Multi-Horizon Explainability in Energy Price Forecasting: Disentangling Market Dynamics and Renewable Intermittency in Romania’s Liberalized Electricity Market

Accurate electricity price forecasting is essential for participants in liberalized electricity markets, yet the black-box nature of machine learning models limits their practical adoption. While explainable artificial intelligence methods such as SHAP (SHapley Additive exPlanations) address interpretability concerns, existing applications present global feature importance rankings that ignore how price drivers vary across forecast horizons and market conditions. This study introduces two novel contributions: a multi-horizon explainability framework analyzing feature importance across immediate (0–6 hours), short-term (6–24 hours), medium- term (1–7 days), and long-term (7–14+ days) horizons, and a market regime analysis examining how explanations shift across low-price, normal, and high-price conditions. We apply Random Forest and XGBoost models to hourly Romanian electricity market data spanning January 2018 to September 2024 (59,160 observations), with features engineered from data encompassing the 2022–2023 energy crisis and summer 2024 heat wave. The selected Random Forest model achieves an R2 of 0.60 with a mean absolute error of 31.04 EUR/MWh on the test set. SHAP analysis reveals that short-term features contribute 41 percent of horizon-attributed predictive contribution, substantially exceeding other horizons. Market regime analysis demonstrates that price momentum becomes 4.25 times more important during high-price events compared to nor- mal conditions (p < 0.001), validated through Mann–Whitney U tests. These findings provide targeted guidance for market participants and offer quantitative signals for identifying market stress periods requiring enhanced risk management.