TipAI - Football Score Predictor

Advanced football score prediction system using an ensemble of XGBoost, LightGBM, and CatBoost models with Optuna hyperparameter optimization.

Features

• Ensemble Stacking: Combines XGBoost, LightGBM, and CatBoost with Ridge meta-learner
• Optuna Optimization: 30 trials per model for hyperparameter tuning
• Elo Rating System: Dynamic team strength tracking (K=32, home advantage=100)
• Dixon-Coles Adjustment: Corrects for correlated low-scoring outcomes
• Advanced Features: Momentum, win/lose streaks, exponential weighted averages, attack/defense efficiency
• Elegant Web App: Streamlit interface with team comparison and batch predictions

Installation

git clone https://github.com/marcoloco23/TipAI.git
cd TipAI
pip install -r requirements.txt

Usage

Full Pipeline

Download data, preprocess, and train models:

python run.py all

Individual Commands

python run.py download    # Download match data from football-data.co.uk
python run.py preprocess  # Clean and normalize data
python run.py train       # Train ensemble models with Optuna
python run.py predict     # Make predictions

Web App

streamlit run app.py

Quick Predictions

python get_predictions.py

Project Structure

TipAI/
├── app.py                  # Streamlit web application
├── run.py                  # CLI pipeline runner
├── get_predictions.py      # Quick prediction script
├── requirements.txt        # Python dependencies
└── src/
    ├── download_data.py    # Downloads historical match data
    ├── preprocess.py       # Data cleaning and normalization
    ├── train_advanced.py   # Optuna-tuned ensemble training
    └── predict.py          # Score prediction with stacked models

Model Architecture

1. Base Models (Poisson regression):

   • XGBoost (home goals, away goals)
   • LightGBM (home goals, away goals)
   • CatBoost (home goals, away goals)

2. Meta-Learner: Ridge regression stacking

3. Score Selection:

   • Bivariate Poisson distribution for score probabilities
   • Dixon-Coles adjustment (ρ=-0.20) for low scores
   • Lambda scaling (1.8x) for prediction variance
   • Smart selection based on most likely outcome

Data

• Source: football-data.co.uk
• Leagues: Premier League, La Liga, Serie A, Bundesliga, Ligue 1, Champions League
• Matches: 18,000+ historical matches
• Features: ~70 engineered features including Elo, form, H2H stats

Requirements

• Python 3.8+
• pandas, numpy, scikit-learn
• xgboost, lightgbm, catboost
• optuna, streamlit

License

MIT