Inner pace: A dynamic exploration and analysis of basketball game pace

Abstract

This study aims to investigate the dynamics of basketball game pace and its influence on game outcomes through a novel intra-game segmentation approach. By employing K-means clustering on possession duration, we categorized possessions from 1,141 NBA games in the 2019-2020 season into high-frequency (HFS), low-frequency (LFS), and normal-frequency segments (NFS). A sliding window method was utilized to identify these segments, revealing distinct temporal patterns within games. To analyze the predictive value of these segments, we applied machine learning models, including Random Forest and Light Gradient Boosting Machine (LightGBM), complemented by SHapley Additive exPlanations (SHAP) for interpretability. Our findings demonstrate that HFS segments increase toward the end of each quarter, driven by rapid transitions and tactical urgency, whereas LFS segments dominate the middle phases, reflecting strategic tempo control. NFS accounts for the majority of game time but decreases as the game progresses. The LightGBM analysis highlighted the importance ranking of key performance indicators (KPIs) across different segments and revealed differences in the importance of these indicators within each segment. Compared to traditional methods, our approach provides a finer-grained analysis of game pace dynamics and offers actionable insights for optimizing coaching strategies. This study not only advances the understanding of basketball game rhythm but also establishes a robust framework for integrating machine learning and statistical models in sports analysis.

Fig 1. Optimal number of clusters determined by the elbow method.

Fig 2. Examples of different frequency segments in games.

(X-axis: Game time, in seconds. Y-axis: Possession duration, in seconds. Dots represent consecutive possession events of games. Red and green backgrounds represent the time periods belonging to HFS and LFS, respectively, while the white background represents NFS.).

Fig 3. Random forest prediction results.

(The X-axis represents 1-16 periods in the game, totaling 48 minutes of a regular match. The Y-axis represents the duration of the segments. The top-left bar chart shows the time proportion changes of the three types of segments across the 16 periods. The top-right, bottom-left, and bottom-right bar charts respectively represent the predicted time changes for HFS, LFS, and NFS.).

Fig 4. Feature importance of HFS, LFS and NFS.

(Fig 4-1, Fig 4-2, and Fig 4-3 represent HFS, NFS, and LFS, respectively. From left to right are HFS, NFS, and LFS. The X-axis shows the magnitude of each feature’s impact on the prediction results. The farther a point is from the centerline (zero), the greater the feature’s impact on the model output. Positive SHAP values indicate a positive impact, while negative SHAP values indicate a negative impact. The Y-axis ranks the features by their influence from top to bottom, with the top features having a greater overall impact on the model output, and the bottom features having a lesser impact. Pink dots represent that the feature value has a positive impact on the model prediction in this observation. Blue dots indicate that the feature value has a negative impact on the model prediction in this observation.).

Fig 5. SHAP heatmap.

(Fig 5-1, Fig 5-2, and Fig 5-3 represent HFS, NFS, and LFS, respectively. The left Y-axis shows the ranking of important features, and the right Y-axis visualizes these features. The color intensity in the image represents the magnitude of the SHAP values: the darker the color, the larger the absolute value of the SHAP, and the greater its impact on the model. The top part visualizes the model’s prediction results under these values. The chart visualizes data from 500 test set samples, as too much data would affect the clarity of the visualization.).