-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA512
In TeX/fs2000_corr_ME a singularity is detected and Dynare goes to the univariate Kalman filter. Unfortunately, in this situation (with correlated measurement errors) the state vector does not have the correct dimension.
If we force Dynare not to use the univariate filter in case of singularity (setting use_univariate_filters_if_singularity_is_detected equal to zero) the estimation goes smoothly. Obviously this is not a solution for this test (though I always use this option).
I fixed the problem by adding a flag called singularity_has_been_detected in dsge_likelihood (at the beginning of section 4) equal to false by default and resetting this variable to true when isinf(LIK) and DynareOptions.use_univariate_filters_if_singularity_is_detected are true. When the flag is true I augment the size of the state vector (a) when needed (if the model has correlated measurement errors). I need to do more testing with other mod files...
I think that this is a long standing bug. Before the last commits by Marco the singularity was not detected (and we never reached this case).
The bugs in kalman/likelihood_from_dynare/*.mod are not related to this one.
Best, Stéphane.
On 14/10/2015 21:58, Michel Juillard wrote:
Some test cases never ran. For example: | * kalman/likelihood_from_dynare/fs2000_corr_ME.mod | * kalman/likelihood_from_dynare/fs2000_corr_ME_missing.mod | * kalman/likelihood_from_dynare/fs2000_uncorr_ME.mod | * kalman/likelihood_from_dynare/fs2000_uncorr_ME_missing.mod
The fixing is not trivial and on my todo list. I don't know if it corresponds to what Johannes just observed.
Best
Michel
Johannes Pfeifer writes:
Try TeX/fs2000_corr_ME.mod I selectively ran some of the failed tests from today's mail to the dev list, not the whole suite
Am 14.10.2015 um 21:26 schrieb Stéphane Adjemian stepan@dynare.org:
Hi Johannes,
Can you send the output of the testsuite (the txt file returned after make check-matlab) ? I didn't observe any problem today. I have just checked that estimation/fs2000.mod runs without any problem.
Best, Stéphane.
On 14/10/2015 19:41, Johannes Pfeifer wrote: Hi Marco,
it seems your recent commits broke the dsge_likelihood routine. Almost all unit tests crash in the Kalman filter routines due to non-conformable matrices. Unfortunately, I could not trace the origin of the problem.
Best,
Johannes
--
Johannes Pfeifer
Friesenwall 104
50672 Köln
Tel.: +49-221-29873852
Mobil.: +49-170-6936820
Germany
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev