Realized volatilities, when observed over time, share the following stylised facts: comovements, clustering, long-memory, dynamic volatility, skewness and heavy-tails. We propose a dynamic factor model that captures these stylised facts and that can be applied to vast panels of volatilities as it does not suffer from the curse of dimensionality. It is an enhanced version of Bai and Ng (2004) in the following respects: i) we allow for longmemory in both the idiosyncratic and the common components, ii) the common shocks are conditionally heteroskedastic, and iii) the idiosyncratic and common shocks are skewed and heavy-tailed. Estimation of the factors, the idiosyncratic components and the parameters is simple: principal components and low dimension maximum likelihood estimations. A Monte Carlo study shows the usefulness of the approach and an application to 90 daily realized volatilities, pertaining to S&P100, from January 2001 to December 2008, evinces, among others, the following findings: i) All the volatilities have long-memory, more than half in the nonstationary range, that increases during financial turmoils. ii) Tests and criteria point towards one dynamic common factor driving the co-movements. iii) The factor has larger long-memory than the assets volatilities, suggesting that long–memory is a market characteristic. iv) The volatility of the realized volatility is not constant and common to all. v) A forecasting horse race against 8 competing models shows that our model outperforms, in particular in periods of stress.
