Ahead of the 2026 FIFA World Cup kickoff on June 11 at Mexico City Stadium, data engineers and sports analysts have developed a machine learning framework to forecast international soccer match outcomes. Leveraging a historical dataset of 49,000 matches spanning 1872 to 2026, the project applies a probabilistic approach to address the sport’s distinct statistical profile, characterized by low scoring volumes and a 22 percent draw rate. The modeling pipeline integrates historical match results, team ratings, and contextual variables. To prevent data leakage, pre-match Elo ratings and their update timestamps were engineered into time-sensitive features. Additional inputs capture team momentum through rolling performance metrics, recent scoring and conceding rates, and match context indicators such as venue type and tournament status. A chronological train-validation-test split reserves all matches from 2018 onward for final evaluation. Several algorithmic families were benchmarked, including baseline frequency models, multinomial logistic regression, and LightGBM. Grid search and logarithmic parameter tuning were employed to optimize the gradient boosting framework for tree complexity, regularization, and learning rates. Despite the computational advantage of tree-based architectures, validation metrics reveal that multinomial regression performs within a margin of 0.002 log-loss points of the LightGBM implementation. The selected LightGBM configuration achieved a test log-loss of 0.873, with multinomial regression slightly outperforming it on macro F1 scores. The findings indicate that for this dataset, simpler linear models remain highly competitive while offering greater interpretability. Calibration analysis demonstrates that the models produce reliable probability distributions across confidence thresholds. Predicted likelihoods for home and away wins align closely with observed frequencies. However, a persistent structural weakness emerges in draw prediction. The models consistently overestimate home-win probabilities, often assigning high confidence to matches that ultimately end in draws. Feature analysis confirms that while the algorithms correctly identify balanced matchups as draw-prone, they fail to elevate draw probabilities sufficiently to shift the final classification. Consequently, the model achieves a home win accuracy of 86 percent, but draw recall remains below one percent on the test set. The project underscores both the utility and the limitations of machine learning in sports forecasting. While the framework reliably quantifies team strength differentials and momentum, the inherent unpredictability of low-margin soccer requires specialized architectural adjustments for draw modeling. Analysts recommend developing a binary classification system dedicated to draw probability and incorporating granular player-level metrics to further refine offensive and defensive assessments. The complete methodology, codebase, and dataset are publicly available for independent validation. As international competition intensifies ahead of the 2026 tournament, this probabilistic approach provides a transparent, data-driven foundation for match forecasting and strategic analysis.