Predictive analytics in eSports remains underexplored at the round level, particularly in fast-paced titles like Counter-Strike: Global Offensive (CS:GO). This study introduces a novel application of a feature-optimised Random Forest model to predict round winners in CS:GO, focusing on both accuracy and robustness across varying gameplay contexts. Unlike previous work, our approach incorporates detailed feature importance analysis to reduce noise and enhance interpretability. The model achieved a cross-validated mean accuracy of 0.87 (SD = 0.00) and demonstrated strong generalisation, reaching 91% accuracy on rounds with high-impact features and 75% in diverse scenarios. Performance was evaluated using a confusion matrix, ROC curve, and precision-recall curve. The study contributes a scalable framework for round-level prediction and offers insights into key variables influencing outcomes. Our results suggest promising avenues for real-time integration of machine learning in strategic support tools for professional and semi-professional eSports teams.
Daka et al. (Thu,) studied this question.