owenlee-5678/happy-badminton-models

Happy Badminton Prediction Models

Trained machine learning models for badminton match prediction.

Models

Simplified Ensemble (simplified_ensemble.pkl)

• Task: Binary classification (predict match winner)
• Framework: Custom stacking ensemble (LightGBM + XGBoost + CatBoost → BayesianRidge meta-learner)
• Features: 35 pre-match features (ranking, form, streak, H2H, nationality)
• Performance:
  • AUC: 0.9608
  • LogLoss: 0.2316
  • Brier Score: 0.0722

Set Count Model (set_count_model.pkl)

• Task: Binary classification (predict if match goes to 2 or 3 sets)
• Framework: StackingEnsemble (same architecture)
• Features: 31 pre-match features + historical 3-set rates
• Performance:
  • AUC: 0.6635
  • LogLoss: 0.5583

Usage

from huggingface_hub import hf_hub_download
import joblib

# Download main model
model_path = hf_hub_download(
    repo_id="owenlee-5678/happy-badminton-models",
    filename="simplified_ensemble.pkl"
)
model = joblib.load(model_path)

# Download set count model
set_count_path = hf_hub_download(
    repo_id="owenlee-5678/happy-badminton-models",
    filename="set_count_model.pkl"
)
set_count_model = joblib.load(set_count_path)

Training Data

• Source: BWF official tournament records (2019-2025)
• Matches: ~15,000 professional matches
• Split: Time-based (70% train, 15% val, 15% test)

Feature Schema

See simplified_results.json for the complete feature list.

Citation

@software{happy_badminton_2026,
  title={Happy Badminton Prediction Models},
  author={OWENLEE},
  year={2026},
  url={https://huggingface.co/owenlee-5678/happy-badminton-models}
}