Energy
Accuracy, shown as a baseline
Each complete UTC hour is predicted from the price exactly 24 hours earlier. Incomplete hours are excluded, not filled. The score uses rolling daily origins and completed delivery hours; historical auction revisions are disclosed.
Backend evaluation · LV
Loading backend data...
Forecast accuracy
Walk-forward backtest since 2022: each model is refitted every quarter on data that was public at the time, then scored on the following months. Lower error is better; skill above 0% means fewer errors than repeating the last known day.
| Model | Mean abs. error | Skill vs persistence | Quantile loss | In 80% range | Cheapest 3 h found | Hours |
|---|
Live track record
Only forecasts frozen before the outcome existed. This is the evidence that cannot be fitted to the past.
No frozen forecast has reached its delivery hour yet. Scores appear the day after the first forecast.
What persistence means
Each complete UTC hour is predicted from the price exactly 24 hours earlier. Incomplete hours are excluded, not filled. The score uses rolling daily origins and completed delivery hours; historical auction revisions are disclosed.
What comes next
The next model is gradient-boosted, with calendar, weather and lagged prices, then an archive of every forecast we stored. This baseline stays on the page so the newer number has to beat something real.