B-VAR with uninformative prior yields dubious results

Dear All,

I'm currently attempting to switch from OLS/ML estimation of VARs to Bayesian VARs. For this purpose, I've used http://www.bankofengland.co.uk/education/Pages/ccbs/technical_handbooks/techbook4.aspx

For instance, the attached code uses a Gibbs sampler to implement an uninformative independent normal Wishart prior. And yet, the results I obtain are completely different from what I find using classical estimation methods. In particular, the shock responses are significantly. Note that lines 106 and 107 actually specify the prior to be un-informative. I get similarly problematic result using different codes.

The two figures attached illustrate the big difference between classical and Bayesian result. The data set is the same. The key plot of interest is the response of variable 4 to variable 1 (lls to g).

I would hugely appreciate any assistance you could provide!!

Kind regards, Lukas