Vector Series (VAR Models)
Vector Auto-regression of order 1 VAR(1)
where \(A_{k\times k}\) is the coefficient matrix and \(a_t\) is a K-dimensional white noise process with time-invariant positive definite covariance matrix \(E(aa')=\Sigma\)
We can repeatedly substitute the VAR(1) above to
For the above VMA(\(\infty\)) process to be stationary. \(A^i\) must converge to zero, i.e. the all K eigenvalues of \(A\) must be less than 1.
Then, the VAR(p) model is
OR the compact form
Then, to be stationary, all the roots of \(\det(I_k - A_1B-...-A_pB^p)=0\) are greater than 1 in absolute value.
Companion form
We can check the stability of a VAR(p) model using its companion form.
where \(\xi_t = \begin{bmatrix} y_t \\ ... \\ y_{t-p+1} \end{bmatrix} , A = \begin{bmatrix} A_1 & A_2 & ... & A_{p-1} & A_p \\ I & 0 & ... & 0 & 0 \\ 0 & I & ... & 0 & 0 \\ ... & ... & ... & ... & ... \\ 0 & 0 & ... & I & 0 \end{bmatrix}, v_t = \begin{bmatrix} a_t \\ 0 \\...\\0\end{bmatrix}\)
If the moduli of the eigenvalues of \(A\) are less than 1, then the process is stable
Example
Then, its companion matrix \(A\) will be
Order selection
- sequential likelihood ratio tests of VAR(p) vs. VAR(p-1)
- BigVAR: dimension reduction methods for multi time series