Evaluating player performance in ice hockey is a long-standing challenge because of the high complexity of the sport and limitations of existing metrics. Traditional statistics, such as goals and assists, provide a partial view of performance, while advanced statistics like Corsi or expected goals remain difficult to interpret and often fail to isolate individual contributions. External ratings, such as those in the EA Sports NHL series, offer standardized measures but lack transparency in how scores are obtained. The aim of this thesis is to investigate whether it is possible to develop an interpretable scoring system for ice hockey players based on season-level statistics, designed to predict standardized player ratings and identify which features contribute most to these predictions. Data from the 2024-2025 season was collected from Eliteprospects and Moneypuck, with ratings collected automatically from NHLratings.net. Additional features, including per-game and per-60 statistics, shooting percentages, and z-scores, were engineered. Machine learning models, including linear regression, Ridge regression, Lasso Regression, random forest, and gradient boosting were applied together with interpretability analysis using feature importance analysis. The results show that season-level statistics alone can approximate standardized player ratings with reasonable accuracy, with offensive production and ice time consistently emerging as the most influential features, while defensive contributions were less well captured by the available metrics.
Mathias Olsson (Thu,) studied this question.