Factor Stochastic Volatility Models for Portfolio Construction

We propose a new factor stochastic volatility model that increases the accuracy of short-term forecasts for financial assets. Our new model, called the Markov-Switching Loadings (MSL) model, extends previous models by including latent processes that control the mean vectors and covariance matrices of random sub-vectors of returns. In addition, we describe our estimation routine, a novel particle Markov chain Monte Carlo algorithm, which allows for efficient estimation of a wide range of models and requires little tuning or model-specific derivations. We give two specifications of the MSL model, and both are estimated and used to generate out-of-sample forecasts for weekly returns of Select Sector SPDR exchange-traded funds over a time window spanning the 2008 financial crisis. We examine these forecasts from a statistical perspective, as well as through a financial lens, by analyzing the returns of a hypothetical investment strategy.