Beyond the Boxscore: Expected Goals Modeling, Latent Player Archetypes, and Lineup Optimization in Ice Hockey

Abstract

This project develops a data-driven framework for evaluating player performance and optimizing hockey lineups by integrating expected goals modeling, player archetyping, and lineup optimization. Using player- and shot-level data from recent NHL seasons, it estimates expected goals through generalized additive and gradient boosting models, then applies principal component analysis and clustering to identify player archetypes and stylistic similarities across the league. These results are incorporated into an optimization framework that evaluates whether lineups with a broader distribution of roles can improve lineup construction despite limited data on defensemen and on-ice player interactions. The results show that a multi-factor approach combining a spatial baseline with a context-boosted interaction model provides stronger explanatory power than traditional single-metric methods. The analysis also finds that dimensionality reduction and clustering reveal meaningful player roles not fully captured by standard box-score statistics. Although the archetyping process exposed important gaps in available defensemen metrics relative to scoring-based measures, the optimization model still produced reasonably successful line combinations, providing a more systematic and interpretable framework for roster evaluation and lineup construction in hockey.