설명가능 인공지능을 활용한 코스피 지수 변동성 예측 연구

Abstract

This study compares the performance of a statistical model (GARCH) and a machine learning model (XGBoost) in predicting the volatility of the KOSPI index, while employing Explainable Artificial Intelligence (XAI) to identify the key volatility drivers. Using daily data from 2000 to 2024, this study finds that XGBoost outperforms GARCH in accuracy metrics. This performance gap widens when comparing different error metrics, with XGBoost showing 1.61 times lower error in RMSE and an even greater 2.88 times improvement in MAPE, suggesting machine learning approaches better capture the complex, non-linear patterns in equity market volatility. Both feature importance analyses using gain and SHAP values consistently identify the previous day's volatility as the most critical predictor, aligning with the volatility clustering in financial theory. This paper highlights how combining machine learning with SHAP enhances both performance and interpretability in volatility forecasting, providing a practical framework for implementing explainable machine learning solutions in financial risk management.