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

The mean-variance model remains the most prevalent investment framework, built on diversification principles. However, it consistently struggles with estimation errors in expected returns and the covariance matrix, its core parameters. To address this concern, this research evaluates the performance of mean variance (MV) and global minimum-variance (GMV) models across various shrinkage estimators designed to improve these parameters. Specifically, we examine five shrinkage estimators for expected returns and eleven for the covariance matrix. To compare multiple portfolios, we employ a super efficient data envelopment analysis model to rank the portfolios according to investors risk-return preferences. Our comprehensive empirical investigation utilizes six real world datasets with different dimensional characteristics, applying a rolling window methodology across three out of sample testing periods. Following the ranking process, we examine the chosen shrinkage based MV or GMV portfolios against five traditional portfolio optimization techniques classical MV and GMV for sample estimates, MiniMax, conditional value at risk, and semi mean absolute deviation risk measures. Our empirical findings reveal that, in most scenarios, the GMV model combined with the Ledoit Wolf two parameter shrinkage covariance estimator (COV2) represents the optimal selection for a broad spectrum of investors. Meanwhile, the MV model utilizing COV2 alongside the sample mean (SM) proves more suitable for return oriented investors. These two identified models demonstrate superior performance compared to traditional benchmark approaches. Overall, this study lays the groundwork for a more comprehensive understanding of how specific shrinkage models perform across diverse investor profiles and market setups.

💡 Research Summary

The paper investigates how shrinkage estimators for the mean vector and covariance matrix affect the performance of mean‑variance (MV) and global minimum‑variance (GMV) portfolio models. Five shrinkage methods for expected returns (James‑Stein, Bayes‑Stein, Quadratic, Bodnar‑optimal linear, and a sample‑mean baseline) and eleven for the covariance matrix (including the Ledoit‑Wolf linear and two‑parameter estimators, several nonlinear variants, Hilbert‑transform based shrinkage, and inverse‑eigenvalue shrinkage) are combined with the two portfolio optimization frameworks.

Six equity markets are examined: three low‑dimensional (Dow Jones 30, NIFTY 50, FTSE 100) where the number of assets p is smaller than the observation window n, and three high‑dimensional (S&P 500, Russell 1000, TOPIX 1500) where p > n. Daily closing prices spanning roughly eleven years (Sept 2012 – Jun 2024) are used. A rolling‑window scheme with a one‑year in‑sample period and three out‑of‑sample horizons (3‑month, 6‑month, 1‑year) generates three testing windows for each market. For each window, every combination of mean‑shrinkage and covariance‑shrinkage is applied to both MV and GMV, yielding 66 portfolios for low‑dimensional markets and 48 for high‑dimensional ones (only eight covariance shrinkage estimators are feasible when p > n).

To rank the resulting portfolios, the authors employ a super‑efficiency Data Envelopment Analysis (DEA) model. The DEA treats each portfolio as a decision‑making unit (DMU) and uses a set of risk‑related inputs (e.g., volatility, CVaR, maximum drawdown) and return‑related outputs (e.g., mean return, Sharpe ratio, reward‑to‑risk ratios). Three investor groups are defined: Group A (balanced risk‑return), Group B (return‑seeking/aggressive), and Group C (risk‑averse). For each market, group, and out‑of‑sample horizon, portfolios are ordered by DEA efficiency; the top ten are identified, and from these the “market best” (most frequently top‑ten across the three horizons for a given market) and the “universal best” (appearing most often across all markets) are selected.

The selected shrinkage‑based portfolios are then benchmarked against five traditional optimization approaches: classical MV and GMV using sample estimates, a MiniMax risk‑minimization model, a Conditional Value‑at‑Risk (CVaR) model, and a semi‑mean‑absolute‑deviation (SMAD) model. Performance is evaluated using out‑of‑sample returns, risk metrics, and DEA efficiency scores.

Key empirical findings:

1. Covariance shrinkage dominates – Across almost all markets and investor groups, the choice of covariance estimator has a larger impact on portfolio efficiency than the choice of mean estimator. The Ledoit‑Wolf two‑parameter shrinkage estimator (COV2) appears in the majority of top‑ranked portfolios.

2. GMV + COV2 is the most robust – For balanced (Group A) and risk‑averse (Group C) investors, the GMV model combined with COV2 consistently yields the highest DEA efficiency and superior out‑of‑sample risk‑adjusted returns.

3. MV + COV2 + sample mean suits return‑seekers – Aggressive investors (Group B) benefit most from the MV framework that uses the sample mean (SM) together with COV2. This configuration delivers higher average returns while maintaining acceptable risk levels.

4. Traditional sample‑based MV/GMV underperform – Portfolios built on raw sample estimates of mean and covariance perform poorly, especially in high‑dimensional settings where over‑fitting and singular covariance matrices are common.

5. Alternative risk‑based models are competitive only locally – MiniMax, CVaR, and SMAD sometimes outperform shrinkage‑based portfolios in specific periods or markets, but on average they lag behind the COV2‑based solutions.

The study concludes that shrinkage estimators, particularly the Ledoit‑Wolf two‑parameter covariance shrinkage, substantially improve portfolio construction in both low‑ and high‑dimensional environments. Moreover, the optimal combination of mean and covariance shrinkage depends on the investor’s risk‑return profile, underscoring the need for a tailored approach. The DEA super‑efficiency framework proves valuable for multi‑criteria ranking, offering a more nuanced assessment than single‑metric comparisons.

Future research directions suggested include extending shrinkage to multi‑target frameworks, incorporating time‑varying or regime‑switching shrinkage intensities, and developing real‑time DEA‑driven rebalancing algorithms to further enhance out‑of‑sample performance.