Add DynamicFactor and a criterion for number of factors

data/README.md

@@ -9,7 +9,7 @@ that generates these files from original data.
 
 | Name | Source | File | Note |
 | :--: | :----: | :--: | :--- |
-| lpw_est_data.mat | [Li, Plagborg-Møller and Wolf (2024)](https://doi.org/10.1016/j.jeconom.2024.105722) | `DFM/Subroutines/SW_DFM_Estimation/data/hom_fac_1.xlsx` | Data are processed with [`lpw_savedata.m`](src/lpw_savedata.m) and used in [Stock and Watson (2016)](https://doi.org/10.1016/bs.hesmac.2016.04.002) |
+| lpw_data.mat | [Li, Plagborg-Møller and Wolf (2024)](https://doi.org/10.1016/j.jeconom.2024.105722) | `DFM/Subroutines/SW_DFM_Estimation/data/hom_fac_1.xlsx` | Data are processed with [`lpw_savedata.m`](src/lpw_savedata.m) and used in [Stock and Watson (2016)](https://doi.org/10.1016/bs.hesmac.2016.04.002) |
 
 ## References
 

data/src/lpw_savedata.m

@@ -262,6 +262,87 @@
   lsout.r2vec = r2vec;   lsout.nobs = nobs;
   lsout.nt = nt;         lsout.ns = ns;
 
+
+%% Content from factor_estimation_ls_full.m
+
+if levels
+  lsout.fac_diff = lsout.fac;
+  lsout.fac = cumsum_nan(lsout.fac(2:end,:)); % If data was differenced, cumulate factors
+end
+
+
+% Compute estimates of factor loadings;
+n_lc = 0;       % number of constraints placed on lambda
+lambda_constraints_full = est_par.fac_par.lambda_constraints_full;
+if size(lambda_constraints_full,2) > 1;
+  lam_c_index = lambda_constraints_full(:,1);        % Which row of lambda: Constraints are then R*lambda = r
+  lam_c_R     = lambda_constraints_full(:,2:end-1);  % R matrix
+  lam_c_r     = lambda_constraints_full(:,end);      % r value
+  n_lc = size(lambda_constraints_full,1);
+end;
+
+lam_mat = NaN(n_series,est_par.fac_par.nfac.total);
+ismpl = smpl(calvec,nfirst,nlast,nper);
+uar_coef_mat = NaN(n_series,n_uarlag);
+uar_ser_mat = NaN(n_series,1);
+uar_resid_mat = NaN(n_series,size(calvec,1));
+res_mat = NaN(n_series,size(calvec,1));
+r2_mat = NaN(n_series,1);                              % R-squared value
+trend_tmp = (1:1:size(calvec,1))';
+for is = 1:n_series;
+  tmp = packr([data(ismpl==1,is) lsout.fac(ismpl==1,:) trend_tmp(ismpl==1)]);
+  itmp = tmp(:,end);
+  tmp = tmp(:,1:end-1+levels); % Include time trend if data is in levels
+  if size(tmp,1) >= ntmin;
+     y = tmp(:,1);
+     x = [tmp(:,2:end), ones(size(tmp,1),1)];
+     xxi = inv(x'*x);
+     bols = xxi*(x'*y);
+     b = bols;
+     % Check for restrictions and impose;
+     if n_lc > 0;
+      ii = lam_c_index == is;
+      if sum(ii) > 0;
+          R = [lam_c_R(ii==1,:), zeros(sum(ii),1)];   % No constraints on constant term
+          r = lam_c_r(ii==1,:);
+          tmp1 = xxi*R';
+          tmp2 = inv(R*tmp1);
+          b = bols - tmp1*tmp2*(R*bols-r);
+      end;
+     end;
+     lam_mat(is,:) = b(1:end-1-levels)';
+     u = y - x*b;
+     % Compute R-squared 
+     ssr = sum(u.^2);
+     ym = y - mean(y);
+     tss = sum(ym.^2);
+     r2_mat(is) = 1-(ssr/tss);
+     % Compute AR model for errors
+     if r2_mat(is) < 0.9999;
+      if levels==1 % If variables are in levels...
+       [~,arcoef,ser2] = VAR_CorrectBias(u(~isnan(u)),n_uarlag); % Bias-corrected AR coefficients
+       arcoef = arcoef(:)';
+       ser = sqrt(ser2);
+       ar_resid = NaN(size(u,1),1); % We don't really need the AR residuals later
+      else
+       [arcoef, ser, ar_resid] = uar(u,n_uarlag);    % AR Coefficients and ser 
+      end
+     else;
+      arcoef = zeros(n_uarlag,1);
+      ser = 0.0;
+      ar_resid = NaN(size(u,1),1);
+     end;
+     uar_coef_mat(is,:) = arcoef'; 
+     uar_ser_mat(is,1) = ser;
+     uar_resid_mat(is,itmp) = ar_resid';
+     res_mat(is,itmp) = u';
+  end;
+end;
+
+varout = varest(lsout.fac,est_par.var_par,est_par.smpl_par,levels,coint_rank);
+
+
 %% Save the data
-save("../lpw_est_data.mat", 'est_data')
+dataout = data(:,inclcode==1);
+save("../lpw_data.mat", 'dataout')
 

src/AutoregressiveModels.jl

@@ -36,7 +36,7 @@ export VARProcess,
        histvar!,
        histvar,
 
-       OLS,
+       VAROLS,
        intercept,
        coefcorrected,
        residvcov,
@@ -51,14 +51,28 @@ export VARProcess,
 
        ARMAProcess,
 
-       Factor
+       AbstractNFactorCriterion,
+       ICp2penalty,
+       BaiNg,
+       criterion,
+       nfactor,
+       Factor,
+
+       AbstractDetrend,
+       NoTrend,
+       FirstDiff,
+       nskip,
+       detrend!,
+       invdetrend!,
+       DynamicFactor
 
 include("lapack.jl")
 include("utils.jl")
-include("process.jl")
-include("estimation.jl")
+include("varprocess.jl")
+include("varestimation.jl")
 include("bootstrap.jl")
 include("arma.jl")
 include("factor.jl")
+include("dfm.jl")
 
 end # module AutoregressiveModels

src/bootstrap.jl

@@ -1,60 +1,3 @@
-# An unsafe in-place version of OLS for bootstrap
-function _fit!(m::OLS{TF}) where TF
-    Y = m.resid
-    X = m.X
-    coef = m.coef
-    mul!(coef, X', Y)
-    mul!(m.crossXcache, X', X)
-    ldiv!(cholesky!(m.crossXcache), coef)
-    if m.intercept === nothing
-        copyto!(m.coefB, coef')
-    else
-        copyto!(m.coefB, view(coef, 2:size(coef,1), :)')
-        copyto!(m.intercept, view(coef, 1, :))
-    end
-    mul!(Y, X, coef, -1.0, 1.0)
-    mul!(m.residvcov, Y', Y)
-    rdiv!(m.residvcov, m.dofr)
-    residchol = m.residchol
-    residcholL = m.residcholL
-    if residchol !== nothing
-        Cfactors = getfield(residchol, :factors)
-        Cuplo = getfield(residchol, :uplo)
-        copyto!(Cfactors, m.residvcov)
-        # residchol could involve changes in immutable objects but it is not used
-        residchol = cholesky!(Cfactors)
-        N = size(m.residvcov, 1)
-        @inbounds for j in 1:N
-            for i in 1:j-1
-                residcholL[i,j] = zero(TF)
-            end
-            for i in j:N
-                residcholL[i,j] = Cuplo === 'U' ? Cfactors[j,i] : Cfactors[i,j]
-            end
-        end
-    end
-    return nothing
-end
-
-# An unsafe version for bootstrap
-function _fitvar!(data::Matrix, m::OLS, nlag, nocons)
-    i0 = nocons ? 0 : 1
-    # Y in OLS is left untouched but use resid in-place
-    Y = residuals(m)
-    X = modelmatrix(m)
-    N = size(Y, 2)
-    # esample is not used here
-    for j in 1:N
-        col = view(data, :, j)
-        Tfull = length(col)
-        copyto!(view(Y, :, j), view(col, nlag+1:Tfull))
-        for l in 1:nlag
-            copyto!(view(X, :, i0+(l-1)*N+j), view(col, nlag+1-l:Tfull-l))
-        end
-    end
-    _fit!(m)
-end
-
 randomindex(r::VectorAutoregression) = sample(1:size(residuals(r),1))
 
 function wilddraw!(out, r::VectorAutoregression)
@@ -103,7 +46,7 @@ function _bootstrap!(ks, stats, r::VectorAutoregression, var, initialindex, draw
             var(view(bootdata, t, :), _reshape(view(bootdata, t-1:-1:t-nlag, :)', N*nlag))
         end
         keepbootdata && (allbootdata[k] = copy(bootdata))
-        estimatevar && _fitvar!(bootdata, m, nlag, !hasintercept(r))
+        estimatevar && fit!(m, bootdata; Yinresid=true)
         for stat in stats
             nt = (out=selectdim(stat[1], ndims(stat[1]), k), data=bootdata, r=rk)
             stat[2](nt)