Each group submitted a point forecast and 19 quantiles (5%, 10%, …, 95%) for the nine days 23–31 October and for the October average. Lower scores are better.
CRPS is the average over the 19 quantiles of the quantile score (the sum of the quantile scores divided by 20): it rewards forecasts that are both close to the outcome and honest about their uncertainty. RMSE is the root mean squared error of the point forecasts. Bias is the average forecast error (positive: the forecasts were too high). 90% hits is how many outcomes fell inside the group's 5%–95% interval (about 9 in 10 is the ideal).
The official rule is that the last submission before the deadline (22 October, 12:00) counts, and the winner is the group with the best combined rank in four categories: daily RMSE, daily CRPS, and the absolute error and CRPS of the October average. This page ranks on the first two while the month is running, using the days known so far. Best submission of each group picks, in hindsight, the submission with the lowest CRPS so far. It is for fun and learning (it shows what the earlier forecasts were worth), and it is not the competition result.
The benchmarks (naive, seasonal naive and 7-day mean) are made on the deadline day with the same data, so you can see who beat them.
The models are re-implemented from the groups' reports (the final model they chose: transformation, regressors, ARIMA/ETS family, length of the training window and the bootstrap simulation of the paths). Run on the prices the group had at the deadline, fifteen of the eighteen groups reproduce their submitted forecast to about two decimals (four of them had prices only to 21 October). The three groups that use weather forecasts have the weather rebuilt from archived Open-Meteo forecasts, so those are approximate: two come close to the submission, one (Gruppe 10) less so.
For a date d the model is fitted on prices up to and including d (a price is public the afternoon before), the remaining days are simulated with resampled residuals (1,000 paths), and the October average of each path is the average of the known days and the simulated days. The grey benchmark is the seasonal naive method. The submitted forecasts that count in the competition are on the Scoreboard tab; this tab shows how the same models would have behaved on every other day.