EViews.com

Hello,

I am having trouble with an apparently simple task: how do you estimate a bivariate markov switching model in EViews? I tried with switchvar but it does not seem possible to estimate a specification without lags.

The model I have in mind: vector of two observations distributed as a bivariate normal with all parameters dependent on latent state; two possible latent states, state process distributed as a markov chain.
For now I've been able to estimate a specification with one lag using
Where y1 and y2 are my dependent variables
Would it be possible to estimate the same specification but without lags? Or to restrict the lag coefficients to be zero?
Thanks

There is an internal explicit error check which is being triggered which may not be necessary. I am investigating whether this may safely be modified.

[edit: to temper expectations, this may take a bit of time...]

Thanks Glenn! Let us know :)

Sorry, just to be sure, is there any other straightforward way to estimate the model I described?

I've finished the modifications of the VAR switching code to permit a 0 0 lag specification. This fix will be included in the next patch for EViews 12.

Hi Glenn,

Conceptually, does it make sense to estimate a VAR model without lags? My understanding is that the core idea of a VAR model (switching or not) is to capture the dynamic relationship between variables over time, including relations with past values. A VAR model inherently relies on past values (lags) of the variables to predict their current or future values.

Regards,
Nwabisa Florence Ndzama

From the point of VAR analysis, a VAR with no lags isn't particularly useful. It is simply a system of equations. As such, much of the standard VAR toolkit like impulse response analysis simply isn't useful in that case. But some of the methods that we support, like Markov switching, may be of interest and are not supposed for general nonlinear systems.