File: imfsim.m

package info (click to toggle)
dynare 5.3-1
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 77,852 kB
  • sloc: cpp: 94,481; ansic: 28,551; pascal: 14,532; sh: 5,453; objc: 4,671; yacc: 4,442; makefile: 2,923; lex: 1,612; python: 677; ruby: 469; lisp: 156; xml: 22
file content (346 lines) | stat: -rw-r--r-- 14,444 bytes parent folder | download | duplicates (8)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
function imfsim(xinput)
% imfsim(xinput)
%        Save the simulated pdfs of impulse responses;
%        Print and save Gelman's measures of B and W for A0's,
%                    only when nstarts (# of starting points) >1.
%        Ref:  Waggoner and Zha "Does Normalization Matter for Inference"
%        See note Forecast (2)
%
% xinput{1}: nfp -- total number of free parameters
% xinput{2}: nvar -- number of variables
% xinput{3}: xhat -- ML estimate of free parameters in A0
% xinput{4}: hess -- Hessian of -logLH
% xinput{5}:Indxv -- index for selected variables of interest; normall first 2 are of our interest
%        to select variables, always check idmat0 to make sure
%        it plots: (1) pdf of 1st v for every buffer, (2) scattered plot of 1st and 2nd for every buffer,
%                  (2) pdf of 1st v for all sequences; (4) scattered plot of 3rd and 4th for all sequences
%                  (5) scattered plot of 1st and 2nd for al sequences.
% xinput{6}: IndxGraph - 1: plot graphs; 0: no graphs
% xinput{7}: idmat0 -- Index for non-zero elements in A0 with column to equation
% xinput{8}: nstarts -- # of starting points in Gibbs sampling
% xinput{9}: ndraws1 -- # of 1st loop to be discarded
% xinput{10}: ndraws2 -- # of 2nd loop for saving A0 draws
% xinput{11}: imndraws=nstarts*ndraws2
% xinput{12}: a0indx -- index number for non-zero elements in A0
% xinput{13}: tdf -- degrees of freedom for t-distribution
% xinput{14}: nbuffer -- interval for printing, plotting, and saving
% xinput{15}: Sbd -- nvar-by-nvar S{1}, ..., S{n} -- kind of covariance matrix for each simultaneous equation
%             Already divided by "fss."
% xinput{16}: nSample -- the original sample size including lags
% xinput{17}: IndxNmlr -- index for which normalization rule to choose
% xinput{18}: IndxGibbs -- index for WZ Gibbs; 1; Gibbs; 0: Metropolis
% xinput{19}: scf -- reduction scale factor for Metropolis jumping kernel
% xinput{20}: H_sr -- square root of the inverse of the covariance matrix
%             for free elements in A0 (nfp-by-nfp)
% xinput{21}: fss -- effective sample size == nSample-lags+# of dummy observations
% xinput{22}: idfile1 -- calls "iden6std."   Save stds. of both data and impulse responses in idfile1
% xinput{23}: xxhpc -- chol(X'X+inv(H_p_tilde)): upper triangular but its transpose
%                                      is lower triagular Choleski
% xinput{24}: ImfErr -- if 1, impulse response simulation; if 0, disable this simulation
% xinput{25}: ninv -- number of bins pre-specified to put each draw of impulse response
%                            into a proper bin (or small interval)
% xinput{26}: imstp -- # of steps for impulse responses
% xinput{27}: forep -- forecast periods (# of steps)
% xinput{28}: yact -- actual data (in log except R, U, etc.)
% xinput{29}: yactqg -- quarterly annualized growth in actual data
% xinput{30}: yactCalyg -- calendar annual growth in actual data
% xinput{31}: imfml -- imstp-by-nvar^2 ML impulse responses
% xinput{32}: forepq -- forecast periods for quarterly growth
% xinput{33}: forepy -- forecast periods for annual growth
% xinput{34}: ncoef -- k: # of coeffients per equation
% xinput{35}: Bhml -- ML reduced form parameter B (nvar-by-k)
% xinput{36}: lags -- # of lags
%------------------
% imfcntmulti: (ninv+2,imstp*nvar^2,nstarts) matrix
%           cnt: count; for impulse responses; multi (nstarts) sequences
% All output is saved in outB_W, including Range5, invc, ninv, and imfcntmulti
%
% Written by T. Zha 1999
%
% Copyright (C) 1997-2012 Tao Zha
%
% This 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.
%
% It 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.
%
% If you did not received a copy of the GNU General Public License
% with this software, see <http://www.gnu.org/licenses/>.
%

nfp=xinput{1}; nvar=xinput{2}; xhat=xinput{3}; hess=xinput{4}; Indxv=xinput{5};
IndxGraph=xinput{6}; idmat0=xinput{7}; nstarts=xinput{8}; ndraws1=xinput{9}; ndraws2=xinput{10};
imndraws=xinput{11}; a0indx=xinput{12}; tdf=xinput{13}; nbuffer=xinput{14}; Sbd=xinput{15};
nSample = xinput{16}; IndxNmlr=xinput{17}; IndxGibbs=xinput{18}; scf=xinput{19}; H_sr=xinput{20};
fss=xinput{21}; idfile1=xinput{22}; xxhpc=xinput{23}; ImfErr=xinput{24}; ninv=xinput{25};
imstp=xinput{26}; forep=xinput{27}; yact=xinput{28}; yactqg=xinput{29}; yactCalyg=xinput{30};
imfml=xinput{31}; forepq=xinput{32}; forepy=xinput{33}; ncoef=xinput{34}; Bhml=xinput{35};
lags=xinput{36};

Avhx_bs = zeros(nfp,1);
Avhx_bm = zeros(nfp,1);
Avhx_bj = zeros(nfp,1);
Avhx_cj = zeros(nfp,1);
Avhx_cm = zeros(nfp,1);
A0_h = zeros(nvar);
A0gbs = A0_h;    % drawn A0 in Gibbs sampling
Avhxm = zeros(nfp,1);
Avhxs = Avhxm;
A0xhat = zeros(nvar);
A0xhat(a0indx) = xhat;
%  A0hatw = zeros(nvar^2,nbuffer);

countJump = zeros(nstarts,1);

imfmean = zeros(imstp,nvar^2);
imfcntmulti = zeros(ninv+2,imstp*nvar^2,nstarts);
          % cnt: count; for impulse responses; multi (nstarts) sequences


%---------------------------------------------------
%  Specify the range for counting the empirical distribution
%
%** load the standard deviations of 6 variables, one for log(y), one for gq, and
%**   the third one for yg
eval(['load ' idfile1 '.prn -ascii']);
eval(['ABstd=' idfile1 ';']);
Range5 = cell(4,1);  % 4: log, qg, yg, and imf

%@@@ Tony's trick to expand the matrix
%
%** In order of log(y), qg, and yg for Range5{i} for i=1:3
Range5{1} =zeros(forep,nvar,2);  % 2: min and max
Range5{1}(:,:,1) = repmat(yact(length(yact(:,1)),:)-10*ABstd(1,:),[forep 1]);  % min, 30 std.
Range5{1}(:,:,2) = repmat(yact(length(yact(:,1)),:)+10*ABstd(1,:),[forep 1]);  % max, 30 std.
%
Range5{2} =zeros(forepq,nvar,2);  % 2: min and max
Range5{2}(:,:,1) = repmat(yactqg(length(yactqg(:,1)),:)-10*ABstd(2,:),[forepq 1]);  % min, 30 std.
Range5{2}(:,:,2) = repmat(yactqg(length(yactqg(:,1)),:)+10*ABstd(2,:),[forepq 1]);  % max, 30 std.
%
Range5{3} =zeros(forepy,nvar,2);  % 2: min and max
Range5{3}(:,:,1) = repmat(yactCalyg(length(yactCalyg(:,1)),:)-10*ABstd(3,:),[forepy 1]);  % min, 30 std.
Range5{3}(:,:,2) = repmat(yactCalyg(length(yactCalyg(:,1)),:)+10*ABstd(3,:),[forepy 1]);  % max, 30 std.
%
Range5{4} =zeros(imstp,nvar^2,2);  % 2: min and max
imfscale = repmat(ABstd(4,:),[1 nvar]);  % because nvar variables to 1, ..., nvar shocks
Range5{4}(:,:,1) = repmat(imfml(1,:)-5*imfscale,[imstp 1]);  % min, 5 std.
Range5{4}(:,:,2) = repmat(imfml(1,:)+5*imfscale,[imstp 1]);  % max, 5 std.
         % Range5(4)(:,:,1): imstp-by-nvar^2.  Column: nvar responses to 1st shock,
         %                        nvar responses to 2nd shock, ...
%**
invc = cell(4,1);     % interval length (used for counting later). 1st 3 cells have each
                      %    forep-by-nvar, and 4th has imstp-by-nvar^2.
for i=1:4
   invc{i} = Range5{i}(:,:,2) - Range5{i}(:,:,1);
end
hbin =  invc{4} ./ ninv;    % bin size for each point of impulse responses kkdf
imfloor = Range5{4}(:,:,1);



%===================================
%  Here begins with the big loop
%===================================
H1 = chol(hess);  % upper triangular so that H1' is a lower triangular decomp
baseW = H_sr;  %inv(H1);  %H_sr;   % covariance matrix without scaling
nswitch=0;  %<<>> total number of sign switches
A0inxhat = inv(A0xhat);   % inverse of ML estimate
a0indx0 = find(idmat0==0);    % index for all zero's in A0;
nn=[nvar lags imstp];

[cT,vR,kdf] = gibbsglb(Sbd,idmat0,nvar,fss);

tic
for starts = 1:nstarts
   starts
   if starts == 1
      A0gbs(a0indx) = xhat;   % from "load ..."
      if ~IndxGibbs   % Metropolist
         Avhx = xhat;
         hAvhx = a0asfun(Avhx,Sbd,fss,nvar,a0indx);
         hAvhx = -hAvhx;      % converted to logLH
      end
   else
      Avhx = baseW*randn(nfp,1);  %H_sr*randn(nfp,1);   % D: discarded sequence
      csq=randn(tdf,1);
      csq=sum(csq .* csq);
      Avhx = xhat+Avhx/sqrt(csq/tdf);
      %** Normalization by the choice of IndxNmlr
      A0gbs(a0indx) = Avhx;
      if ~IndxNmlr(5)
         [A0gbs,jnk] = nmlzvar(A0gbs,A0xhat,A0inxhat,IndxNmlr,nswitch,[]);
      else
         A0ingbs = inv(A0gbs);
         [A0gbs,jnk,jnk1] = nmlzvar(A0gbs,A0xhat,A0inxhat,IndxNmlr,nswitch,A0ingbs);
      end
      %
      if ~IndxGibbs   % Metropolist
         Avhx = A0gbs(a0indx);
         hAvhx = a0asfun(Avhx,Sbd,fss,nvar,a0indx);
         hAvhx = -hAvhx;      % converted to logLH
      end
   end
   %
   Avhxmm = zeros(nfp,1);
   Avhxss = zeros(nfp,1);
   cJump = 0;
   imfcnt = zeros(ninv+2,imstp*nvar^2);   % cnt: count; for impulse responses

   for draws = 1:ndraws1
      if IndxGibbs
         A0gbs = gibbsvar(A0gbs,cT,vR,nvar,fss,kdf);
      else     % Metropolis
         [Avhx,hAvhx,cJump] = smtplis(Avhx,hAvhx,tdf,cJump,scf,...
                       baseW,nfp,Sbd,fss,nvar,a0indx);
      end
   end

   wdraws=(starts-1)*ndraws2+0;
   for draws = 1:ndraws2
      drawsc = (starts-1)*ndraws2+draws;
      if IndxGibbs
         A0gbs = gibbsvar(A0gbs,cT,vR,nvar,fss,kdf);
         A0gbs(a0indx0) = 0; % set all zeros in A0gbs clean to avoid possible cumulative round-off errors
      else        % Metropolis
         [Avhx,hAvhx,cJump] = smtplis(Avhx,hAvhx,tdf,cJump,scf,...
                       baseW,nfp,Sbd,fss,nvar,a0indx);
         A0gbs(a0indx) = Avhx;
      end

      %*** call normalization so that A0_h is normalized
      if ~IndxNmlr(5)
         [A0_h,nswitch] = nmlzvar(A0gbs,A0xhat,A0inxhat,IndxNmlr,nswitch,[]);
         A0_hin = inv(A0_h);
      else
         A0ingbs = inv(A0gbs);
         [A0_h,nswitch,A0_hin] = nmlzvar(A0gbs,A0xhat,A0inxhat,IndxNmlr,nswitch,A0ingbs);
      end
      Avhx_norm = A0_h(a0indx);

      %
      % *** normal draws for posterior Aplus conditional on A0h
      %
      %** the mean is Aplushm, and the covariance is inv(xxhp)=Lxxhpc*Lxxhpc'
      Apindm = randn(ncoef,nvar);
      %
      if ~all(all(finite(Bhml)))
         Aplushm=zeros(ncoef,nvar);
         for i=1:nvar
            Aplushm(:,i)=Gb{i}*A0_h(:,i);    % see Zha's forecast (1) p.9
                   % Here, Gb is used to compute A+ where A+(i) = Gb(i)*a0(i)
         end
         Bh_h = (Aplushm + xxhpc\Apindm)*A0_hin;
      else
         Bh_h = Bhml + (xxhpc\Apindm)*A0_hin;
      end

      if ImfErr
         swish_h = A0_hin';     % Switching back to the form A0*y(t) = e(t)
         imf_h = zimpulse(Bh_h,swish_h,nn);   % in the form that is congenial to RATS
         %  imf3_h=reshape(imf_h,size(imf_h,1),nvar,nvar);
         %      % imf3: row--steps, column--nvar responses, 3rd dimension--nvar shocks
         imfmean = imfmean + imf_h;  % posterior mean
         imfcnt = empdfsort(imfcnt,imf_h,imfloor,hbin,ninv);
                       % sorted counts (prob.) in bins
      end

      Avhxm = Avhxm + Avhx_norm;      % 1st step to overall mean of parameter
      Avhxs = Avhxs + Avhx_norm.^2;   % 1st step to overall 2nd moment of parameter

      %* compute the mean and 2nd moment
      %** Getting average of variances W and variance of means B/n -- B_n
      %*   see Gelman, p.331, my Shock(0), 12-13, and my Forecast (2), 28-31
      if (nstarts>1)
         Avhxmm = Avhxmm + Avhx_norm;      % n*(phi_.j)
         Avhxss = Avhxss + Avhx_norm.^2;   % 1st step to (phi_ij) for fixed j
      end

      %  A0hatw(:,drawsc-wdraws) = A0_h(:);
      if ~mod(draws,nbuffer)
         starts
         draws
         wdraws=drawsc
         %  fwriteid = fopen('outA0.bin','a');
         %  count = fwrite(fwriteid,A0hatw,'double');
         %  status = fclose('all');
      end
   end
   %
   imfcntmulti(:,:,starts) = imfcnt;

   if ~IndxGibbs
      countJump(starts,1) = cJump;
   end
   %
   %** Getting average of variances W and variance of means B/n -- B_n
   %**   see Gelman, p.331, my Shock(0), pp.12-13, and my Forecast (2), pp.28-31
   if (nstarts>1)
      Avhx_aj = Avhxmm/ndraws2;         %  (phi_.j)
      Avhx_bs = Avhx_bs + Avhx_aj.^2;  %  1st step to 2nd moment of (phi_.j)
      Avhx_bm = Avhx_bm + Avhx_aj;     % 1st step to (phi_..)
      %
      Avhx_bj = Avhxss/ndraws2;          % 2nd moment of (phi_ij) for fixed j
      Avhx_cj = Avhx_bj - Avhx_aj.^2;   %  ((n-1)/n)*S^2_j
      Avhx_cm = Avhx_cm + Avhx_cj;     %  sum of ((n-1)/n)*S^2_j across j
   end
end
timend = toc
timeminutes=timend/60

if ~IndxGibbs
   countJump = countJump/ndraws2
end

Avhxm = Avhxm/(imndraws);
Avhxs = Avhxs/(imndraws);
Avhxv = Avhxs - Avhxm.^2;
Avhxs = sqrt(Avhxv);   % stardard deviation
A0hm = zeros(nvar);
A0hm(a0indx) = Avhxm   % mean
A0hv = zeros(nvar);
A0hv(a0indx) = Avhxv;  % varaince matrix
A0hs = zeros(nvar);
A0hs(a0indx) = Avhxs;   % standar deviation

imfmean = imfmean/(imndraws);

%**** Getting Within-Sequence W and Between-Sequence B_n
%        see Gelman, p.331, my Shock(0), pp.12-13, and my Forecast (2), pp.28-31
if (nstarts>1)
   AvhxW = (ndraws2/(ndraws2-1))*Avhx_cm/nstarts;
                                 % W: average of j within-sequence variances
   AvhxB_n = (nstarts/(nstarts-1)) * ( Avhx_bs/nstarts - (Avhx_bm/nstarts).^2 );
                                 % B/n:  variance of J within-sequence means
   AvhxB = ndraws2*AvhxB_n;    % B
   %
   B_W1 = AvhxB ./ AvhxW;
   B1 = AvhxB;
   W1 = AvhxW;
   GR1 = sqrt((ndraws2-1)/ndraws2 + B_W1/ndraws2);
      % measure of Gelman reduction; need not be 1 to be accurate,
      %               contrary to what Gelman claims
   save outB_W B_W1 B1 W1 GR1 nstarts ndraws2 imndraws timeminutes Avhxs ...
                  A0xhat A0hm A0hs A0hv IndxGibbs countJump
   if ImfErr
      save outB_W nswitch imfml imfmean imfcntmulti Range5 ninv invc imstp nvar -append
   end

   titstr = ['J ' num2str(nstarts) ' n1 ' num2str(ndraws1) ...
            ' n2 ' num2str(ndraws2) ' timend minutes ' num2str(timend/60)];
   disp(' ')
   disp(titstr)
   disp('B/W sqrt(B) sqrt(W) Std(A0) GR')
   format short g
   [B_W1 sqrt(B1) sqrt(W1) Avhxs GR1]
else
   save outB_W nstarts ndraws1 ndraws2 imndraws timeminutes Avhxs ...
                  A0xhat A0hm A0hs A0hv IndxGibbs countJump
   if ImfErr
      save outB_W nswitch imfml imfmean imfcntmulti Range5 ninv invc imstp nvar -append
   end
end