All methods
Win Probability
Validated modelTested on data the model never fit, with the accuracy published.
What it means
The chance the home team wins from this exact game state.
How it's computed
Gradient-boosted model over score, time remaining and manpower.
What's validated - and what's not
Validated on held-out games it never trained on: log-loss 0.504 vs 0.688 for guessing the base rate, with a published calibration table.
Coverage
Trained on 19 seasons of game states.
Provenance
Framing after Bernier, “Forecasting Real-Time Win Probability in NHL Games” (2018).
Our own gradient-boosted model, validated on held-out games (log-loss 0.504 vs 0.688 baseline).
Our research ledger - every source we tested, adopted, or rejected →Tiers are a promise, not a style. Reconciled is checked against the official NHL record; Validated is tested on data the model never fit; Experimental is a derived construct without out-of-sample validation - read it as informed opinion. When in doubt we take the lower tier.