Portfolio Efficiency with High-Dimensional Data as Conditioning Information

In this paper, we build efficient portfolios using different frameworks proposed in the literature with several datasets containing an increasing number of predictors as conditioning information. We carry an extensive empirical study to investigate several approaches to impose sparsity and dimensionality reduction, as well as possible latent factors driving the returns of the risky assets. In contrast to previous studies that made use of naive OLS and low-dimension information sets, we find that (i) accounting for large conditioning information sets, and (ii) the use of variable selection, shrinkage methods and factors models, such as the principal component regression and the partial least squares provide better out-of-sample results as measured by Sharpe ratios.