SVAR with SR non-recursive matrix "A"

[philipecon]

Dear all,

I am trying to estimate a SVAR model with SR restrictions which do not follow the usual recursive structure, that would give a triangular matrix "A".
Once I try to estimate it, eviews gives me the following error message: "Hessian of SVAR likelihood is singular at starting values. Reset starting values or respecify restrictions to ensure that the model is (locally) identified"

My model is just-identified as I imposed sufficient restrictions on the "A" matrix and assumed that the "B" matrix is diagonal.

Could someone please suggest what have gone wrong?

Cheers.

[trubador]

It does not necessarily mean that you have done something wrong. The message is pretty clear and also tells you what you should do next. There is an "Optimization Control" tab inside the SVAR dialog box. You can try other alternatives for starting values.

[philipecon]

Thanks for your reply! However optimization control does not help.
I thought that maybe if I outline my identification structure, someone could suggest something?

I am running a 7 variable macro model with commodity prices, output, inflation, money balances, interest rate, exchange rate and stock prices. My contemporaneous "A" matrix takes the following form:

e(com.prices) = 1 0 0 0 0 0 0
e(output) = 0 1 0 0 0 0 0
e(inflation) = x x 1 0 0 0 0
e(money) = 0 x x 1 x 0 0
e(int. rate) = x 0 0 x 1 x x
e(ex. rate) = x x x x x 1 x
e(stocks) = x x x x x x 1

Could someone please suggest why this cannot be estimated?

Cheers

[trubador]

There is nothing inherently wrong here. Unfortunately, imposing correct number of restrictions is a necessary, but not sufficient condition for exact identification. And there is no easy way to do that, I am afraid.

[philipecon]

Thanks for your reply!

What actions can I take in order to be able to estimate my model? What are the sufficient conditions for this?

[trubador]

I do not think there is an easy way out. Even if you manage to find a solution, there will probably be alternative factorizations that maximize the likelihood function. You can try different restrictions, but if you don't then you'll have to explicitly model the covariance matrix and use a more powerful optimization routine.