Abstract
This paper presents a systematic benchmarking study of ensemble methods for Day-Ahead (DA) Electricity Price Forecasting (EPF), an area where rigorous evaluation remains scarce despite the well-established benefits of forecast combination. Unlike existing studies that typically propose and benchmark a single ensemble technique, we evaluate over 100 ensemble variants — spanning simple averaging, rule-based heuristics, and data-driven weighting schemes — applied to a diverse suite of statistical, machine learning, and deep learning base models. Our framework jointly addresses three dimensions: unified evaluation across combination rules, ensemble sizes, and weight update schemes; systematic comparison at both daily and per-period (quarter-hourly) granularity; and joint assessment of forecast accuracy and stability. All methods are applied at 15-minute price resolution, in line with the Italian day-ahead market’s transition to intraday settlement granularity completed in October 2025. Our results show that the best ensemble improves upon the best individual model by 5.81% in MAE — a statistically significant gain — while also reducing error variance and large-error incidence. Averaging-based methods deliver more uniform and statistically robust improvements than rule-based heuristics, which yield larger peak gains but with lower consistency. Component analysis confirms that all model families contribute to the best per-period ensemble, with deep learning architectures receiving greater weight during volatile midday periods. Overall, data-driven averaging ensembles emerge as practical and robust tools for operational DA-EPE.