Anatomy of a miss: the Sunday our models lost to a free baseline

We sell forecasts and publish their scores every day. On Sunday, July 12, 2026, those scores were bad: of the 42 zones our production models were live-scored in that day, 17 lost to the free naive-persistence baseline — replaying Saturday's curve would have beaten us. Most vendors would never mention a day like this. We think it's the most instructive thing we can publish.

What happened

A price regime shift rolled through the Nordics over the weekend. Bergen (NO5) tells it cleanly — daily average day-ahead prices:

Models trained on hundreds of days in the old regime kept forecasting the old level. Persistence — which simply assumes tomorrow looks like today — got the new level for free, because by Saturday it was already standing in it. That is the structural weakness of every learned model on regime days and the structural strength of the dumbest baseline in the industry. It is exactly why we score against persistence and not against something easier to beat.

The damage, zone by zone

What we changed

Three things, all visible on the public board. First, the Nordic zones moved to newer model generations (volt-4 and the pooled volt-3n family) that carry hydrology and grid drivers — the flip was evidence-gated on walk-forward backtests, not on embarrassment. Second, live records under three days are now labeled burn-in on /accuracy and excluded from headlines: one bad (or lucky) day is noise in either direction. Third, this article — because the honest account of a miss is worth more than ten victory laps, and because if the models lose to the baseline for a month in your zone, that zone costs you nothing. The scorecard that enforces that promise is the same one that produced these numbers.