What matters
Goals are rare, so short-run results are dominated by variance and goaltending. Shot volume and ice time are far more stable. Strength state — even strength, power play, penalty kill, empty net — changes scoring rates sharply within a game.
Important data
| Input | Why it matters |
|---|---|
| Shots on goal and attempts | Stable volume signal |
| Time on ice and power-play role | Opportunity |
| Starting goaltender | A major pregame uncertainty |
| Strength state | The core live condition |
Common metrics
| Metric | What it measures | How to read it |
|---|---|---|
| Shots on goal | Shots requiring a save | Count stat, often Poisson-like |
| Save percentage | Saves / shots faced | Very noisy in small samples |
| Expected goals (xG) | Goal probability per shot | Only meaningful once validated |
| Strength state | 5v5, 5v4, 6v5 | Live scoring-rate switch |
Model concepts
Criteria were frozen before the first fit: beat distance-only and constant baselines on log loss and Brier, stay inside calibration bounds, hold up season by season, show zero leakage findings and score identically in production. Even a pass keeps the model in shadow; shadow models never appear as picks. Danger buckets are labelled as a geometric heuristic, not xG.
Live intelligence
PBE Cast follows score, period, strength and shots. A power play is the biggest live shift in scoring probability short of an empty net.
PropBetEdge NHL tools
Worked example
- A skater averages 3.0 shots on goal per game. If shots were Poisson with λ = 3.0: P(≥3) = 57.7%.
- If his ice time varies a lot game to game, real shots are overdispersed and P(≥3) is lower than the Poisson figure, with more 0–1 shot nights.
Glossary
Related lessons
Poisson
How do you turn "0.35 expected home runs" into a probability — and when is that math wrong?
Overfitting & holdout
How do I know a model’s published numbers are not flattering itself?
Confidence & gates
What do Lean, LIMITED, Insufficient and "gated" each mean?
Game state
What exactly does a live model know that a pregame model does not?
Sample size
How many results before a record means something?