‹ Thesis in 5 minutes
Who wins this round?
One game state from the test set, and what a logistic regression says about it as feature groups are added.
CT T
Core model M0, 4 features. The CT side won this round.
Add one feature group to the core
0.702ROC-AUC on the test set
0.608log loss on the test set
The core: players alive on each side, round time and the round timer.
Nine models on the full feature set
Gradient boosting wins on log loss. Walk-forward validation by quarter keeps ROC-AUC around 0.81–0.82. The production XGBoost model has a calibration error of 0.0086, against 0.0183 for logistic regression.
Explained: a 2 v 0 after the plant
Another test state: 108 s, bomb planted, CT 2 alive against 0, equipment +20,750. The production XGBoost gives CT 0.992. The three largest SHAP contributions:
alive_players_count_t = 0+2.221delta_equip_val = 20,750+1.540avg_hp_per_alive_t = 0+0.911MSc thesis, University of Zielona Góra, 2026. 9,196 professional maps, 187,361 rounds, 16.8 million game states.