in this thesis is very versatile, it is applicable to a broad range of data types as well as research fields, and in particular to economic time series.

Cite this

Essays on time series analysis and statistical machine learning. / Meier, Johanna.
Hannover, 2023. 57 p.

Research output: Thesis › Doctoral thesis

Meier, J 2023, 'Essays on time series analysis and statistical machine learning', Doctor rerum politicarum, Leibniz University Hannover, Hannover. https://doi.org/10.15488/14896
Meier, J. (2023). Essays on time series analysis and statistical machine learning. [Doctoral thesis, Leibniz University Hannover]. https://doi.org/10.15488/14896
Meier J. Essays on time series analysis and statistical machine learning. Hannover, 2023. 57 p. doi: 10.15488/14896
Meier, Johanna. / Essays on time series analysis and statistical machine learning. Hannover, 2023. 57 p.
Download
@phdthesis{bac4022d6ccb428f90ed3969aaad35f9,
title = "Essays on time series analysis and statistical machine learning",
abstract = "This thesis encompasses three research articles contributing to the fields of time series analysis and statistical machine learning. Firstly, we develop a peaks-over- threshold approach, which captures both short- and long-term correlations in the underlying time series in order to model the clustering behaviour in high-threshold exceedances. The suggested model is motivated by and applied to oceanographic data. Secondly, we propose an efficient discrepancy-based inference approach for intractable generative models based on quasi-Monte Carlo methods. We demonstrate that this method substantially reduces the computational cost of estimating the model parameters in various applications of academic and practical interest. Thirdly, we suggest training methods for deep sequential models, which improve the forecast precision when facing structural breaks in the in-sample period. These mitigation strategies are examined in an extensive simulation study and utilised to forecast energy data. As the developed theory in this thesis is very versatile, it is applicable to a broad range of data types as well as research fields, and in particular to economic time series.",
author = "Johanna Meier",
year = "2023",
doi = "10.15488/14896",
language = "English",
school = "Leibniz University Hannover",

}

Download

TY - BOOK

T1 - Essays on time series analysis and statistical machine learning

AU - Meier, Johanna

PY - 2023

Y1 - 2023

N2 - This thesis encompasses three research articles contributing to the fields of time series analysis and statistical machine learning. Firstly, we develop a peaks-over- threshold approach, which captures both short- and long-term correlations in the underlying time series in order to model the clustering behaviour in high-threshold exceedances. The suggested model is motivated by and applied to oceanographic data. Secondly, we propose an efficient discrepancy-based inference approach for intractable generative models based on quasi-Monte Carlo methods. We demonstrate that this method substantially reduces the computational cost of estimating the model parameters in various applications of academic and practical interest. Thirdly, we suggest training methods for deep sequential models, which improve the forecast precision when facing structural breaks in the in-sample period. These mitigation strategies are examined in an extensive simulation study and utilised to forecast energy data. As the developed theory in this thesis is very versatile, it is applicable to a broad range of data types as well as research fields, and in particular to economic time series.

AB - This thesis encompasses three research articles contributing to the fields of time series analysis and statistical machine learning. Firstly, we develop a peaks-over- threshold approach, which captures both short- and long-term correlations in the underlying time series in order to model the clustering behaviour in high-threshold exceedances. The suggested model is motivated by and applied to oceanographic data. Secondly, we propose an efficient discrepancy-based inference approach for intractable generative models based on quasi-Monte Carlo methods. We demonstrate that this method substantially reduces the computational cost of estimating the model parameters in various applications of academic and practical interest. Thirdly, we suggest training methods for deep sequential models, which improve the forecast precision when facing structural breaks in the in-sample period. These mitigation strategies are examined in an extensive simulation study and utilised to forecast energy data. As the developed theory in this thesis is very versatile, it is applicable to a broad range of data types as well as research fields, and in particular to economic time series.

U2 - 10.15488/14896

DO - 10.15488/14896

M3 - Doctoral thesis

CY - Hannover

ER -