Predição de resultados de partidas de profissionais de Counter-Strike 2 com aprendizado de máquina

Abstract

This work investigates win prediction in professional Counter-Strike 2 matches using only information available before each match begins. Each observation is represented by differences between the two teams, computed from recent statistics, performance indicators and a pre-match strength measure. The task is framed as binary classification, with the aim of estimating the win probability of the analyzed team. The methodology separates training, selection and evaluation in temporal order: older matches are used for training and internal selection, while the 2026_s1 window is reserved as a later-period holdout evaluation. The experimental evaluation started from 38 pre-match attributes and registered 71 feature sets, 58 possible classifier configurations and 2,582 combinations actually evaluated. The selected solution is an L2-regularized logistic regression with C = 0.3, no artificial class balancing and 11 variables chosen by a univariate selection criterion. On the holdout set, this logistic regression achieved ROC-AUC of 0.6828 and an accuracy of 0.6293. The prospective evaluation over later events was kept as complementary evidence from operational use, without changing the selected model. The contribution is a reproducible workflow for pre-match prediction in Counter-Strike 2, connecting data, variables, training, classifier comparison, interpretation and complementary validation.