Goalie model performance

Starter accuracy

How often our top pick actually started. Each team counts separately, so one NHL game can produce two results.

Contested starts

Accuracy when there was more than one eligible goalie to choose from. This leaves out cases labeled “Only eligible candidate.”

Recent accuracy

How often we were right across the latest 20 evaluated team-games, or however many are available.

Log loss and Brier score

These measure the quality of our percentages, beyond simply picking the right goalie. Lower is better for both; zero would mean perfect predictions.

Log loss strongly penalizes giving the actual starter a very low chance. Brier score measures how far all the goalie chances were from what happened. For example, a wrong pick made with 95% confidence is penalized more than a wrong pick made with 55% confidence. These scores are not percentages and use different scales.

Confidence versus results

If we give a goalie about a 70% chance, those selections should be right about 70% of the time across many games. The paired bars help you check that. A small number of games is too little to judge reliably.

The simple choice

Our comparison picks the eligible goalie who had started most often for that team earlier in the season. Both approaches are scored on the same team-games.

Which predictions count?

We use the last forecast saved before the scheduled game start. Games without a usable forecast or a clear recorded starter are listed in coverage but left out of accuracy. If the actual starter was missing from our candidate list, it counts as a miss. All forecast times are shown in Winnipeg time.