Applying SMOTE to Machine Learning Models for NBA Player Performance Evaluation

Abstract

The most critical events in sports analytics, such as championship winners and player awards, are necessarily rare; therefore, attempts to generate predictive models have not been successful. Data imputation algorithms such as the synthetic minority oversampling technique (SMOTE) can successfully generate samples in contexts where data is limited. In this study, I applied SMOTE to create samples to predict the Most Valuable Player (MVP) award in the National Basketball Association (NBA), improving model accuracy by up to 27%. My work illustrates that ensemble learning methods, which combine multiple smaller models to form larger models, generate superior predictive performance: improving accuracy by a further 7%. This work suggests that imputation techniques and ensemble learning methods could predict rare sporting events beyond MVP prediction, such as World Cup winners and Olympic medalists. These events are rare because of the frequency in which they happen, such as every four years for the World Cup and Olympics.