NBA Machine Learning for Game Outcome Prediction

This thesis studies the problem of predicting National Basketball Association (NBA) gameoutcomes using only information available prior to tip-off. While team-level rating systems such as Elo provide a strong baseline for game prediction, they do not fully account for game-to-game variation in roster composition, player availability, and lineup strength. To address this limitation, this thesis develops a player-informed pre-game prediction framework that combines traditional team Elo with a player-level Elo construction designed to reflect the expected strength of the active rotation. Player ratings are aggregated to the team context using recent information on playing time and contribution, allowing the model to incorporate variation in lineup quality while preserving a strict pre-tip-off constraint. The modeling pipeline is built from publicly available NBA box score data and engineered to avoid look-ahead bias. Several feature sets are compared, including a baseline team Elo logistic regression, an expanded model incorporating player-level Elo, and a final specification that adds rolling box score summaries and contextual features such as team record and scheduling effects. Model assessment is conducted using a walk-forward out-of-sample evaluation scheme that mirrors a realistic deployment setting, with performance measured by accuracy, area under the receiver operating characteristic curve, log loss, and Brier score. The results show that player-level Elo contributes meaningful predictive signal beyond team Elo alone, improving performance across all major evaluation metrics. Further gains are obtained by incorporating a small set of contextual and rolling-performance variables, with the preferred expanded logistic model achieving the strongest overall out-of-sample results. These findings suggest that pre-game NBA prediction can be improved by combining team strength, expected lineup quality, and recent team context in a unified and interpretable framework. The thesis also discusses limitations related to imperfect lineup availability information and the use of aggregate box score data, and it outlines future directions involving richer availability modeling, nonlinear extensions such as generalized additive models, and more granular possession-level or tracking-based data.

Level thesis:degree_level

Master’s Degree

Year dc:date.issued

2026

Author dc:creator

• Bradley, James Benjamin

Advisor dc:contributor.advisor

• Rojas-Gonzalez, Raul

Committee members dc:contributor.committeemember

• Hurtado, Paul
• Tavakkoli, Alireza

• Basketball
• Data Science
• Logistic Regression
• Machine Learning

Language dc:language.iso

en_US, English

Repository record dc:identifier.uri

https://scholarwolf.unr.edu/handle/11714/11854

OAI identifier oai:identifier

oai:scholarwolf.unr.edu:11714/11854