Art der Arbeit

Dissertation Betriebswirtschaftslehre

Verfasser

Berrisch, Jonathan

Gutachter

Prof. Dr. Florian Ziel

Volltext

https://doi.org/10.17185/duepublico/83826

Kurzfassung

This thesis contributes to the field of probabilistic forecasting in energy markets through novel methodological developments and empirical applications. The liberalization of energy markets over the past 25 years, combined with the transition towards carbon-neutral energy systems, has created increasingly complex and interconnected markets characterized by high uncertainty from renewable energy sources, demand fluctuations, and geopolitical risks. Traditional point forecasting methods are inadequate for capturing this uncertainty, necessitating probabilistic approaches that enable better risk management and decision-making.
The thesis makes several methodological contributions to the field of probabilistic online learning. First, we introduce CRPS learning, a novel aggregation method that combines probabilistic forecasts using flexible weight functions across quantiles rather than static weights. This approach accounts for varying forecast performance within different parts of the predictive distribution. Second, we extend this framework to multivariate settings, developing a method that exploits dependencies between combination weights of marginal distributions using penalized splines. Third, we propose a regularized probabilistic online learning algorithm that combines online coordinate descent with the framework of generalized additive models for location, scale, and shape (GAMLSS), enabling direct prediction of distributional parameters in high-dimensional settings.
The thesis also addresses specific forecasting challenges in energy markets through three empirical applications. We develop probabilistic forecasting models for day-ahead and month-ahead natural gas prices, accounting for heavy tails and asymmetric effects while considering the influence of European Emission Allowances (EUA) prices. We propose a multivariate VECM-Copula-GARCH model to jointly forecast EUA, natural gas, coal, and oil prices, capturing long-term equilibrium relations, conditional heteroscedasticity, and time-varying dependence structures through one-step estimation. Additionally, we explore the prediction of high-resolution electricity peak demand from low-resolution data using both statistical and machine learning approaches.
Complementing these methodological and empirical contributions, we address the computational challenges of implementing probabilistic forecasting methods by developing an R package for benchmarking C++ code. This tool provides overlapping and nested timers with OpenMP support and nanosecond-level precision, facilitating the development of efficient implementations for online learning algorithms.
The proposed methods are particularly well-suited for streaming data applications in energy markets, where models must update rapidly with new observations. Through simulation studies and empirical applications using real energy market data, we demonstrate the practical utility and superior performance of our approaches compared to established benchmarks. This work advances both the theoretical understanding of probabilistic forecasting and its practical implementation in energy market applications, contributing to more informed decision-making in an increasingly uncertain energy landscape.