1
$\begingroup$

I am trying to compute the likelihood function for the following model, where the state $ \begin{bmatrix} \pi_t \\ x_t \\ i_t \end{bmatrix}$ is measured directly:

$$ \begin{bmatrix} \pi_t \\ x_t \\ i_t \end{bmatrix} = \begin{bmatrix} a_{\pi} \\ a_x \\ a_i \end{bmatrix} + \begin{bmatrix} b_{11} & b_{12} & b_{13} \\ b_{21} & b_{22} & b_{23} \\ b_{31} & b_{32} & b_{33} \end{bmatrix} \begin{bmatrix} \pi_{t-1} \\ x_{t-1} \\ i_{t-1} \end{bmatrix} + \begin{bmatrix} \varepsilon_{\pi} \\ \varepsilon_x \\ \varepsilon_i \end{bmatrix} $$

All instructions for using the Kalman filter to estimate the parameters of this model have required the following state and measurement equations: $$ \xi_{t+1} = A_0\xi_t + Cw_{t+1} \\ y = G x_t + v_t $$

To this end, I have then re-written the above model in the following state-space form:

$$ \begin{bmatrix} 1 \\ \pi_t \\ x_t \\ i_t \end{bmatrix} = \ \begin{bmatrix} 1 & 0 & 0 & 0 \\ a_{\pi} & b_{11} & b_{12} & b_{13} \\ a_x & b_{21} & b_{22} & b_{23} \\ a_i & b_{31} & b_{32} & b_{33} \end{bmatrix} \begin{bmatrix} 1 \\ \pi_{t-1} \\ x_{t-1} \\ i_{t-1} \end{bmatrix} + \begin{bmatrix} 0 & 0 & 0 \\ 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \end{bmatrix} \begin{bmatrix} \varepsilon_{\pi} \\ \varepsilon_x \\ \varepsilon_i \end{bmatrix} $$

However any time I try to compute the likelihood function, I am told that there is no way to compute the variance of the forecast of the state because the variance-covariance matrix is singular. The vcov matrix $\Gamma_0$ is defined by:

$$ vec(\Gamma_0) = (I - A_0 \otimes A_0)^{-1}(C\otimes C)vec(\Sigma) $$

How does one elicit the likelihood function of the first system then?

$\endgroup$
3
  • $\begingroup$ Hi: are the a's and the b's known ? If so, then I don't see why you get singularity of the variance covariance matrix ? Also, I don't see why you have kronecker products for the vcov matrix ? I think there's a simpler formula but it's been a while since I looked at the kalman fitler updates. Is there special about this KF that makes you need kronecker products ? $\endgroup$
    – mlofton
    Commented Jun 20, 2020 at 12:15
  • $\begingroup$ The a's and b's are what I am trying to estimate by computing the likelihood function. stats.stackexchange.com/questions/251067/… This defines the variance of a stationary VAR process and why I use kronecker products. The Kalman filter requires an an initial distribution of the states, defined by an expectation $\begin{pmatrix} \pi_0 \\ x_0 \\ i_0 \end{bmatrix}$ and a variance-covariance of that state. $\endgroup$ Commented Jun 20, 2020 at 23:22
  • $\begingroup$ Hi Jacob: Maybe the singularity has something to do with your updating equation where you're trying to mimic the variance of a VAR. It seems like a regular KF ( that's why I asked about a's and b's ) so can't you just use the update for the system variance in the KF updating equations. For estimates of $v_t$ and $w_t$, you can use the likelihood decomposition. See Harvey's blue book ( structural models and the kalman filter, 1990 ) for details of how that works. $\endgroup$
    – mlofton
    Commented Jun 22, 2020 at 1:46

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.