Retraining Matters More Than the Model: Day-Ahead Price Forecasting for Germany with Grid-Operator

German day-ahead electricity prices are increasingly driven by wind and solar generation,which makes them a natural test case for comparing time-series models with supervisedmodels that use forecasts of the physical drivers. Such comparisons are often made withobserved instead of forecast weather, or with a model trained once and evaluated over aperiod in which the market changes. We compare naïve baselines, Prophet, Ridge regressionand gradient-boosted trees on the year from October 2025 to October 2026 under a strictday-ahead information set, using the grid operators’ day-ahead forecasts of load, wind andsolar generation and treating observed weather only as an upper bound. Gradient boostingwith lagged prices and these forecasts, expressed as shares of the forecast load and retrainedeverytwoweeksonarolling180-daywindow, reachesameanabsoluteerrorof15.98€/MWh,48.6% lower than repeating yesterday’s price (p < 0.001) and lower in every month. Anywindow between 90 and 270 days gives the same result within the uncertainty of a pairedbootstrap (maximum paired difference to the 180-day window of 0.24 €/MWh). Trainedonce, the same model is only 18% better and falls behind the naïve forecast from July 2026;a drift analysis shows that the price response to residual load shifted while the inputs did not.Retraining also reverses the ranking of inputs and algorithms: observed weather, absoluteforecasts and Ridge regression win under a single fit and lose with retraining, after whichperfect weather knowledge adds at most 0.3 €/MWh. TreeSHAP explanations recover aseasonal merit-order curve. Which model ranks highest depends on the training scheme;the training scheme itself is the first-order decision. A supervised model needs time-seriesmemory in its inputs and time-series practice in its training.