Live Site: fplbangers.com It provides predictions based on custom-trained machine learning models, updated weekly. The model identifies high-potential "underdog" players ((10% Own, £7m Cost) with 'Fail-Upward' logic) with strong form or underlying stats.
Built with Antigravity, Opus, and time :)
A machine learning-powered Fantasy Premier League (FPL) assistant that identifies high-potential "underdog" players—those with less than 10% ownership and under £8.0m—who have strong form or underlying stats. Predictions are updated weekly using custom-trained models.
- Component-Based Predictions: Predicts goals, assists, and clean sheets separately using LightGBM classifiers
- Confidence Scoring: Color-coded predictions showing model certainty (green = high, yellow = medium, orange = low)
- Dynamic Thresholds: Percentile-based player selection adapts to seasonal ownership/cost shifts
- Live Match Scores: Real-time points updates during active gameweeks
- Feature Importance Dashboard: Transparent view of which stats drive predictions
- Prediction History: Track past picks with actual points backfill
- Fantasy Premier League Official API – Player metadata, fixtures, and per-gameweek points
- Understat API – Advanced metrics (xG, xA) and team defensive strength (xGA)
See run_instructions.md for detailed setup and deployment instructions.
# Install dependencies
py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
# Run full pipeline
python update_pipeline.py
# Or quick launch (app only)
python update_pipeline.py --quick- run_instructions.md – Setup, running, and deployment guide
- architectural_decisions.md – Technical design decisions and model architecture
- claude_analysis.md – System audit and improvement tracking
- Backend: Flask (Python)
- ML Models: LightGBM with TimeSeriesSplit cross-validation
- Frontend: Responsive HTML/CSS/JS with tab-based navigation
- Deployment: Northflank with Cloudflare DNS
- CI/CD: GitHub Actions for automated weekly updates
MIT