We propose a new method for multivariate forecasting which combines the Generalized Dynamic Factor Model (GDFM) and the multivariate Generalized Autoregressive Conditionally Heteroskedastic (GARCH) model. We assume that the dynamic common factors are conditionally heteroskedastic. The GDFM, applied to a large number of series, captures the multivariate information and disentangles the common and the idiosyncratic part of each series; it also provides a first identification and estimation of the dynamic factors governing the data set. A time-varying correlation GARCH model applied on the estimated dynamic factors finds the parameters governing their covariances' evolution. A method is suggested for estimating and predicting conditional variances and covariances of the original data series. We suggest also a modified version of the Kalman filter as a way to get a more precise estimation of the static and dynamic factors' in-sample levels and covariances in order to achieve better forecasts. Simulation results on different panels with large time and cross sections are presented. Finally, we carry out an empirical application aiming at comparing estimates and predictions of the volatility of financial asset returns. The Dynamic Factor GARCH model outperforms the univariate GARCH.