SVAR model

[Caipirinha]

Hey guys!

I'm still pretty new to Eviews and VAR estimations and can't solve a few problems which appear running my factorization. Please excuse my basic questions.

The idea of my SVAR is to measure the influence of ECB's asset purchase programs (APP) on real estate markets (RE). In this context I used six variables (APP, inflation, interest rate (IR), mortgage volume (CV), GDP and a real estate index. APP, Inflation and GDP are used on levels. IR, CV and RE as percentage changes to avoid non-stationary. I use quarterly data from 2008Q1 to 2019Q2.

Further I want to use a recursive factorization with some changes on the lower triangular matrix A. B should remain diagonal to isolate the APP shock on all other variables.

My Questions are:

1.) Will the Cholesky decomposition for IRF automatically use a lower triangular A and diagonal B for the estimated VAR? (Sorry, somehow I couldn't find that)

2.) When I try to apply the structural factorization there is always an error message ("Optimization may be unreliable (first or second order conditions
not met)"). I've changed meanwhile every parameter in the optimization control but couldn't fix it. Any idea what could cause this?


I would really appreciate every help. Once again, sorry for that basic questions. I'm meanwhile completely confused.

Thx!

[EViews Matt]

Hello,

(1) Yes in the sense that they both produce a recursive factorization. If you're asking whether using the Cholesky decomposition versus estimating an A-B SVAR decomposition will produce exactly the same results, that's a more subtle issue. Theoretically, the two techniques can match, and likely will in many cases, but in practice they can numerically diverge for any number of reasons.

(2) Chances are the SVAR is under-specified, i.e., there are not enough restrictions on the coefficients of A and B. If your basic model is a lower triangular A (not unit lower triangular) and a diagonal B, then with six variables you'd ideally specify six additional restrictions to produce a unique factorization. Without sufficient restrictions, the factorization is not unique and EViews will generate the error message you describe. You can force EViews to report it's results anyway by setting the "Ignore Errors" option on the SVAR "Optimization Control" tab (or the command-line equivalent). Just be forewarned that the resulting factorization might be unappealing, e.g., the matrix coefficients may be very large.