Evaluating Time Series Foundation Models for Electricity Price Forecasting: Contamination Risk, Distributional Shifts, and Covariate Dependence

Short summary

This research benchmarks time series foundation models for electricity price forecasting using a systematic two-dataset framework to avoid contamination bias. While TSFMs are competitive zero-shot performers, they critically depend on covariate support and don't consistently outperform domain-specific methods. Ensemble combinations of TSFMs and domain-specific approaches show significant potential, suggesting the two methods capture complementary information.

• •TSFMs competitive for electricity price forecasting but heavily dependent on covariate support
• •Domain-specific methods still necessary; TSFMs alone don't guarantee better performance
• •Ensemble approaches combining both methods show promise and capture complementary signals

Generated with AI, which can make mistakes.