Forensic Hockey
Forensic Takes
The most striking things our models found this week - who's riding luck, who got robbed, which goalie is running hot, the flukiest Cups. Auto-surfaced and ranked by punch. Tap any card to dig into the numbers.
No model can pick a single NHL game much better than ~64%
The spread of team win% is the tell: across 19 seasons it sits at a standard deviation of just 0.098 around .500 — the salary cap squeezes the talent so tight that one night is mostly chance. Simulating the league shows it behaves like a sport where only ~28% of games are decided by skill and ~72% by a coin flip, so even a perfect forecaster tops out near 64%. And the sliver that IS predictable is almost all context, not analytics: just picking the home team is right about 54% of the time — a third of the way to that ceiling — and 'which team is the better club' gets you most of the rest, while shot share and the fancier metrics add almost nothing on top. Even clear pre-game favorites only win about 65% of the time, and our own pre-game rating lands around 58% — close to a wall no amount of analytics can climb over. Most of a single game is luck.
Explore the numbers →Takes are generated from our own models - luck-adjusted standings, the goaltending verdict, finishing luck, and the Monte-Carlo deserved bracket - and ranked by how far each is from the norm. They refresh as the season's data updates.