Beyond Average Error: SpikeBench-PJM for Real-Time Electricity Price Forecasting

Real-time (RT) electricity price forecasting is more challenging than day-ahead forecasting because RT markets operate closer to physical delivery, with the difficulty particularly pronounced during price spikes. Although numerous forecasting models have been developed for electricity prices, existing studies primarily emphasize average predictive accuracy and provide limited evaluation of extreme-price regimes. We introduce SpikeBench-PJM, a unified benchmark for multi-horizon PJM RT price forecasting with explicit spike-aware evaluation. SpikeBench-PJM defines causally aligned information regimes and systematically evaluates lightweight forecasting models, pretrained time-series foundation models, and multimodal models. Our results show that the value of exogenous information depends strongly on both the information type and the evaluation regime, and that model rankings can differ between average-error and spike-focused evaluation. Pretrained time-series foundation models provide competitive performance but do not consistently outperform lightweight baselines, while fine-tuning on PJM data generally improves over zero-shot forecasting. Multimodal experiments further indicate that operational text can provide additional predictive value, particularly after task-specific adaptation and for spike-focused forecasting. These findings highlight the importance of evaluating real-time electricity price forecasting beyond average error.