TheDataAthlete

LightGBM-powered football betting predictor – Over/Under, 1X2, BTTS, and clean sheet probabilities from historical match data.

---

📌 Overview

TheDataAthlete is a machine learning project that builds a unified betting model for football matches. Using a single feature set (shots, corners, Elo ratings, rolling form, etc.), the model predicts:

• Total expected goals (regression) → derived Over/Under probabilities for 0.5, 1.5, 2.5, 3.5, 4.5 markets.
• Binary outcomes: 1X, X2, 12 (no draw), BTTS, Home Clean Sheet, Away Clean Sheet.

The model was trained on the Club Football Match Data (2000–2025) and ELO ratings, achieving reliable accuracy on test sets (2023+ matches). All code is provided as a Google Colab notebook, and trained models can be downloaded and used locally.

---

✨ Features

• Single regression model for total goals → all Over/Under thresholds derived via Poisson distribution.
• Six binary LightGBM classifiers for key betting markets.
• Time‑based train/test split (no look‑ahead bias).
• Rolling feature engineering – last‑5 goals scored/conceded per team.
• Easy local inference – load the saved .pkl models and run predictions on any match.
• Visual outputs – clear probability tables and the model’s best pick.

---

🧠 Model Training

Data Sources

• MATCHES.csv – match statistics (goals, shots, corners, cards, form, results) from 2000 to 2025.
• ELO.csv – team Elo ratings over time (merged into matches).

Feature Set (example)

• Basic stats: HomeShots, AwayShots, HomeTarget, AwayTarget, HomeCorners, AwayCorners, HomeFouls, AwayFouls, cards, reds.
• Form columns: Form3Home, Form5Home, Form3Away, Form5Away.
• Elo ratings: HomeElo, AwayElo.
• Rolling averages: HomeGoalsScoredLast5, HomeGoalsConcededLast5, AwayGoalsScoredLast5, AwayGoalsConcededLast5.

Targets

• Regression: TotalGoals (FTHome + FTAway)
• Binary:
  • 1X (Home win or draw)
  • X2 (Draw or away win)
  • 12 (Home or away win – no draw)
  • HomeCleanSheet (Away goals = 0)
  • AwayCleanSheet (Home goals = 0)
  • BTTS (Both teams scored)

Training Pipeline

All steps are in the included Colab notebook (TheDataAthlete_Training.ipynb):

1. Load and merge MATCHES.csv with ELO.csv.
2. Feature engineering (rolling averages, etc.).
3. Chronological split (train < 2023, test ≥ 2023).
4. Train LightGBM regression model for total goals.
5. Train six LightGBM binary classifiers.
6. Save models and feature list via joblib.
7. Evaluate on test set (MAE, AUC, accuracy).

---

📁 Project Structure

TheDataAthlete/
├── TheDataAthlete_Training.ipynb   # Full training pipeline (Google Colab)
├── models/                         # Trained models (downloaded from Drive)
│   ├── lgb_total_goals_model.pkl
│   ├── lgb_binary_models.pkl
│   └── feature_cols.pkl
├── assets/                         # Screenshots of winning predictions
│   ├── real_madrid_bayern.png
│   ├── barca_atletico.png
│   └── psg_liverpool.png
├── predict_match.py                # Local inference script
├── requirements.txt
└── README.md

---

🚀 Usage

1. Clone the repository

git clone https://github.com/SupaTx/TheDataAthlete.git
cd TheDataAthlete

2. Install dependencies

pip install -r requirements.txt

3. Download the trained models

The models are saved in Google Drive after training. Place the three .pkl files into the models/ folder.

4. Run a prediction for a new match

Edit predict_match.py with your match’s statistics (shots, corners, Elo, etc.), then run:

python predict_match.py

Example output:

==================================================
MATCH: Barcelona vs Atlético Madrid
==================================================
Predicted total goals: 3.85

OVER/UNDER MARKETS (highest probability first):
  Over_0.5: 0.979 (97.9%)
  Over_1.5: 0.896 (89.6%)
  Over_2.5: 0.738 (73.8%)
  ...

BINARY MARKETS:
  1X: YES (0.752)
  12: YES (0.806)
  BTTS: YES (0.740)

🎯 MODEL'S BEST PICK: Over_0.5 with 97.9% probability

---

📊 Results

The model has been tested on real matches from 2023–2025. Example winning predictions (screenshots in assets/):

• Real Madrid vs Bayern Munich – Over 0.5 ✅ (97.9%)
• Barcelona vs Atlético Madrid – Over 0.5 ✅ (98.6%)
• PSG vs Liverpool – Over 0.5 ✅ (98.6%)

Always bet on the market with the highest probability – the model is calibrated to maximize confidence in that single outcome.

---

🛠 Requirements

• Python 3.8+
• pandas
• numpy
• lightgbm
• scikit-learn
• scipy
• joblib

See requirements.txt for exact versions.

---

🔮 Future Improvements

• Add expected goals (xG) data from Understat or StatsBomb.
• Implement rolling win/draw/loss streaks as features.
• Build a simple web dashboard (Streamlit) to input match stats and get predictions.
• Backtest betting ROI using historical odds.

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Data: Club Football Match Data (2000–2025)
• LightGBM developers
• Poisson distribution methodology inspired by Dixon‑Coles model

---

Made with ⚽ by TheDataAthlete