Feature Importance Analysis To Predict Winning Percentage In The NBA Utilizing Machine Learning

PURPOSE:To determine the period of performance strategy trends and the key indicators for winning NBA games by analyzing times-series data between 2000 and 2023 by applying machine learning. METHODS: A total of 716 NBA team performance records for each season were collected from NBA.com. The data analysis was conducted in three steps. First, a Random Forest Regressor model was simultaneously trained and evaluated by dividing the seasons into 2- to 10-year increments. For feature importance analysis, Random Forest, a representative bagging model that is more stable and robust for overfitting was selected than other boosting models. Second, the model was trained on data divided into the appropriate periods identified in the first step, and feature importance scores were calculated for each period. Feature importance is calculated by tree-based models. If a particular feature is heavily used for segmentation or results in a significant decrease in impurity, it will have a high score. Finally, a Pearson correlation analysis was performed between the top indicators for each period and winning percentage. All analyses were conducted using Python. RESULTS: As a result of the first step of machine learning, the model trained on the data divided into 9 years had the lowest values of minimum and maximum RMSE, respectively. Therefore, the model was trained on the data split into 9 years increments. During the two periods (2000-2005, 2006-2014), two-point field goal percentage was identified as the most influential factor in predicting team performance (2000-2005, 2P% = 0.36; 2006-2014, 2P% = 0.46). In contrast, during the period (2015-2023), three-point field goal percentage had the highest feature importance score (2015-2023, 3P% = 0.38). The Pearson correlation coefficients between the variable with the highest feature importance in each period and the winning percentage ranged from 0.53 to 0.68, indicating a strong correlation. CONCLUSIONS:A Random Forest Regressor model was efficient to determine the optimal periods of NBA performance trends as well as distinguished the key feature importance of winning the game. Supported by: National Research Foundation of Korea[NRF-2021R1A2C2013651]