Page 1 of 1
Estimation problem in large-scale model
Posted:
Wed Aug 27, 2014 9:26 amby AS90
This time:
- the model is stationary by construction;
- the data is one sided HP filtered;
- Bayesian estimation functions only with algorithm #6, although, delivering inaccurate results.
Any insights would be appreciated.
PS. Any news on the SMM front?
Thanks.
Re: Usual Bayesian estimation problem
Posted:
Wed Jun 17, 2015 4:09 pmby Oriana
My model is rather complex and I am also having some difficulties in finding the posterior mode for all my parameters.
Suggestions to overpass these difficulties are mostly welcome!
Re: Usual Bayesian estimation problem
Posted:
Sun Jun 21, 2015 8:40 amby Oriana
In order to obtain the following results I already replaced chol(hh) with hh=1e-4*eye(size(hh)). How reliable these results could be?
The identification strength of most of parameters is also 0. Could it mean some model misspecification?
Re: Usual Bayesian estimation problem
Posted:
Thu Jun 25, 2015 5:42 pmby jpfeifer
Are you sure you solved the identification problem? It still shows up for me with the posted files.
Re: Usual Bayesian estimation problem
Posted:
Fri Jun 26, 2015 11:08 pmby Oriana
Please, if possible, try with the following files again.
Re: Usual Bayesian estimation problem
Posted:
Sun Jul 05, 2015 4:02 pmby jpfeifer
Put
- Code: Select all
identification(ar=10);
before the estimation command and you will see
WARNING !!!
The rank of H (model) is deficient!
c1phimc is not identified in the model!
[dJ/d(c1phimc)=0 for all tau elements in the model solution!]
c1phioc is not identified in the model!
[dJ/d(c1phioc)=0 for all tau elements in the model solution!]
c1telast is not identified in the model!
[dJ/d(c1telast)=0 for all tau elements in the model solution!]
c1oelast is not identified in the model!
[dJ/d(c1oelast)=0 for all tau elements in the model solution!]
c1kelast is not identified in the model!
[dJ/d(c1kelast)=0 for all tau elements in the model solution!]
c1less is not identified in the model!
[dJ/d(c1less)=0 for all tau elements in the model solution!]
[c1gamdpc2,c1gamrs2] are PAIRWISE collinear (with tol = 1.e-10) !
WARNING !!!
The rank of J (moments) is deficient!
c1phimc is not identified by J moments!
[dJ/d(c1phimc)=0 for all J moments!]
c1phioc is not identified by J moments!
[dJ/d(c1phioc)=0 for all J moments!]
c1telast is not identified by J moments!
[dJ/d(c1telast)=0 for all J moments!]
c1oelast is not identified by J moments!
[dJ/d(c1oelast)=0 for all J moments!]
c1kelast is not identified by J moments!
[dJ/d(c1kelast)=0 for all J moments!]
c1less is not identified by J moments!
[dJ/d(c1less)=0 for all J moments!]
[c1gamdpc2,c1gamrs2] are PAIRWISE collinear (with tol = 1.e-10) !
[c1scalepitarerr,c1gamrs2] are PAIRWISE collinear (with tol = 1.e-10) !
[c1scalepitarerr,c1gamdpc2] are PAIRWISE collinear (with tol = 1.e-10) !
c1omegal is collinear w.r.t. all other params!
c1chi is collinear w.r.t. all other params!
c1thetap is collinear w.r.t. all other params!
c1thetaw is collinear w.r.t. all other params!
c1tauw is collinear w.r.t. all other params!
c1taup is collinear w.r.t. all other params!
c1scalethetapvarerr is collinear w.r.t. all other params!
c1scalethetawvarerr is collinear w.r.t. all other params!
Re: Usual Bayesian estimation problem
Posted:
Tue Jul 07, 2015 7:14 pmby Oriana
Thank you, for the previous observation.
I already fixed the identification problems and got ...
All parameters are identified in the model (rank of H).
All parameters are identified by J moments (rank of J)
Nevertheless, the graph shock decomposition c1y (in the zip file below), raises some suspicion on some kind of misspecification in my code.
Could you, please, refute or confirm it ?
Re: Usual Bayesian estimation problem
Posted:
Wed Jul 08, 2015 3:36 amby jpfeifer
Yes, that looks weird.
Re: Usual Bayesian estimation problem
Posted:
Fri Jul 10, 2015 5:17 pmby Oriana
Thanks. I already found the problem. It seems the discount factor "beta" parameter was too high.
I only changed c1beta = c1muzss/1.01 to c1beta = c1muzss/1.05 (in the steadystate file) and the graphs of shocks decomposition appeared quite fine .
Unfortunately, I had another problem. It is related with my data frequency of oil production (Japanese version). I have quarterly data after 1984, but not for previous years. My period study begins in 1970.
A mixed frequency approach, as you suggested once, seems the most convenient choice, however, I would lose the quarterly data information after 1984.
In your view a better solution can be applied?
Re: Usual Bayesian estimation problem
Posted:
Sat Jul 11, 2015 8:14 amby jpfeifer
The mixed frequency approach does not mean that the series has to have the same frequency the whole time. Dynare should treat unobserved values as missing values. After 1984, there should be no more missing values.
Re: Usual Bayesian estimation problem
Posted:
Sun Jul 12, 2015 8:22 amby Oriana
Thank you for the previous information. The subject was not well clarified in your document ( A Guide to Specifying Observation Equations for the Estimation of DSDE Models).