Hybrid Baseball Game Result Prediction using KcBERT and TabNet Stacking Ensemble: Evidence from the KBO League

Key Points

• To propose a new game outcome prediction framework integrating structured and unstructured data.
• Integrated KBO League data from 2020-2024 with 1 million posts from the MLBPARK community.
• Utilized KcBERT for quantifying fan sentiment and developed a TabNet stacking ensemble model.
• Generated lag variables for public sentiment and conducted seed initialization experiments to address model sensitivity.
• Achieved 70.2% accuracy with the proposed model, outperforming the sentiment-excluded model's 58.60% accuracy by 11.6%.
• Pre-game public sentiment showed statistically significant influence on prediction outcomes, suggesting it as a valuable predictor.