Portfolio Choice with Subset Combination of Characteristics

Esfandiar Essie Maasoumi,Guoshi Tong,Xudong Wen,Ke Wu

DOI: https://doi.org/10.2139/ssrn.4051796

2022-01-01

SSRN Electronic Journal

Abstract:This paper proposes a novel portfolio strategy over a large number of asset characteristics. This compares with high dimensional "hedonic'' predictive regressions, but with model uncertainty. We consider aggregation strategies over subsets of characteristics similar, in spirit, to forecast combination and shrinkage. We avoid some shortcomings of model and feature selection common in machine learning (ML) approaches. The problem of biased over elimination of closely related features in naive LASSO type approaches is avoided by combination of subsets of characteristics. There is no "true" subset model, as is assumed in penalization and shrinkage methods. Empirical application to US stocks with 92 characteristics shows that our aggregation strategies outperform popular characteristics-sparse strategies. The subset combination strategy can be adapted to any utility function, as demonstrated with mean-variance and CRRA preferences. Its superiority is robust to inclusion of transaction costs and post-publication decay of anomalies.