Forecasting the conditional covariance matrix of a portfolio under long-run temporal dependence

Long-range persistence in volatility is widely modelled and forecast in terms of the so-called fractional integrated models. These models are mostly applied in the univariate framework, since the extension to the multivariate context of assets portfolios, while relevant, is not straightforward. We discuss and apply a procedure which is able to forecast the multivariate volatility of a portfolio including assets with long memory. The main advantage of this model is that it is feasible enough to be applied on large-scale portfolios, solving the problem of dealing with extremely complex likelihood functions which typically arises in this context. An application of this procedure to a portfolio of five daily exchange rate series shows that the out-of-sample forecasts for the multivariate volatility are improved under several loss functions when the long-range dependence property of the portfolio assets is explicitly accounted for.