We propose a new dimension reduction procedure for portfolio optimization. The vanilla principal component analysis (PCA) restricts the portfolio on the linear subspace spanned by the PCs, which often requires many PCs to performance well. In our framework, the linear subspace is based on a most...