Time series foundation models struggle with electricity price forecasting, ensembles show promise

A new paper from Zhenghua Pan and Ahmed Aziz Ezzat at Rutgers University, presented at the ICML 2026 Foundation Models for Structured Data Workshop, systematically evaluates time series foundation models (TSFMs) for electricity price forecasting (EPF). EPF is a challenging testbed because of non-stationary data, price spikes, and strong dependence on covariates like weather and grid conditions. The authors propose a two-dataset benchmarking framework to mitigate data contamination and ensure fair evaluation. They test multiple TSFMs on point and probabilistic forecasting, tail behavior, and spike prediction.

Key finding: TSFMs are highly competitive and outperform general-purpose baselines, but their performance critically depends on covariate support—without proper inputs they degrade. Moreover, they do not consistently surpass domain-specific methods tailored for EPF. The most striking result: simple ensembles of TSFMs and domain-specific models achieve the best performance, indicating the two approaches capture complementary predictive signals. This suggests the future of electricity forecasting lies in hybrid systems that leverage both foundation model versatility and domain expertise.