BSc Economics · Aarhus University · May 2025
ELECTRICITY PRICE FORECASTING IN THE ERA OF FOUNDATION MODELS
A benchmark of twelve forecasting models across six European bidding zones, covering classical econometric methods, machine learning, and time series foundation models.
Abstract
This thesis provides a comparative analysis of classical time series models, machine learning and zero-shot foundation models for forecasting European Day-Ahead electricity prices. Driven by the increasing integration of renewable energy, modern electricity markets exhibit extreme volatility and non-linear dynamics that have challenged traditional forecasting methods. Using hourly data from 2020 to 2024 across six European bidding zones, this study benchmarks twelve models. The empirical results demonstrate that classical linear models are insufficient for capturing volatile merit-order dynamics. While state-of-the-art multivariate foundation models achieve highly competitive point and probabilistic forecasts, though they present a memorization bias risk. Localized machine learning models consistently demonstrate a robust and competitive performance, particularly outperforming foundation models in the highly volatile bidding zones.
Conclusion
This thesis empirically evaluated the predictive performance of classical time series models, machine learning algorithms and zero shot foundation models across six heterogeneous European Day-Ahead electricity markets. The results prove that the ongoing integration of renewable sources has changed the market dynamics, resulting in classical time series models such as SARIMAX, ETS, MSTL-ARIMAX, uncapable of capturing the highly complex nonlinear relationships. The transition to nonlinear architectures is therefore necessary. Machine learning models such as CatBoost and NHITS demonstrated statistically significant predictive gains over classical models. Furthermore, the analysis revealed that even though state of the art foundation models are the most complex, this is not sufficient to offset the need for covariates, hence univariate zero shot forecasting is inadequate. However, when multivariate foundation models like Chronos V2 and Times FM are supplied with covariates, they deliver highly competitive forecasts. Nevertheless, because pre-trained foundation models carry an unquantifiable risk of data leakage, inflicted by the memorization bias, localized machine learning algorithms remain the most trustworthy. Concluding this thesis, multivariate foundation models perform competitively, but localized machine learning models remain the most robust and reliable tools for electricity price forecasting, not only because they excel in highly volatile bidding zones, but also because their localized training guarantees validity in their out-of-sample performance.