lsqnonlin is an optimization algorithm that is specifically well suited for minimizing quadratic objectives, so I would like to have it as default. However, we fallback to csminwel if the optimization toolbox for Matlab or optim for Octave are not available.
As we have a working interface now, this commit improves the provisional handling of finding duplicate moments. Previously, indices for GMM were created, but this is not really needed. This commit cleans this up and similar to SMM makes use of the matched_moments block.
As a by-product of the previous provisonal handling higher-order moments for SMM where not correctly displayed as no labels were created. This is now fixed. The comparison of data moments and estimated model moments is also in the same ordering as the inputed orthogonality conditions in matched_moments.
Signed-off-by: Willi Mutschler <willi@mutschler.eu>
The mean(…, 'omitnan') syntax does not exist before R2015a. Use nanmean
instead (either provided by the statistics toolbox, or by our fallback under
matlab/missing/stats/).
The TeX option crashes MATLAB R2014a run with "-nodisplay" option
(as is done from the testsuite). We disable the TeX option in that case.
Since we can’t directly test whether "-nodisplay" has been passed, we test for
the "TOP_TEST_DIR" environment variable, which is set by the testsuite.
Note that it was not tested whether the crash happens with more recent MATLAB
versions, so when OLD_MATLAB_VERSION is increased, one should make a test
before removing this workaround.
- lsqnonlin takes less time in Andreasen toolbox due to different handling of bounds; our approach is better (objective is smaller for RBC model)
- user-specified weightning matrix does work
- qz_criterium is set to 1+e-6 to allow for unit-roots (but observables need to be stationary)
- informative message is printed that NaN values in data moments are replaced by mean
- estimated_params_bounds work as well as informative message when parameters are not calibrated
- SMM at any order without pruning works
- SMM with pruning works for orders 1,2,3; for orders above 3 simult_.m throws an error that pruning is not available
- Analytical standard errors for GMM and Bayesian estimation is included
- dirname is not needed
- Remaining to dos are updated
- Set pruning to false as default, for GMM we turn it on for order>2 and display a message for the user that we do so
- Provide error if users try GMM with order > 3
- For SMM, order > 2 and no pruning, simult_.m uses Dynare++ routines which require a seed
Stéphane Adjemian (Charybdis)
Willi Mutschler
Johannes Pfeifer
Sébastien Villemot