Document details

Whether electricity price forecasts generate operational value in industrial settings — beyond improving predictive accuracy — remains an open question, particularly under extended 168-hour horizons where forecast skill degrades, and tariff-facing decisions must be made against a plausible fixed-tariff baseline. This work evaluated a leakage-safe hybrid model of OMIE gross hourly prices in a backtest of 15 audited Portuguese weeks, balanced across low-, average-, and high-volatility regimes (5/5/5), and compared a fixed BAU tariff (0.0829 €/kWh) with a forecast-informed dynamic hourly tariff, under fixed production conditions. The electricity price forecast is assessed with MAE, RMSE, bias, empirical 90% interval coverage, and Diebold-Mariano tests against naïve lag-24 and lag-168 persistence. Then, the forecast's cost ranking quality was assessed, along with realized proxy-cost gains and downside risk. The results show the electricity price model with mean MAEs of 24.1, 26.4, and 43.0 €/MWh across low-, average-, and high-volatility weeks, and empirical coverage of 66.4%, 58.5%, and 39.8%. The model beat lag-24 persistence in all regimes (p ≤ 0.001) but beat lag-168 persistence only in average-volatility weeks (p < 0.001). However, the broad tariff ordering remained stable (Spearman ρ 0.67-0.69, pairwise accuracy 0.747-0.749). Mean realized gains relative to BAU were €559.3, €458.0, and €311.2, with positive outcomes in 60%, 80%, and 80% of weeks, and underperformance probabilities of 40%, 20%, and 20%. The results indicate a conditional decision value driven more by tariff-order preservation than by point accuracy alone. This study contributes an applied, regime-aware backtesting framework that evaluates forecast value under bounded industrial flexibility — a setting underrepresented in the existing forecast-value literature, which predominantly addresses trading and market-operator contexts.