WorksEconomics, Econometrics and Finance
Exploiting Common Volatility Dynamics in High-Dimensional Portfolio Selection
Abstract
We study how commonalities in monthly realized stock volatilities can be exploited for high-dimensional covariance forecasting and portfolio selection. We compare latent factor models with ARMA models that pool dynamic parameters within pre-specified groups of stocks. We show that pooled ARMA models admit an exact dynamic factor representation with group-specific dynamics and stock-specific loadings. The implied static factor representation involves group-specific factors that need not be pervasive. Combining volatility forecasts with realized dynamic conditional correlations and nonlinear shrinkage, we evaluate out-of-sample minimum variance portfolios of S&P 500 constituents. Pooling-based approaches consistently outperform the alternatives across portfolio constraints and model specifications.