NBA Most Valuable Player Projection model: Based on Traditional and Advanced Metrics

People often like to make predictions about winning or losing before NBA games, much like predicting a lottery win. So, can the Most Valuable Player (MVP) of each season also be predicted? This study proposes a model that can predict MVP based on machine learning. In this study, Random Forest, LightGBM, XGBoost, CatBoost and NGBoost are mainly tested and compared. The results show that the CatBoost model has a mean square error of 0.3527 and a coefficient of determination of 0.9622 on the test set. Therefore, the CatBoost model is finally selected. Finally, the training features were interpreted by visual analysis and SHAP values. Among them, Value of Replacement of Players (VORP), Playing Percentage and Player Efficiency (PER) have key roles in player value prediction. The model also successfully predicted Nikola Jokić as the MVP player in 2024 NBA to empirically increase the reliability of the model predictions.