function ylevf = recmf_g(y,nlag,nfor,begf,prior,ndraw,nomit,r) % PURPOSE: Gibbs sampling forecasts for Bayesian error correction % model using Random-walk averaging prior % dy = A(L) DY + E, E = N(0,sige*V), % V = diag(v1,v2,...vn), rval/vi = ID chi(rval)/rval, rval = Gamma(m,k) % c = R A(L) + U, U = N(0,Z), Random-walk averaging prior %--------------------------------------------------- % USAGE: yfor = recmf_g(y,nlag,nfor,begf,prior,ndraw,nomit,r) % where: y = an (nobs x neqs) matrix of y-vectors in levels % nlag = the lag length % nfor = the forecast horizon % begf = the beginning date of the forecast % prior = a structure variable % prior.rval, rval prior hyperparameter, default=4 % prior.m, informative Gamma(m,k) prior on rval % prior.k, informative Gamma(m,k) prior on rval % prior.w, an (neqs x neqs) matrix containing prior means % (rows should sum to unity, see below) % prior.freq = 1 for annual, 4 for quarterly, 12 for monthly % prior.sig = prior variance hyperparameter (see below) % prior.tau = prior variance hyperparameter (see below) % prior.theta = prior variance hyperparameter (see below) % ndraw = # of draws % nomit = # of initial draws omitted for burn-in % r = # of cointegrating relations to use % (optional: this will be determined using % Johansen's trace test at 95%-level if left blank) % priors for important variables: N(w(i,j),sig) for 1st own lag % N( 0 ,tau*sig/k) for lag k=2,...,nlag % priors for unimportant variables: N(w(i,j) ,theta*sig/k) for lag 1 % N( 0 ,theta*sig/k) for lag k=2,...,nlag % e.g., if y1, y3, y4 are important variables in eq#1, y2 unimportant % w(1,1) = 1/3, w(1,3) = 1/3, w(1,4) = 1/3, w(1,2) = 0 % typical values would be: sig = .1-.3, tau = 4-8, theta = .5-1 % --------------------------------------------------- % NOTES: - estimation is carried out in annualized growth terms because % the prior means rely on common (growth-rate) scaling of variables % hence the need for a freq argument input. % - constant term included automatically % - x-matrix of exogenous variables not allowed % - error correction variables are automatically % constructed using output from Johansen's ML-estimator % --------------------------------------------------- % RETURNS % yfor = an nfor x neqs matrix of level forecasts for each equation %--------------------------------------------------- % SEE ALSO: becmf_g, rvarf_g, bvarf_g, recm_g %--------------------------------------------------- % References: LeSage and Krivelyova (1998) % ``A Spatial Prior for Bayesian Vector Autoregressive Models'', % forthcoming Journal of Regional Science, (on http://www.econ.utoledo.edu) % and % LeSage and Krivelova (1997) (on http://www.econ.utoledo.edu) % ``A Random Walk Averaging Prior for Bayesian Vector Autoregressive Models'' % written by: % James P. LeSage, Dept of Economics % University of Toledo % 2801 W. Bancroft St, % Toledo, OH 43606 % jpl@jpl.econ.utoledo.edu [nobs neqs] = size(y); % find # observations up to forecast period nmin = min(nobs,begf-1); nx = 0; if nargin == 8 % user is specifying the # of error correction terms to % include -- get them using johansen() jres = johansen(y,0,nlag); % recover error correction vectors ecvectors = jres.evec; index = jres.ind; % construct r-error correction variables x = mlag(y(:,index),1)*ecvectors(:,1:r); [nobs2 nx] = size(x); elseif nargin == 7 % we need to find r jres = johansen(y,0,nlag); % find r = # significant co-integrating relations using % the trace statistic output trstat = jres.lr1; tsignf = jres.cvt; r = 0; for i=1:neqs; if trstat(i,1) > tsignf(i,2) r = i; end; end; % recover error correction vectors ecvectors = jres.evec; index = jres.ind; % construct r error correction variables if r > 0 x = mlag(y(:,index),1)*ecvectors(:,1:r); [junk nx] = size(x); end; else error('Wrong # of arguments to recmf_g'); end; % do error checking here, even though it is redundant since % recm_g will do the same error checking. BUT, we avoid % confusing the poor user who will get error messages from % this routine that he called, rather than recm_g fields = fieldnames(prior); nf = length(fields); mm = 0; rval = 4; % rval = 4 is default nu = 0; d0 = 0; % default to a diffuse prior on sige for i=1:nf if strcmp(fields{i},'rval') rval = prior.rval; elseif strcmp(fields{i},'m') mm = prior.m; kk = prior.k; rval = gamm_rnd(1,1,mm,kk); % initial value for rval elseif strcmp(fields{i},'tau') tau = prior.tau; elseif strcmp(fields{i},'w') w = prior.w; [wchk1 wchk2] = size(w); if (wchk1 ~= wchk2) error('non-square w matrix in recmf_g'); elseif wchk1 > 1 if wchk1 ~= neqs error('wrong size w matrix in recmf_g'); end; end; elseif strcmp(fields{i},'theta') theta = prior.theta; elseif strcmp(fields{i},'sig') sig = prior.sig; elseif strcmp(fields{i},'freq') freq = prior.freq; end; end; if nlag < 1 error('Lag length less than 1 in recmf_g'); end; % truncate to begf-1 for estimation ytrunc = y(1:nmin,:); % call rvar_g with input information if r > 0 result = recm_g(ytrunc,nlag,prior,ndraw,nomit,r); else result = recm_g(ytrunc,nlag,prior,ndraw,nomit); end; % all we really care about is: % result(eq).bdraw = bhat draws for equation eq for j=1:neqs; b = mean(result(j).bdraw); bmat(:,j) = b'; end; yfor = zeros(nfor,neqs); ylev = zeros(nfor,neqs); % given bmat values generate future % growth rate forecasts dy = growthr(y,freq); ylevf = zeros(nfor,neqs); % storage for level forecasts % 1-step-ahead forecast xtrunc = [dy(nmin-(nlag):nmin,:) zeros(1,neqs)]; xfor = mlag(xtrunc,nlag); [xend junk] = size(xfor); xobs = xfor(xend,:); if nx > 0 ecterm = y(begf-1,index)*ecvectors(:,1:r); % add ec variables xvec = [xobs ecterm 1]; else xvec = [xobs 1]; end; % loop over equations for i=1:neqs; bhat = bmat(:,i); yfor(1,i) = xvec*bhat/100; % growth rate forecast ylevf(1,i) = (1+yfor(1,i))*y(begf-freq,i); % construct level forecasts end; xnew = zeros(nlag+1,neqs); % 2 through nlag-step-ahead forecasts for step=2:nlag; if step <= nfor xnew(1:nlag-step+1,:) = dy(nmin-nlag+step:nmin,:); xnew(nlag-step+2:nlag,:) = yfor(1:step-1,:); xnew(nlag+1,:) = zeros(1,neqs); xfor = mlag(xnew,nlag); [xend junk] = size(xfor); xobs = xfor(xend,:); if nx > 0 % add ec variables based on past level forecasts ecterm = ylevf(step-1,index)*ecvectors(:,1:r); xvec = [xobs ecterm 1]; else xvec = [xobs 1]; end; % loop over equations for i=1:neqs; bhat = bmat(:,i); yfor(step,i) = xvec*bhat/100; if freq < step, % here we can use past level forecasts ylevf(step,:) = (1 + yfor(step,:)).*ylevf(step-freq,:); else % case of freq > step, use past actual levels ylevf(step,:) = (1 + yfor(step,:)).*y(begf+step-freq-1,:); end; % end of if freq <= step end; end; end; % nlag through nfore-step-ahead forecasts for step=nlag:nfor-1; if step <= nfor; cnt = step-(nlag-1); for i=1:nlag; xnew(i,:) = yfor(cnt,:); cnt = cnt+1; end; xfor = mlag(xnew,nlag); [xend junk] = size(xfor); xobs = xfor(xend,:); if nx > 0 % add ec variables based on past level forecasts ecterm = ylevf(step,index)*ecvectors(:,1:r); xvec = [xobs ecterm 1]; else xvec = [xobs 1]; end; % loop over equations for i=1:neqs; bhat = bmat(:,i); yfor(step+1,i) = xvec*bhat/100; if freq < step+1, % here we can use past level forecasts ylevf(step+1,:) = (1 + yfor(step+1,:)).*ylevf(step+1-freq,:); else % case of freq > step, use past actual levels ylevf(step+1,:) = (1 + yfor(step+1,:)).*y(begf+step-freq,:); end; % end of if freq <= step end; end; % end of if step end;