Who’s In by Game 15? Multi-Model Per-Game Forecasting of NHL Playoff Qualification

Abstract

Most published NHL playoff forecasting systems operate at pre-season or season-end resolution, yet the practical question is not who qualifies but when a forecast becomes trustworthy. We retrain nine heterogeneous classifiers and a uniform-mean ensemble at all 82 regular-season checkpoints, across 14 full seasons from a 40,600 team-game corpus (2008–09 through 2024–25) with 632 cumulative features. Tree-based learners reach ROC-AUC near 0.78–0.82 by game 10, exceed 0.90 by mid-season, and identify 13 of 16 playoff teams by game 15; the tuned random forest peaks at 0.970, while aggressive ℓ₁ regularization collapses the tuned logistic regression to a degenerate flat predictor. Against Win% and Pts% baselines the lift concentrates at 20 of 82 checkpoints. In a per-game betting test, no fixed model or public-metric picker profits over the season, but a Win%-anchored hybrid that resolves the contested bubble seats with the ensemble returns a 0.9% edge against a 7.1% best-model ceiling. The right model is a function of the game checkpoint.