When everything happens
The rules
- 01
One task, one clock
On each issue day D-1, every model forecasts all quarter-hours of delivery day D from data public at 11:30 Brussels time, and the forecast is locked before the 12:00 SDAC day-ahead gate closure, exactly when a day-ahead bid would be due. A live run that finishes after the gate is never scored. The forecast covers 96 values, or 92 and 100 on clock-change days. Targets are day-ahead price (EUR/MWh), wind generation (on- plus offshore, MW) and solar generation (MW) for BE, NL and FR.
- 02
Data: ENTSO-E Transparency Platform
Prices and actual generation come from the ENTSO-E Transparency Platform, parsed directly from its XML: day-ahead prices (document A44, the 15-minute SDAC prices), and actual generation per production type (A75), wind as onshore (B19) plus offshore (B18) and solar (B16). Quarter-hours ENTSO-E has not published are filled from the TSOs' own open data: Elia Open Data for Belgium and RTE éCO2mix for France, then Energy-Charts (Fraunhofer ISE) as a last resort; the Netherlands has no TSO service without registration. ENTSO-E alone lacked 12% of Belgian wind quarter-hours (13 of 27 days below the 95% needed to score) and 2.6% of French wind; Elia and RTE fill all of them and agree with ENTSO-E without bias (mean difference +0.2 MW for Belgian wind). Every snapshot and every set of actuals records how many points came from each source; the explorer and the Inputs table show it. The ENTSO-E day-ahead total load forecast (A65) is one of the covariates. Issue days up to and including 3 Oct 2026 used Energy-Charts only.
- 03
No look-ahead
Each model gets the last 28 days of history, cut off at what was public at issue time: prices through the end of D-1, which the previous day's auction published on D-2; wind and solar up to the cut-off minus one hour of publication lag. The input snapshot is stored with every run, so each forecast can be reproduced exactly.
- 04
Scored on accuracy
Point forecasts are scored with MAE and RMSE; quantile forecasts with pinball loss over the 10th to 90th percentiles. Rel. MAE divides a model's MAE by that of the naive_weekly baseline, the same quarter-hour one week earlier, on the same day and series: below 1 beats the baseline. Prices are scored on D-1 once the day-ahead auction for D has cleared, since the cleared price is the outcome a day-ahead forecast targets; it is final. Wind and solar are scored at 07:00 on D+1 if every quarter-hour of D has been published, otherwise at the next scoring run once 95% has; until then the explorer shows the values metered so far as provisional actuals, which are never scored. For the following three days their actuals are fetched again at every scoring run: every check is logged, and if a series has moved by more than 0.5% of its daily energy, the actuals are replaced and that day is scored again. Before 3 Oct 2026 wind and solar were scored on D+2; actuals frozen then changed by at most 0.75% afterwards.
- 05
Covariates, frozen at the cut-off
Covariate models also get weather forecasts from three independent models, ICON-EU (DWD), GFS (NOAA) and IFS HRES (ECMWF), through the Open-Meteo single-runs archive. For each model only a run published before the cut-off counts (initialisation at most cut-off minus its publication delay plus a margin), which at 11:30 means the 06 UTC ICON-EU run and the 00 UTC GFS and IFS runs. Their training history uses the same rule day by day, so models train on forecasts with the same lead time as the ones they forecast with. Added: ENTSO-E's day-ahead total load forecast, which TSOs must publish two hours before the gate, and fuel cost when configured. Live, all of it is fetched at the cut-off and archived; in the backtest, weather runs come from the same archive, but the load forecast is today's published version, which may include later revisions.
- 06
Known caveats
NL solar as published by TenneT on ENTSO-E covers metered plants only, peaking at a few hundred MW against roughly 25 GW installed, so it is not national output. Models that publish no quantiles have no pinball score and sit out rankings on pinball.