Held-out backtest · Final September forecast shown against observed demand · Five days of actuals · Axes: UTC / MW.
* September 2026 held-out test, 30 daily origins, 0–24h. Model settings selected using 12 monthly expanding validation folds. Bands use July residuals, before the test month. Added value is the percentage reduction in mean absolute error (MAE) versus the weighted seasonal baseline: 100 × (1 − model MAE / baseline MAE), using matched forecasts and actuals. Positive means less error; negative means more. These scores describe the backtest, not the future outlook.