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
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
|
function [yhat,Estr,rcon,Rcon,u,v,d] = fn_fcstidcnd2(valuecon,stepcon,varcon,nstepsm,...
nconstr,eq_ms,nvar,lags,phil,Aband,Sband,yfore_h,imf3s_h,A0_h,Bh_h,...
forep,TLindx,TLnumber,nCms,eq_Cms,nconadd,eq_oth,radd,Radd)
% [yhat,Estr,rcon,Rcon,u,v,d] = fn_fcstidcnd2(valuecon,stepcon,varcon,nstepsm,...
% nconstr,eq_ms,nvar,lags,phil,Aband,Sband,yfore_h,imf3s_h,A0_h,Bh_h,...
% forep,TLindx,TLnumber,nCms,eq_Cms,nconadd,eq_oth,radd,Radd)
%
% 5/2/99: This one is much, much slower than fidcnderr.m when eq_ms=[] and
% nconstr>0 and length(eq_ms)*nstepsm>nconstr. Seems to me that fcstidcnd2.m
% is designed for the situation which eq_ms=2. The slow speed is due to
% "Radd" added. When I have time, I need to check to see if I can speed up
% this M file.
%
% Note that the case Aband=1 is not finished yet.
%
% Conditional forecasting in the identified model with or without error bands
% It handles conditions on average values as well, so "valuecon" must be
% expressed at average (NOT sum) level. Including unconditional forecasts
% when nconstr=nCms=0. Maybe slower than "fcstidcnd.m" when length(eq_ms)*nstepsm=nconstr
% because of Radd added. But "fsctidcnd.m" is incorrect when length(eq_ms)
% *nstepsm>nconstr (situations like the average annual FFR constraint). 3/22/99
% Ref.: Zha Forecast (I), pp. 17-17c.
%
% valuecon: vector of values conditioned
% stepcon: sequence (cell) of steps conditioned; if length(stepcon{i}) > 1, the condition
% is then an arithmetic average of log(y) over the stepcon{i} period.
% varcon: vector of variables conditioned
% nconstr: number of DLS constraints
% nstepsm: maximum number of steps in all DLS constraints
% nvar: number of variables in the BVAR model
% lags: number of lags in the BVAR model
% phil: the 1-by-(nvar*lags+1) data matrix where k=nvar*lags+1
% (last period plus lags before the beginning of forecast)
% eq_ms: identified equation or shock (in terms of number). If equ_ms=[], then "fidencond" or
% 'fcstidcond' is, similar to RATS, to compute forecasts with *all* shocks.
% Aband: 1: draws from A0 and Bh; 0: no draws
% Sband: 1: draws from random shocks E; 0: no random shocks
% yfore_h: uncondtional forecasts: forep-by-nvar. Never used when nconstr=0.
% In this case, set it to [];
% imf3s_h: 3-dimensional impulse responses matrix: impsteps-by-nvar shocks-by-nvar responses
% Never used when nconstr=0. In this case, set it to [];
% A0_h: A0 contemporaneous parameter matrix
% Bh_h: reduced-form parameter matrix: k-by-nvar, y(t) = X(t)*Bh+e(t)
% where X(t) is k-by-nvar and y(t) is 1-by-nvar
% forep: # of forecast periods (e.g., monthly for a monthly model)
% TLindx: 1-by-nCms vector of 1's and 0's, indicating tight or loose; 1: tighter, 0: looser
% Used only when /* (MS draws) is activated. Right now, MS shocks are deterministic.
% TLnumber: 1-by-nCms vector, lower bound for tight and upper bound for loose
% nCms: # of LZ conditions
% eq_Cms: equation location of MS shocks
% nconadd: number of DLS added constraints DIRECTLY on structural shocks
% for identified version where eq_ms is not empty. Note
% nconadd=0 when eq_ms is empty.
% eq_oth: index for other structural shocks than those in eq_ms.
% If eq_ms=[], eq_oth=[]. Note length(eq_oth)+length(eq_ms)=nvar
% radd: sparse nconadd-by-1; nconadd values added to rcon in the text later
% Radd: sparce nvar*nstepsm-by-nconadd; added to Rcon in the text later
% ------
% yhat: conditional forecasts: forep-by-nvar
% Estr: backed-out structural shocks (from N(0,1))
% rcon: vector - the difference between valuecon and log(yfore) (unconditional forecasts)
% Rcon: k-by-q (q constranits and k=nvar*max(nsteps)) so that
% Rcon'*e = rcon where e is k-by-1
% [u,d,v]: svd(Rcon,0)
%
% Written by Tao Zha March 1998. Revised November 1998;
% 3/20/99 Disenabled draws of MS shcoks. To enable it, activate /* part
% 3/20/99 Added A0_h and forep and deleted Cms as input argument.
% 3/21/99 Added nconadd and eq_oth.
% Previous programs may not be compatible.
%
% This version differs from "fn_fcstidcnd.m" because nconadd, radd, and Radd are added to allow for
% both DLS-WZ conditions and structural-shock conditions to exist at the same time. My recognition
% is that it does not work well at all. So one should use the version "fn_fcstidcnd.m."" April, 2009 by TZ.
%
% 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/>.
%
%
%% See Zha's note "Forecast (1)" p. 5, RATS manual (some errors in RATS), etc.
%% See also Zha Forecast (I), pp. 17-17c.
%
%% Some notations: y(t+1) = y(t)B1 + e(t+1)inv(A0). e(t+1) is 1-by-n.
%% Let r(t+1)=e(t+1)inv(A0) + e(t+2)C + .... where inv(A0) is impulse
%% response at t=1, C at t=2, etc. The row of inv(A0) or C is
%% all responses to one shock.
%% Let r be q-by-1 (such as r(1) = r(t+1)
%% = y(t+1) (constrained) - y(t+1) (forecast)).
%% Use impulse responses to find out R (k-by-q) where k=nvar*nsteps
%% where nsteps the largest constrained step. The key of the program
%% is to creat R using impulse responses
%% Optimal solution for shock e where R'*e=r and e is k-by-1 is
%% e = R*inv(R'*R)*r and k>=q
%
DLSIdShock = ~isempty(eq_ms); % if not empty, the MS shock is identified as in DLS
impsteps=size(imf3s_h,1);
if (forep<nstepsm) | (impsteps<nstepsm)
disp('Increase # of forecast or impulse steps!!')
disp('Or decrease # of constraints (nconstr) or constrained steps (stepcon(i))!!')
error('Maximum of conditional steps > # of forecast or impulse steps!!')
end
kts = nvar*nstepsm; % k -- ts: total shocks some of which are restricted and others
% are free.
%*** initializing
Rcon = zeros(kts,nconstr); % R: k-by-q where q include possible additional
% constraints directly on structural shocks for identified models.
% These addtional constraints will be added later.
Econ = sparse(zeros(kts,1)); % E: k-by-1
rcon = zeros(nconstr,1); % r: q-by-1
%rcon=valuecon-diag(yfore(stepcon,varcon)); % another way is to use "loop" below.
tcwc = nvar*lags; % total coefficients without constant
phi=phil;
%----------------------------------------------------
% Form rcon, Rcon, and Econ (the mean of structural shocks)
%----------------------------------------------------
if nconstr
A0in = reshape(imf3s_h(1,:,:),nvar,nvar); % nvar shocks-by-nvar responses
for i=1:nconstr
rcon(i)=length(stepcon{i})*valuecon(i) - ...
sum(yfore_h(stepcon{i},varcon(i)),1); %<<>>
Rmat = zeros(nstepsm,nvar);
% Must be here inside the loop because it's matrix of one column of Rcon
for j=1:length(stepcon{i})
if DLSIdShock % Assuming the Fed can't see all other shocks within a month
Rmat(1:stepcon{i}(j),:) = Rmat(1:stepcon{i}(j),:) + ...
imf3s_h(stepcon{i}(j):-1:1,:,varcon(i));
% Rmat: row--nstepsm, column--nvar shocks (here all shocks except
% the identified one are set to zero) for a particular
% endogenous variable 'varcon(i)'. See Zha Forcast (1), pp.6-7
else % Rcon random with (A0,A+)
Rmat(1:stepcon{i}(j),:) = Rmat(1:stepcon{i}(j),:) + ...
imf3s_h(stepcon{i}(j):-1:1,:,varcon(i));
% Rmat: row--nstepsm, column--nvar shocks (here all shocks are
% *not* set to zero) for a particular endogenous
% variable 'varcon(i)'. See Zha Forcast (1), pp.6-7
end
end
Rmatt = Rmat'; % Now, nvar-by-nstepsm. I think here is where RATS has an error
% i.e. "OVERR" is not transposed when overlaid to "CAPR"
Rcon(:,i)=Rmatt(:); % Rcon: k-by-q where q=nconstr
end
rcon = [rcon;radd]; % added nconadd constrained values for identified model
% sparse because radd is sparse
Rcon = [Rcon Radd]; % added constraints on shocks for identified version
% sparse because Radd is sparse
Rcont=Rcon';
rinvrtr=Rcon/(Rcont*Rcon);
Econ = rinvrtr*rcon;
%/* Too slow, I believe, when q is large. 3/21/99
% [u d v]=svd(Rcon,0); %trial
% %???? Can we reduce the time by computing inv(R'*R) directly?
% % rtr = Rcon'*Rcon; %trial
% % rtrinv = inv(Rcon'*Rcon); %trial
% vd=v.*(ones(size(v,2),1)*diag(d)'); %trial
% dinv = 1./diag(d); % inv(diag(d))
% vdinv=v.*(ones(size(v,2),1)*dinv'); %trial
% rtr=vd*vd'; % R'*R
% rtrinv = vdinv*vdinv'; % inv(R'*R)
%
% Econ = Rcon*rtrinv*rcon; % E = R*inv(R'R)*r; the mean of structural shocks
else
Econ = sparse(zeros(kts,1)); % the mean of shocks is zero under no variable condition
Rcon = NaN;
rcon = NaN;
u = NaN;
d = NaN;
v = NaN;
end
%---------------------------------------
% No uncertainty at all or only random (A0,A+)
% In other words, no future shocks
%---------------------------------------
if (~Sband) | (nconstr & length(eq_ms)*nstepsm==nconstr)
% length(eq_ms)==1 implies one-one mapping between MS shocks and, say, FFR
% if nstepsm==nconstr. If this condition does not hold, this procedure
% is incorrect. I don't have time to fix it now (3/20/99). So I use
% this as a proximation
Estr = reshape(Econ,nvar,nstepsm);
Estr = Estr'; % transpose so that
% Estr: structural shocks. Row--steps, Column--n shocks
Estr = [Estr;zeros(forep-nstepsm,nvar)];
% Now, forep-by-nvar -- ready for forecasts
Estr(1:nCms,eq_Cms) = TLnumber(:);
Ures = Estr/A0_h; % nstepsm-by-nvar
% Ures: reduced-form residuals. Row--steps; Column--n shocks
% ** reconstruct x(t) for y(t+h) = x(t+h-1)*B
% ** where phi = x(t+h-1) with last column being constant
%
yhat = zeros(forep,nvar);
for k=1:forep
yhat(k,:) = phi*Bh_h + Ures(k,:);
phi(1,nvar+1:tcwc) = phi(1,1:tcwc-nvar);
phi(1,1:nvar) = yhat(k,:);
%
end
%---------------------------------------
% With random future shocks and possibly (A0,A+) depending
% on if imf3s_h is random
%---------------------------------------
else
%--------------
% Condition on variables and A random
%--------------
if nconstr & Aband
warning(' ')
disp('This situation (both E and A random) is still under construction')
disp('It is closely related to Waggoner and Zha ReStat Gibbs sampling method')
disp('and related to "if DLSIdShock" in the following')
disp('Please press ctrl-c to abort')
pause
elseif nconstr
%--------------
% Condition on variables and DLS MS shock, no A random but S random
%--------------
if DLSIdShock % other shocks are indepedent of the eq_ms shock
Ome=speye(kts)-rinvrtr*Rcont;
Ome(find(abs(Ome)<1e-10))=0; % tighter the tolerance is, the longer it
% takes for svds to compute. E.g., size(Ome,1)*eps is perhaps very
% tight. After all, when the number in Ome is very small relative to
% 1 (which is the variance of structural shocks Estr), we can treat it
% as zero.
[u1 d1 v1] = svds(Ome,kts-nconstr-nconadd);
% We have exactly nconstr+nconadd zero singular values
Stdcon = u1*diag(sqrt(diag(abs(d1)))); % find a square root of Ome
% see Zha's forecast (1), p.17
Stdcon = [Stdcon sparse(zeros(kts,kts-size(d1,1)))];
Estr1 = Econ + Stdcon*randn(kts,1); % We must have rand(kts,1). Assigning
% some of randn(kts,1) to be zero according to Radd is WRONG.
% For this argument, see Zha Forecast (I) p.17a
Estr2 = reshape(Estr1,nvar,nstepsm);
% nvar-by-nstepsm; Needed to be transposed later
Estr = [Estr2';randn(forep-nstepsm,nvar)]; % transpose so that
% Estr2: structural shocks. Row--nstepsm, Column--n shocks
% Now, forep-by-nvar -- ready for forecasts
else
Ome=speye(kts)-rinvrtr*Rcont;
Ome(find(abs(Ome)<1e-10))=0; % tighter the tolerance is, the longer it
% takes for svds to compute. E.g., size(Ome,1)*eps is perhaps very
% tight. After all, when the number in Ome is very small relative to
% 1 (which is the variance of structural shocks Estr), we can treat it
% as zero.
[u1 d1 v1] = svds(Ome,kts-nconstr-nconadd);
% We have exactly nconstr+nconadd zero singular values
Stdcon = u1*diag(sqrt(diag(abs(d1)))); % find a square root of Ome
% see Zha's forecast (1), p.17
Stdcon = [Stdcon sparse(zeros(kts,nconstr+nconadd))];
%*** The following very inefficient
% Ome = eye(kts) - u*u'; % note, I-u*u' = I - R*inv(R'*R)*R'
% %[u1 d1 v1] = svd(Ome); % too slow
% [u1 d1] = eig(Ome);
% Stdcon = u1*diag(sqrt(diag(abs(d1)))); % lower triagular chol of conditional variance
% % see Zha's forecast (1), p.17
%--------------
% Condition on variables and LZ MS shock, no A random but S random
% This section has not be tested yet, 10/14/98
%--------------
if nCms
Estr1 = Econ + Stdcon*randn(kts,1);
Estr2 = reshape(Estr1,nvar,nstepsm);
Estr2 = Estr2'; % transpose so that
% Estr2: structural shocks. Row--nstepsm, Column--n shocks
Estr = [Estr2;randn(forep-nstepsm,nvar)];
% Now, forep-by-nvar -- ready for forecasts
Estr(1:nCms,eq_Cms) = TLnumber(:);
%/* draw MS shocks
% for k=1:nCms
% if TLindx(k) % tighter
% while (Estr(k,eq_Cms)<TLnumber(k))
% Estr(k,eq_Cms) = randn(1,1);
% end
% else % looser
% while (Estr(k,eq_Cms)>TLnumber(k))
% Estr(k,eq_Cms) = randn(1,1);
% end
% end
% end
%--------------
% Condition on variables only, no A random but S random
%--------------
else
Estr1 = Econ + Stdcon*randn(kts,1);
Estr2 = reshape(Estr1,nvar,nstepsm);
Estr2 = Estr2'; % transpose so that
% Estr2: structural shocks. Row--nstepsm, Column--n shocks
Estr = [Estr2;randn(forep-nstepsm,nvar)];
% Now, forep-by-nvar -- ready for forecasts
end
end
%--------------
% Condition on LZ MS shocks only, S random and possibly A random depending on
% if A0_h and Bh_h are random
%--------------
else
if nCms
Estr = randn(forep,nvar);
% Now, forep-by-nvar -- ready for forecasts
Estr(:,eq_Cms)=0;
Estr(1:nCms,eq_Cms) = TLnumber(:);
%/* draw MS shocks
% for k=1:nCms
% if TLindx(k) % tighter
% while (Estr(k,eq_Cms)<TLnumber(k))
% Estr(k,eq_Cms) = randn(1,1);
% end
% else % looser
% while (Estr(k,eq_Cms)>TLnumber(k))
% Estr(k,eq_Cms) = randn(1,1);
% end
% end
% end
else
Estr = randn(forep,nvar); % Unconditional forecast
% Now, forep-by-nvar -- ready for forecasts
end
end
%
Ures = Estr/A0_h; % nstepsm-by-nvar
% Ures: reduced-form residuals. Row--steps; Column--n shocks
% ** reconstruct x(t) for y(t+h) = x(t+h-1)*B
% ** where phi = x(t+h-1) with last column being constant
%
yhat = zeros(forep,nvar);
for k=1:forep
yhat(k,:) = phi*Bh_h + Ures(k,:);
phi(1,nvar+1:tcwc) = phi(1,1:tcwc-nvar);
phi(1,1:nvar) = yhat(k,:);
end
end
|
