How scoring works

Every forecast is compared with the official day-ahead prices, period by period, once they are published. The standard error metrics are always shown, and the leaderboard ranks by them today.

How models are ranked today

Pick a market and a resolution on the leaderboard to compare like with like. With all markets or all resolutions selected, a model's results across them are counted together. The leaderboard orders by MAE, lowest first, with RMSE breaking ties; equal MAE and RMSE share a place. Over several days each day counts by its number of periods, so a 25-hour day counts a little more than a 23-hour one. Only scored forecasts count: replaced, cancelled and pending versions never do.

The benchmark

Every forecast is also compared with a simple, public benchmark in the same market and resolution:

• Tuesday to Friday: the same period's price the previous day.
• Monday, Saturday and Sunday: the same period's price on the same weekday last week.
• On a national holiday, the most recent Sunday that was not a holiday.

Periods are matched by local clock time. When the clocks go back, both 02:00 hours use the reference day's single 02:00; a period the reference day skipped in spring has no benchmark, so the relative metrics stay empty for that forecast.

The standard metrics

MAE

Mean absolute error: the average size of a miss, in €/MWh. The leaderboard ranks by it, lowest first.

RMSE

Root mean squared error: like MAE, but large misses weigh more. It breaks ties in MAE.

Bias

The average of forecast minus actual. Positive means the forecast ran high, negative that it ran low.

MedAE

Median absolute error: the middle miss, which a few spikes cannot move.

P90 absolute error

Nine in ten misses are this size or smaller.

Benchmark MAE

The benchmark's own MAE over the same periods.

rMAE

Your MAE divided by the benchmark's. Below 1 beats the benchmark.

Skill Score

1 − rMAE: the share of the benchmark's error you removed. 25 % means a quarter less error.