FIFA World Cup 2026 Winner Prediction

A machine learning system that predicts the winner of the FIFA World Cup 2026 by training a Random Forest classifier on 74 years of international football results and running a Monte Carlo simulation of the expanded 48-team tournament.

---

Table of Contents

• Overview
• Methodology
• Project Structure
• Setup
• Usage
• Outputs
• 2026 World Cup Format
• Technical Details
• Reproducibility
• License

---

Overview

This project predicts the FIFA World Cup 2026 champion using a data-driven approach:

1. Data Collection -- Downloads the International Football Results dataset (1872-2024), containing over 50,000 international matches from all FIFA-recognised nations.
2. Feature Engineering -- Computes rolling ELO ratings, recent form indicators, goal-scoring averages, head-to-head records, and tournament importance weights for every match.
3. Model Training -- Trains a RandomForestClassifier (300 trees, max depth 8) with a time-based train/test split to prevent look-ahead bias.
4. Monte Carlo Simulation -- Simulates the full 2026 tournament 10,000 times, sampling match outcomes from the model's probability distribution and tracking each team's advancement rate.

---

Methodology

Feature Engineering

Each match in the training set is transformed into a 14-dimensional feature vector computed exclusively from data available before the match date:

Feature	Description
ELO rating difference	Difference in computed ELO ratings between the two teams
Home form	Average points per match from the last 10 matches (3 for win, 1 for draw, 0 for loss)
Away form	Same for the opponent
Goals scored average	Average goals scored in the last 10 matches (per team)
Goals conceded average	Average goals conceded in the last 10 matches (per team)
Head-to-head win rate	Win percentage in the last 10 meetings between the two teams
Venue flags	Binary indicators for home or neutral venue
Tournament weight	Importance weight (1-4) based on competition type (friendly to World Cup)
Days since last match	Days elapsed since each team's previous match (fatigue proxy)

Model

A RandomForestClassifier is configured with:

• 300 decision trees
• Maximum depth of 8 to prevent overfitting
• Balanced class weights to handle the natural class imbalance (wins are more common than draws)
• Time-based 85/15 train/test split (no random shuffling)

Tournament Simulation

The 2026 World Cup features an expanded format:

• 48 teams split into 12 groups of 4
• Top 2 from each group advance automatically (24 teams)
• The 8 best third-placed teams also advance
• 32-team knockout bracket: Round of 32 -> Round of 16 -> Quarter-Finals -> Semi-Finals -> Final

For each group match, the model produces Win/Draw/Loss probabilities. A Monte Carlo sample determines the result, and goal counts are sampled from a Poisson distribution conditioned on the outcome. Group standings are determined by points, then goal difference, then goals scored, then team strength.

Knockout matches use normalised win probabilities (draws are redistributed proportionally) to determine the winner, simulating the effect of extra time and penalty shootouts.

---

Project Structure

fifa-2026-predictor/
├── data/
│   ├── raw/                      # Original CSV datasets (downloaded automatically)
│   ├── processed/                # Cleaned parquet files and feature matrices
│   └── fixtures/                 # 2026 groups and qualified teams configuration
├── src/
│   ├── __init__.py
│   ├── data_loader.py            # Download and clean raw match data
│   ├── feature_engineering.py    # Compute rolling features and ELO ratings
│   ├── train_model.py            # Train and evaluate RandomForest classifier
│   ├── simulate.py               # Monte Carlo tournament simulator
│   └── utils.py                  # ELO computation, constants, helpers
├── models/
│   └── rf_model.pkl              # Trained RandomForest model (joblib)
├── outputs/
│   ├── win_probabilities.csv     # Per-team advancement probabilities
│   ├── bracket_simulation.png    # Bar chart of top favorites
│   └── calibration_curve.png     # Model calibration visualisation
├── notebooks/
│   └── eda.ipynb                 # (optional) Exploratory data analysis
├── requirements.txt
├── main.py                       # CLI entry point
└── README.md

---

Setup

Prerequisites

• Python 3.11 or later
• pip (Python package manager)

Installation

# Navigate to the project directory
cd fifa-2026-predictor

# Install Python dependencies
pip install -r requirements.txt

The dependencies are:

• pandas and numpy for data manipulation
• scikit-learn for the RandomForest classifier
• matplotlib and seaborn for visualisation
• joblib for model serialisation
• requests for dataset download

---

Usage

Quick Start

Run the full pipeline end-to-end with a single command:

python main.py

This will:

1. Download the international football results dataset (first run only, ~1 minute)
2. Clean and process the data, compute ELO ratings and feature vectors
3. Train and evaluate the RandomForest model
4. Run 10,000 tournament simulations
5. Display the predicted champion and top contenders
6. Save results to the outputs/ directory

Command-Line Options

Option	Description	Default
--steps	Comma-separated steps to run: load, train, simulate, or all	all
--simulations	Number of Monte Carlo simulations	10000
--force-rebuild	Force re-download and reprocess all data	(flag)
--top-n	Number of top teams to display	10

Examples

# Faster run with fewer simulations for testing
python main.py --simulations 1000

# Rerun only the simulation step (keep existing model)
python main.py --steps simulate

# Full rebuild from scratch
python main.py --force-rebuild

# Show more contenders
python main.py --top-n 20

---

Outputs

After running the pipeline, the following files are produced in the outputs/ directory:

File	Description
win_probabilities.csv	Per-team probabilities for each knockout round (R32 through champion)
bracket_simulation.png	Horizontal bar chart showing the top N teams by win probability
calibration_curve.png	Three-panel calibration plot showing model reliability for Win/Draw/Loss

The win_probabilities.csv file contains the following columns:

• team -- National team name
• pct_r32 -- Percentage of simulations in which the team reached the Round of 32
• pct_r16 -- Percentage reaching the Round of 16
• pct_qf -- Percentage reaching the Quarter-Finals
• pct_sf -- Percentage reaching the Semi-Finals
• pct_final -- Percentage reaching the Final
• pct_win -- Percentage of simulations won (primary prediction metric)

---

2026 World Cup Format

Qualified Teams

The 48 qualified nations (as of March 2026) span all six confederations, including debutants Cape Verde, Curacao, Jordan, and Uzbekistan.

Group Stage

Group	Teams
A	Mexico, South Africa, South Korea, Czech Republic
B	Canada, Bosnia and Herzegovina, Qatar, Switzerland
C	Brazil, Morocco, Haiti, Scotland
D	USA, Paraguay, Australia, Turkey
E	Germany, Curacao, Ivory Coast, Ecuador
F	Netherlands, Japan, Sweden, Tunisia
G	Belgium, Egypt, Iran, New Zealand
H	Spain, Cape Verde, Saudi Arabia, Uruguay
I	France, Senegal, Iraq, Norway
J	Argentina, Algeria, Austria, Jordan
K	Portugal, DR Congo, Uzbekistan, Colombia
L	England, Croatia, Ghana, Panama

Advancement Rules

• The top 2 teams from each group advance to the knockout stage (24 teams)
• The 8 best third-placed teams across all groups also advance
• Third-placed teams are ranked by: points, goal difference, goals scored, team strength proxy
• The knockout stage is single elimination: Round of 32, Round of 16, Quarter-Finals, Semi-Finals, and Final

---

Technical Details

Data Source

The project uses the International Football Results from 1872 to Present dataset maintained by Mart JuriSOO. This dataset is downloaded automatically from GitHub on first run. Matches before 1950 are excluded to ensure stable ELO ratings.

ELO Rating System

A custom ELO rating system is implemented for each team with:

• Initial rating: 1500
• K-factor: 32 (adjusted by goal difference margin)
• Home advantage: 100 ELO points (not applied for neutral venues)
• Ratings are updated iteratively match-by-match in chronological order

Random Seed

All stochastic processes are seeded with 42 for fully reproducible results.

---

Reproducibility

The project is designed to produce identical results across runs when given the same input data:

• Random seed is set to 42 at the start of every run
• The RandomForest classifier uses random_state=42
• Monte Carlo simulations use NumPy's seeded random number generator

---

License

Apache 2.0

---

Built with Python, scikit-learn, and 74 years of international football history.