sur gibbs sampling: first pass

matlab/surgibbs.m

@@ -0,0 +1,134 @@
+function surgibbs(ds, A, ndraws, varargin)
+%function sur(ds)
+% Implements Gibbs Samipling for SUR
+%
+% INPUTS
+%   ds      [dseries]    data
+%   A       [matrix]     prior distribution variance
+%   ndraws  [int]        number of draws
+%
+% OUTPUTS
+%   none
+%
+% SPECIAL REQUIREMENTS
+%   none
+
+% Copyright (C) 2017 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 <http://www.gnu.org/licenses/>.
+
+global M_ oo_
+
+%% Check input
+assert(nargin == 1 || nargin == 3, 'Incorrect number of arguments passed to sur');
+
+jsonfile = [M_.fname '_original.json'];
+if exist(jsonfile, 'file') ~= 2
+    error('Could not find %s! Please use the json option (See the Dynare invocation section in the reference manual).', jsonfile);
+end
+
+%% Get Equations
+jsonmodel = loadjson(jsonfile);
+jsonmodel = jsonmodel.model;
+[lhs, rhs, lineno] = getEquationsByTags(jsonmodel, varargin{:});
+
+m = length(lhs);
+if m <= 1
+    error('SUR estimation requires the selection of at least two equations')
+end
+
+%% Construct regression matrices
+Y = dseries();
+Xi =  cell(m, 1);
+pnamesall = [];
+vwlagsall = [];
+for i = 1:m
+    Y = [Y ds{lhs{i}}];
+
+    rhs_ = strsplit(rhs{i}, {'+','-','*','/','^','log(','exp(','(',')'});
+    rhs_(cellfun(@(x) all(isstrprop(x, 'digit')), rhs_)) = [];
+    vnames = setdiff(rhs_, cellstr(M_.param_names));
+    regexprnoleads = cell2mat(strcat('(', vnames, {'\(\d+\))|'}));
+    if ~isempty(regexp(rhs{i}, regexprnoleads(1:end-1), 'match'))
+        error(['olseqs: you cannot have leads in equation on line ' ...
+            lineno{i} ': ' lhs{i} ' = ' rhs{i}]);
+    end
+    regexpr = cell2mat(strcat('(', vnames, {'\(-\d+\))|'}));
+    vwlags = regexp(rhs{i}, regexpr(1:end-1), 'match');
+
+    % Find parameters
+    pnames = cell(1, length(vwlags));
+    for j = 1:length(vwlags)
+        regexmatch = regexp(rhs{i}, ['(\w*\*?)?' strrep(strrep(vwlags{j}, '(', '\('), ')', '\)') '(\*?\w*)?'], 'match');
+        regexmatch = strsplit(regexmatch{:}, '*');
+        assert(length(regexmatch) == 2);
+        if strcmp(vwlags{j}, regexmatch{1})
+            pnames{j} = regexmatch{2};
+        else
+            pnames{j} = regexmatch{1};
+        end
+    end
+    pnamesall = [pnamesall pnames];
+    vwlagsall = [vwlagsall vwlags];
+    Xi{i} = cellfun(@eval, strcat('ds.', vwlags), 'UniformOutput', false);
+end
+
+fp = Y.firstobservedperiod;
+lp = Y.lastobservedperiod;
+for i = 1:m
+    X = dseries();
+    for j = 1:length(Xi{i})
+        X = [X dseries(Xi{i}{j}.data, Xi{i}{j}.dates, ['V' num2str(i) num2str(j)])];
+    end
+    Xi{i} = X;
+    fp = max(fp, X.firstobservedperiod);
+    lp = min(lp, X.lastobservedperiod);
+end
+Y = Y(fp:lp).data(:);
+X = [];
+for i = 1:m
+    Xi{i} = Xi{i}(fp:lp).data;
+    ind = size(X);
+    X(ind(1)+1:ind(1)+size(Xi{i}, 1), ind(2)+1:ind(2)+size(Xi{i},2)) = Xi{i};
+end
+
+%% Estimation
+nobs = length(fp:lp);
+nvars = size(X, 2);
+
+[q, r] = qr(X, 0);
+resid = Y - X * (r\(q'*Y));
+resid = reshape(resid, nobs, m);
+S = resid'*resid/nobs;
+tmp = kron(inv(S), eye(nobs));
+beta0 = (X'*tmp*X)\X'*tmp*Y;
+beta = beta0;
+
+oo_.surgibbs.betadraws = zeros(ndraws, nvars);
+for i = 1:ndraws
+    % Draw S, given X, Y, Beta
+    resid = reshape(Y - X*beta, nobs, m);
+    Omega = rand_inverse_wishart(m, nobs, (resid'*resid)/nobs);
+
+    % Draw beta, given X, Y, S
+    tmp = kron(inv(Omega), eye(nobs));
+    tmp1 = X'*tmp*X;
+    Omegabar = inv(tmp1 + A);
+    betabar = Omegabar*(tmp1*(tmp1\X'*tmp*Y)+A*beta0);
+    Sigma_upper_chol = chol(Omegabar, 'upper');
+    beta = rand_multivariate_normal(betabar', Sigma_upper_chol, nvars)';
+    oo_.surgibbs.betadraws(i, :) = beta';
+end

tests/ecb/contribution-plots/var.mod

@@ -32,4 +32,6 @@ olseqs(ds1, 'eqnum', {'ffr', 'cpi'});
 
 sur(ds1);
 
+surgibbs(ds1, randn(17,17), 1000);
+
 plot_contributions('eqnum', 'ffr', ds1, ds0);