CVModel

[TG81]

Dear Tom,

I am trying to replicate Kim and Roubini's paper "Exchange rate anomalies in the industrial countries: A solution with a structural VAR approach" for Canada. Attached are my codes. As I gathered from RATS user's guide and forum that we should use cvmodel with method=bfgs,pmethod=genetic before moving on to purely, method=bfgs(as is shown in my program).

1. How do I pick the initial values? Can we take the estimates from the method cvmodel(method=bfgs,pmethod=genetic) as the initial values for cvmodel(method=bfgs)?

2. Is there any thumb rule for doing cvmodel? Eventhough I am getting convergence with the process cvmodel(method=bfgs,pmethod=genetic) but not with cvmodel(method=bfgs) for many initial values. What is the next step?

3. For the Canadian data(attached), I am getting exchange rate puzzle and price puzzle with my program. This is wrong as the authors Kim and Roubini seemed to have eliminated these puzzles through their SVAR model. Clearly, I am not able to achieve global identification. Even if I do, how would I know that it is not local but global. Please help me figure out the problem.

Thanks in advance!

[TomDoan]

For a model like that (DMATRIX=IDENTITY), I would probably start with 0's on the off-diagonal and the square roots of the diagonals on the diagonal:

compute g21=g31=g41=g43=g51=g56=g57=g63=g64=g65=g71=g72=g73=g74=g75=g76=0.0
compute g11=sqrt(v(1,1)),g22=sqrt(v(2,2)),g33=sqrt(v(3,3)),g44=sqrt(v(4,4)),g55=sqrt(v(5,5)),g66=sqrt(v(6,6)),g77=sqrt(v(7,7))

0's across the board is unlikely to work well for BFGS.

I don't see anything wrong with your estimates. For your data set, the MS and MD equations in effect switch places from what they were in the K&R paper, which probably results in the "wrong" responses. That just means that the K&R result isn't stable across time periods. You wouldn't be the first person to discover that. The SVAR with the MS and MD equations both having R and M has been problematical from the day it was introduced---I'm not sure I've seen any SVAR which definitively is able to separate the two empirically with any high degree of confidence. Where the off-diagonal has the "right" sign (as in the original K&R paper) it's usually the right sign but not statistically significant. Yours has the "wrong" signs but they are quite insignificant.

[TG81]

Hi,

I have two questions regarding the same model.

First, I want to generate confidence bands for the impulse responses for the SVAR model and attached is the program for Monte Carlo integration. Factor matrix "g" is generated through the cvmodel. Is the insertion of the factor matrix "g" inside the Monte Carlo Integration done correctly.

Second, I want all the variables to respond to its shocks positively and for which I use "flippers". Is it done correctly in my program?

I will be really grateful if you can please confirm this. Thanks in advance!

[TomDoan]

What that's doing is drawing coefficients while leaving the structural model at the maximum likelihood estimates. While not the worst idea, it's far from what someone would expect from an error band calculation.

This is an overidentified model. There is no simple way to do Monte Carlo integration with an overidentified SVAR. You either have to do importance sampling or Metropolis-Hastings, either of which requires some experimentation to come up with a good sampling method. See the MONTESVAR.RPF example.