About the model

NHLPredict estimates the probability that the home team wins a regular-season NHL game. The current version uses official NHL schedule/results JSON and a chronological walk-forward Elo rating with a home-ice adjustment. Glicko-2 state is also updated and blended in the prediction helper for uncertainty-aware probabilities.

Data

Ingest pulls regular-season schedule/results from api-web.nhle.com/v1/schedule/{date}, normalizes teams and games into data/raw/nhl/, then writes model artifacts under data/model/nhl/, data/ratings/nhl/, and data/performance.json. It does not bypass paywalls or anti-bot systems, and it does not fabricate missing games.

Model tuning

The first-pass backtest tried a small grid over Elo K-factor, home advantage, and off-season carryover. The selected configuration is K=12, home advantage=35 Elo, and carryover=0.65.

Performance

Walk-forward scoring over 6,560 completed games: Brier 0.2383, log loss 0.6692, directional accuracy 58.6%. Brier/log loss are the primary health checks; accuracy is secondary context.

These are model probabilities for entertainment and analysis, not betting advice.