Loading...
When a model says a team has a 65% chance to win, what does that actually mean? Here's how pre-game win probabilities work, what drives them, and how to use them.
TL;DR
Win probability is a model's estimate of each team's chance of winning before the game starts. It's based on team strength, recent form, goaltending, home ice, and schedule context. A 65% probability doesn't mean a guaranteed win. It means if you played that game 100 times, that team wins roughly 65.
Overall rating, power index, and goal differential. The foundation of any prediction — how good is this team relative to the opponent?
Starting goalie's recent form, save percentage, and workload. A hot goalie can swing a game by 5-10 percentage points.
Home teams win roughly 54% of NHL games historically. Last change, familiar surroundings, and crowd energy all contribute.
Back-to-back games, travel distance, and rest days. A team on the second night of a back-to-back is measurably worse.
Performance over the last 10 games — momentum, lineup changes, and whether results match underlying shot quality.
Not all probabilities are created equal. Here's a practical guide to what different ranges actually mean on game night.
Slight edge at best. Could go either way — don't read too much into it.
Meaningful edge, but upsets happen all the time in this range. Think of it as winning 6 out of 10.
The model sees a significant gap between the teams. Still, the underdog wins roughly 3 out of 10.
Rare in hockey. Usually requires a massive talent gap plus favorable schedule and goaltending matchup.
Even a 70% favorite loses 3 out of 10 times. That's not a miss — that's how probability works.
PuckCast assigns a letter grade to every prediction based on how confident the model is in the outcome. Higher grades mean the model sees a larger edge.
Model is very confident. Large gap between the teams across multiple factors. Historically hits at the highest rate.
Solid edge. The favorite is clearly stronger, but the game isn't a foregone conclusion.
Low confidence. Close to a coin flip. The model doesn't see a meaningful separation between the teams.
PuckCast uses an ensemble model trained on 16 seasons of NHL data. It combines logistic regression and gradient boosting to produce calibrated win probabilities for every game.
Calibrated means the probabilities are honest. When the model says 60%, teams in that range actually win about 60% of the time. Predictions are updated daily before puck drop using the latest stats, lineups, and goalie confirmations.
Under the hood
A well-calibrated model will match its stated probabilities over time. If it says 60% for a set of games, roughly 60% of those teams should win. PuckCast's ensemble model has been validated across 16 seasons of NHL data using walk-forward testing, achieving ~62% accuracy on out-of-sample predictions.
Hockey has more parity than almost any other major sport. A single bad bounce, a hot goalie, or an early power play goal can flip a game. Even a 70% favorite loses 3 out of 10 times. That's not a flaw in the model — it's the nature of the sport.
It means the model sees a slight edge for one team, but it's close to a coin flip. If you replayed that game 100 times under identical conditions, the favored team wins about 55 and loses about 45. These are the hardest games to predict and the ones most likely to go either way.
Betting odds include a built-in margin (the vig) that ensures the sportsbook profits regardless of the outcome. A model's win probability is a pure estimate of each team's chance of winning. Odds of -150 imply about 60% probability, but the true implied probability after removing the vig might be closer to 57-58%.