Spectral Covariance Approximation for Faster Portfolio Optimization

Summary

The report presents a way to speed up portfolio optimization while retaining a covariance estimate built with statistical shrinkage. It starts with a linearly shrunk covariance matrix, decomposes it spectrally, keeps the components associated with the largest eigenvalues, and approximates the residual component with a diagonal matrix. This produces a factor-like representation that can reduce the computational burden in optimization.

The report tests index-enhancement strategies within constituent universes and across the full market for the CSI 300 and CSI 500. It says that retaining at least 40 components yielded results broadly similar to using the original shrunk covariance estimate, while optimization with CVXPY and ECOS was about two orders of magnitude faster. The number of components trades approximation error against optimization speed. These findings are limited to the reported tests; the authors flag model failure and extreme market conditions as risks, and do not establish performance across other universes or periods.

Key ideas

• A shrunk statistical covariance matrix can be decomposed into leading spectral components and a diagonal residual.
• The resulting factor-like form can make covariance-based portfolio optimization faster.
• Keeping more leading components reduces approximation error but slows optimization.
• The reported CSI 300 and CSI 500 tests found similar outputs above 40 components and substantially faster solver runs.
• The evidence is specific to the tested portfolios and does not remove model or extreme-market risk.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.