Weighted Ensemble을 활용한 다중 예측 기간 다중 자산 실현분산 예측

초록

This study designs and evaluates a machine learning system for multi-horizon RV forecasting across asset classes. We organize 35 features from daily OHLCV and the VIX term structure into four groups (HAR-based, HF Proxy, IV Surface, Alternative), and propose WEns, a weighted ensemble of Ridge regression and XGBoost, applied to 11 ETFs spanning equities, bonds, and commodities. Forecasting horizons cover eight trading-day intervals: 1, 5, 22, 60, 90, 120, 180, and 252. Under identical feature sets and validation procedures, WEns achieves the best performance over 1–22 and 180–252 trading days and Ridge over 60–120 trading days, ranking first in the eight horizons. Reliability and interpretability are examined through feature importance consensus, statistical significance tests, and robustness experiments on maturity mismatch, missing-value handling, and asset-specific indicator substitution.

키워드