When everything happens

1The task

What is forecast?

Every quarter-hour of the next delivery day D (96 values; 92 or 100 on clock-change days) for four targets in six zones: day-ahead price (EUR/MWh), wind (onshore plus offshore, MW), solar (MW) and total load (MW) in BE, NL, FR, DE-LU, DK1 and DK2. That is 24 series a day.

When?

From the data public at 11:30 CET/CEST on D-1, locked before the 12:00 SDAC day-ahead gate, exactly when a day-ahead bid is due. A live run that finishes after the gate is never scored.

What may a model see?

  • History up to what was public at the cut-off: prices through the end of D-1, wind, solar and load up to the cut-off minus one hour.
  • Weather forecasts from ICON-EU (DWD), GFS (NOAA) and IFS (ECMWF), only runs published before the cut-off: at 11:30 the 06 UTC ICON-EU and the 00 UTC GFS and IFS runs, at capacity-weighted points per zone.
  • ENTSO-E's day-ahead load forecast, due two hours before the gate (an input for price models).
  • Fuel cost: TTF gas and EU carbon daily closes, each counting from the end of its date.

Everything is frozen at the cut-off and stored with the run, so every forecast can be reproduced. Models that need training are trained outside the gate window, on history that ends before the issue day; the 11:30 run only predicts.

Are forecasts corrected afterwards?

No. A forecast is scored exactly as the model produced it: nothing is clipped, calibrated or post-processed on its behalf.

Why gate-aligned?

Every forecast is one that could have been bid: the same delivery day, the same deadline and the information a trader had at that moment. Benchmarks built on pre-registration, such as TS-Arena, are leak-free too, but forecast each series from its own last published value, so the lead time follows each source's publication delay rather than a market deadline.

2Data and scoring times

Where do the actuals come from?

PriceENTSO-E day-ahead prices (A44), 15-minute SDAC
Wind, solarENTSO-E actual generation per type (A75); gaps filled from Elia Open Data (BE), RTE éCO2mix (FR), then Energy-Charts. NL: NED.nl national estimates. DK wind: Energinet real-time production.
LoadENTSO-E total load (A65 actual); NL: NED.nl gross consumption; DK published hourly, so DK load is scored on hourly averages.

Every snapshot records how many quarter-hours came from each source; the Forecast Explorer shows it.

When is a forecast scored?

  • Price: on D-1, once the day-ahead auction for D has cleared. The cleared price is final.
  • Wind, solar and load: at 07:00 on D+1 if the whole day is published, otherwise once 95% of it is. Until then the explorer shows provisional actuals, which are never scored.
  • Revisions: for three more days the actuals are fetched again; a series that moved by more than 0.5% of its daily energy is replaced and the day scored again.

3Scores

Point forecastsMAE, RMSE and their ratio to the baseline
Quantile forecastsWIS, Winkler 80%, coverage 80%
RankingElo from daily head-to-head races

What is the baseline?

naive_weekly: the same quarter-hour one week earlier. Its quantiles are a normal band around that value, one width all day, set by the typical week-on-week change, so it is an easy baseline for interval scores.

What is Rel. MAE?

Per series (zone and target), all of a model's errors in the period are pooled first, then divided by the baseline's on the same days. Lower is better; below 1 beats the baseline.

What is Rel. RMSE?

The same with squared errors, so large misses (price spikes, wind ramps) weigh more. Lower is better.

What is WIS?

qτ: the forecast's τ-quantile; [l, u]: a central interval; (x)+ = max(x, 0).

The weighted interval score of the 10–90%, 20–80%, 30–70% and 40–60% intervals and the median (K = 4); with these quantiles it equals twice the average pinball loss. It rewards a band that is both narrow and calibrated. Rel. WIS divides it by the baseline's, per series, like Rel. MAE. Lower is better.

What is the Winkler score?

The interval score of the 10–90% interval [l, u]: its width plus ten times any miss outside it. Lower is better.

What is coverage?

A calibrated forecast covers 80%; less means overconfident, more means too wide. A diagnostic, not a ranking. Closest to 80% is best.

How are several series combined?

When the filters select more than one zone or target, the per-series ratios are combined by their mean, median or geometric mean (the Across series switch). The mean is moved by a series where a model is very good or very bad, the median says whether it wins on most series, the geometric mean treats halving and doubling alike.

What is the Elo score?

A race is one issue day and one series; in each race every pair of models plays one game, won by the lower Rel. MAE (a tie splits the game). All games of the period are fitted at once with a Bradley–Terry model, as on the Chatbot Arena leaderboard, and shown on the Elo scale: 1000 is the field's average, and 400 points more means 10:1 odds of winning a game. A win over a strong model counts more than a win over a weak one, and the order of the races does not matter. Higher is better.

What does the 95% interval mean?

How much the Elo score would move with a different set of days: the fit is repeated 200 times on issue days drawn with replacement (whole days, since all races of a day share the weather). Overlapping intervals mean the order of two models is not settled yet. On the leaderboard it is the whisker on each Elo bar.

What is the win rate?

The average win rate of fev-bench: the share of head-to-head games a model won, over the same races and games as the Elo score (a tie counts half). Unlike the Elo score, every opponent counts the same. Its 95% interval comes from the same 200 resamples of issue days. Higher is better.

What is the skill score?

The skill score of fev-bench: 1 minus the geometric mean, over the model's races, of its error divided by the baseline's (naive_weekly, the seasonal naive) in the same race, each ratio clipped to 0.01–100. 0% matches the baseline; 30% means 30% less error on a typical race; below 0% is worse than it. Every race counts the same, and the geometric mean keeps a single bad day from dominating. Higher is better.

What is the average rank?

In every series the models are ranked by the day's error, ties sharing the place; ranks are averaged over the series, then over the days. Lower is better.

4Leaderboard

What does Rated over mean?

Which days count: the live runs of the last 7 days (the default) or the last 30 days up to the latest scored issue day, all live runs, or the backtest. The charts and the rank-stability table follow the same choice.

What is the backtest?

Every model re-run on past issue days (from 14 Sep 2026) with the data as it was public at their cut-off: the same history rules and the archived weather runs. One difference: the load forecast is today's published version, which may include later revisions. The Forecast Explorer shows live runs only.

What happens when a run fails?

The model sits out that race; a failure is never counted as a loss, but the run tracker shows it. Models with fewer than 12 races, or a fifth of the most any model has, are provisional and listed after the ranked ones. Under All targets, a model with no scored race yet in the period on one of the targets it forecasts (a newcomer whose wind, solar and load actuals are not in yet, say) is left out until it has one, so it is never rated on its best target alone.

What are TSO_cutoff and TSO_final?

  • TSO_cutoff, ranked only where it covers every selected series: the grid operators' own forecasts as they stood before the gate. Under All targets, or a zone or target it forecasts only in part, it has no Elo score and is listed below the table with its errors on its own series, since a rating on fewer series than the rest of the field would not be comparable. BE wind and solar: Elia's 11:00 forecast; NL: NED.nl's forecast fetched at the cut-off; load in every zone: the TSO's day-ahead load forecast. FR, DE-LU, DK1 and DK2 wind and solar have none, because their TSOs publish only at 16:15 or 18:00, after the gate.
  • TSO_final, a reference, never ranked: the final day-ahead wind and solar forecasts published after the gate (Elia 18:00, RTE 16:15, ENTSO-E for DE-LU and DK, NED in the evening for NL).

What are the ensembles?

Three entries, marked ens, that combine other models' forecasts of the same issue day: per quarter-hour, the median of the members' point forecasts and, separately, of each quantile (then sorted).

  • ens_fm: five history-only foundation models (chronos2, timesfm3, t0_gamma, tirex2, timestral).
  • ens_fm_cov: their native-covariate variants (chronos2_cov, timesfm3_cov, t0_gamma_cov, tirex2_cov, timestral_cov).
  • ens_hybrid: those five plus lgbm_prod, a gradient-boosted model of a different kind trained on the same weather.

The members are the newest model of each family that has a native-covariate variant, fixed by that rule and not picked by score; the median keeps one member's bust from moving the result. A series is combined when all members but at most one delivered it. Ensembles use only forecasts made before the gate and run in the same live run, after every other model.

Excluded by default. With the Ensembles switch on Exclude, they also leave the races, so every other model's score is what it would be without them; Include adds them to the field. The Forecast Explorer and the Forecast Value page always show them.

Added on 19 Oct 2026 and computed back from the members' stored pre-gate forecasts: live record from 16 Oct, backtest on every backtest day.

When is a model relegated?

The live run has half an hour, from the 11:30 cut-off to the 12:00 gate. To keep it inside, a model that costs at least 30 s per run stops running live when a model with the same kind of input (history only, or covariates) that costs at most half as much beats it clearly over the last 28 live days: pooled Rel. MAE at least 10% better and a lower MAE in at least 60% of the races both ran. Both need 10 scored live days. Light models are never relegated, however they score. Baselines, the grid operators' forecasts, the ensembles and their members, lgbm_cov (the weather-driven gradient-boosting benchmark) and the last live model of each family are never relegated, so every model family stays in the live run. The check runs weekly. A relegated model keeps its live record and backtest on the site, with the date and the reason under Models in the field.

What do the filters change?

Zone and target choose which races count. The Covariates switch only chooses which models are shown; positions stay those of the whole field. The Ensembles switch is the one filter that changes the field: on Exclude (the default) the ensembles leave the races as well.

5Forecast value

What does the Forecast Value page measure?

What a wind or solar forecast earns the producer who sells on it. A virtual 100 MW asset in each zone produces like the zone's whole fleet, scaled down by installed capacity, so every zone counts equally. Before the gate it sells a fixed percentile of a model's forecast at the day-ahead price.

ct: quantity sold day-ahead; yt: actual output; pDAt: day-ahead price; costt: what the error costs under the chosen settlement.

How is the error settled?

  • Error as missed income at DA (default): each MWh off costs |imbalance price − day-ahead price|, whichever way it goes, so an error never earns and the score is at most 100%.
  • Error at imbalance price: the error is paid as the zone would pay it, a surplus at the long price, a shortfall at the short price (two prices in NL and FR, one elsewhere; German reBAP). An error can then earn when the price sits on its side, which over a few weeks rests on a handful of spikes.

The producer is a price taker, cannot curtail and does not trade intraday. Percentiles between the published deciles are interpolated; a quantity below zero is sold as zero. Imbalance prices: ENTSO-E 17.1.G.

6Known caveats

NL: national estimates from NED.nl

NL wind, solar and load are NED.nl's national series (Nationaal Energie Dashboard, backed by TenneT), not ENTSO-E's: TenneT reports to ENTSO-E only the plants it meters, about 1% of the solar fleet, and a load net of rooftop solar (about 6 GW at a sunny midday against 17 GW gross). NED upscales monitored systems to the whole fleet, so it is an estimate rather than a meter reading. Offshore wind is NED's metered output from 13 Oct 2026; earlier days keep their scores on NED's park model, which ran 7–27% higher. Live NL forecasts made from the ENTSO-E series (wind and solar up to 10 Oct, load up to 11 Oct 2026) are voided: shown in the run tracker, never scored.

DK: Energinet wind, ENTSO-E load

DK1 and DK2 wind are Energinet's real-time production from 18 Oct 2026: since July 2026 ENTSO-E's DK1 offshore series has left out 0.3–0.5 GW of new parks. Earlier days keep their ENTSO-E scores; live DK wind forecasts of 17 and 18 Oct 2026 are voided. Danish solar stays on ENTSO-E, which matches Energinet's settlement including self-consumption. Danish load stays on ENTSO-E (A65, hourly) and is scored on hourly averages. That is the quantity Energinet's own day-ahead load forecast, the grid-operator benchmark, predicts: in 2026 it is 1.00 times that series, with an hourly error of 40–89 MW in DK1, against 226–287 MW to Energinet's real-time production plus net imports. Since 10 Oct 2026, Energinet has sent ENTSO-E only 12–20 of the 24 hourly values a day, so Danish load days stay unscored until those gaps close.

FR and DK solar at night

The French and Danish solar actuals carry a baseline through the night (about 200 MW in FR, 25–78 MW in DK1 and DK2), where BE, NL and DE-LU are zero. It is a defect of the feeds, left untouched; it slightly penalises forecasts that correctly go to zero.

Load forecast as an input

ENTSO-E's load forecast is an input for price models only since issue day 16 Oct 2026, so load models compete with the grid operators rather than learn from them. TenneT's NL load forecast is net of embedded solar, unlike the gross NL actuals.

History and data sources before October 2026

Issue days up to 3 Oct 2026 used Energy-Charts only, and wind and solar were scored on D+2 (later changes stayed below 0.75%). DE-LU, DK1 and DK2 run live from 15 Oct 2026 and are backtested like the other zones; their weather points come from the Marktstammdatenregister and Energinet's capacity per municipality.

Where the weather is sampled

Models in the field