%
 -------------------------------------------------------------------------
% Credit Risk Assessment Toolbox
%
% Author: D. N. Sotiropoulos (2019)
%
 -------------------------------------------------------------------------
% This script file provides fundamental computational functionality
 for
% training a fitnet type network for the problem of function
% approximation. In fact, we are going to approximate the
 fico_range_low
% and fico_range_high feature values as a function of the remaining
% features. Therefore, the loan status related features will be
 ignored.
% This neural netwrok simulator uses gpu computation assistance.
% Clear command window and workspace.
clc
clear
% Load dataset from file "LendingClub.mat" which creates the table
 type
% variable lending_club_data_table. This table will be loaded into
 table T.
load('LendingClub.mat');
T = lending_club_data_table;
clear lending_club_data_table
% The following table columns will be excluded from the available set
 of
% features upon which training will be conducted: =
% ColumnID = 14 ==> VariableName = 'fico_range_low'
% ColumnID = 15 ==> VariableName = 'fico_range_high'
% ColumnID = 36 ==> VariableName = 'last_fico_range_high'
% ColumnID = 37 ==> VariableName = 'last_fico_range_low'
% ColumnID = 39 ==> VariableName = 'mths_since_last_derog'
% ColumnID = 121 ==> VariableName = 'loan_status1'
% ColumnID = 122 ==> VarialeName = 'loan_status2'
% Set the variable names to be removed from training:
RemoveVariableNames = {'fico_range_low','fico_range_high',...

 'last_fico_range_high','last_fico_range_low',...

 'mths_since_last_major_derog','loan_status1',...
 'loan_status2'};
% Set the feature data.
Data = removevars(T,RemoveVariableNames);
1
% Convert the feature data table to array.
Data = table2array(Data)';
% Set the feature labels.
Targets = T.fico_range_low';
% Set the network object.
net = fitnet(10,'trainscg');
% Configure the network object.
net = configure(net,Data,Targets);
% View network.
view(net);
% Initialize the network object.
net = init(net);
% Enable verbal output during training.
net.trainParam.showCommandLine = true;
% Enable visual output during training.
net.trainParam.showWindow = true;
% Devide the data randomly.
net.divideFcn= 'dividerand';
% Use the 70% of the available data for training purposes.
net.divideParam.trainRatio= 0.7;
% Use the 20% of the available data for validations purposes.
net.divideParam.valRatio= 0.2;
% Use the 10% of the available data for testing purposes.
net.divideParam.testRatio= 0.1;
% Before training, the network’s tansig layers can be converted to
% elliotsig layers.
for i=1:net.numLayers
 if strcmp(net.layers{i}.transferFcn,'tansig')
 net.layers{i}.transferFcn = 'elliotsig';
 end
end
% Convert MATLAB arrays to GPU arrays.
DataGPU = nndata2gpu(Data);
TargetsGPU = nndata2gpu(Targets);
% Train the network object while keeping track of the training,
 validation
% and testing processes.
net = train(net,Data,Targets);
% Compute the estimated targets.
EstimatedTargets = net(Data);
% Compute the corresponding error values per training pattern.
Errors = Targets - EstimatedTargets;
% Compute the associated mean squared error.
MeanSquaredError = mse(net,Targets,EstimatedTargets);
fprintf('Neural network performacne in terms of mse: %f
\n',MeanSquaredError);
% Plot the associated error histogram in 30 bins.
2
figure('Name','Error Histogram in 30 bins');
ploterrhist(Errors,'bins',30);
% Plot the histogram for actual and estimated targets.
figure('Name','Estimated vs Actual Targets Histogram');
hold on
histogram(Targets,'Normalization','countdensity','BinMethod','scott');
histogram(EstimatedTargets,'Normalization','countdensity','BinMethod','scott');
hold off;
grid on
% Plot the regression results between the actual and the estimated
 targets.
plotregression(Targets,EstimatedTargets);