Can Old and New Models Work Together? Predicting Volatility across U.S. Sectors with GARCH and Machine Learning

This study investigates the relative predictive importance of global risk factors for the conditional volatility of the eleven S&P 500 GICS sectors. We ask whether global risk signals contain common predictive information across sectors or whether their relevance is sector-specific and changes across market conditions. Using daily total-return sector indices, we first estimate sector-specific EGARCH models to recover conditional volatility. We then forecast next-day volatility with XGBoost using six cross-market indicators that capture equity-market uncertainty, Treasury-market volatility, global dollar conditions, energy-price shocks, safe-haven dynamics, and corporate credit stress. Forecast performance is evaluated out of sample against conventional econometric and persistence benchmarks. The empirical design treats each sector as a distinct forecasting problem rather than imposing predefined sector classifications. TreeSHAP is used to decompose model forecasts into factor-specific contributions, construct comparable sectoral risk profiles, and identify nonlinearities, threshold effects, and interactions among global risk signals. We further examine whether these profiles remain stable across market regimes and rolling estimation windows. The analysis is designed to reveal which global risks dominate the volatility forecasts of individual sectors, whether the same predictors matter uniformly across the market, and when sectoral exposures diverge from aggregate risk conditions. This sector-level perspective can uncover predictive heterogeneity that is obscured by market-wide models and clarify the extent to which broad indicators such as VIX provide an incomplete representation of sectoral risk. By combining asymmetric conditional volatility modelling, nonlinear forecasting, and time-varying explainability, the study provides a unified framework for mapping cross-market risk signals into sector-specific volatility predictions, with implications for forecasting, portfolio allocation, stress testing, and risk management.