Demo
This forecast explores a data-parsimonious approach to German day-ahead electricity price modeling. Instead of depending on external load or renewable generation forecasts, the model works directly with recent prices, commodity context, calendar features, and raw ECMWF AIFS numerical weather forecasts.
Attention Mechanism
Methodologically, the interesting part is temporal alignment. Price observations, market fundamentals, and weather forecasts live on different grids, with different horizons, publication rhythms, and information sets. The model uses learned time encodings and attention mechanisms to relate discontinuous price history, forecast weather fields, and the target delivery window without forcing every signal into a single hand-engineered interpolation scheme.
Limitations
At this stage, the live forecast is a point-forecast baseline: deliberately lean, inspectable, and useful for exercising the full data-to-forecast pipeline. One area to refine is the averaging and smoothing tendency of the attention mechanism, to capture sharper price movements and leverage autoregressive behaviour more effectively. Another direction I am considering is probabilistic forecasting, for example productionizing an ensemble prototype based on a neural network with a quantile head paired with copulas to model dependency across the forecast horizon.