Interpretable Machine Learning for Volatility Forecasting Under Realistic Walk-Forward Constraints

Forecasting market volatility matters for risk management, portfolio allocation, and financial monitoring [1]. This paper studies whether interpretable machine-learning methods can improve forecasts of future realized volatility under a realistic walk-forward design. Using SPY as the benchmark asset and cross-asset predictors from QQQ, IWM, TLT, GLD, and VXX, I compare HAR, GARCH(1,1), GJR-GARCH(1,1), Elastic Net, Random Forest, and XGBoost at the 5-, 10-, and 21-day horizons. Forecasts are evaluated with RMSE, MAE, QLIKE, and prediction–realization correlation under rolling re-estimation with training-sample preprocessing only. Across all three horizons, the tree-based models outperform the linear and GARCH benchmarks on the main loss metrics, with Random Forest ranking first overall and XGBoost remaining close behind. Feature-importance and diagnostic results show that these gains are tied to economically plausible predictors, especially measures of volatility persistence, Treasury-market conditions, and market stress. A stylized volatility-targeting exercise suggests that the forecasting gains also have practical value, although the best statistical model is not always the best under every economic criterion. Performance is less uniform during high-volatility episodes, and the largest realized-volatility spikes remain difficult to predict. Overall, the results suggest that interpretable machine-learning methods can improve volatility forecasting in a disciplined out-of-sample setting.