World Cup 2026 predictor

Predicting a World Cup winner is a tiny-data problem dressed up as a big one: there are 150 years of matches, but only one winner per tournament. I wanted to see which pre-tournament signals actually go with winning, without letting the tournament itself leak into the features.

• A star-schema warehouse. Match results from 1872 to 2024 and FIFA rankings go into SQLite: a match fact table with team, tournament and date dimensions.
• Six features, from the two years before each tournament: win rate, goals scored and conceded per game, FIFA rank, host status and confederation. Matches from the tournament itself are excluded to prevent leakage.
• A temporal split, not a random one. Train on 1998–2018, validate on 2022 (won by Argentina).
• One winner per ~32 teams is heavy class imbalance, handled with balanced class weights. I compared Logistic Regression, Decision Tree, Naive Bayes, Random Forest and XGBoost on accuracy, precision, recall and F1. Logistic Regression generalised best.

With one positive example per tournament, any single prediction is fragile. Next I'd add Elo ratings and squad-level data, and evaluate with leave-one-tournament-out across every World Cup instead of one validation year.

Code on GitHub