- Added a
`NEWS.md`

file to track changes to the package. - New exported functions: uncond_moments, cond_moments, get_regime_means, get_regime_autocovs, get_omega_eigens, get_soc.
- simulateGMVAR now provides better tools for forecasting. This update includes non-backward compatible changes for the return values if the argument ntimes is set to be larger than one. In additional to the samples, it now returns a list containing the mixing weights and component that was used to generate each observation.
- diagnostic_plot normality plot now plots histograms instead of kernel estimates.
- In the predict method arguments “ci” and “ci_type” were changed to “pi” and “pi_type” to signify “prediction interval”" as it’s more correct expression than “confidence interval”. Also the default prediction method is now median, and not mean.
- Generally more functionality for conditional and unconditional moments.
- Increased numerical tolerance for considering models to be stationary to attain better numerical stability.
- Updates on documentation, fixed typos.

- Changed the default number of CPU cores employed by the estimation function fitGMVAR to be at most two due to CRAN policy.
- Added argument “seeds” to fitGMVAR allowing one to set the random number generator seed for each call to the genetic algorithm.
- New exported function: alt_gmvar
- Updated the vignette to include the definition of the GMVAR model

- New exported functions: ‘profile_logliks’ (plot profile log-likelihood functions), ‘get_foc’ (gradient at the estimates, recall also ‘get_soc’)
- Now standard errors are printed correctly for models imposing all kinds of constraints (in earlier versions standard errors for such constrained AR parameters that involved sums or multiplication were incorrect).
- Minor update on the genetic algorithm.
- Minor update on the print and summary-print methods.
- Improvements on the comments and documentation.

- Updated the predict method to include forecasts for the mixing weights. Also updated the return value.
- The headlines in ‘profile_logliks’ are now correct for mean-parametrized models (there was phi parameter instread of mu).
- Improved the vignette.
- The default population size in the genetic algorithm is now 50
*ceiling(sqrt(npars)), was 10*npars. - In the function quantile_residual_tests the default argument ‘nsimu’ is now 1 so that the tests are based on the given data only (and not on simulation).
- Corrected an error-causing bug in the predict method when the argument ‘nt’ was not specified.

- A structural GMVAR model is introduced! The SGMVAR model can be estimated with the function ‘fitGMVAR’ or constructed by hand with the function ‘GMVAR’, while also other existing functions such as ‘profile_logliks’, ‘quantile_residual_tests’, ‘diagnostic_plot’, etc. work with the SGMVAR model as well. See the cited paper by Virolainen (2020) or the vignette for the model definition and identification conditions.
- New exported function: ‘GIRF’ for estimating generalized impulse function for the variables in a SGMVAR model.
- New exported function: ‘Wald_test’ for conducting a Wald test testing validity of parameter constraints.
- New exported function: ‘LR_test’ for conducting a likelihood ratio test testing validity of parameter constraints.
- New exported function: ‘gmvar_to_sgmvar’ for building a structural GMVAR model based on a two-regime reduced form GMVAR model.
- New exported function: ‘reorderd_W_column’ for reordering the columns of W matrix of a structural GMVAR model.
- New exported function: ‘swap_W_signs’ for swapping all signs the pointed columns of W matrix (and also B-matrix) of a structural GMVAR model.
- Bug fix: the prediction intervals for mixing weights were incorrect when calculating upper or lower prediction intervals with only one level of significance.
- Minor computation speed improvements.
- Non-backward-compatible change: the functions ‘get_boldA_eigens’ and ‘get_omega_eigens’ now return matrices and not lists.
- Included the structural GMVAR model in a vignette.

- Fixed CRAN issues concerning certain unit tests (not visible to users).

- The plot method of class gmvar objects now also plots marginal stationary densities along with kernel density estimates of the individual time series.
- New exported function: ‘cond_moment_plot’ for plotting in-sample conditional means and variances
- Changed ‘diagnostic_plot’ to automatically show all four figures by default and improved the quantile residual time series plot.
- Added the argument ‘which_largest’ to the function ‘alt_gmvar’
- In GIRF the default shock size is now one, which is then amplified by the B-matrix according to the conditional standard deviation of the model.
- Fixed a bug that caused error when trying to estimate a structural model with more than two regimes and no zero restrictions.
- Fixed a bug that caused error when estimating GIRF with only one initial regime that is not the first regime.
- Fixed a bug that caused error when printing standard errors of a SGMVAR model with constraints on the lambda parameters.

- The function GIRF now enables to directly calculate cumulative impulse responses.
- Updated some of the examples, readme, and vignette.
- Fixed a bug that caused error in estimation of GIRF with very large shock size.
- Fixed a bug that caused error in estimation of AR constrained models when initial population is used in the genetic algorithm.

- Major speed improvement!
- Changing the numerical tolerances for stationarity of the AR parameters and positive definiteness of the error term covariance matrices is now possible everywhere expect in the main estimation function and genetic algorithm (to avoid errors in the estimation). This might help in the analysis (std errors, quantile residual tests, further iterative estimation, etc) of models with parameter values that are very close to the border of the parameter space, but too small numerical tolerance may also cause errors.
- In the function GIRF, the argument ‘variables’ was renamed as ‘which_shocks’.
- New exported function: “update_numtols” which enables one to update the numerical tolerances for stationarity of the AR parameters and positive definiteness of the error term covariance matrices in existing class ‘gmvar’ model.
- Log-likelihood function should not return -Inf anymore but with extremely bad estimates (for some t) the value is not precise (before in such cases it was -Inf). This should lead to better estimation results in some cases.
- Fixed a problem in the estimation algorithm that occurred when estimating a structural model with zero constraints in the W matrix.
- Increased the default maxit from 300 to 500 in fitGMVAR, the default maxlag from 10 to 12 in diagnostic_plot, and decreased the default initial ar_scale from 1 to 0.2 in GAfit.
- There is now randomly varying “ar_scale” in the genetic algorithm but initial population ar_scale can still be adjusted.
- There is now possibility to set random preliminary smart mutations in the genetic algorithm.
- The package Brobdingnag is not imported anymore.
- The default lags used in quantile residual tests are now 1, 3, 6, and 12.
- Updated the documentation of fitGMVAR: more instructions on what to do when the algorithm fails to create an initial population were added.
- Some updates in the vignette, examples, and other documentation.

- Bug fix which required that Brobdingnag was added back to the imported packages. There was in some rare cases a problem with the exact log-likelihood function that was introduced in the version 1.3.0 (the fix was introduced in the same day as the bug, however).

- Major speed improvement (when d*p > 12)!
- New feature: possibility to constrain the (unconditional) mean parameters to the same among some regimes with the argument ‘same_means’. This feature is available for mean-parametrized models only (due to technical reasons).
- The default number of estimation rounds was increased to ’floor(10 + 30*log(M))’.
- Added more adjustable parameters to the genetic algorithm.
- Some changes related to estimation with the genetic algorithm (the estimation results might be different with this version, if AR or lambda constraints are employed, or in some cases when there are no overidentifying constraints in the W matrix).
- Fixed a bug in the predict method that appeared in the very rare cases when one uses the exact one-step conditional mean without confidence intervals as the the prediction method with mean-parametrization.
- In this version, the estimation results with a given seed are different to those in the previous one due to a hidden problem that was introduced.

- Fixed CRACN check issues and examples regarding LR and Wald tests.
- Fixed a problem in the estimation procedure that was introduced in the previous version.
- Updated some of the examples so that the running time fpr all of them is now shorter.
- The genetic algorithm now sorts regimes of the structural models by mixing weight parameters to decreasing order by redecomposing the error term covariance matrices if the first regime changes (before, only regimes 2,..,M were sorted). As a result, the MLE is now found with higher probability in each estimation round.
- In this version, the estimation results with a given seed are different to those in the previous versions due to the updates (see above).