Abstract
Accurate electricity price forecasting (EPF) is crucial for effectivedecision-making in power trading on the spot market. While recent advances ingenerative artificial intelligence (GenAI) and pre-trained large languagemodels (LLMs) have inspired the development of numerous time series foundationmodels (TSFMs) for time series forecasting, their effectiveness in EPF remainsuncertain. To address this gap, we benchmark several state-of-the-artpretrained models--Chronos-Bolt, Chronos-T5, TimesFM, Moirai, Time-MoE, andTimeGPT--against established statistical and machine learning (ML) methods forEPF. Using 2024 day-ahead auction (DAA) electricity prices from Germany,France, the Netherlands, Austria, and Belgium, we generate daily forecasts witha one-day horizon. Chronos-Bolt and Time-MoE emerge as the strongest among theTSFMs, performing on par with traditional models. However, the biseasonal MSTLmodel, which captures daily and weekly seasonality, stands out for itsconsistent performance across countries and evaluation metrics, with no TSFMstatistically outperforming it.