Realized Volatility Forecasting: A Comparative Study of Machine Learning Models and Econometric Benchmarks

Sammanfattning

Accurate volatility forecasting is essential for financial risk management, derivative pricing, and portfolio optimization. This thesis examines whether machine learning models improve one-day-ahead realized volatility forecasts relative to standard econometric benchmarks. Using high-frequency bid-ask data for the iShares S&P 500 Value ETF (2013–2025), we construct daily realized variance from five-minute returns and evaluate forecasts in a rolling-window out-of-sample framework. We compare HAR-RV and GARCH(1,1) with Random Forest, XGBoost, LightGBM, and LSTM, using QLIKE loss and Diebold–Mariano tests. Tree-based models consistently outperform econometric benchmarks: XG-Boost achieves the lowest QLIKE (0.1219), followed by LightGBM (0.1224) and Random Forest (0.1288), versus 0.1482 for HAR-RV, with Diebold–Mariano tests confirming statistical significance at the 1% level. LSTM does not improve on a Naive benchmark, while GARCH(1,1) performs worst. Robustness checks across volatility regimes and training windows confirm that these conclusions are stable.