Foundation Models' Role in Electricity Price Forecasting Examined.

The promise of foundation models to deliver accurate forecasts with minimal task-specific training is appealing, but their effectiveness in specialized domains like electricity price forecasting remains an open question. This research investigates whether these general-purpose models can truly replace highly optimized, market-specific forecasting tools. The study evaluated nine variants from five foundation model families in a zero-shot setting against two state-of-the-art electricity price forecasting benchmarks across German, Polish, and Spanish markets from 2021-2025. Performance was assessed based on point and probabilistic forecasting accuracy, as well as economic value in battery energy storage arbitrage. Results showed that only TabPFN models consistently and significantly outperformed the benchmarks statistically. However, this statistical edge did not always translate into economic superiority. TabPFN excelled under unlimited bids and riskier quantile-based strategies, while a Distributional Deep Neural Network benchmark proved more profitable with lower risk tolerance. The conclusion is that foundation models cannot universally replace market-specific models, as their value is highly dependent on the specific model architecture and the decision-making context.