A Latent Factor Model for Forecasting Realized Variances[Stock Returns and Volatility: Pricing the Short-Run and Long-Run Components of Market Risk]

This article proposes a parsimonious model to forecast large vectors of realized variances (RVar) by exploiting their common dynamics within a latent factor structure. Their long persistence is captured by aggregating latent factors with AR(1) dynamics. The model has obvious advantages over standard autoregressive models not only in terms of parametrization, but also in terms of efficiency, when increasing the dimension of the vector, as it provides more information on the commonality of the series’ dynamics. The model easily accommodates further empirical features of RVars, such as conditional heteroskedasticity. For estimation purposes, we use the maximum likelihood method based on Kalman filter and the efficient method of moments, both being easy to implement and providing accurate estimates. Our empirical illustration on real data shows that the model we propose often outperforms standard models, most of which are, for vectors of RVar series, only implementable under heavy parametric restrictions.