Regular Season Predictors of NHL Playoff Success: An Empirical Analysis Using Advanced Team-Level Statistics

Abstract

The NHL playoffs are among the most unpredictable events in professional sports, where top seeded teams are frequently eliminated early and statistical favorites often fall short. This thesis investigates whether regular season team-level statistics contain meaningful predictive information about playoff success, using betting market odds as a benchmark to evaluate model calibration and predictive accuracy.

Using team-level data from the 2007–08 through 2021–22 NHL seasons, sourced from MoneyPuck and the official NHL website, four modeling frameworks are evaluated out-ofsample: ordinary least squares regression, logistic regression, random forest, and elastic net. Models are trained on historical seasons and tested on future data to simulate a realistic prediction setting. The primary outcome is playoff rounds won, reformulated as a binary indicator of winning at least one playoff round for classification models. Across all models, Fenwick percentage emerges as the most consistent predictor of playoff advancement, with power play percentage and save percentage providing additional signal. However, out-of-sample performance remains modest, with AUC values between 0.63 and 0.70, reflecting the inherent difficulty of predicting outcomes in a low scoring, high variance sport.

A betting simulation applied to first round matchups compares model implied probabilities to sportsbook odds. While some models generate positive returns, these results are not robust and are likely driven by small sample variation. A simple underdog strategy outperforms all models, suggesting potential favorite overpricing in playoff markets. Overall, the results indicate that regular season metrics contain limited predictive signal, while playoff outcomes remain largely driven by variance, goaltending, and matchup dynamics not captured by season-level data.