Identity-target covariance shrinkage

Calculation contract

S_new = (1 - a)*S + a*(trace(S)/2)*I

Method and interpretation

The target uses the average of the two variances on both diagonal entries and zero off-diagonal entries. At intensity zero the matrix is unchanged; at one it becomes the target. Intermediate settings reduce the off-diagonal covariance and move each variance toward the shared mean.

The default half-strength example changes the variances to 0.0325 and 0.0175, with covariance 0.003. Preserving the trace makes the target easy to reconcile, but is not an argument that this target best represents any particular investment universe.

Where the reference stops

This is a fixed-intensity transform, not the Ledoit-Wolf estimator. It does not select a parameter, repair an indefinite starting matrix or prove better out-of-sample risk estimates. Review window, data cleaning and target suitability separately.

Published synthetic inputs

{
  "covariance": [
    [
      0.04,
      0.006
    ],
    [
      0.006,
      0.01
    ]
  ],
  "intensity": 0.5
}

Frozen calculation output

{
  "covariance": [
    [
      0.0325,
      0.003
    ],
    [
      0.003,
      0.0175
    ]
  ],
  "target_variance": 0.025
}

User-specified convex shrinkage to mean-variance identity. Intensity is not estimated by Ledoit-Wolf or selected by validation.

Verification coverage

• Both endpoints, trace preservation and symmetry tested.
• Invalid intensity and non-PSD inputs fail before calculation.

Unit and independent-reference checks test arithmetic and rejection rules, not future investment performance.

Method references

• scikit-learn: covariance estimation and shrinkage

Read the assumptions in context