7.7 Sampling distributions of \(\hat{\mu}\) and \(\hat{\Sigma}\)

Consider the GWN model in matrix form (7.1). The estimator of the \(N \times 1\) matrix \(\mu\) is the sample mean vector (7.16) and the estimator of the \(N \times N\) matrix \(\Sigma\) is the sample covariance matrix (7.17). In this section we give the sampling distributions for the mean vector \(\hat{\mu}\) and for the elements of the covariance matrix \(\hat{\Sigma}\).

7.7.1 Sampling distribution of \(\hat{\mu}\)

The following proposition gives the sampling distribution of the \(N \times 1\) vector \(\hat{\mu}\).

Proposition 7.8 (Sampling distribution of \(\hat{\mu}\)) Under the GWN model and for large enough T, the sampling distribution for \(\hat{\mu}\) is multivariate normal:

\[\begin{equation} \hat{\mu} \sim N\left(\mu, ~ \frac{1}{T}\hat{\Sigma}\right). \end{equation}\]

This result shows that the elements of the vector \(\hat{\mu}\) are correlated, with the correlations being the same as the asset return correlations.

Example 3.9 (Joint distribution of \(\hat{\mu}_1\) and \(\hat{\mu}_2\))

Consider the bivariate case (\(N=2\)). Proposition 7.8 tells us the joint distribution between \(\hat{\mu}_1\) and \(\hat{\mu}_2\) is bivariate normal:

\[\begin{equation} \left( \begin{array}{c} \hat{\mu}_1 \\ \hat{\mu}_2 \end{array} \right) \sim N\left( \left(\begin{array}{c} \mu_1 \\ \mu_2 \end{array} \right), ~ \frac{1}{T} \left( \begin{array}{cc} \hat{\sigma}_1^2 & \hat{\sigma}_{12} \\ \hat{\sigma}_{12} & \hat{\sigma}_{2}^{2} \end{array} \right) \right). \end{equation}\]

Hence, \(\widehat{\mathrm{cov}}(\hat{\mu}_1, \hat{\mu}_2) = \frac{1}{T}\hat{\sigma}_{12}\) and \(\widehat{\mathrm{cor}}(\hat{\mu}_1, \hat{\mu}_2)=\hat{\rho}_{12}\).

The fact that the elements of \(\hat{\mu}\) are correlated and the correlations are equal to the correlations between the corresponding assets is a useful result. For example, if assets 1 and 2 are strongly positively (negatively) correlated, then \(\hat{\mu}_1\) and \(\hat{\mu}_2\) are strongly positively (negatively) correlated. As a result, a positive (negative) estimate of \(\hat{\mu}_1\) will tend to be associated with a positive (negative) estimate of \(\hat{\mu}_2\).


7.7.2 Sampling distribution of the elements of \(\hat{\Sigma}\)

We can also determine the sampling distribution for the vector of unique elements of \(\hat{\Sigma}\). This requires a bit of new matrix notation because the joint distribution of a \(N \times N\) symmetric matrix of random variables is defined as the joint distribution of the \(N(N+1)/2\) elements of the matrix stacked into a \(N(N+1)/2 \times 1\) vector.

Define \(\hat{\theta}\) as the \(N(N+1)/2 \times 1\) vector of unique elements of \(\Sigma\). These elements can be extracted using the \(\mathrm{vech}(\cdot)\) operator, which extracts the elements on and below the main diagonal of \(\hat{\Sigma}\) column-wise and stacks them into a \(N(N+1)/2\) vector. For example, if \(N=3\) then \(\hat{\theta}\) is the \(6 \times 1\) vector

\[\begin{equation} \hat{\theta} = \mathrm{vech}(\hat{\Sigma}) = \mathrm{vech} \left(\begin{array}{ccc} \hat{\sigma}_1^2 & \hat{\sigma}_{12} & \hat{\sigma}_{13} \\ \hat{\sigma}_{12} & \hat{\sigma}_2^2 & \hat{\sigma}_{23} \\ \hat{\sigma}_{13} & \hat{\sigma}_{23} & \hat{\sigma}_{3}^2 \end{array} \right) = \left(\begin{array}{c} \hat{\sigma}_1^2 \\ \hat{\sigma}_{12}\\ \hat{\sigma}_{13}\\ \hat{\sigma}_2^2\\ \hat{\sigma}_{23}\\ \hat{\sigma}_{3}^2 \end{array} \right). \end{equation}\]

Proposition 7.9 (Sampling distribution of the unique elements of \(\hat{\Sigma}\)) Under the GWN model and for large enough T, the sampling distribution for the \(N(N+1)/2 \times 1\) vector \(\hat{\theta} = \mathrm{vech}(\hat{\Sigma})\) is multivariate normal:

\[\begin{equation} \hat{\theta} \sim N\left(\theta, ~ \frac{1}{T}\hat{\Omega}\right). \end{equation}\]

where \(\frac{1}{T}\hat{\Omega}\) is an \(N(N+1)/2 \times N(N+1)/2\) matrix giving the estimated variances and covariances of the unique elements of \(\hat{\Sigma}\). The element of \(\frac{1}{T}\hat{\Omega}\) corresponding to \(\widehat{\mathrm{cov}}(\hat{\sigma}_{ij},\hat{\sigma}_{lm})\) is given by \(T^{-1}(\hat{\sigma}_{il}\hat{\sigma}_{jm} + \hat{\sigma}_{im}\hat{\sigma}_{jl})\) for all \(i,j,l,m = 1,2,\ldots,N\), including \(i=j=l=m.\) If \(i=j\) and \(l=m\) then

\[\begin{align*} \widehat{\mathrm{cov}}(\hat{\sigma}_{ij},\hat{\sigma}_{lm}) & = \widehat{\mathrm{cov}}(\hat{\sigma}_{ii},\hat{\sigma}_{ll}) = \widehat{\mathrm{cov}}(\hat{\sigma}_i^2,\hat{\sigma}_l^2) \\ & =\frac{\hat{\sigma}_{il}\hat{\sigma}_{il} + \hat{\sigma}_{il}\hat{\sigma}_{il}}{T} = \frac{\hat{\sigma}_{il}^2 + \hat{\sigma}_{il}^2}{T} = \frac{2\hat{\sigma}_{il}^2}{T} \end{align*}\].

If \(i=j=l=m\) then

\[\begin{align*} \widehat{\mathrm{cov}}(\hat{\sigma}_{ij},\hat{\sigma}_{lm}) & = \widehat{\mathrm{cov}}(\hat{\sigma}_i^2,\hat{\sigma}_i^2) = \widehat{\mathrm{var}}(\hat{\sigma}_i^2) \\ & =\frac{\hat{\sigma}_i^2\hat{\sigma}_i^2 + \hat{\sigma}_i^2\hat{\sigma}_i^2}{T} = \frac{2\hat{\sigma}_i^4}{T} \end{align*}\].

In Proposition 7.9, it is not easy to see what are the elements of \(\hat{\Omega}\). A simple example can make things more clear. Consider the bivariate case \(N=2\). Then the result in Proposition 7.9 gives:

\[\begin{align*} \hat{\theta} = \left(\begin{array}{c} \hat{\sigma}_1^2 \\ \hat{\sigma}_{12} \\ \hat{\sigma}_2^2 \end{array} \right) & \sim N\left( \left(\begin{array}{c} \sigma_1^2\\ \sigma_{12} \\ \sigma_2^2 \end{array} \right),~ \frac{1}{T}\left( \begin{array}{ccc} \hat{\omega}_{11} & \hat{\omega}_{12} & \hat{\omega}_{13} \\ \hat{\omega}_{21} & \hat{\omega}_{22} & \hat{\omega}_{23} \\ \hat{\omega}_{31} & \hat{\omega}_{32} & \hat{\omega}_{33} \end{array} \right) \right) \\ & \sim N\left( \left(\begin{array}{c} \sigma_1^2\\ \sigma_{12} \\ \sigma_2^2 \end{array} \right),~ \left( \begin{array}{ccc} \widehat{\mathrm{var}}(\hat{\sigma}_1^2) & \widehat{\mathrm{cov}}(\hat{\sigma}_1^2, \hat{\sigma}_{12}) & \widehat{\mathrm{cov}}(\hat{\sigma}_1^2, \hat{\sigma}_2^2) \\ \widehat{\mathrm{cov}}(\hat{\sigma}_{12}, \hat{\sigma}_1^2) & \widehat{\mathrm{var}}(\hat{\sigma}_{12}) & \widehat{\mathrm{cov}}(\hat{\sigma}_{12}, \hat{\sigma}_2^2) \\ \widehat{\mathrm{cov}}(\hat{\sigma}_1^2, \hat{\sigma}_2^2) & \widehat{\mathrm{cov}}(\hat{\sigma}_2^2, \hat{\sigma}_{12}) & \widehat{\mathrm{var}}(\hat{\sigma}_2^2) \end{array} \right) \right) \\ & \sim N\left( \left(\begin{array}{c} \sigma_1^2\\ \sigma_{12} \\ \sigma_2^2 \end{array} \right),~ \frac{1}{T}\left( \begin{array}{ccc} 2\hat{\sigma}_1^4 & 2\hat{\sigma}_1^2\hat{\sigma}_{12} & 2\hat{\sigma}_{12}^2 \\ 2\hat{\sigma}_1^2 \hat{\sigma}_{12} & \hat{\sigma}_1^2 \hat{\sigma}_2^2 + \hat{\sigma}_{12}^2 & 2\hat{\sigma}_{12}\hat{\sigma}_2^2 \\ 2\hat{\sigma}_{12} & 2\hat{\sigma}_{12} \hat{\sigma}_2^2 & 2\hat{\sigma}_2^4 \end{array} \right) \right). \end{align*}\]

Notice that the results for the diagonal elements of \(\frac{1}{T}\hat{\Omega}\) match those given earlier (see equations (7.20) and (7.22)). What is new here are formulas for the off-diagonal elements.

7.7.3 Joint sampling distribution between \(\hat{\mu}\) and \(\mathrm{vech}(\hat{\Sigma})\)

When we study hypothesis testing in Chapter 9 and the statistical analysis of portfolios in Chapter 15 we will need to use results on the joint sampling distribution between \(\hat{\mu}\) and \(\mathrm{vech}(\hat{\Sigma})\). The follow proposition states the result:

Proposition 7.10 (Joint sampling distribution between \(\hat{\mu}\) and \(\mathrm{vech}(\hat{\Sigma})\)) Under the GWN model and for large enough T, the joint sampling distribution for the \(N+N(N+1)/2 \times 1\) vector \((\hat{\mu}', \hat{\theta}')'\), where \(\hat{\theta} = \mathrm{vech}(\hat{\Sigma})\) is multivariate normal \[\begin{equation} \left(\begin{array}{c} \hat{\mu} \\ \hat{\theta} \end{array} \right) \sim N \left( \left( \begin{array}{c} \mu \\ \theta \end{array} \right), ~ \frac{1}{T} \left( \begin{array}{cc} \hat{\Sigma} & 0 \\ 0 & \hat{\Omega} \end{array} \right) \right). \end{equation}\]

Proposition 7.10 tells us that the vectors \(\hat{\mu}\) and \(\hat{\theta}\) are jointly multivariate normally distributed and that all elements of \(\hat{\mu}\) are independent of all the elements of \(\hat{\theta}\).

Example 2.35 (Joint distribution of \((\mu,\sigma)'\))

Consider the GWN model for a single asset and suppose we are interested in the joint distribution of the vector \((\hat{\mu}, \hat{\sigma}^2)'\). Then Proposition 7.10 gives us the result:

\[\begin{equation} \left( \begin{array}{c} \hat{\mu} \\ \hat{\sigma}^2 \end{array} \right) \sim N \left( \left( \begin{array}{c} \mu \\ \sigma^2 \end{array} \right), ~ \left( \begin{array}{cc} \frac{\hat{\sigma}^2}{T} & 0 \\ 0 & \frac{2\hat{\sigma}^4}{T} \end{array} \right) \right). \tag{7.44} \end{equation}\]

Here, we see that \(\mathrm{cov}(\hat{\mu}, \hat{\sigma}^2)=0\), which implies that \(\hat{\mu}\) and \(\hat{\sigma}^2\) are independent since they are jointly normally distributed. Furthermore, since \(\hat{\mu}\) and \(\hat{\sigma}^2\) are independent and function of \(\hat{\mu}\) is independent of any function of \(\hat{\sigma}^2\). For example, \(\hat{\sigma} = \sqrt{\hat{\sigma}^2}\) is independent of \(\hat{\mu}\). Using (7.34) we can then deduce the joint distribution of \((\hat{\mu}, \hat{\sigma}^2)'\):

\[\begin{equation} \left( \begin{array}{c} \hat{\mu} \\ \hat{\sigma} \end{array} \right) \sim N \left( \left( \begin{array}{c} \mu \\ \sigma \end{array} \right), ~ \left( \begin{array}{cc} \frac{\hat{\sigma}^2}{T} & 0 \\ 0 & \frac{\hat{\sigma}^2}{2T} \end{array} \right) \right). \tag{7.45} \end{equation}\]