Exponential smoothing of realized portfolio weights

The model-free exponential smoothing (ES) approach is a simple and robust way to make forecasts of random vectors. In this paper we investigate ES predictors for weights of high-dimensional realized global minimum variance portfolio (GMVP) which depend only on a realized covariance matrix of financial risky assets. We contrast a direct ES prediction of realized GMVP proportions and an indirect ES forecast, where smoothing is applied to realized covariance matrices and the GMVP composition is computed afterwards. We show analytically that either direct or indirect ES predictors of GMVP proportions could be advantageous but neither of them dominates. For this reason we suggest a dynamic time series approach in order to combine them. We illustrate our findings in an empirical study for GMVPs based on 100 risky assets and report that the proposed ES forecast combination is suitable for GMVP prediction.