Prediction of NHL Game Outcomes Using Machine Learning Methods

Due to the continuous nature of ice hockey and a high degree of randomness, predictive modeling of NHL game outcomes presents a specific analytical challenge. The main objective of the thesis is to design, train, and optimize machine learning models capable of predicting home or away wins with the highest possible accuracy, and subsequently verify the practical applicability of the winning algorithm in a simulated betting environment. The input dataset covering the 2019/20 to 2024/25 seasons was extracted from the official NHL API. From the original broad dataset enriched with the Elo rating system, a final set of 35 key predictors was defined using a correlation matrix and feature selection algorithms. Within the scope of the experiments, Logistic Regression, Random Forest, XGBoost, and neural network models were compared. Logistic Regression emerged as the most robust model in this environment. It achieved an accuracy of 60.60 % on unseen data from the test season, thereby successfully outperforming the comparative baseline. Finally, the practical potential of the winning model was confirmed by a betting simulation guided by a defensive Kelly criterion, in which the algorithm achieved a return on investment of 30.75 %.