World Cup 2026 Champion Forecaster ⚽🏆

A machine learning pipeline that forecasts the probability of each remaining team winning the 2026 FIFA World Cup, using ~49,000 real international football matches played between 1872 and 2026.

---

Overview

This project uses real historical match data to:

1. Compute running Elo ratings for every national team, updated match-by-match through the present day.
2. Engineer match-level features — Elo difference, recent form, head-to-head history, rest days, and neutral-venue flag — using only information available before each match (no data leakage).
3. Train a calibrated Gradient Boosting classifier to predict Home / Draw / Away outcomes, with isotonic probability calibration (essential for feeding probabilities into a Monte Carlo simulation).
4. Validate the model using a time-based split (train on the past, test on the most recent slice), reporting log loss and Brier score.
5. Monte Carlo-simulate the live knockout bracket (50,000 tournaments) to estimate P(champion) for each of the 16 teams still alive.

---

Data Source

File	Description
international_results.csv	~49,000 international matches (1872–2026), sourced from martj42/international_results — a widely-used mirror of the Kaggle dataset “International football results from 1872 to 2026”.

The dataset already includes the 2026 World Cup fixtures. As of the current run, two Round-of-16 matches have been played and are treated as hard results:

Date	Match	Score
2026-07-04	France vs Paraguay	1–0
2026-07-04	Morocco vs Canada	3–0

The remaining six R16 matches and the entire knockout tree (QF → SF → Final) are forecast by the model.

---

Why Gradient Boosting?

Tabular data with tens of thousands of rows is the sweet spot for tree-based ensemble methods. Deep learning offers no advantage here without event-level tracking data (player positions, possession chains, etc.). Gradient Boosting is fast, interpretable, handles mixed feature types well, and — with isotonic calibration — produces the reliable probability estimates needed for bracket simulation.

---

Project Structure

.
├── world_cup_forecast.py              # Main forecasting pipeline (5 steps)
├── world_cup_forecast_analysis.ipynb  # Jupyter notebook with full analysis & visualizations
├── international_results.csv          # Raw match data (~49K matches, 1872–2026)
├── .gitignore
└── README.md

---

Requirements

• Python 3.8+
• numpy
• pandas
• scikit-learn

Install dependencies:

pip install numpy pandas scikit-learn

---

Quick Start

# 1. Clone the repository
git clone https://github.com/andresveraf/World-Cup-2026.git
cd World-Cup-2026

# 2. (Optional) Create a virtual environment
python3 -m venv venv
source venv/bin/activate

# 3. Install dependencies
pip install numpy pandas scikit-learn

# 4. Run the forecaster
python world_cup_forecast.py

The script will print:

• Model validation metrics (log loss, Brier score)
• Current Elo ratings for the 16 teams still alive
• A ranked table of champion probabilities from 50,000 Monte Carlo simulations

---

Pipeline Details

Step	What it does
1. Load data	Parses international_results.csv, separates played matches from upcoming fixtures
2. Feature engineering	Builds Elo diff, form (last 10 matches), H2H win rate, rest days, neutral flag — all leak-free
3. Train + calibrate	Gradient Boosting (250 trees) wrapped in CalibratedClassifierCV (isotonic, 3-fold)
4. Validate	Time-based 90/10 split; reports log loss and Brier score
5. Monte Carlo	Precomputes pairwise win probabilities, then simulates 50,000 complete tournaments through the real live bracket

---

Notes

• Host nations (USA, Mexico, Canada) receive a small Elo bonus (+30) even at neutral venues to account for crowd support.
• Knockout-stage draws are resolved via a penalty-shootout model: a near coin-flip nudged by the Elo difference between the two teams.
• The random seed is fixed (42) for reproducibility.

---

License

This project is for educational and analytical purposes. The match data is sourced from the public martj42/international_results dataset.