Mobile Legends: Bang Bang is a widely played Multiplayer Online Battle Arena game in Southeast Asia, and its competitive ecosystem has driven the need for accurate match outcome prediction. Most existing studies analyze either the draft pick phase or the in game phase in isolation, limiting their ability to capture the full progression of a match. To address this limitation, this study evaluates the performance of Random Forest and Extreme Gradient Boosting (XGBoost) in predicting match outcomes across both phases using data from the MSC 2025 tournament. The dataset was collected from Liquipedia’s official API and match replay recordings. Draft pick features represent team composition factors such as synergy, hero strength, and patch impact, while in game features consist of statistical indicators including gold, kills, turrets, and objectives extracted from multiple time based snapshots. Both models were trained using qualification stage matches and tested on the main event. A phase separated hybrid feature engineering approach was employed to represent strategic differences between the draft pick and in game phases. Evaluation metrics include accuracy, precision, recall, F1 score, and ROC AUC. Results show that the draft pick models achieved a maximum accuracy of 57%, whereas the in game models reached 88% for Random Forest and 84% for XGBoost, with both achieving a ROC AUC of 0.94. These findings indicate that snapshot based in game features provide stronger predictive signals than draft pick composition features, which reflect only the initial strategic potential rather than actual match conditions.