Probabilistic Prediction of the 2026 FIFA World Cup Using Gradient Boosting and Monte Carlo Simulation

Abstract

A comprehensive end-to-end  system  is  presented  to  estimate  the  championship  probability  of  each of the 48 national teams participating in the 2026 FIFA World Cup under the new 12-group format. The pipeline integrates heterogeneous data sources, including international match results from 1872–2026, event data (goalscorers and penalty shootouts), Expected Goals (xG) from StatsBomb, Transfermarkt market values, and historical FIFA rankings extended with a 2026 snapshot, to construct nine features— including derived features calculated strictly as-of-date—which are used to train four multiclass models: Logistic Regression, XGBoost, LightGBM, and Platt-calibrated XGBoost. The models are trained using exponential temporal decay weighting and class rebalancing. The best-performing model, Platt-calibrated XGBoost, achieves a Log-Loss of 0.855 and a Brier score of 0.166 on a temporal test set and is coupled with a 10,000-iteration Monte Carlo engine that complies with the 2026 competition regulations (top two teams from each group + eight best third-placed teams, tiebreakers based on points–goal difference–goals scored, and penalty shootouts). Clopper–Pearson confidence intervals place a small group of historically strong national teams among the favorites. A dedicated section discusses the model’s limitations, including debutant teams with no historical record, partial xG coverage, and the absence of structural injury modeling.