% Various tests within Maximum Likelihood estimation
% with skew normally distributed shocks
% -------------------------------------------------------------------------
% Copyright © 2025 Dynare Team
%
% This file is part of Dynare.
%
% Dynare is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% Dynare is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with Dynare. If not, see .
@#include "ireland2004_model.inc"
estimated_params;
stderr eta_a, 3.01673, 0, 10;
stderr eta_e, 0.02476, 0, 10;
stderr eta_z, 0.88648, 0, 10;
stderr eta_r, 0.27903, 0, 10;
OMEGA, 0.05809, 0, 1;
RHO_PI, 0.38648, 0, 1;
RHO_G, 0.39601, 0, 1;
RHO_X, 0.16539, 0, 1;
RHO_A, 0.90480, 0, 1;
RHO_E, 0.99067, 0, 1;
end;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TEST 1: Gaussian and pruned skewed kalman filter should run without error %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% use Gaussian Kalman filter
estimation(datafile = 'ireland2004_post1980.csv'
,mode_compute = 8
,silent_optimizer
,lik_init = 1 % initialize Kalman filter at Gaussian steady-state distribution
,kalman_algo = 1 % use Gaussian Kalman filter
);
struct0 = cell2struct([struct2cell(oo_.mle_mode.parameters); struct2cell(oo_.mle_mode.shocks_std); oo_.posterior.optimization.log_density], ...
[fieldnames(oo_.mle_mode.parameters); [strcat('stderr_', fieldnames(oo_.mle_mode.shocks_std)); {'loglik'}]]);
smoothed_shocks_0 = oo_.SmoothedShocks;
% use Pruned Skewed Kalman filter as Gaussian distribution is special case of CSN distribution
estimation(datafile = 'ireland2004_post1980.csv'
,mode_compute = 8
,silent_optimizer
,lik_init = 1 % initialize Kalman filter at Gaussian steady-state distribution
,kalman_algo = 5 % use pruned skewed Kalman filter routine even in Gaussian case for comparability;
% note that Dynare's implementation is much faster because it switches to the steady-state Kalman filter which we have not implemented yet for the PSKF
,skewed_kalman_mvnlogcdf = 'mvncdf' % test whether the option is passed correctly, in Gaussian case there is no curse of increasing skewness dimension (as there is no need to compute Gaussian cdfs)
);
struct5 = cell2struct([struct2cell(oo_.mle_mode.parameters); struct2cell(oo_.mle_mode.shocks_std); oo_.posterior.optimization.log_density], ...
[fieldnames(oo_.mle_mode.parameters); [strcat('stderr_', fieldnames(oo_.mle_mode.shocks_std)); {'loglik'}]]);
smoothed_shocks_5 = oo_.SmoothedShocks;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TEST 2: skewed_kalman_mvnlogcdf option should have been set to 'mvncdf' %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
if ~strcmp(options_.skewed_kalman.mvnlogcdf,'mvncdf')
error('option ''skewed_kalman_mvnlogcdf'' was not set correctly')
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TEST 3: Estimation of Gaussian model with either Gaussian %
% or Pruned Skewed Kalman filter should be numerically %
% equal to each other %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
fprintf('Comparing estimation results:\n');
fprintf('%-15s %15s %15s %15s %10s\n', 'Parameter', 'Gaussian KF', 'Skewed KF', 'Max Abs Dev', 'Status');
fprintf('%s\n', repmat('-', 1, 85));
status = nan(length(fieldnames(struct0)),1);
for i = 1:length(fieldnames(struct0))
field_name = fieldnames(struct0);
max_abs_dev = max(abs(struct0.(field_name{i}) - struct5.(field_name{i})));
status(i) = (max_abs_dev > 1e-4);
if status(i); status_str = 'FAIL'; else status_str = 'PASS'; end
fprintf('%-15s %15f %15f %15.6e %10s\n', field_name{i}, struct0.(field_name{i}), struct5.(field_name{i}), max_abs_dev, status_str);
end
if any(status)
error('Estimates with Gaussian KF (kalman_algo=0) are not numerically equal to Pruned Skewed KF (kalman_algo=5), something is wrong!')
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TEST 4: Smoothed shock values of Gaussian model with either Gaussian %
% or Pruned Skewed Kalman smoother should be numerically %
% equal to each other %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
if ~isoctave % array2table is not available in Octave
fprintf('\nSmoothed Gaussian eta_r\n');
disp(array2table([smoothed_shocks_0.eta_r smoothed_shocks_5.eta_r (smoothed_shocks_0.eta_r-smoothed_shocks_5.eta_r)], ...
'RowNames', "t="+string(1:options_.nobs), ...
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"]));
fprintf('\nSmoothed Gaussian eta_z\n');
disp(array2table([smoothed_shocks_0.eta_z smoothed_shocks_5.eta_z (smoothed_shocks_0.eta_z-smoothed_shocks_5.eta_z)], ...
'RowNames', "t="+string(1:options_.nobs), ...
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"]));
fprintf('\nSmoothed Gaussian eta_e\n');
disp(array2table([smoothed_shocks_0.eta_e smoothed_shocks_5.eta_e (smoothed_shocks_0.eta_e-smoothed_shocks_5.eta_e)], ...
'RowNames', "t="+string(1:options_.nobs), ...
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"]));
fprintf('\nSmoothed Gaussian eta_a\n');
disp(array2table([smoothed_shocks_0.eta_a smoothed_shocks_5.eta_a (smoothed_shocks_0.eta_a-smoothed_shocks_5.eta_a)], ...
'RowNames', "t="+string(1:options_.nobs), ...
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"]));
end
if any(abs(smoothed_shocks_0.eta_r - smoothed_shocks_5.eta_r) > 1e-3) || ...
any(abs(smoothed_shocks_0.eta_z - smoothed_shocks_5.eta_z) > 1e-3) || ...
any(abs(smoothed_shocks_0.eta_e - smoothed_shocks_5.eta_e) > 1e-3) || ...
any(abs(smoothed_shocks_0.eta_a - smoothed_shocks_5.eta_a) > 1e-3)
error('Smoothed values should be numerically equal to each other')
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TEST 3: Estimation of skewness coefficients can be successfully done with %
% Pruned Skewed Kalman filter, estimation should run without error %
% and estimation options should have be correctly changed %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
OMEGA = 0.15356;
RHO_PI = 0.28761;
RHO_G = 0.33795;
RHO_X = 0.28325;
RHO_A = 0.91704;
RHO_E = 0.98009;
shocks;
var eta_a; stderr 2.5232;
var eta_e; stderr 0.021228;
var eta_z; stderr 0.79002;
var eta_r; stderr 0.28384;
skew eta_z = -0.9950;
end;
estimated_params(overwrite);
OMEGA, 0.15356;
stderr eta_e, 0.021228;
skew eta_a, -0.1948;
skew eta_e, -0.2140;
skew eta_r, 0.8128;
end;
% should run without error
estimation(datafile = 'ireland2004_post1980.csv'
,mode_compute = 5
,silent_optimizer
,kalman_algo = 5 % use pruned skewed Kalman filter
,lik_init = 1 % initialize Kalman filter at Gaussian steady-state distribution
,skewed_kalman_prune_tol = 0.02
,skewed_kalman_rank_deficiency_transform
,skewed_kalman_mvnlogcdf = 'gaussian_log_mvncdf_mendell_elston'
,tex
);
% check whether options have been correctly passed by the preprocessor
if (options_.skewed_kalman.prune_tol ~= 0.02) || ...
(options_.skewed_kalman.rank_deficiency_transform ~= true) || ...
~strcmp(options_.skewed_kalman.mvnlogcdf, 'gaussian_log_mvncdf_mendell_elston')
error('skewed kalman options were not passed correctly')
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% TEST 4: Likelihood with skew normal model should be higher than with Gaussian model %
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
loglik_csn = oo_.posterior.optimization.log_density; % expected to be approx 1215.68
if (loglik_csn-1215.6815424) > 1e-5
error('Something wrong with CSN likelihood')
end
loglik_gauss = struct5.loglik; % expected to be approx 1207.56
if (loglik_gauss-1207.5618755) > 1e-5
error('Something wrong with Gaussian likelihood')
end
lr_stat = 2 * (loglik_csn - loglik_gauss);
lr_pval = 1 - chi2cdf(lr_stat, 4); % 4 additional shock parameters with CSN (only if one would have estimated all parameters, here we don't do this to speed up testsuite)
lr_str = sprintf('test statistic of $%.2f$ and a p-value of $%.4f$.', lr_stat, lr_pval);
fprintf('\n%s\n* LIKELIHOOD RATIO TEST *\n%s\n', repmat('*',1,25), repmat('*',1,25));
fprintf('- unconstrained log-likelihood (CSN): %.4f\n', loglik_csn);
fprintf('- constrained log-likelihood (Gaussian): %.4f\n', loglik_gauss);
fprintf('- %s\n\n',lr_str);