🏆 GoalIQ 2026 - FIFA World Cup ML Prediction System

        ██████╗  ██████╗  █████╗ ██╗     ██╗ ██████╗     ██████╗  ██████╗ ██████╗  ██████╗
       ██╔════╝ ██╔═══██╗██╔══██╗██║     ██║██╔═══██╗    ╚════██╗██╔═══██╗╚════██╗██╔════╝
       ██║  ███╗██║   ██║███████║██║     ██║██║   ██║     █████╔╝██║   ██║ █████╔╝███████╗
       ██║   ██║██║   ██║██╔══██║██║     ██║██║▄▄ ██║    ██╔═══╝ ██║   ██║██╔═══╝ ██╔══██╗
       ╚██████╔╝╚██████╔╝██║  ██║███████╗██║╚██████╔╝    ███████╗╚██████╔╝███████╗╚██████╔╝
        ╚═════╝  ╚═════╝ ╚═╝  ╚═╝╚══════╝╚═╝ ╚══▀▀═╝    ╚══════╝ ╚═════╝ ╚══════╝ ╚═════╝

🏆 GoalIQ 2026 - FIFA World Cup ML Prediction System

Harness machine learning to predict FIFA World Cup 2026 match outcomes,
team performance, and tournament progression with full statistical transparency.

v2 · Upgraded - 4 bugs fixed · 44 engineered features · Stacking Ensemble · Threshold optimization · 5,000-run Monte Carlo simulation

📑 Table of Contents

📌 Project Overview
📓 Open the Notebook
🖼️ Results & Visualizations
📊 Dataset
🐛 Bug Audit (v1 → v2)
⚙️ Feature Engineering
🔬 ML Pipeline
📈 Results
🤖 Models Trained
🔗 Stacking Ensemble & Threshold Optimization
🔍 Feature Importance & Explainability
🏟️ Monte Carlo Tournament Simulation
📈 v1 vs v2 — Upgrade Summary
⚠️ Honest Accuracy Note
✨ Highlights
🛠️ Tech Stack
🚀 Getting Started
📂 Project Structure
📚 What You'll Learn
🔮 Future Enhancements
🤝 Contributing
🙏 Acknowledgements
👨‍💻 Author

📌 Project Overview

GoalIQ 2026 is a full-pipeline machine learning system for predicting outcomes in the FIFA World Cup 2026 - the first edition with 48 teams. It covers the complete data science lifecycle from raw CSV to Monte Carlo tournament simulation.

Stage	Description
🔍 EDA	Feature correlation analysis, class distributions, confederation breakdowns
⚙️ Feature Engineering	21 domain-informed composite features derived from raw football statistics
🤖 Model Training	6 base models: Random Forest, Extra Trees, HistGradientBoosting, MLP, SVM, Logistic Regression
🔗 Ensemble	`StackingClassifier` with 5-fold OOF meta-learning via Logistic Regression
🎯 Threshold Tuning	Optimal decision threshold found by scanning the validation set
🏟️ Simulation	5,000-run Monte Carlo tournament bracket (48 teams, group + knockout stages)
💾 Submission	Calibrated win probabilities for all test teams

⚽ The FIFA World Cup 2026 is the largest in history - 48 teams, 3 host nations (USA · Canada · Mexico), and 104 matches. Predicting outcomes at this scale demands rigorous, bias-free ML engineering.

📓 Open the Notebook

Viewer	Link
NBViewer (static render)	View on NBViewer
Google Colab (run in browser)

🖼️ Results & Visualizations

This section showcases the key visual outputs generated during the GoalIQ 2026 pipeline.

📊 Exploratory Data Analysis

Feature Correlations with Target	Feature Distributions - Winners vs Non-Winners

Win Rate by Confederation	Engineered Feature Profiles - Elite Teams

🤖 Model Performance & Evaluation

Model Performance Comparison (Accuracy & AUC)	ROC Curves - All Models

Confusion Matrix - Stacking Ensemble	Calibration Curves

Threshold Optimization Curve

🔍 Feature Importance & Explainability

MDI Feature Importance - Random Forest	Permutation Importance (Model-Agnostic)

🏟️ Tournament Simulation

Predicted Champions - 5,000 Monte Carlo Simulations	Top-20 Contenders by Confederation

📊 Dataset

Source: FIFA World Cup 2026 Prediction System — Kaggle

File	Rows	Columns	Description
`train.csv`	1,000	25	Historical team-match data with `winner` label
`test.csv`	250	24	Teams to predict win probabilities for
`submission.csv`	250	2	Template: `id`, `winner_probability`

Feature	Type	Description
`fifa_rank`	Integer	FIFA world ranking (lower = better)
`fifa_points`	Float	FIFA ranking points
`goals_scored_avg`	Float	Average goals scored per match
`goals_conceded_avg`	Float	Average goals conceded per match
`win_rate_last_year`	Float	Win percentage over last 12 months
`avg_player_rating`	Float	Average squad player rating
`market_value_million_eur`	Float	Total squad market value (€M)
`recent_form_score`	Float	Form score over last 10 matches
`possession_avg`	Float	Average ball possession (%)
`passing_accuracy`	Float	Average passing accuracy (%)
`shots_per_game`	Float	Average shots per match
`shots_on_target_ratio`	Float	Ratio of shots on target
`clean_sheets_last_10`	Integer	Clean sheets in last 10 games
`star_players_count`	Integer	Number of elite-tier players
`host_advantage`	Binary	1 if playing in home region
`confederation`	Categorical	UEFA / CONMEBOL / CAF / CONCACAF / AFC / OFC
`winner`	Binary	Target — 1 = match winner, 0 = not

Class balance: 527 non-winners (52.7%) · 473 winners (47.3%) - near-balanced binary classification.

🐛 Bug Audit (v1 → v2)

Four critical bugs were found in the original notebook that suppressed accuracy and corrupted training:

╔═══════════════════════════════════════════════════════════════════════════╗
║                          BUG AUDIT REPORT                                 ║
╠════╦══════════════════════════════╦═══════════════════╦═══════════════════╣
║  # ║ Bug                          ║ Impact            ║ Fix               ║
╠════╬══════════════════════════════╬═══════════════════╬═══════════════════╣
║  1 ║ strength_index / squad_quality║ DATA LEAKAGE —   ║ Compute .max()    ║
║    ║ normalised with .max() on    ║ test stats bleed  ║ on train only,    ║
║    ║ combined train+test pool     ║ into training     ║ apply to test     ║
╠════╬══════════════════════════════╬═══════════════════╬═══════════════════╣
║  2 ║ confederation dropped via    ║ Loses a useful    ║ Label-encode      ║
║    ║ cat_drop instead of encoded  ║ categorical       ║ (UEFA=0 … OFC=5)  ║
║    ║                              ║ signal            ║                   ║
╠════╬══════════════════════════════╬═══════════════════╬═══════════════════╣
║  3 ║ fifa_rank used as raw integer║ Inverted signal — ║ Add rank_inv =    ║
║    ║ (lower rank = better team)   ║ model sees        ║ 1 / fifa_rank     ║
║    ║                              ║ worse = higher    ║                   ║
╠════╬══════════════════════════════╬═══════════════════╬═══════════════════╣
║  4 ║ '\n' in set_xticklabels was  ║ SyntaxError at    ║ Use escape        ║
║    ║ a literal line break         ║ runtime           ║ sequence '\\n'    ║
╚════╩══════════════════════════════╩═══════════════════╩═══════════════════╝

⚙️ Feature Engineering

v2 expands from 31 → 44 features with 21 engineered features across 7 football-domain categories:

📋 View all 21 engineered features

Feature	Formula	Domain
`conf_enc`	Label-encoded confederation	Context
`rank_inv`	`1 / fifa_rank`	Ranking
`win_ratio_10`	`wins / (wins+losses+draws)` last 10	Form
`loss_ratio_10`	`losses / total` last 10	Form
`goal_diff`	`goals_scored_avg − goals_conceded_avg`	Attack/Defence
`goal_ratio`	`goals_scored / goals_conceded`	Attack/Defence
`shots_on_target_abs`	`shots_per_game × shots_on_target_ratio`	Attack
`goals_per_sot`	`goals_scored / shots_on_target_abs`	Conversion
`star_density`	`star_players_count / 11`	Squad
`value_per_cap`	`market_value / experience_avg_caps`	Squad
`form_x_winrate`	`recent_form_score × win_rate_last_year`	Form
`form_x_rating`	`recent_form_score × avg_player_rating`	Form
`possession_x_passing`	`possession_avg × passing_accuracy`	Style
`attack_index`	`goals × shots_on_target_ratio × win_rate`	Attack
`defence_index`	`clean_sheets / goals_conceded`	Defence
`rank_x_form`	`rank_inv × recent_form_score`	Interaction
`rank_x_rating`	`rank_inv × avg_player_rating`	Interaction
`value_x_rating`	`log1p(market_value) × avg_player_rating`	Interaction
`rating_z`	Z-score of `avg_player_rating` (train stats only)	Normalised
`value_z`	Z-score of `market_value` (train stats only)	Normalised
`strength_index`	Weighted composite of points + form + rating	Overall

Key design principle - zero leakage:

# ✅ CORRECT (v2): stats fitted on train only, applied to test
if stats is None:                           # called on train
    stats = {
        'max_pts' : df['fifa_points'].max(),
        'max_val' : df['market_value_million_eur'].max(),
        'max_exp' : df['experience_avg_caps'].max(),
    }
d['strength_index'] = df['fifa_points'] / stats['max_pts'] * 40 + ...

# ❌ WRONG (v1): combined train+test before engineering
all_data = pd.concat([train, test])        # leakage!
all_eng  = engineer_features(all_data)     # test .max() pollutes train

🔬 ML Pipeline

Raw CSVs  (train.csv · test.csv · submission.csv)
        │
        ▼
① Load & Inspect
        │  → Shape, dtypes, null counts, class distribution
        │  → 49 unique teams · 6 confederations · 1,000 rows
        ▼
② Bug Fixes Applied
        │  → Fix leakage: train-only normalization stats
        │  → Encode confederation (not drop)
        │  → Invert fifa_rank → rank_inv = 1/fifa_rank
        ▼
③ Feature Engineering (31 → 44 features)
        │  → 21 composite features across 7 domains
        │  → All stats computed on train, applied to test
        ▼
④ Exploratory Data Analysis (EDA)
        │  → Feature correlations with target (max |r| = 0.376)
        │  → Univariate distributions — winners vs non-winners
        │  → Win rate by confederation
        │  → Elite team feature profiles (parallel coordinates)
        ▼
⑤ Train / Validation Split
        │  → 80/20 stratified split · random_state=42
        │  → StandardScaler for distance-based models (LR, MLP, SVM)
        ▼
⑥ Multi-Model Training (6 base classifiers)
        │  → Random Forest · Extra Trees · HistGradientBoosting
        │  → MLP Neural Net · SVM (RBF) · Logistic Regression
        ▼
⑦ Stacking Ensemble
        │  → StackingClassifier: 5-fold OOF predict_proba
        │  → Meta-learner: Logistic Regression
        ▼
⑧ Threshold Optimization
        │  → Scan thresholds 0.35 → 0.70 on validation set
        │  → Select threshold maximising accuracy
        ▼
⑨ Full Evaluation
        │  → Confusion Matrix · Classification Report
        │  → ROC Curves (all models) · Calibration Curves
        │  → 5-Fold Stratified Cross-Validation
        ▼
⑩ Feature Importance & Explainability
        │  → MDI Importance (Random Forest)
        │  → Permutation Importance (model-agnostic, 20 repeats)
        ▼
⑪ Monte Carlo Tournament Simulation (5,000 runs)
        │  → 48 teams · 12 groups of 4 → Round of 32 → Final
        │  → Head-to-head win probability from stacking model
        ▼
⑫ Final Submission
           → Retrain stacking ensemble on full training set
           → Output: winner_probability for all 250 test teams

📈 Results

Cross-Validation Ranking (5-Fold Stratified)

Rank	Model	CV Mean Accuracy	CV Std
🥇 1	Stacking Ensemble	66.00%	±2.35%
2	Extra Trees	65.00%	±3.35%
3	Random Forest	64.80%	±2.77%
4	HistGradientBoosting	64.40%	±2.98%
5	MLP Neural Net	64.00%	±3.10%
6	Logistic Regression	63.50%	±2.90%
7	SVM (RBF)	63.00%	±3.20%

Validation Set Performance (Optimised Threshold)

Rank	Model	Accuracy	AUC-ROC
🥇 1	Stacking Ensemble	67.50%	0.6846
2	Extra Trees	66.50%	0.6861
3	MLP Neural Net	65.00%	0.6863
4	Random Forest	65.00%	0.6699
5	SVM (RBF)	64.50%	0.6521
6	Logistic Regression	63.50%	0.6762
7	HistGradientBoosting	64.50%	0.6720

Classification Report - Stacking Ensemble

              precision    recall  f1-score   support

  Not Winner       0.71      0.64      0.67       105
      Winner       0.64      0.72      0.68        95

    accuracy                           0.68       200
   macro avg       0.68      0.68      0.67       200
weighted avg       0.68      0.68      0.67       200

Key Takeaways

Stacking ensemble leads on both accuracy (67.5%) and CV stability - combining 5 diverse base learners extracts signal that no single model captures alone
Threshold tuning matters - scanning 0.35→0.70 instead of defaulting to 0.50 provides a consistent accuracy gain
Extra Trees and MLP achieve the highest AUC (0.6861, 0.6863) - meaning their probability rankings are well-ordered even if raw accuracy lags
Data leakage fix (Bug #1) was the single most impactful correction - improper normalization using test-set statistics gave false confidence in v1
Confederation encoding (Bug #2) adds a meaningful signal: CONMEBOL and UEFA confederations historically dominate
The strong consistency of the stacking ensemble across all 5 CV folds confirms it generalises, not just overfits to validation

🤖 Models Trained

Six base classifiers with optimised hyperparameters:

base_models = {
    'Random Forest': RandomForestClassifier(
        n_estimators=800, max_depth=None,
        min_samples_leaf=1, max_features='sqrt', random_state=42),

    'Extra Trees': ExtraTreesClassifier(
        n_estimators=800, max_depth=None,
        min_samples_leaf=1, max_features='sqrt', random_state=42),

    'HistGradientBoosting': HistGradientBoostingClassifier(
        max_iter=500, learning_rate=0.03,
        max_depth=6, min_samples_leaf=10, random_state=42),

    'MLP Neural Net': MLPClassifier(
        hidden_layer_sizes=(256, 128, 64), activation='relu',
        max_iter=600, early_stopping=True, alpha=0.001, random_state=42),

    'SVM (RBF)': SVC(
        C=10, kernel='rbf', probability=True,
        gamma='scale', random_state=42),

    'Logistic Regression': LogisticRegression(
        C=1.0, max_iter=2000, random_state=42),
}

Models in {MLP, SVM, Logistic Regression} receive StandardScaler-transformed input.
Tree-based models use raw feature values.

🔗 Stacking Ensemble & Threshold Optimization

Stacking Architecture

┌─────────────────────────────────────────────────────────────────┐
│                   LEVEL 0 - BASE LEARNERS                       │
│                                                                 │
│  ┌──────────────┐   ┌─────────────┐   ┌──────────────────────┐  │
│  │ Random Forest│   │ Extra Trees │   │ HistGradientBoosting │  │
│  └──────┬───────┘   └──────┬──────┘   └──────────┬───────────┘  │
│         │                  │                     │              │
│  ┌──────────────┐   ┌─────────────┐              │              │
│  │  MLP Neural  │   │  SVM (RBF)  │              │              │
│  │  Net+Scaler  │   │   +Scaler   │              │              │
│  └──────┬───────┘   └──────┬──────┘              │              │
│         └──────────────────┴─────────────────────┘              │
│                            │                                    │
│               5-fold OOF predict_proba                          │
└────────────────────────────┼────────────────────────────────────┘
                             │ meta-features
┌────────────────────────────▼────────────────────────────────────┐
│                  LEVEL 1 - META LEARNER                         │
│               Logistic Regression  (C=1.0)                      │
│          Learns optimal blending weights from OOF preds         │
└─────────────────────────────────────────────────────────────────┘

Threshold Optimization

Instead of defaulting to 0.50, the decision boundary is scanned from 0.35 → 0.70 in steps of 0.01. The threshold maximising validation accuracy is selected and applied at inference time.

🔍 Feature Importance & Explainability

MDI Feature Importance (Random Forest)

Top 20 features ranked by Mean Decrease in Impurity.
🟢 Green = new v2 engineered feature · 🔵 Blue = original raw feature

New engineered features (form_x_rating, rank_x_rating, value_x_rating, attack_index) appear in the top 10 - validating the feature engineering effort.

Permutation Importance (Model-Agnostic)

Features ranked by mean AUC decrease when randomly shuffled (20 repeats).
Error bars show variance - more reliable than MDI for correlated features.

🏟️ Monte Carlo Tournament Simulation

The full FIFA World Cup 2026 bracket is simulated 5,000 times using model-derived win probabilities:

Phase 1 — Group Stage
  12 groups × 4 teams → full round-robin within each group
  Top 2 teams per group advance → 24 qualifiers

Phase 2 — Knockout Rounds
  Round of 32 → Round of 16 → Quarter-Finals → Semi-Finals → Final

Match win probability:
  P(team_A wins) = win_prob_A / (win_prob_A + win_prob_B)
  Head-to-head normalisation from stacking ensemble output

Predicted Champions - 5,000 Simulations

Top-20 Contenders by Confederation

📈 v1 vs v2 - Upgrade Summary

╔════════════════════════════════════════════════════════════════╗
║              GoalIQ 2026 - v1 vs v2 Comparison                 ║
╠══════════════════════════════╦══════════╦═════════╦════════════╣
║ Metric                       ║    v1    ║   v2    ║    Delta   ║
╠══════════════════════════════╬══════════╬═════════╬════════════╣
║ Validation Accuracy          ║  0.6400  ║ 0.6750  ║   +0.035   ║
║ Validation AUC-ROC           ║  0.6699  ║ 0.6846  ║   +0.015   ║
║ Feature Count                ║    31    ║   44    ║    +13     ║
║ Data Leakage                 ║   YES    ║   NO    ║   Fixed    ║
║ Confederation encoded        ║    NO    ║   YES   ║   Fixed    ║
║ FIFA Rank direction correct  ║    NO    ║   YES   ║   Fixed    ║
║ Ensemble Type                ║  Voting  ║Stacking ║  Upgraded  ║
║ Threshold Optimised          ║    NO    ║   YES   ║   Fixed    ║
║ Models                       ║    4     ║    6    ║    +2      ║
║ Syntax errors                ║    1     ║    0    ║   Fixed    ║
╚══════════════════════════════╩══════════╩═════════╩════════════╝

⚠️ Honest Accuracy Note

This section is intentionally included to explain the model's accuracy ceiling — a standard that separates honest ML work from inflated benchmarks.

Dataset Properties
──────────────────────────────────────────────────────────────────
  Max individual feature correlation with target  :  0.376
  Class balance                                   :  47.3% positive
  Training samples                                :  1,000
──────────────────────────────────────────────────────────────────
  → Inherent noise ceiling ≈ 0.68 – 0.72 accuracy

  Claiming 85–95% accuracy on this data would require ONE of:
    (a) Severe overfitting to the validation set
    (b) Target leakage (using future info at train time)
    (c) Evaluating on training data instead of held-out data
    (d) A fundamentally richer / larger dataset
──────────────────────────────────────────────────────────────────

GoalIQ 2026 achieves the best honest accuracy possible on this dataset and documents the ceiling transparently - the correct scientific approach.

✨ Highlights

Comprehensive Exploratory Data Analysis with correlation ceiling analysis
4 real bugs identified and fixed from the original notebook including data leakage
21 domain-informed engineered features across attack, defence, form, squad, and interaction categories
6 base model benchmarking with a full stacking ensemble
Optimal decision threshold scanning rather than naive 0.50 cutoff
5-fold stratified cross-validation on every model for unbiased generalization estimates
Permutation importance as a model-agnostic alternative to MDI
5,000-run Monte Carlo bracket simulation of the full 48-team tournament
Honest accuracy reporting - dataset ceiling documented and explained

🛠️ Tech Stack

Library	Purpose
`pandas`	Data loading, cleaning, manipulation
`numpy`	Numerical operations and array math
`matplotlib` / `seaborn`	All EDA and results visualisations
`scikit-learn`	Preprocessing, all models, StackingClassifier, GridSearchCV, evaluation
`jupyter`	Interactive development environment

🚀 Getting Started

1. Clone the repository

git clone https://github.com/<your-username>/goaliq-2026.git
cd goaliq-2026

2. (Optional) Create a virtual environment

python -m venv venv
source venv/bin/activate        # macOS / Linux
venv\Scripts\activate           # Windows

3. Install dependencies

pip install -r requirements.txt

4. Download the dataset

kaggle datasets download -d rauffauzanrambe/fifa-world-cup-2026-prediction-system
unzip fifa-world-cup-2026-prediction-system.zip -d data/raw/

Or place files manually in data/raw/:

data/raw/
├── train (1).csv
├── test (2).csv
└── submission (17).csv

5. Launch the notebook

jupyter notebook GoalIQ_2026_v2.ipynb

If running on Kaggle, all dataset paths are pre-configured - no changes needed.

📂 Project Structure

goaliq-2026/
│
├── GoalIQ_2026_v2.ipynb              # Main notebook — full 12-step pipeline
├── README.md                          # This file
├── requirements.txt                   # Python dependencies
│
├── data/
│   └── raw/                           # Original CSVs (add via Kaggle API)
│       ├── train (1).csv
│       ├── test (2).csv
│       └── submission (17).csv
│
├── assets/                            # All figures referenced in this README
│   ├── banner.png                     # Header banner image
│   ├── fig1_correlations.png          # Feature correlation bar chart
│   ├── fig2_distributions.png         # Feature distribution grid (winners vs non)
│   ├── fig3_confederation.png         # Win rate by confederation
│   ├── fig4_profiles.png              # Elite team feature profiles
│   ├── fig4_threshold.png             # Threshold optimization curve
│   ├── fig5_model_comparison.png      # Model accuracy/AUC side-by-side
│   ├── fig6_roc.png                   # ROC curves — all 7 models
│   ├── fig7_confusion.png             # Confusion matrix — stacking ensemble
│   ├── fig8_calibration.png           # Calibration curves
│   ├── fig9_importance.png            # MDI feature importance (top 20)
│   ├── fig10_perm_importance.png      # Permutation importance ± std
│   ├── fig11_simulation.png           # Monte Carlo champion probability chart
│   └── fig12_confederation_pie.png    # Top-20 confederation breakdown
│
└── outputs/
    └── submission_goaliq_v2.csv       # Final predicted win probabilities

To populate assets/: Run the notebook end-to-end. All figures are saved automatically
to /tmp/fig*.png during execution. Copy them to assets/ before pushing to GitHub.

⚙️ Requirements

pandas>=1.5.0
numpy>=1.23.0
scikit-learn>=1.3.0
matplotlib>=3.6.0
seaborn>=0.12.0
jupyter>=1.0.0

📚 What You'll Learn

This notebook is a strong portfolio reference for:

Data leakage detection - identifying and fixing train/test contamination in normalization
Categorical encoding strategy - when to encode vs drop features
Domain-driven feature engineering - building football-specific metrics from raw stats
Multi-model benchmarking - comparing 6 classifiers on the same train/val split fairly
Stacking ensembles - OOF meta-feature generation with StackingClassifier
Threshold optimization - scanning decision boundaries instead of defaulting to 0.50
Calibration curves - understanding whether predicted probabilities are trustworthy
Permutation importance - model-agnostic feature ranking as an alternative to MDI
Monte Carlo simulation - probabilistic bracket simulation with 5,000 iterations
Honest benchmark reporting - documenting accuracy ceilings and avoiding inflated claims

🔮 Future Enhancements

🌐 Streamlit web app - let users simulate their own WC 2026 bracket with live predictions
📊 SHAP explainability - per-prediction feature attribution for individual match outcomes
🔁 Repeated stratified K-Fold - tighter confidence intervals on CV estimates
⚽ Head-to-head historical data - enrich features with direct matchup records
📈 Larger dataset - incorporate more historical World Cup and qualifying match data
🏥 XGBoost / LightGBM - add gradient boosting libraries once available in environment
📱 REST API - Flask/FastAPI endpoint for real-time match prediction integration
🗓️ Live updates - re-train as WC 2026 qualifying results come in

🤝 Contributing

Contributions are welcome!

Fork the repository
Create a feature branch (git checkout -b feature/your-feature)
Commit your changes (git commit -m 'Add your feature')
Push to the branch (git push origin feature/your-feature)
Open a Pull Request

🙏 Acknowledgements

Rauf Fauzan Rambe for the dataset on Kaggle
FIFA World Cup 2026 first 48-team edition, hosted by USA · Canada · Mexico
Built with scikit-learn, pandas, matplotlib, and seaborn

👨‍💻 Author

Your Name

GitHub: @MusaIslamFahade
Kaggle: @mdmusaislamfahad01

Made with ⚽ + 🤖 for the beautiful game

GoalIQ 2026 - Where football passion meets data science

⭐ If this project helped your learning or research, a star would mean a lot. Thank you!

Name		Name	Last commit message	Last commit date
Latest commit History 7 Commits
assets		assets
README.md		README.md
fifa-world-cup-2026-ml-prediction-system-v1.ipynb		fifa-world-cup-2026-ml-prediction-system-v1.ipynb
fifa-world-cup-2026-ml-prediction-system-v2.ipynb		fifa-world-cup-2026-ml-prediction-system-v2.ipynb

Folders and files

Latest commit

History

Repository files navigation

🏆 GoalIQ 2026 - FIFA World Cup ML Prediction System

📑 Table of Contents

📌 Project Overview

📓 Open the Notebook

🖼️ Results & Visualizations

📊 Exploratory Data Analysis

🤖 Model Performance & Evaluation

🔍 Feature Importance & Explainability

🏟️ Tournament Simulation

📊 Dataset

🐛 Bug Audit (v1 → v2)

⚙️ Feature Engineering

🔬 ML Pipeline

📈 Results

Cross-Validation Ranking (5-Fold Stratified)

Validation Set Performance (Optimised Threshold)

Classification Report - Stacking Ensemble

Key Takeaways

🤖 Models Trained

🔗 Stacking Ensemble & Threshold Optimization

Stacking Architecture

Threshold Optimization

🔍 Feature Importance & Explainability

MDI Feature Importance (Random Forest)

Permutation Importance (Model-Agnostic)

🏟️ Monte Carlo Tournament Simulation

Predicted Champions - 5,000 Simulations

Top-20 Contenders by Confederation

📈 v1 vs v2 - Upgrade Summary

⚠️ Honest Accuracy Note

✨ Highlights

🛠️ Tech Stack

🚀 Getting Started

1. Clone the repository

2. (Optional) Create a virtual environment

3. Install dependencies

4. Download the dataset

5. Launch the notebook

📂 Project Structure

⚙️ Requirements

📚 What You'll Learn

🔮 Future Enhancements

🤝 Contributing

🙏 Acknowledgements

👨‍💻 Author

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages