How often the band holds

D+1 and D+2 are not the same thing

D+1 is produced after tomorrow's auction closes — the day-ahead price is already set, so it is validation, not a decision. D+2 is issued while the auction for the day after tomorrow has not been held yet: it is the prediction you actually operate on. When the two diverge, D+2 is the figure that matters, and it is the harder one to hit.

What does not hold yet

The D+2 band covers below the 80% target on Bulgaria (BG), 61.9%, Greece (GR), 36.8%, Hungary (HU), 46.5% and Romania (RO), 67.0% — but the headline figure hides something more specific than “too narrow”, and we say so. Broken down by hour over 45 days, the miss is not symmetric: in the evening the band misses almost entirely upward (hour 18, Romania: 26.7% of slots above P90, against 10% expected), while the lower edge catches nothing at all — at hours 21 and 22, 0.0% of prices fell below P10, where the definition calls for 10%. At midday it is the opposite: both edges miss, so there the band genuinely is too tight.

This matters because it changes the fix. Today's correction adds the same amount to both sides; in the evening hours everything added below is wasted, and what is added above is not enough. A wider band would make the ratio worse, not better. We are working on a correction that calibrates the two edges separately, per hour. Until then: anyone sizing risk on P10–P90 at D+2 underestimates the upper tail at peak hours, and pays for a lower tail that does not materialise overnight.

One of them sits below target for a different reason than the others, and it is worth saying: in Greece (GR), renewables cover 53.7% of demand on average, against 22.1% across the other zones. The model rests on residual load — demand minus renewables — and at that penetration it becomes the difference between two large numbers: the same absolute error in the inputs produces a far larger relative error in the output. In 15.7% of slots renewables actually exceed forecast demand, so residual load turns negative — a regime the training history barely contains (in RO, 0.4%). This is not a tuning problem, it is a property of the input; that is why we do not fix it with a recalibration, and do not hide it until we do.

Coverage, month by month

The whole scored history, with no periods removed. February 2026 is on the chart exactly like July: publish only the good months and you have not published a track record.

Frozen forecasts

Every row here was frozen in the database before the answer was known, then scored once the realised price arrived. Nobody can touch them any more — not even us. This is the evidence the page stands on.

P50→P90 and P10→P50 are the band's two margins, each as a percentage of the period's average price. They sit next to coverage because coverage alone can always be raised by widening the band — a wide enough one covers 100% and says nothing. And they sit apart, not summed: when the miss is one-directional, the right correction moves one margin only. On a single total, a targeted repair and lazy widening look identical. Here you see which one moved. A “—” means that edition predates margin measurement; we do not backfill it where it cannot be recomputed.

Series marked “thin sample” have fewer than 60 scored days. We publish them because they exist, not because they support a conclusion.

Validation, not forecast

The same days, but the forecast was issued in the evening, after the auction closed — when the next day's price was already published. It could not be traded on, and we do not call it a forecast. It is useful to know how well the model reads a day it already almost has in front of it, but that is a question about the model, not an instrument.

And how it would have behaved over history

A reconstruction: for each past day, a model trained strictly on the days before it, compared against the realised price. The training is causal — but the inputs are not, and that changes how the figure must be read. The model reads weather from our database, and there the realised weather archive has since replaced the forecast that was available at the time. The reconstruction therefore “forecasts” a day whose weather it knows.

What that advantage is worth, measured on pairs: for each series below, exactly the days that also have a frozen forecast, reconstructed now and compared one by one. It is not seasonality — the period is identical by construction. Days on which the frozen forecast ran on a different model from the one the reconstruction picks today are excluded, not blended in.

Measured on 2026-08-28. Recomputed as the series grows; never typed by hand.

The two right-hand columns are read together. When the mean advantage is positive but the reconstruction covers better on fewer than half the days, the advantage sits in a handful of days rather than in ordinary behaviour — and the mean alone would say otherwise. Part of the gap also comes from code, since the frozen forecasts ran on the version of the day; we cannot yet separate the two causes, and we do not claim to.

Plainly: the figure below does not predict where the product will land. It stays published because it is the only measurement covering a full year — including February 2026, the worst month in the whole record — and because a hidden backtest is worse than an explained one. It does not compare row by row with the tables above.

How it is measured

• The truth is the day-ahead price as reported on the ENTSO-E Transparency Platform, in each market's own currency; for Romania, converted to RON at the official National Bank of Romania rate. We do not compare a forecast against another forecast.
• Coverage is the fraction of 15-minute intervals where the realised price fell between P10 and P90. The target, by construction, is 80%.
• rMAE is the mean absolute error on P50, relative to the day's mean realised price. A high rMAE in a low-price month is not necessarily a bad month.
• The model re-trains from scratch every day, with no accumulated weights. There is no moment at which “the model was calibrated”; there is only how it behaves, day after day.