function ylevf = bvarf(y,nlag,nfor,begf,tight,weight,decay,x,transf); % PURPOSE: Estimates a Bayesian vector autoregression of order n % and produces f-step-ahead forecasts (Minnesota prior) %--------------------------------------------------------------- % USAGE: yfor = bvarf(y,nlag,nfor,begf,tight,weight,decay,x,transf) % 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 % tight = Litterman's tightness hyperparameter % weight = Litterman's symmetric weight (scalar) % decay = Litterman's lag decay = lag^(-decay) % x = an optional matrix of deterministic variables % transf = 0, no data transformation % = 1, 1st differences used to estimate the model % = freq, seasonal differences used to estimate % = cal-structure growth rates used to estimate % e.g., cal(1982,1,12) [see cal() function] %--------------------------------------------------------------- % NOTE: - use bvarf(y,nlag,nfor,begf,tight,weight,decay,[],transf) % for a transformation model with no x's (deterministic variables) % - includes constant term automatically %--------------------------------------------------------------- % RETURNS: % yfor = an nfor x neqs matrix of level forecasts for each equation %--------------------------------------------------------------- % SEE ALSO: bvar, plt_var, prt_var %--------------------------------------------------------------- % written by: % James P. LeSage, Dept of Economics % University of Toledo % 2801 W. Bancroft St, % Toledo, OH 43606 % jpl@jpl.econ.utoledo.edu if nargin == 9 % user wants us to transform the data [nobs2 nx] = size(x); if isstruct(transf) % a growth rates transform tform = 2; freq = transf.freq; elseif transf == 0 % no transform tform = 0; freq = 0; elseif transf == 1 % 1st difference transform tform = 1; freq = 0; elseif (transf == 1) | (transf == 4) | (transf == 12) tform = 3; % seasonal differences transform freq = transf; end; elseif nargin == 8 [nobs2 nx] = size(x); tform = 0; freq = 0; elseif nargin == 7 nx = 0; tform = 0; freq = 0; else error('Wrong # of arguments to bvarf'); end; % flag an error where x-variables exist but not enough forecast values % are supplied for these variables if nx > 0 if nobs2 < begf-1+nfor error('bvarf: not enough observations in x to forecast'); end; end; [nobs neqs] = size(y); % adjust nobs to feed the lags nmin = min(nobs,begf-1); % error checking on inputs if nlag < 1 error('Lag length less than 1 in bvarf'); end; if nlag > nobs error('Lag length exceeds observations in bvarf'); end; if tight < 0.01 warning('Tightness less than 0.01 in bvarf'); end; if tight > 1.0 warning('Tightness greater than unity in bvarf'); end; if decay < 0 error('Negative lag decay in bvarf'); end; [wchk1 wchk2] = size(weight); if (wchk1 ~= wchk2) error('non-square weight matrix in bvarf'); elseif wchk1 > 1 if wchk1 ~= neqs error('wrong size weight matrix in bvarf'); end; end; % check for zeros in weight matrix if wchk1 == 1 if weight == 0 error('bvarf: must have weight > 0'); end; elseif wchk1 > 1 zip = find(weight == 0); if length(zip) ~= 0 error('bvarf: must have weights > 0'); end; end; % nvar adjusted for constant term and deterministic variables k = neqs*nlag + nx + 1; ndiff = 0; switch tform case 1 % 1st differences transform % transform data dy = y - mlag(y,1); ndiff = 1; % generate lagged rhs matrix xlag = mlag(dy,nlag); % constant term iota = ones(nobs,1); % truncate variables to feed lags and 1st diff and end at begf-1 iota = trimr(iota,nlag+1,nobs-begf+1); dys = trimr(dy,nlag+1,nobs-begf+1); xlag = trimr(xlag,nlag+1,nobs-begf+1); % add x-matrix and constant to x-matrix if nx > 0 xmat = [xlag x(nlag+2:nmin,:) iota]; else xmat = [xlag iota]; end; % end of 1st difference transformation case case 2 % growth rates transformation % transform data dy = growthr(y,freq); % generate lagged rhs matrix xlag = mlag(dy,nlag); % constant term iota = ones(nobs,1); % truncate variables to feed lags and freq diff's and end at begf-1 iota = trimr(iota,nlag+freq,nobs-begf+1); dys = trimr(dy,nlag+freq,nobs-begf+1); xlag = trimr(xlag,nlag+freq,nobs-begf+1); % add x-matrix and constant to x-matrix if nx > 0 xmat = [xlag x(nlag+freq+1:nmin,:) iota]; else xmat = [xlag iota]; end; % end of growth-rates transform case case 3 % seasonal differences transform % transform data dy = y - lag(y,freq); % generate lagged rhs matrix xlag = mlag(dy,nlag); % constant term iota = ones(nobs,1); % truncate variables to feed lags and freq diff's and end at begf-1 iota = trimr(iota,nlag+freq,nobs-begf+1); dys = trimr(dy,nlag+freq,nobs-begf+1); xlag = trimr(xlag,nlag+freq,nobs-begf+1); % add x-matrix and constant to x-matrix if nx > 0 xmat = [xlag x(nlag+freq+1:nmin,:) iota]; else xmat = [xlag iota]; end; otherwise % case of no transformation % generate lagged rhs matrix xlag = mlag(y,nlag); % constant term iota = ones(nobs,1); % truncate to feed lags and to end at begf-1 for estimation dys = trimr(y,nlag,nobs-begf+1); dy = y; xlag = trimr(xlag,nlag,nobs-begf+1); iota = trimr(iota,nlag,nobs-begf+1); % add x-matrix and constant to x-matrix if nx > 0 xmat = [xlag x(nlag+1:nmin,:) iota]; else xmat = [xlag iota]; end; end; % end of data transformation cases % do scaling here % determine scale factors using univariate AR model % Doan uses the full vector whereas we truncate the % first lags, so we will get slightly difference estimates scale = zeros(neqs,1); scale2 = zeros(neqs,neqs); ytmp = zeros(nmin,1); for j=1:neqs ytmp = dy(freq+ndiff+1:nmin,j); scale(j,1) = scstd(ytmp,length(ytmp),nlag); end; for j=1:neqs; for i=1:neqs; scale2(i,j) = scale(j)/scale(i); end; end; % form xpx only once to save time xpx = xmat'*xmat; % pull out each y-vector and run regressions for j=1:neqs; yvec = dy(nlag+freq+ndiff+1:nmin,j); xpy = xmat'*yvec; reslt = theilbf(xpy,xpx,nlag,neqs,j,tight,weight,decay,scale2,scale,nx); bmat(:,j) = reslt.beta; end; % given bmat values generate future forecasts % These may be levels, 1st-differences, growth rates or seas diff's % we worry transforming back to levels later % 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 xvec = [xobs x(begf,:) 1]; else xvec = [xobs 1]; end; % loop over equations for i=1:neqs; bhat = bmat(:,i); yfor(1,i) = xvec*bhat; 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 xvec = [xobs x(begf+step-1,:) 1]; else xvec = [xobs 1]; end; % loop over equations for i=1:neqs; bhat = bmat(:,i); yfor(step,i) = xvec*bhat; 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 xvec = [xobs x(begf+step,:) 1]; else xvec = [xobs 1]; end; % loop over equations for i=1:neqs; bhat = bmat(:,i); yfor(step+1,i) = xvec*bhat; end; end; end; % we now worry about transforming the forecasts back % to levels switch tform case 1 % 1st differences forecasts % convert 1st difference forecasts to levels ylevf = zeros(nfor,neqs); % 1-step-ahead forecast ylevf(1,:) = yfor(1,:) + y(begf-1,:); % add change to actual from time t; % 2-nfor-step-ahead forecasts for i=2:nfor % ylevf(i,:) = yfor(i,:) + ylevf(i-1,:); end; % end of 1st differences case case 2 % growth rates forecasts % convert growth rate forecasts to levels ylevf = zeros(nfor,neqs); yfor = yfor/100; for step=1:nfor; 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 of for step loop case 3 % seasonal difference forecasts % convert seasonal difference forecasts to levels for step=1:nfor; if freq < step, % here we use past level forecasts ylevf(step,:) = yfor(step,:) + ylevf(step-freq,:); else % case of freq > step, use past actual levels ylevf(step,:) = yfor(step,:) + y(begf+step-freq-1,:); end; % end of if freq <= step end; % end of for step loop otherwise % no transformation, so we have level forecasts already ylevf = yfor; end;