Predicting the in-game status in soccer with machine learning using spatiotemporal player tracking data.

Area of Science:

• Sports Science
• Data Science
• Machine Learning

Background:

• Accurate in-game status (in-play vs. interrupted) is crucial for soccer match analysis.
• Calculating performance indicators requires effective playing time, necessitating reliable in-game status determination.
• Spatiotemporal data offers potential for automated in-game status prediction.

Purpose of the Study:

• To investigate the feasibility of determining soccer match in-game status using only time-continuous player positions.
• To evaluate the performance of various machine learning models for this prediction task.
• To assess the practical utility of automated in-game status prediction for performance diagnostics.

Main Methods:

• Utilized spatiotemporal player position data from 102 German Bundesliga matches.
• Trained and evaluated four machine learning algorithms: logistic regression, decision trees, random forests, and AdaBoost.
• Compared model accuracy, precision, time shift error, and impact on performance indicator calculations.

Main Results:

• Achieved up to 92% accuracy in predicting in-game status on a frame level.
• AdaBoost demonstrated 81% precision in detecting stoppages longer than 2 seconds.
• Time shift errors for stoppages were within 2 seconds for a majority of predictions.
• Predicted in-game status resulted in a minimal error (1.3%) for player distance performance indicators.

Conclusions:

• Machine learning models can reliably predict soccer match in-game status from player position data.
• The AdaBoost model shows high accuracy and precision, suitable for practical applications.
• Automated in-game status prediction is valuable for performance diagnostics, especially for teams lacking manual annotations or ball tracking data.