Combination Portfolio: A Cocktail Therapy for Training Portfolio Selection

Abstract

The quantitative practice of portfolio selection aims to select the in-sample optimal portfolio that is robust out of sample. However, at each estimation period, the conventional method is selection by solving a given objective function, without a learning mechanism, or training. This paper designs a method to train portfolio selection. Each optimal portfolio is a combination of three basic elements: strategy, covariance matrix, and risk type; therefore, like the cocktail therapy, we propose a selection method by first augmenting the combination to 250 optimal portfolios at each estimation period, and then we propose a score to select the best portfolio to hold in the next period. In the machine learning literature, combining multiple models together in some way has proven to be useful. Such combinations of models are sometimes called committees. We show that the combination portfolio exhibits superior performance and robustness, and the alternative way to train portfolio selection is useful.