A 500 by 500 sample covariance matrix estimated from 1000 returns has eigenvalues spanning 0.0858 to 2.914. You shrink it toward the identity as S* = 0.6*S + 0.4*I. What is the new condition number, and what does the shrinkage intensity trade off?

A 500 by 500 sample covariance matrix estimated from 1000 returns has eigenvalues spanning 0.0858 to 2.914. You shrink it toward the identity as S* = 0.6*S + 0.4*I. What is the new condition number, and what does the shrinkage intensity trade off?

Approach: Shrinkage toward the identity is a linear map on eigenvalues, so apply it to the two extremes. Then describe what the intensity is chosen to minimise.

4.76. Shrinkage toward the identity acts on eigenvalues alone, so 2.914 becomes 0.6*2.914 + 0.4 = 2.148 and 0.0858 becomes 0.6*0.0858 + 0.4 = 0.4515, giving a condition number of 2.148/0.4515 = 4.76 against 33.97 before. The small eigenvalues move by far the most in relative terms, which is exactly where the inversion did its damage. The Ledoit-Wolf estimator picks the intensity that minimises the expected squared Frobenius distance to the true matrix, a bias variance trade-off: the identity target carries bias and no estimation error, the sample matrix carries no bias and a great deal of error. The optimal intensity rises with N/T and falls with sample length, and it has a closed form, so no cross validation is required.

Follow-up: Why does shrinking toward a constant correlation target usually beat shrinking toward the identity for equities?

Key concepts: shrinkage intensity, ledoit-wolf estimator, bias variance trade-off, eigenvalue.