Hybrid Electricity Price Forecasting in Swedish Bidding Zones

Sammanfattning

Electricity prices spike sharply, cluster in volatility, and respond to weather, demand, and grid conditions in ways no single framework captures cleanly. This thesis examines whether machine-learning residual corrections add forecast value on top of a transparent GARCH-X baseline across Sweden's four bidding zones (SE1-SE4). Using hourly ENTSO-E, SMHI, and SCB data for 2024-2025, we evaluate GARCH-X and three residual learners (XGBoost, Random Forest, LSTM) in rolling 90-day out-of-sample backtests over the full 2025 calendar year at the intraday (h=1) and day-ahead (h=24) horizons. The findings are conditional on a constrained exogenous-information design. Moving from a plain GARCH(1,1) benchmark to GARCH-X, with autoregressive lags and exogenous fundamentals, substantially improves accuracy in all four zones. Beyond that, machine-learning corrections add value selectively. At h=1, the best residual-learning hybrid significantly improves on GARCH-X in every zone under Newey-West-adjusted Diebold-Mariano comparisons, with the largest gains from XGBoost in SE3 and SE4 and smaller gains from Random Forest in SE1 and XGBoost in SE2. At h=24, ARX is best in SE1-SE3 while Random Forest is best in SE4; only the SE4 day-ahead gain is statistically significant. The practical message is diagnostic rather than universal: GARCH-X is the main econometric baseline, and residual-learning overlays are useful only where it leaves stable, forecastable residual structure. The findings offer actionable model guidance for producers, traders, and system operators operating across Sweden's structurally distinct bidding zones.