Multi-Day Forecasting of Day-Ahead Electricity Prices under Input Forecast Uncertainty

Electricity price forecasts are increasingly important for trading and procurement decisions in European power markets with growing renewable generation and increasing price volatility. When forecasting prices for future delivery periods, key drivers such as load and renewable generation are only available as forecasts rather than realised values.This paper investigates how the use of realised versus forecasted inputs in model training and prediction affects multi-day-ahead forecasts of hourly German day-ahead market prices. A feed-forward neural network is evaluated in a rolling training framework under three input configurations: training on realised inputs and predicting with forecasted inputs, training and predicting with forecasted inputs, and an ex post perfect-input benchmark using realised inputs at prediction time. Forecast errors increase with prediction horizon, and input forecast errors account for a substantial share of overall prediction error. Training on forecasted inputs does not systematically improve forecasting accuracy in this setting.