Page 1 of 1

error with svd

PostPosted: Mon Dec 08, 2014 11:01 pm
by francesca90
Hello,

I'm new to dynare and i'm working on my master thesis, trying to estimate gali-monacelli small open economy model (the nber working paper version).
I've done an estimation with results that's have absolutley no sense, so I start seraching for some problem about identification of the parameters, using the command "identification" and I met this error:

Monte Carlo Testing
??? Error using ==> svd
Input to SVD must not contain NaN or Inf.

Error in ==> rank at 15
s = svd(A);

Error in ==> identification_checks at 56
rankJ = rank(JJ./norm(JJ),1.e-10);

Error in ==> identification_analysis at 312
[ide_moments.cond, ide_moments.ind0, ide_moments.indno, ide_moments.ino,
ide_moments.Mco, ide_moments.Pco, ide_moments.jweak, ide_moments.jweak_pair] = ...

Error in ==> dynare_identification at 376
[dum1, ideJ, ideH, ideGP, dum2 , info] = ...

Error in ==> gmmio_est at 476
dynare_identification(options_ident);

Error in ==> dynare at 180
evalin('base',fname) ;

I read some post of people who had similar problems and I understood that usually it's linked to problem with the steady state values but my model is linearized, and have a steady state which is zero for all variables!! I really don't know what could be the problem here and I'll be glad to anyone could give me a little help.
Thank's in advance.

Re: error with svd

PostPosted: Tue Dec 09, 2014 1:32 pm
by jpfeifer
We will look into the crash, but before the crash, Dynare reports
WARNING !!!
The rank of H (model) is deficient!

eta is not identified in the model!
[dJ/d(eta)=0 for all tau elements in the model solution!]
theta is not identified in the model!
[dJ/d(theta)=0 for all tau elements in the model solution!]
theta_r is not identified in the model!
[dJ/d(theta_r)=0 for all tau elements in the model solution!]

You should fix this.
You are neglecting the parameter dependence in your estimation. For example, you need to set omega_a in each iteration, because you are estimating eta. Define it (and the other ones) as model-local variables. See Pfeifer(2013): "A Guide to Specifying Observation Equations for the Estimation of DSGE Models" https://sites.google.com/site/pfeiferecon/Pfeifer_2013_Observation_Equations.pdffor more information.

Re: error with svd

PostPosted: Tue Dec 09, 2014 4:03 pm
by francesca90
Thank's Professor Pfeifer,
I fix that part of my mod file and now I've no more warning about rank deficient of the H or J matrix and all my parameters seems to be identified. Nevertheless, using the identification command I'm always meeting the same error about presence of Nan or Inf in the svd:

??? Error using ==> svd
Input to SVD must not contain NaN or Inf.

Error in ==> rank at 15
s = svd(A);

Error in ==> identification_checks at 56
rankJ = rank(JJ./norm(JJ),1.e-10);

Error in ==> identification_analysis at 312
[ide_moments.cond, ide_moments.ind0, ide_moments.indno, ide_moments.ino,
ide_moments.Mco, ide_moments.Pco, ide_moments.jweak, ide_moments.jweak_pair] = ...

Error in ==> dynare_identification at 376
[dum1, ideJ, ideH, ideGP, dum2 , info] = ...

Error in ==> gmmio_est1 at 438
dynare_identification(options_ident);

Error in ==> dynare at 180
evalin('base',fname) ;

What could be the problem? It's possible that depends on the trasformation I made on my data?

My (fixed) file mod is in attachment.

Re: error with svd

PostPosted: Thu Dec 11, 2014 6:29 am
by jpfeifer
For one particular parameter draw the Lyapunov equation cannot be solved. This leads to an exception in Dynare that is not correctly handled. We are working on this.

But given that the basic identification checks now do not spit out any complaints about rank deficiency of the H or J matrix, you fixed your identification problem. Does estimation still result in crazy estimates?

Re: error with svd

PostPosted: Thu Dec 11, 2014 9:44 am
by francesca90
My estimation now is quite better than the one I did a few days ago: the most strange thing was the non-identification of same parameters (now identified) and the estimated habit, wich was about 0.15 (or less) for Italy and now raised to 0.4. I'm not an expert of bayesian estimation of italian dynamics but I think that it could be higher than this in a country of savers. Nevertheless 0.4 could be realistic beacause the data I used is 1999:Q1 2013:Q4, which is a thight sample, when we observe a great fall of consumption due to the financial crisis. Today I will speak with my Professor to eliminate any doubt about it.
Thank's again mr Pfeifer, for your time and your passion in the dynare implementation.