Forecasting day-ahead electricity market results

Kurzfassung

Rising price volatilities in the German electricity market exposes the participants to risks from renewables intermittency, underscoring the need for accurate price forecasts for operational optimization and trading. This work generates deterministic and probabilistic price forecasts for the Single Day Ahead Coupling (SDAC) using XGBoost and Quantile XGBoost (QXGB), benchmarked against a weekly Naive model. The objective is to comprehend the gains and limitations of the selected algorithms in the context realistic uncertainty-aware electricity production planning and market participation. The model incorporates core grid features, calendar encodings, commodity prices, autoregressive lags, seasonality and volatility components across 1-to-4-year lookbacks. Probabilistic outputs are calibrated via conformal adjustment and sorting. With an 85%-15% train-validation split and 2025 as test set, deterministic model achieves an MAE of 18.6€/MWh with an optimal lookback of 2 years. Point forecaster has a significantly lower MAE than the Naive forecaster, confirmed by block bootstrapping and Diebold-Mariano tests. Probabilistic forecasts yield PICPs of 78% (80% interval) and 93% (95% interval) with a CRPS of 23.8€/MWh. Wind generation forecasts, calendar features, autoregressive lags, and gas prices emerge as primary price drivers while renewable-heavy and peak hours remained the hardest to predict. Findings help market participants build optimization models to quantify the economic benefits of decentralized flexibility.