| 1 |
% Various tests within Maximum Likelihood estimation |
| 2 |
% with skew normally distributed shocks |
| 3 |
% ------------------------------------------------------------------------- |
| 4 |
|
| 5 |
% Copyright © 2025 Dynare Team |
| 6 |
% |
| 7 |
% This file is part of Dynare. |
| 8 |
% |
| 9 |
% Dynare is free software: you can redistribute it and/or modify |
| 10 |
% it under the terms of the GNU General Public License as published by |
| 11 |
% the Free Software Foundation, either version 3 of the License, or |
| 12 |
% (at your option) any later version. |
| 13 |
% |
| 14 |
% Dynare is distributed in the hope that it will be useful, |
| 15 |
% but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 |
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 17 |
% GNU General Public License for more details. |
| 18 |
% |
| 19 |
% You should have received a copy of the GNU General Public License |
| 20 |
% along with Dynare. If not, see <https://www.gnu.org/licenses/>. |
| 21 |
|
| 22 |
@#include "ireland2004_model.inc" |
| 23 |
|
| 24 |
estimated_params; |
| 25 |
stderr eta_a, 3.01673, 0, 10; |
| 26 |
stderr eta_e, 0.02476, 0, 10; |
| 27 |
stderr eta_z, 0.88648, 0, 10; |
| 28 |
stderr eta_r, 0.27903, 0, 10; |
| 29 |
OMEGA, 0.05809, 0, 1; |
| 30 |
RHO_PI, 0.38648, 0, 1; |
| 31 |
RHO_G, 0.39601, 0, 1; |
| 32 |
RHO_X, 0.16539, 0, 1; |
| 33 |
RHO_A, 0.90480, 0, 1; |
| 34 |
RHO_E, 0.99067, 0, 1; |
| 35 |
end; |
| 36 |
|
| 37 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 38 |
% TEST 1: Gaussian and pruned skewed kalman filter should run without error % |
| 39 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 40 |
|
| 41 |
% use Gaussian Kalman filter |
| 42 |
estimation(datafile = 'ireland2004_post1980.csv' |
| 43 |
,mode_compute = 8 |
| 44 |
,silent_optimizer |
| 45 |
,lik_init = 1 % initialize Kalman filter at Gaussian steady-state distribution |
| 46 |
,kalman_algo = 1 % use Gaussian Kalman filter |
| 47 |
); |
| 48 |
struct0 = cell2struct([struct2cell(oo_.mle_mode.parameters); struct2cell(oo_.mle_mode.shocks_std); oo_.posterior.optimization.log_density], ... |
| 49 |
[fieldnames(oo_.mle_mode.parameters); [strcat('stderr_', fieldnames(oo_.mle_mode.shocks_std)); {'loglik'}]]); |
| 50 |
smoothed_shocks_0 = oo_.SmoothedShocks; |
| 51 |
|
| 52 |
% use Pruned Skewed Kalman filter as Gaussian distribution is special case of CSN distribution |
| 53 |
estimation(datafile = 'ireland2004_post1980.csv' |
| 54 |
,mode_compute = 8 |
| 55 |
,silent_optimizer |
| 56 |
,lik_init = 1 % initialize Kalman filter at Gaussian steady-state distribution |
| 57 |
,kalman_algo = 5 % use pruned skewed Kalman filter routine even in Gaussian case for comparability; |
| 58 |
% 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 |
| 59 |
,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) |
| 60 |
); |
| 61 |
struct5 = cell2struct([struct2cell(oo_.mle_mode.parameters); struct2cell(oo_.mle_mode.shocks_std); oo_.posterior.optimization.log_density], ... |
| 62 |
[fieldnames(oo_.mle_mode.parameters); [strcat('stderr_', fieldnames(oo_.mle_mode.shocks_std)); {'loglik'}]]); |
| 63 |
smoothed_shocks_5 = oo_.SmoothedShocks; |
| 64 |
|
| 65 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 66 |
% TEST 2: skewed_kalman_mvnlogcdf option should have been set to 'mvncdf' % |
| 67 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 68 |
if ~strcmp(options_.skewed_kalman.mvnlogcdf,'mvncdf') |
| 69 |
error('option ''skewed_kalman_mvnlogcdf'' was not set correctly') |
| 70 |
end |
| 71 |
|
| 72 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 73 |
% TEST 3: Estimation of Gaussian model with either Gaussian % |
| 74 |
% or Pruned Skewed Kalman filter should be numerically % |
| 75 |
% equal to each other % |
| 76 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 77 |
fprintf('Comparing estimation results:\n'); |
| 78 |
fprintf('%-15s %15s %15s %15s %10s\n', 'Parameter', 'Gaussian KF', 'Skewed KF', 'Max Abs Dev', 'Status'); |
| 79 |
fprintf('%s\n', repmat('-', 1, 85)); |
| 80 |
status = nan(length(fieldnames(struct0)),1); |
| 81 |
for i = 1:length(fieldnames(struct0)) |
| 82 |
field_name = fieldnames(struct0); |
| 83 |
max_abs_dev = max(abs(struct0.(field_name{i}) - struct5.(field_name{i}))); |
| 84 |
status(i) = (max_abs_dev > 1e-4); |
| 85 |
if status(i); status_str = 'FAIL'; else status_str = 'PASS'; end |
| 86 |
fprintf('%-15s %15f %15f %15.6e %10s\n', field_name{i}, struct0.(field_name{i}), struct5.(field_name{i}), max_abs_dev, status_str); |
| 87 |
end |
| 88 |
if any(status) |
| 89 |
error('Estimates with Gaussian KF (kalman_algo=0) are not numerically equal to Pruned Skewed KF (kalman_algo=5), something is wrong!') |
| 90 |
end |
| 91 |
|
| 92 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 93 |
% TEST 4: Smoothed shock values of Gaussian model with either Gaussian % |
| 94 |
% or Pruned Skewed Kalman smoother should be numerically % |
| 95 |
% equal to each other % |
| 96 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 97 |
if ~isoctave % array2table is not available in Octave |
| 98 |
fprintf('\n<strong>Smoothed Gaussian eta_r</strong>\n'); |
| 99 |
disp(array2table([smoothed_shocks_0.eta_r smoothed_shocks_5.eta_r (smoothed_shocks_0.eta_r-smoothed_shocks_5.eta_r)], ... |
| 100 |
'RowNames', "t="+string(1:options_.nobs), ... |
| 101 |
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"])); |
| 102 |
fprintf('\n<strong>Smoothed Gaussian eta_z</strong>\n'); |
| 103 |
disp(array2table([smoothed_shocks_0.eta_z smoothed_shocks_5.eta_z (smoothed_shocks_0.eta_z-smoothed_shocks_5.eta_z)], ... |
| 104 |
'RowNames', "t="+string(1:options_.nobs), ... |
| 105 |
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"])); |
| 106 |
fprintf('\n<strong>Smoothed Gaussian eta_e</strong>\n'); |
| 107 |
disp(array2table([smoothed_shocks_0.eta_e smoothed_shocks_5.eta_e (smoothed_shocks_0.eta_e-smoothed_shocks_5.eta_e)], ... |
| 108 |
'RowNames', "t="+string(1:options_.nobs), ... |
| 109 |
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"])); |
| 110 |
fprintf('\n<strong>Smoothed Gaussian eta_a</strong>\n'); |
| 111 |
disp(array2table([smoothed_shocks_0.eta_a smoothed_shocks_5.eta_a (smoothed_shocks_0.eta_a-smoothed_shocks_5.eta_a)], ... |
| 112 |
'RowNames', "t="+string(1:options_.nobs), ... |
| 113 |
'VariableNames', ["kalman_algo=0", "kalman_algo=5", "dev"])); |
| 114 |
|
| 115 |
end |
| 116 |
|
| 117 |
if any(abs(smoothed_shocks_0.eta_r - smoothed_shocks_5.eta_r) > 1e-3) || ... |
| 118 |
any(abs(smoothed_shocks_0.eta_z - smoothed_shocks_5.eta_z) > 1e-3) || ... |
| 119 |
any(abs(smoothed_shocks_0.eta_e - smoothed_shocks_5.eta_e) > 1e-3) || ... |
| 120 |
any(abs(smoothed_shocks_0.eta_a - smoothed_shocks_5.eta_a) > 1e-3) |
| 121 |
error('Smoothed values should be numerically equal to each other') |
| 122 |
end |
| 123 |
|
| 124 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 125 |
% TEST 3: Estimation of skewness coefficients can be successfully done with % |
| 126 |
% Pruned Skewed Kalman filter, estimation should run without error % |
| 127 |
% and estimation options should have be correctly changed % |
| 128 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 129 |
OMEGA = 0.15356; |
| 130 |
RHO_PI = 0.28761; |
| 131 |
RHO_G = 0.33795; |
| 132 |
RHO_X = 0.28325; |
| 133 |
RHO_A = 0.91704; |
| 134 |
RHO_E = 0.98009; |
| 135 |
|
| 136 |
shocks; |
| 137 |
var eta_a; stderr 2.5232; |
| 138 |
var eta_e; stderr 0.021228; |
| 139 |
var eta_z; stderr 0.79002; |
| 140 |
var eta_r; stderr 0.28384; |
| 141 |
skew eta_z = -0.9950; |
| 142 |
end; |
| 143 |
|
| 144 |
estimated_params(overwrite); |
| 145 |
OMEGA, 0.15356; |
| 146 |
stderr eta_e, 0.021228; |
| 147 |
skew eta_a, -0.1948; |
| 148 |
skew eta_e, -0.2140; |
| 149 |
skew eta_r, 0.8128; |
| 150 |
end; |
| 151 |
|
| 152 |
% should run without error |
| 153 |
estimation(datafile = 'ireland2004_post1980.csv' |
| 154 |
,mode_compute = 5 |
| 155 |
,silent_optimizer |
| 156 |
,kalman_algo = 5 % use pruned skewed Kalman filter |
| 157 |
,lik_init = 1 % initialize Kalman filter at Gaussian steady-state distribution |
| 158 |
,skewed_kalman_prune_tol = 0.02 |
| 159 |
,skewed_kalman_rank_deficiency_transform |
| 160 |
,skewed_kalman_mvnlogcdf = 'gaussian_log_mvncdf_mendell_elston' |
| 161 |
,tex |
| 162 |
); |
| 163 |
|
| 164 |
% check whether options have been correctly passed by the preprocessor |
| 165 |
if (options_.skewed_kalman.prune_tol ~= 0.02) || ... |
| 166 |
(options_.skewed_kalman.rank_deficiency_transform ~= true) || ... |
| 167 |
~strcmp(options_.skewed_kalman.mvnlogcdf, 'gaussian_log_mvncdf_mendell_elston') |
| 168 |
error('skewed kalman options were not passed correctly') |
| 169 |
end |
| 170 |
|
| 171 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 172 |
% TEST 4: Likelihood with skew normal model should be higher than with Gaussian model % |
| 173 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% |
| 174 |
loglik_csn = oo_.posterior.optimization.log_density; % expected to be approx 1215.68 |
| 175 |
if (loglik_csn-1215.6815424) > 1e-5 |
| 176 |
error('Something wrong with CSN likelihood') |
| 177 |
end |
| 178 |
loglik_gauss = struct5.loglik; % expected to be approx 1207.56 |
| 179 |
if (loglik_gauss-1207.5618755) > 1e-5 |
| 180 |
error('Something wrong with Gaussian likelihood') |
| 181 |
end |
| 182 |
lr_stat = 2 * (loglik_csn - loglik_gauss); |
| 183 |
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) |
| 184 |
lr_str = sprintf('test statistic of $%.2f$ and a p-value of $%.4f$.', lr_stat, lr_pval); |
| 185 |
fprintf('\n%s\n* LIKELIHOOD RATIO TEST *\n%s\n', repmat('*',1,25), repmat('*',1,25)); |
| 186 |
fprintf('- unconstrained log-likelihood (CSN): %.4f\n', loglik_csn); |
| 187 |
fprintf('- constrained log-likelihood (Gaussian): %.4f\n', loglik_gauss); |
| 188 |
fprintf('- %s\n\n',lr_str); |