Portfolio Construction
Covariance Shrinkage for Allocation
- Published
- 02 MAR 2026
- Reading
- 1 MIN
- Status
- DEMONSTRATION ENTRY
Sample covariance matrices are noisy when assets are many and observations are few. Shrinkage trades a little bias for a lot of variance.
DEMONSTRATION ENTRY — synthetic correlation structure for illustration.
Why the sample estimate misleads
With p assets and n observations, eigenvalues of the sample covariance are dispersed beyond their true values. Optimisers amplify exactly that error.
The idea
Blend the sample estimate with a structured target. The blend weight controls the bias–variance trade-off, and can be chosen analytically or by cross-validation.
Placeholder
Add your own estimator comparison, out-of-sample protocol and references here.
Solid = positive · Hatched = negative · Hover a cell