Dynamic Modeling of Large Dimensional Covariance Matrices

Modelling and forecasting the covariance of financial return series has always been a challange due to the so-called "curse of dimensionality". This paper proposes a methodology that is applicable in large dimensional cases and is based on a time series of realized covariance matrices. Some solutions are also presented to the problem of non-positive definite forecasts. This methodology is then compared to some traditional models on the basis of its forecasting performance employing Diebold-Mariano tests. We show that our approach is better suited to capture the dynamic features of volatilities and covolatilities compared to the sample covariance based models.