Shrinkage Estimators for Mean and Covariance: Evidence on Portfolio Efficiency Across Market Dimensions

Battle Count: 117

Highly relevant for quantitative portfolio construction. The paper provides practical guidance on which shrinkage estimators to use for different investor profiles and market dimensions. The finding that GMV+COV2 (Ledoit-Wolf two-parameter) is optimal for most investors and COV2+SM for return-seekers is directly actionable. The high-dimensional analysis (S&P 500, RUSSELL 1000, TOPIX 1500) is particularly relevant for institutional quant strategies. However, it focuses on portfolio construction rather than trading signals or execution.

Moderate complexity. Requires solving multiple quadratic programs (MV/GMV) with various shrinkage estimators, implementing 16 different shrinkage estimators (5 mean + 11 covariance), solving DEA super-efficiency linear programs, and managing rolling window backtesting across 6 markets with 3 out-of-sample periods. The Ledoit-Wolf estimators are well-documented and available in R (HDShOP package) and Python. The main complexity lies in the comprehensive evaluation framework rather than individual components.

Intermediate results and efficiency scores are publicly available on GitHub. Data sourced from LSEG Workspace (formerly Refinitiv Eikon/Datastream), which is a commercial database. MATLAB YALMIP with MOSEK solver used for optimization. Rolling window methodology and all model specifications are clearly described. However, the raw price data is not freely available.