2006-09-13 14:39:23 +02:00
|
|
|
|
function [fval,cost_flag,ys,trend_coeff,info,PHI,SIGMAu,iXX] = DsgeVarLikelihood(xparam1,gend)
|
|
|
|
|
% stephane.adjemian@ens.fr
|
2007-01-26 15:50:30 +01:00
|
|
|
|
global bayestopt_ estim_params_ M_ options_
|
2006-09-13 14:39:23 +02:00
|
|
|
|
|
|
|
|
|
nvx = estim_params_.nvx;
|
|
|
|
|
nvn = estim_params_.nvn;
|
|
|
|
|
ncx = estim_params_.ncx;
|
|
|
|
|
ncn = estim_params_.ncn;
|
|
|
|
|
np = estim_params_.np;
|
|
|
|
|
nx = nvx+nvn+ncx+ncn+np;
|
|
|
|
|
ns = nvx+nvn+ncx+ncn;
|
|
|
|
|
|
2007-07-25 15:25:38 +02:00
|
|
|
|
NumberOfObservedVariables = size(options_.varobs,1);
|
|
|
|
|
NumberOfLags = options_.varlag;
|
|
|
|
|
k = NumberOfObservedVariables*NumberOfLags ;
|
|
|
|
|
|
2006-09-13 14:39:23 +02:00
|
|
|
|
mYY = evalin('base', 'mYY');
|
|
|
|
|
mYX = evalin('base', 'mYX');
|
|
|
|
|
mXY = evalin('base', 'mXY');
|
|
|
|
|
mXX = evalin('base', 'mXX');
|
|
|
|
|
|
|
|
|
|
fval = [];
|
|
|
|
|
cost_flag = [];
|
|
|
|
|
ys = [];
|
|
|
|
|
trend_coeff = [];
|
|
|
|
|
|
|
|
|
|
xparam1_test = xparam1;
|
|
|
|
|
|
|
|
|
|
cost_flag = 1;
|
|
|
|
|
nobs = size(options_.varobs,1);
|
|
|
|
|
|
|
|
|
|
if options_.mode_compute ~= 1 & any(xparam1 < bayestopt_.lb)
|
2007-01-26 15:50:30 +01:00
|
|
|
|
k = find(xparam1 < bayestopt_.lb);
|
|
|
|
|
fval = bayestopt_.penalty*min(1e3,exp(sum(bayestopt_.lb(k)-xparam1(k))));
|
2007-07-25 15:25:38 +02:00
|
|
|
|
info = 41;
|
2007-01-26 15:50:30 +01:00
|
|
|
|
cost_flag = 0;
|
|
|
|
|
return;
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
|
|
|
|
|
if options_.mode_compute ~= 1 & any(xparam1 > bayestopt_.ub)
|
2007-01-26 15:50:30 +01:00
|
|
|
|
k = find(xparam1 > bayestopt_.ub);
|
2007-07-25 15:25:38 +02:00
|
|
|
|
fval = bayestopt_.penalty*min(1e3,exp(sum(xparam1(k)- bayestopt_.ub(k))));
|
|
|
|
|
info = 42;
|
2007-01-26 15:50:30 +01:00
|
|
|
|
cost_flag = 0;
|
|
|
|
|
return;
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
|
|
|
|
|
Q = M_.Sigma_e;
|
|
|
|
|
for i=1:estim_params_.nvx
|
2007-01-26 15:50:30 +01:00
|
|
|
|
k = estim_params_.var_exo(i,1);
|
|
|
|
|
Q(k,k) = xparam1(i)*xparam1(i);
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
offset = estim_params_.nvx;
|
|
|
|
|
if estim_params_.nvn
|
2007-01-26 15:50:30 +01:00
|
|
|
|
disp('DsgeVarLikelihood :: Measurement errors are not implemented!')
|
|
|
|
|
return
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
if estim_params_.ncx
|
2007-07-25 15:25:38 +02:00
|
|
|
|
disp('DsgeVarLikelihood :: Correlated structural innovations are not yet implemented!')
|
2007-01-26 15:50:30 +01:00
|
|
|
|
return
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
2007-07-25 15:25:38 +02:00
|
|
|
|
|
2006-09-13 14:39:23 +02:00
|
|
|
|
M_.params(estim_params_.param_vals(:,1)) = xparam1(offset+1:end);
|
|
|
|
|
M_.Sigma_e = Q;
|
2007-07-25 15:25:38 +02:00
|
|
|
|
|
2006-09-13 14:39:23 +02:00
|
|
|
|
%% Weight of the dsge prior:
|
|
|
|
|
dsge_prior_weight = M_.params(strmatch('dsge_prior_weight',M_.param_names));
|
2007-07-25 15:25:38 +02:00
|
|
|
|
if dsge_prior_weight<(k+NumberOfObservedVariables)/nobs;
|
|
|
|
|
fval = bayestopt_.penalty*min(1e3,(k+NumberOfObservedVariables)/nobs-dsge_prior_weight);
|
2007-07-26 10:21:54 +02:00
|
|
|
|
info = 51
|
2007-07-25 15:25:38 +02:00
|
|
|
|
cost_flag = 0;
|
|
|
|
|
return;
|
|
|
|
|
end
|
2007-07-06 16:43:00 +02:00
|
|
|
|
|
2006-09-13 14:39:23 +02:00
|
|
|
|
%------------------------------------------------------------------------------
|
|
|
|
|
% 2. call model setup & reduction program
|
|
|
|
|
%------------------------------------------------------------------------------
|
2006-09-22 17:58:43 +02:00
|
|
|
|
[T,R,SteadyState,info] = dynare_resolve(bayestopt_.restrict_var_list,...
|
|
|
|
|
bayestopt_.restrict_columns,...
|
|
|
|
|
bayestopt_.restrict_aux);
|
2006-09-13 14:39:23 +02:00
|
|
|
|
if info(1) == 1 | info(1) == 2 | info(1) == 5
|
2007-01-26 15:50:30 +01:00
|
|
|
|
fval = bayestopt_.penalty;
|
|
|
|
|
cost_flag = 0;
|
|
|
|
|
return
|
2006-09-13 14:39:23 +02:00
|
|
|
|
elseif info(1) == 3 | info(1) == 4 | info(1) == 20
|
2007-01-26 15:50:30 +01:00
|
|
|
|
fval = bayestopt_.penalty*min(1e3,exp(info(2)));
|
|
|
|
|
cost_flag = 0;
|
|
|
|
|
return
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
if options_.loglinear == 1
|
2007-01-26 15:50:30 +01:00
|
|
|
|
constant = log(SteadyState(bayestopt_.mfys));
|
2006-09-13 14:39:23 +02:00
|
|
|
|
else
|
2007-01-26 15:50:30 +01:00
|
|
|
|
constant = SteadyState(bayestopt_.mfys);
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
if bayestopt_.with_trend == 1
|
2007-01-26 15:50:30 +01:00
|
|
|
|
disp('DsgeVarLikelihood :: Linear trend is not yet implemented!')
|
|
|
|
|
return
|
2006-09-13 14:39:23 +02:00
|
|
|
|
end
|
|
|
|
|
|
|
|
|
|
%------------------------------------------------------------------------------
|
|
|
|
|
% 3. theorretical moments (second order)
|
|
|
|
|
%------------------------------------------------------------------------------
|
|
|
|
|
tmp = lyapunov_symm(T,R*Q*R');% I compute the variance-covariance matrix
|
|
|
|
|
% of the restricted state vector.
|
2006-09-22 17:58:43 +02:00
|
|
|
|
bayestopt_.mf = bayestopt_.mf1;
|
2006-09-13 14:39:23 +02:00
|
|
|
|
mf = bayestopt_.mf1;
|
|
|
|
|
|
2007-07-06 16:43:00 +02:00
|
|
|
|
|
2006-09-13 14:39:23 +02:00
|
|
|
|
TheoreticalAutoCovarianceOfTheObservedVariables = ...
|
|
|
|
|
zeros(NumberOfObservedVariables,NumberOfObservedVariables,NumberOfLags+1);
|
|
|
|
|
TheoreticalAutoCovarianceOfTheObservedVariables(:,:,1) = tmp(mf,mf);
|
|
|
|
|
for lag = 1:NumberOfLags
|
|
|
|
|
tmp = T*tmp;
|
|
|
|
|
TheoreticalAutoCovarianceOfTheObservedVariables(:,:,lag+1) = tmp(mf,mf);
|
|
|
|
|
end
|
|
|
|
|
GYX = zeros(NumberOfObservedVariables,k);
|
|
|
|
|
for i=1:NumberOfLags
|
|
|
|
|
GYX(:,(i-1)*NumberOfObservedVariables+1:i*NumberOfObservedVariables) = ...
|
|
|
|
|
TheoreticalAutoCovarianceOfTheObservedVariables(:,:,i+1);
|
|
|
|
|
end
|
|
|
|
|
GXX = kron(eye(NumberOfLags), ...
|
|
|
|
|
TheoreticalAutoCovarianceOfTheObservedVariables(:,:,1));
|
|
|
|
|
for i = 1:NumberOfLags-1
|
|
|
|
|
tmp1 = diag(ones(NumberOfLags-i,1),i);
|
|
|
|
|
tmp2 = diag(ones(NumberOfLags-i,1),-i);
|
|
|
|
|
GXX = GXX + kron(tmp1,TheoreticalAutoCovarianceOfTheObservedVariables(:,:,i+1));
|
|
|
|
|
GXX = GXX + kron(tmp2,TheoreticalAutoCovarianceOfTheObservedVariables(:,:,i+1)');
|
|
|
|
|
end
|
|
|
|
|
|
|
|
|
|
GYY = TheoreticalAutoCovarianceOfTheObservedVariables(:,:,1);
|
|
|
|
|
|
|
|
|
|
assignin('base','GYY',GYY);
|
|
|
|
|
assignin('base','GXX',GXX);
|
|
|
|
|
assignin('base','GYX',GYX);
|
|
|
|
|
|
|
|
|
|
if ~isinf(dsge_prior_weight)
|
|
|
|
|
SIGMAu = dsge_prior_weight*gend*TheoreticalAutoCovarianceOfTheObservedVariables(:,:,1) + mYY ;
|
|
|
|
|
tmp1 = dsge_prior_weight*gend*GYX + mYX;
|
|
|
|
|
tmp2 = inv(dsge_prior_weight*gend*GXX+mXX);
|
|
|
|
|
SIGMAu = SIGMAu - tmp1*tmp2*tmp1';
|
2007-07-25 15:25:38 +02:00
|
|
|
|
if ~ispd(SIGMAu)
|
|
|
|
|
v = diag(SIGMAu);
|
|
|
|
|
k = find(v<0);
|
|
|
|
|
fval = bayestopt_.penalty*min(1e3,exp(abs(v(k))));
|
2007-07-26 10:21:54 +02:00
|
|
|
|
info = 52;
|
2007-07-25 15:25:38 +02:00
|
|
|
|
cost_flag = 0;
|
|
|
|
|
return;
|
|
|
|
|
end
|
2006-09-13 14:39:23 +02:00
|
|
|
|
SIGMAu = SIGMAu / (gend*(dsge_prior_weight+1));
|
|
|
|
|
PHI = tmp2*tmp1';
|
|
|
|
|
prodlng1 = sum(gammaln(.5*((1+dsge_prior_weight)*gend- ...
|
|
|
|
|
NumberOfObservedVariables*NumberOfLags ...
|
|
|
|
|
+1-(1:NumberOfObservedVariables)')));
|
|
|
|
|
prodlng2 = sum(gammaln(.5*(dsge_prior_weight*gend- ...
|
|
|
|
|
NumberOfObservedVariables*NumberOfLags ...
|
|
|
|
|
+1-(1:NumberOfObservedVariables)')));
|
|
|
|
|
lik = .5*NumberOfObservedVariables*log(det(dsge_prior_weight*gend*GXX+mXX)) ...
|
|
|
|
|
+ .5*((dsge_prior_weight+1)*gend-k)*log(det((dsge_prior_weight+1)*gend*SIGMAu)) ...
|
|
|
|
|
- .5*NumberOfObservedVariables*log(det(dsge_prior_weight*gend*GXX)) ...
|
|
|
|
|
- .5*(dsge_prior_weight*gend-k)*log(det(dsge_prior_weight*gend*(GYY-GYX*inv(GXX)*GYX'))) ...
|
|
|
|
|
+ .5*NumberOfObservedVariables*gend*log(2*pi) ...
|
|
|
|
|
- .5*log(2)*NumberOfObservedVariables*((dsge_prior_weight+1)*gend-k) ...
|
|
|
|
|
+ .5*log(2)*NumberOfObservedVariables*(dsge_prior_weight*gend-k) ...
|
|
|
|
|
- prodlng1 + prodlng2;
|
|
|
|
|
else % cod<6F> par SM (s<>rement pas exact... Que font ici les moments empiriques ?).
|
|
|
|
|
tmp1 = GYX;
|
|
|
|
|
tmp2 = inv(GXX);
|
|
|
|
|
PHI = tmp2*tmp1';
|
|
|
|
|
SIGMAu = GYY - tmp1*tmp2*tmp1;
|
|
|
|
|
% <20> finir de corriger...
|
|
|
|
|
lik = -.5*sum(diag(inv(tmp2)*(mYY-2*tmp1'*mYX'+tmp1'*mXX*tmp1))) ...
|
|
|
|
|
-(gend/2)*log(det(tmp2));
|
|
|
|
|
end
|
|
|
|
|
|
|
|
|
|
lnprior = priordens(xparam1,bayestopt_.pshape,bayestopt_.p1,bayestopt_.p2,bayestopt_.p3,bayestopt_.p4);
|
|
|
|
|
fval = (lik-lnprior);
|
|
|
|
|
iXX = tmp2;
|