Portfolio Choice with Subset Combination of Characteristics

This paper proposes a novel portfolio strategy over individual stocks based on subset combination of a large number of characteristics documented to predict return. Akin to the forecast combination literature, we exploit all characteristics by combining parametric rules that include a particular subset of characteristics holding fixed the number of inclusion. The choice of subset dimension governs the shrinkage of estimated parametric rule that includes all characteristics, and trades off the efficiency and robustness of portfolio decision. Empirical application to US individual stocks using 92 characteristics shows that subset combination strategy achieves desirable return properties. It outperforms characteristics-sparse strategies based on machine learning and alternative strategies based on principal component analysis. The subset combination strategy is adapted to both mean-variance and CRRA utility and its portfolio value remains in the presence of transaction costs and post-publication decay of anomalies