Noise-Proofing Universal Portfolio Shrinkage

Battle Count: 135

Directly addresses a core problem in quantitative portfolio construction: estimating covariance matrices for Markowitz-type optimization in high-dimensional settings. The proposed AvgUPSA-AO estimator achieves a mean annualized Sharpe ratio of 2.493 vs 2.077 for UPSA, with significantly lower turnover (0.25 vs 1.71) and higher diversification (17.7 vs 6.7). Highly relevant for institutional portfolio managers, factor investors, and systematic trading desks dealing with large cross-sections of assets.

Requires eigenvalue decomposition of large covariance/correlation matrices, implementation of leave-one-out cross-validation with ridge penalties, computation of oracle eigenvalues across multiple calibration/test sub-windows, exponentially weighted moving averages, and constrained quadratic optimization for ridge weights. The mathematical framework is well-defined but involves multiple interacting components. Moderate computational cost for 153 assets but scales with portfolio size.

The methodology is described in full mathematical detail with explicit equations. The dataset (JKP US factors) is publicly available at jkpfactors.com. Key hyperparameters (penalty grid, calibration window, half-life for EWMA) are specified. However, no code repository is provided, and some implementation details (e.g., exact LOO CV fold structure) require careful reading.