A Latent Factor Model for Forecasting Realized Volatilities

This paper proposes a new method of forecasting realized volatilities by exploiting their common dynamics within a latent factor model. The main idea is to use an additive component structure to describe the long-persistence in their autocorrelation function, where the components, extracted from high-dimensional vectors of realized volatilities, follow stationary autoregressive processes of order 1. The model we propose allows also for autoregressive structures in the idiosyncratic noises and conditional hetersokedasticity. Differently from HAR and ARFIMA, our factor model profits from the high-dimensionality of the system that provides more information of the commonality of their dynamics with direct efficiency gains in the estimates and forecasts. For estimation purposes, we use the indirect inference method that is easy to implement and provides accurate estimates. We apply the new models to vectors of up to 30 daily realized volatility series of stocks composing the Dow Jones Industrial Average index and show that they outperform standard long-memory models both in-sample and out-of-sample