| 1 |
function compare_kalman_routines(experience) |
| 2 |
%% This function compares the kalman filter routines. |
| 3 |
%% |
| 4 |
%% |
| 5 |
%% stephane [DOT] adjemian [AT] ens [DOT] fr |
| 6 |
|
| 7 |
pp = experience.Number0fObservedVariables; |
| 8 |
mm = experience.SizeOfTheStateVector; |
| 9 |
rr = experience.NumberOfStructuralShocks; |
| 10 |
measurement_error_flag = experience.MeasurementErrors; |
| 11 |
gend = experience.NumberOfPeriods; |
| 12 |
PercentageOfMissingObservations = experience.PercentageOfMissingObservations; |
| 13 |
PeriodsWithMissingObservations = experience.PeriodsWithMissingObservations; |
| 14 |
|
| 15 |
|
| 16 |
|
| 17 |
%% SET VARIOUS PARAMETERS |
| 18 |
kalman_tol = 1e-12; |
| 19 |
riccati_tol =1e-9; |
| 20 |
start = 1; |
| 21 |
|
| 22 |
|
| 23 |
|
| 24 |
%% SET THE STATE SPACE MODEL: |
| 25 |
|
| 26 |
% I randomly choose the mm eigenvalues of the transition matrix. |
| 27 |
TransitionEigenvalues = rand(mm,1)*2-1; |
| 28 |
% I randomly choose the mm eigenvectors of the transition matrix |
| 29 |
tmp = rand(mm,mm*100); |
| 30 |
TransitionEigenvectors = tmp*tmp'/(mm*100); |
| 31 |
TransitionEigenvectors = rand(mm,mm); |
| 32 |
% I build the transition matrix |
| 33 |
T = TransitionEigenvectors*diag(TransitionEigenvalues)*inv(TransitionEigenvectors); |
| 34 |
% I randomly choose matrix R |
| 35 |
R = randn(mm,rr); |
| 36 |
% I randomly choose the covariance matrix of the structurtal innovations |
| 37 |
E = randn(rr,20*rr); |
| 38 |
Q = E*transpose(E)/(20*rr); |
| 39 |
% If needed I randomly choose the covariance matrix of teh measurement errors |
| 40 |
if measurement_error_flag == 0 |
| 41 |
H = zeros(pp,1); |
| 42 |
elseif measurement_error_flag == 1 |
| 43 |
H = rand(pp,1); |
| 44 |
elseif measurement_error_flag == 2 |
| 45 |
E = randn(pp,20*pp); |
| 46 |
H = E*transpose(E)/(20*pp); |
| 47 |
else |
| 48 |
disp('compare_kalman_routines:: Unknown option!') |
| 49 |
end |
| 50 |
% Set the selection vector (mf) |
| 51 |
MF = transpose(randperm(mm)); |
| 52 |
mf = MF(1:pp); |
| 53 |
|
| 54 |
P = lyapunov_symm(T,R*Q*R',1.000001); |
| 55 |
|
| 56 |
|
| 57 |
%% BUILD DATA SET (zero mean): |
| 58 |
|
| 59 |
if measurement_error_flag == 0 |
| 60 |
Y = simul_state_space_model(T,R,Q,mf,gend); |
| 61 |
elseif measurement_error_flag == 1 |
| 62 |
H = rand(pp,1); |
| 63 |
Y = simul_state_space_model(T,R,Q,mf,gend,diag(H)); |
| 64 |
elseif measurement_error_flag == 2 |
| 65 |
E = randn(pp,20*pp); |
| 66 |
H = E*transpose(E)/(20*pp); |
| 67 |
Y = simul_state_space_model(T,R,Q,mf,gend,H); |
| 68 |
else |
| 69 |
disp('compare_kalman_routines:: Unknown option!') |
| 70 |
end |
| 71 |
|
| 72 |
if PeriodsWithMissingObservations |
| 73 |
data_1 = Y(:,1:PeriodsWithMissingObservations)'; |
| 74 |
if PeriodsWithMissingObservations<gend |
| 75 |
data_2 = Y(:,PeriodsWithMissingObservations+1:end); |
| 76 |
else |
| 77 |
data_2 = []; |
| 78 |
end |
| 79 |
tmp = randperm(1:PeriodsWithMissingObservations*pp); |
| 80 |
tmp = tmp(1:round(PercentageOfMissingObservations*length(tmp))); |
| 81 |
data_1(tmp) = NaN; |
| 82 |
Y = [ transpose(data_1) ; data_2 ]; |
| 83 |
end |
| 84 |
|
| 85 |
[data_index,number_of_observations,no_more_missing_observations] = describe_missing_data(Y,gend,pp); |
| 86 |
|
| 87 |
if PeriodsWithMissingObservations==0 |
| 88 |
|
| 89 |
% kalman_filter.m |
| 90 |
if measurement_error_flag==0 |
| 91 |
HH = 0; |
| 92 |
elseif measurement_error_flag==1 |
| 93 |
HH = diag(H); |
| 94 |
elseif measurement_error_flag==2 |
| 95 |
HH = H; |
| 96 |
end |
| 97 |
instant0 = clock; |
| 98 |
[LIK1,lik1] = kalman_filter(T,R,Q,HH,P,Y,start,mf,kalman_tol,riccati_tol); |
| 99 |
T1 = etime(clock, instant0); |
| 100 |
disp(['kalman_filter = ' num2str(T1)]) |
| 101 |
|
| 102 |
% missing_observations_kalman_filter.m |
| 103 |
if measurement_error_flag==0 |
| 104 |
HH = zeros(pp,pp); |
| 105 |
elseif measurement_error_flag==1 |
| 106 |
HH = diag(H); |
| 107 |
elseif measurement_error_flag==2 |
| 108 |
HH = H; |
| 109 |
end |
| 110 |
instant0 = clock; |
| 111 |
[LIK2,lik2] = missing_observations_kalman_filter(T,R,Q,HH,P,Y,start,mf,kalman_tol,riccati_tol,data_index,number_of_observations,no_more_missing_observations); |
| 112 |
T2 = etime(clock, instant0); |
| 113 |
disp(['missing_observations_kalman_filter = ' num2str(T2)]) |
| 114 |
|
| 115 |
|
| 116 |
% univariate_kalman_filter.m |
| 117 |
if measurement_error_flag==0 |
| 118 |
HH = zeros(pp,pp); |
| 119 |
elseif measurement_error_flag==1 |
| 120 |
HH = H; |
| 121 |
elseif measurement_error_flag==2 |
| 122 |
HH = H; |
| 123 |
end |
| 124 |
if measurement_error_flag<2 |
| 125 |
instant0 = clock; |
| 126 |
[LIK3,lik3] = univariate_kalman_filter(T,R,Q,HH,P,Y,start,mf,kalman_tol,riccati_tol,data_index,number_of_observations,no_more_missing_observations); |
| 127 |
T3 = etime(clock, instant0); |
| 128 |
else |
| 129 |
instant0 = clock; |
| 130 |
[LIK3,lik3] = univariate_kalman_filter_corr(T,R,Q,HH,P,Y,start,mf,kalman_tol,riccati_tol,data_index,number_of_observations,no_more_missing_observations); |
| 131 |
T3 = etime(clock, instant0); |
| 132 |
end |
| 133 |
|
| 134 |
disp(['univariate_kalman_filter = ' num2str(T3)]) |
| 135 |
|
| 136 |
disp(' ') |
| 137 |
disp(' ') |
| 138 |
|
| 139 |
if abs(LIK2/LIK1-1)<1e-9 |
| 140 |
disp('missing data version is Ok') |
| 141 |
else |
| 142 |
disp('missing data version is wrong') |
| 143 |
disp(['percentage dev. = ' num2str((LIK2/LIK1-1)*100)]) |
| 144 |
end |
| 145 |
if abs(LIK3/LIK1-1)<1e-9 |
| 146 |
disp('univariate version is Ok') |
| 147 |
else |
| 148 |
disp('univariate version is wrong') |
| 149 |
disp(['percentage dev. = ' num2str((LIK3/LIK1-1)*100)]) |
| 150 |
%[lik1,lik3,lik1-lik3] |
| 151 |
%[lik1-lik3] |
| 152 |
end |
| 153 |
else |
| 154 |
% missing_observations_kalman_filter.m |
| 155 |
if measurement_error_flag==0 |
| 156 |
HH = zeros(pp,pp); |
| 157 |
elseif measurement_error_flag==1 |
| 158 |
HH = diag(H); |
| 159 |
elseif measurement_error_flag==2 |
| 160 |
HH = H; |
| 161 |
end |
| 162 |
instant0 = clock; |
| 163 |
LIK2 = missing_observations_kalman_filter(T,R,Q,HH,P,Y,start,mf,kalman_tol,riccati_tol,data_index,number_of_observations,no_more_missing_observations); |
| 164 |
T2 = etime(clock, instant0); |
| 165 |
disp(['missing_observations_kalman_filter = ' num2str(T2)]) |
| 166 |
|
| 167 |
% univariate_kalman_filter.m |
| 168 |
if measurement_error_flag==0 |
| 169 |
HH = zeros(pp,pp); |
| 170 |
elseif measurement_error_flag==1 |
| 171 |
HH = H; |
| 172 |
elseif measurement_error_flag==2 |
| 173 |
error('The univariate approach cannot handle correlated measurement errors') |
| 174 |
end |
| 175 |
instant0 = clock; |
| 176 |
LIK3 = univariate_kalman_filter(T,R,Q,HH,P,Y,start,mf,kalman_tol,riccati_tol,data_index,number_of_observations,no_more_missing_observations); |
| 177 |
T3 = etime(clock, instant0); |
| 178 |
disp(['univariate_kalman_filter = ' num2str(T2)]) |
| 179 |
|
| 180 |
disp(' ') |
| 181 |
disp(' ') |
| 182 |
|
| 183 |
if abs(LIK3/LIK2-1)<1e-9 |
| 184 |
disp('univariate version is Ok') |
| 185 |
else |
| 186 |
disp('univariate version is wrong') |
| 187 |
disp(['percentage dev. = ' num2str((LIK3/LIK2-1)*100)]) |
| 188 |
end |
| 189 |
end |