The Empirical Similarity Approach for Combining Predictions of Portfolio Weights

We consider prediction of the realized global minimum variance portfolio (GMVP) weights by pairwise combining the benchmark forecast with several alternative prediction rules. Our approach to model combination relates the ideas behind the empirical similarity approach with those of the logistic threshold autoregressive (LSTAR) approach. It allows to extract in a data-driven way the proportions of optimal forecast combinations. The empirical results are based on the GMVP constructed from 100 stocks referring to the S&P 500 index.