Design decisions
Absolute error over squared error
Section titled “Absolute error over squared error”Electricity prices are skewed: most of the time they sit in a moderate range, and occasionally they spike far above it. A model trained to minimise squared error is pulled hard by those spikes. Four hours with errors of 2, 3, 1.5 and 50 €/MWh give an average absolute error of 14 but an average squared error of 629, almost all of it from the one spike. Training on squared error would sacrifice accuracy on ordinary hours to chase rare ones.
EPF trains on an absolute-error objective. It aims at the middle of the likely outcomes, stays stable when spikes occur, and its errors read directly in €/MWh. Spikes still matter, so they are tracked separately with spike recall and spread capture.
Direct forecasting over recursive
Section titled “Direct forecasting over recursive”Feeding a model’s own predictions back as inputs lets errors compound across a seven-day horizon. EPF predicts each horizon directly from what is known at forecast time, so an error on one day cannot contaminate the next. The cost is training several models instead of one, which is small. See Forecasting approach.
Gradient-boosted trees
Section titled “Gradient-boosted trees”Gradient-boosted trees work well on tabular data that mixes very different kinds of input (prices, calendars, weather, fuel), handle missing inputs without imputation, train quickly, and can be inspected. Each market runs one model family per product. More complex architectures were tested; none beat this one fairly enough to justify the extra complexity and risk in production.
Native 15-minute resolution
Section titled “Native 15-minute resolution”The market clears in quarter-hours, so EPF forecasts in quarter-hours. Splitting an hourly forecast into four equal parts would miss the fast moves around sunrise and sunset that solar now creates, and that is where much of the value for flexible assets lies.
Two products instead of one
Section titled “Two products instead of one”What is known at 10:00 UTC and at 15:00 UTC is different: by the afternoon, tomorrow’s prices are public. One model serving both moments would be wrong for one of them. Two products, each with its own model trained for its own moment, avoid that.
Naive first
Section titled “Naive first”Every forecast is scored next to three naive rules, everywhere it is shown, in every market. A model change is judged on a year of walk-forward testing and on wins day by day, not on one favourable window, so the comparison shows exactly where EPF adds value.
Uncertainty from measured error
Section titled “Uncertainty from measured error”The band in the app comes from the forecast’s own measured error at each time of day, not from a textbook distribution. Uncertainty is reported the way it was observed.