Merge pull request #18 from fipelle/dev

Implemented `compute_scaling_factor` in validation.jl

src/models/dfm.jl

@@ -732,6 +732,26 @@ Model validation
 --------------------------------------------------------------------------------------------------------------------------------
 =#
 
+"""
+    compute_scaling_factors(estim::DFMSettings)
+
+Compute scaling factors for each series in `estim.Y`.
+"""
+function compute_scaling_factors(estim::DFMSettings)
+
+    # Initialise `scaling_factors`
+    scaling_factors = zeros(estim.n, 1); # the final `, 1` is needed to run rescale_estim_params!(...) since it expects a matrix of floats
+
+    # Compute scaling factors
+    for i in axes(estim.Y, 1)
+        coordinates_trends = findall(view(estim.trends_skeleton, i, :) .!= 0);
+        max_order = maximum(view(estim.drifts_selection, coordinates_trends)); # either 1 (smooth trend) or 0 (random walk)
+        scaling_factors[i] = compute_scaling_factor(estim.Y[i, :], max_order==0);
+    end
+
+    return scaling_factors;
+end
+
 """
     rescale_estim_params!(coordinates_params_rescaling::VectorsArray{Int64}, estim::DFMSettings, std_presample::FloatMatrix)
 

src/validation.jl

@@ -80,27 +80,31 @@ function fc_err(validation_settings::ValidationSettings, p::Int64, λ::Number, 
         data = validation_settings.Y;
     end
 
-    # Standardise data
-    data_presample = @view data[:, 1:t0];
-
     # Stop if the estimation sample is entirely missing
-    if count(ismissing, data_presample) == validation_settings.n*t0
+    if count(ismissing, view(data, :, 1:t0)) == validation_settings.n*t0
         return [0.0, 1.0];
 
     # Standard run
+    # TBD: Generalise further the standardisation process for future models whose DataTypes arguments are transformations of the data
     else
+        estim = validation_settings.model_struct(data[:, 1:t0], validation_settings.model_args, validation_settings.model_kwargs, p, λ, α, β);
+
+        # Standardise stationary data
         if validation_settings.is_stationary
-            mean_presample = mean_skipmissing(data_presample);
-            std_presample = std_skipmissing(data_presample);
+            mean_presample = mean_skipmissing(estim.Y);
+            std_presample = std_skipmissing(estim.Y);
             Y = (data .- mean_presample) ./ std_presample;
+            estim.Y .-= mean_presample; # first step of the estimation sample standardisation
+
+        # Standardise non-stationary data
         else
-            std_presample = std_skipmissing(diff(data_presample, dims=2));
+            std_presample = compute_scaling_factors(estim);
             Y = data ./ std_presample;
         end
 
-        estim = validation_settings.model_struct(Y[:, 1:t0], validation_settings.model_args, validation_settings.model_kwargs, p, λ, α, β);
+        estim.Y ./= std_presample; # final step of the data estimation sample standardisation
         rescale_estim_params!(validation_settings.coordinates_params_rescaling, estim, std_presample);
-
+        
         # Estimate model
         sspace = ecm(estim, output_sspace_data=Y);