Comparison of Methods for Covariance Estimation

Mean-variance portfolio optimization is a common technique in portfolio allocation. It usually requires an estimate of the covariance matrix to compute the portfolio weights. Many methods can be used for covariance estimation. The traditional covariance estimate is computed using cov. However, in practice the standard covariance estimate is noisy. Since the optimal solution of the portfolio weights is very sensitive to the mean and covariance estimates, noise in the covariance matrix increases the portfolio weights estimation error leading to large turnover and transaction costs. Here are some methods that mitigate the estimation error:

• robustcov computes a covariance estimate from a data sample. This function reduces the noise in the estimator by identifying the outliers in the sample and removing them from the final covariance estimate.

• covarianceShrinkage computes a covariance estimate from a data sample. This function reduces the noise in the estimator generated from using a finite sample. Shrinkage estimators are used to reduce the mean squared error (MSE) of an estimate. MSE consists of two parts, the bias and the noise. Shrinkage methods reduce the noise in the estimator by pulling the standard covariance estimate towards a target matrix. Thus, the shrinkage estimator results in a biased estimate, but with a lower variance than that of the standard covariance estimate. The method implemented in covarianceShrinkage considers a multiple of the identity as the target matrix.

  In covariance shrinkage, the intensity parameter determines how much the standard covariance estimate is pulled towards the target matrix. Because computing the optimal intensity parameter depends on the sample size, covarianceDenoising runs slower when the sample size significantly increases.

• covarianceDenoising computes a covariance estimate from a data sample or from an initial covariance estimate (usually the standard covariance estimate from cov). The covarianceDenoising function reduces the noise in the estimator generated from using a finite sample. Unlike covarianceShrinkage, the covarianceDenoising function differentiates between the noise and the signal in the data to pull only the eigenvalues associated with noise towards a target value. This technique results in an estimate that decreases noise while increasing the signal.

  The computation of the covarianceDenoising estimate depends on the number of assets (variables). If the number of assets is too small, there is not enough information to identify between the noisy and meaningful factors. On the other hand, if the number of assets is too large, the underlying optimization problem used to identify the noisy eigenvalues grows, and it takes longer to compute the estimate.

The following table summarizes the information needed to compute the different covariance estimates.

Function	Number of Assets	Sample Size	Sample
cov ( MATLAB®)	Any	Any	Required
robustcov ( Statistics and Machine Learning Toolbox™)	Any	≥ 2 × numAssets	Required
covarianceShrinkage ( Financial Toolbox™)	Any (≥ 3 recommended)	Any (≤ 25,000 recommended)	Required
covarianceDenoising ( Financial Toolbox)	Any (≥ 10 recommended)	Any	Not required

Another useful tool for covariance estimation is nearcorr. Use nearcorr to compute the nearest correlation matrix by minimizing the Frobenius distance to an initial estimate. This is particularly useful to get a positive definite covariance matrix when the number of observations is less than the number of assets.