A machine learning system that predicts NHL game outcomes using advanced AI architectures. Built with transformer neural networks, quantum-inspired ensembles, and comprehensive feature engineering for professional-grade accuracy.
- Transformer Neural Networks: Multi-head attention mechanisms for complex pattern recognition
- Quantum-Inspired Ensembles: Advanced ensemble methods with LightGBM, XGBoost, and CatBoost
- 261 Advanced Features: Comprehensive feature engineering from team statistics to venue analysis
- Interactive Web Interface: Beautiful, responsive UI for team vs team predictions
- Real-time Analytics: Live NHL API integration with intelligent caching
- Venue Intelligence: Home/away advantages and playoff mode considerations
This system employs a sophisticated multi-layer approach:
- Core ML Framework: PyTorch transformers with 512-dimensional embeddings and 16 attention heads
- Ensemble Pipeline: LightGBM, XGBoost, CatBoost with intelligent weight optimization
- Feature Engineering: 43 → 261 features including rolling averages, momentum indicators, and venue analytics
- Data Pipeline: Robust NHL API integration with CSV fallback (62,890+ historical games)
- Web Framework: Flask-based interactive interface with real-time predictions
pip install -r requirements.txtpython quick_demo.pypython web_interface.pyThen open: http://localhost:5000
python advanced_nhl_predictor.pyOutput should look like this:
After installing the required prerequisites, the model initiates its self-training process, monitoring the Epoch and loss metrics as displayed below:
Actively in development! Currently resolving input encoding issues and refining preprocessing steps. Neural net architecture and training loop are functional but unvalidated. Please feel free to contribute as needed! Submit an issue or pull request.
GNU AFFERO GENERAL PUBLIC LICENSE
See the license for more details