Predicting the 2026 World Cup with Machine Learning: A Practical Guide to Football Match Modeling Using XGBoost and Poisson Distribution

The open-source project introduced in this article is wc2026-match-predictor, released by HaykDanghyan on GitHub in June 2026. Its core idea is to use XGBoost Poisson regression to predict the expected number of goals for each team, combined with Monte Carlo simulation to forecast match results, qualification probabilities, and group stage advancement for the 2026 World Cup. It aims to address the poor performance of traditional classification models in predicting draws in football matches.

The project tested four mainstream classification algorithms: logistic regression, random forest, gradient boosting, and XGBoost. The results showed an overall accuracy of only 57-60%, with extremely low recall rates for draws (ranging from 2% to 12%). The root cause of the problem is that draws do not have an independent region in the feature space; when teams of similar strength play against each other, classifiers struggle to correctly estimate the probability of a draw. This bias is particularly critical in the group stage, as qualification calculations rely on accurate draw probability estimates.

The project's core innovation uses the Poisson regression method: two XGBoost regressors are trained to predict the expected number of goals for the home and away teams (λ_home and λ_away, respectively). Then, the probability mass function of the Poisson distribution is used to calculate the probability of any score combination (P(X=i,Y=j) = (λ_home^i × e^(-λ_home)/i!) × (λ_away^j × e^(-λ_away)/j!)). By iterating over all score combinations, a complete score probability matrix is constructed, and the probabilities of win, draw, and loss are then aggregated from this matrix.

The model's input features include: 1. Elo Rating System: Calculated based on approximately 23,000 international matches since 2002, with an update coefficient K=20 and a home advantage of +60 points. The Elo rating difference is a strong predictive signal. 2. Recent Form Indicators: Average number of goals scored (home_form_gf/away_form_gf) and conceded (home_form_ga/away_form_ga) over the last 5 matches, which compensates for the slow update of Elo ratings. 3. Other Factors: Match importance (friendly/regional tournament, etc.) and whether the match is played on a neutral venue.

The 2026 World Cup has a complex format (12 groups of 4 teams each; the top two teams from each group qualify directly, plus the 8 best third-placed teams). Qualification probabilities cannot be calculated analytically. The project uses the Monte Carlo method: each group is simulated 5000 times. In each simulation, match scores are randomly sampled (based on the Poisson distribution), and rankings are determined according to World Cup rules using points, goal difference, and goals scored. The frequency of a team qualifying across the 5000 simulations is counted to obtain a reliable estimate of qualification probability.

Model Performance: The Mean Absolute Error (MAE) for home team goals is 1.057, and for away team goals is 0.862, which is within the acceptable range of the 53-58% benchmark accuracy in the football prediction field. Limitations: It is based only on historical match data and does not include player-level information (injuries, suspensions, squad depth) or off-field factors (weather, travel, etc.). The results are statistical estimates rather than deterministic predictions.

The project uses a Python tech stack with key dependencies: pandas/numpy (data processing), XGBoost/scikit-learn (model building), matplotlib/seaborn (visualization), and Streamlit (interactive web interface). After training, the model is serialized into a pickle file. The Streamlit application loads the pre-trained model to provide real-time prediction services, allowing users to select teams to view expected goals, most likely scores, win/draw/loss probabilities, and the score probability matrix.

The value of this project lies in demonstrating the combination of statistics and machine learning to solve practical problems: Poisson distribution has been used in football modeling for decades, and when combined with modern gradient boosting frameworks and careful feature engineering, the results are more robust. For data science practitioners, the insight is: when directly predicting labels is difficult, trying to predict the underlying mechanism that generates the labels (such as the goal rate in this case) often yields better results.