About This Tool

Model Architecture

• LightGBM multi-class (result & handicap)
• Bivariate Poisson (simplified Dixon-Coles)
• Ensemble: weighted average of both models
• Optuna hyperparameter tuning
• Rolling window backtest (RPS / Brier / Accuracy)

Data Sources

• Historical matches: synthetic (supports football-data.co.uk)

Tech Stack

• Python 3.11+ / SQLite
• LightGBM / scikit-learn / Optuna / statsmodels
• Flask web server

Evaluation Metrics

Metric	Full Name	Description
RPS	Ranked Probability Score	Probability calibration — lower is better
Brier	Brier Score	Mean squared error of probabilities
Accuracy	Accuracy	Direction accuracy
Sharpe	Sharpe Ratio	Risk-adjusted virtual return
MaxDD	Maximum Drawdown	Max virtual capital drawdown

Disclaimer

This tool is for research purposes only. All predictions are statistical simulations — no accuracy is guaranteed.

Users must comply with local laws. Any actions based on predictions are at the user's own risk.