How accurate is the model?
Scored with log loss, which punishes confident misses hardest. Lower is better; a 50/50 guess scores 0.693. Each market is compared with a simple baseline that knows nothing about the teams.
This season · live
No finished games recorded yet. Live results are added the morning after each game, scoring only previews locked at puck drop.
Backtest · 2025–26 season
Every game of the 2025–26 regular season was replayed in order. Before each game day, team ratings were rebuilt from only the games already played, exactly as the live site would have seen them, and each game was simulated 5,000 times.
Log loss by market
1,312 games. Skill = how much lower the model's log loss is than the baseline's. Baselines use only results from earlier in the season: the home-win rate, the regulation win/tie/loss rates, and the share of games going over 5.5.
984 games. Skill = how much lower the model's log loss is than the baseline's. Baselines use only results from earlier in the season: the home-win rate, the regulation win/tie/loss rates, and the share of games going over 5.5.
Calibration
When the model gives the home team a certain chance, how often does it actually win? Bars are what happened; the orange mark is what the model said. Close together = well calibrated.
Month by month
Moneyline skill vs the home-win baseline. Early months are weaker because ratings are built from only a few games.
What the backtest can't show. Historical starting goalies aren't in the data, so goaltending uses each team's season-to-date saves above expected instead of the actual starter. It also isn't compared with betting odds, which is the real benchmark for any model: matching the closing line is hard, beating it is rare.