Skip to content
All library documents

Constructing a Ledoit–Wolf Constant-Correlation Shrinkage Target

Article Quant Q&A · Author: Kkk
Read the original

Summary

This note explains how to build a constant-correlation covariance target from a sample covariance matrix, then blend that target with the sample estimate using a chosen shrinkage weight. It first converts each off-diagonal covariance into a correlation, averages those correlations, and uses the average to reconstruct target covariances while retaining the sample variances on the diagonal. The final estimate is an elementwise weighted average of the sample matrix and target.

A three-asset numerical example illustrates the steps and shows the target matrix and a 50/50 blend. The calculation clarifies that the shrinkage parameter controls the relative contribution of the sample estimate and the structured prior. It assumes the shrinkage weight is already known and does not explain how to estimate an optimal weight from data. The example is calculated by hand, and its author explicitly cautions that arithmetic errors are possible; it should therefore be treated as an illustration of the procedure rather than a verified numerical result.

Key ideas

  • Convert sample covariances to pairwise correlations before computing their average.
  • Use the average pairwise correlation to form off-diagonal entries of the constant-correlation target.
  • Keep the sample variances on the target matrix diagonal.
  • Blend the sample covariance matrix and target element by element according to the shrinkage weight.
  • The example assumes the shrinkage weight is given and warns that its hand calculations may contain errors.

Tags

StatisticsPortfolio constructionRisk management
Full text
# Ledoit & Wolf (2003) shrinkage approach


# Ledoit & Wolf (2003) shrinkage approach












http://www.ledoit.net/honey.pdf In the case where you have one sample covariance matrix S and an optimal shrinkage parameter and you want to estimate the covariance matrix resulting from the Ledoit and Wolf approach how do you estimate this matrix given that you want to use the constant (average) correlation model as a prior?

for example:

```
delta = 0.5;
S = [0.0309 0.0056 0.0011; 
     0.0056 0.0739 0.0148; 
     0.0011 0.0148 0.0489]
```

## Answer by nbbo2 (score 1)

https://quant.stackexchange.com/a/63573

The calculations are explained in Appendix A on page 12.

The three estimates of correlation are:

$r_{12}=\frac{s_{12}}{\sqrt{s_{11} s_{22}}}=\frac{0.0056}{\sqrt{0.0309 \times 0.0739}}=0.117189$

$r_{23}=\frac{0.0148}{\sqrt{0.0739 \times 0.0489}}=0.246198$

$r_{13}=\frac{0.0011}{\sqrt{0.0309 \times 0.0489}}=0.028298$

Next we will impose that the common correlation $\bar{r}$ is the average of these 3 values.

$\bar{r}=\frac{1}{3}(0.117189+0.246198+0.028298)=0.130562$

Now I can show you what the matrix F (the constant correlation covariance matrix) looks like. The diagonal entries $f_{ii}$ are taken unchanged from your sample matrix, the off diagonal are found from $f_{ij}=\bar{r}\sqrt{\sigma_{ii} \sigma_{jj}}$

```
F=[0.0309,0.006239,0.005075;
   0.006239,0.0739,0.007849;
   0.005075,0.007849,0.0489]
```

This the shrinkage target.

To find the final result of the Ledoit Wolf approach we have to take a 50/50 compromise (element by element) between your matrix $S$ and the matrix $F$.

$\delta S + (1-\delta) F = 0.5 S + 0.5 F$

(Apologies, I am doing these calculations by hand so there is the possibility of an error).

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.