Journal of Intelligent Learning Systems and Applications, 2011, 3, 103-112
doi:10.4236/jilsa.2011.32012 Published Online May 2011 (http://www.SciRP.org/journal/jilsa)
An Artificial Neural Network Approach for Credit
Risk Management
Vincenzo Pacelli1*, Michele Azzollini2
1 Faculty of Economics, University of Foggia, Foggia, Italy; 2BancApulia S.p.A. San Severo, Italy.
Received July 10th, 2010; revised October 1st, 2010; accepted December 1st, 2010
ABSTRACT
The objective of the research is to analyze the ability of the artificial neural network model developed to forecast the
credit risk of a panel of Italian manufacturing companies. In a theoretical point of view, this paper introduces a litera-
ture review on the application of artificial intelligence systems for credit risk management. In an empirical point of
view, this research compares the architecture of the artificial neural network model developed in this research to an-
other one, built for a research conducted in 2004 with a similar panel of companies, showing the differences between
the two neural network models.
Keywords: Credit Risk, Forecasting, Artificial Neural Networks
1. Introduction
The credit risk has long been an important and widely
studied topic in banking. For lots of commercial banks,
the credit risk remains the most important and difficult
risk to manage and evaluate. In the last years the advances
in information technology have lowered the costs of ac-
quiring, managing and analyzing data, in an effort to
build more robust and efficient techniques for credit risk
management.
In recent years, a great number of the largest banks
have developed sophisticated systems in an attempt to
make more efficient the process of credit risk manage-
ment. The objective of the credit scoring models is to
evaluate the risk profile of the companies and then to
assign different credit scores to companies with different
probability of default. Therefore credit scoring problems
are basically in the scope of the more general and widely
discussed discrimination and classification problems [1-5].
Nowadays there is a need to automate the credit ap-
proval decision process, in order to improve the effi-
ciency of credit risk management processes. This need in
bank lending processes is emphasized by the regulatory
framework of Basel 2 by giving banks a range of in-
creasingly sophisticated options for calculating capital
charges. Banks will be expected to employ the capital
adequacy method most appropriate to the complexity of
their transactions and risk profiles. For credit risk, the
range of options begins with the standardized approach
and extends to the internal rating-based (IRB) ap-
proaches.
The crisis, that hit three years ago the international fi-
nancial system and still has adverse effects on the global
economy, has been considered essential to an overall re-
thinking of prudential regulation. Although the crisis was
the result of many contributing factors, certainly the reg-
ulatory environment and supervision of the financial sec-
tor have not been able to prevent excessive expansion of
risk of harnessing the transmission of financial turmoil.
The set of measures proposed by the Basel Committee,
which will form the basis of the new Capital Accord of
Basel 3, aims to redefine the important aspects of regula-
tory, in line with the ambitious targets set by the G20.
The new Accord of Basel 3 essentially confirms the
basic philosophy of Basel 2, but it notes some limitations
of the framework and introduces the necessary corrective
measures to increase the capital adequacy of banks, es-
pecially about the liquidity risk. In particular, the first ele-
ment of strengthening the prudential rules is the increase
in the quantity and quality of the regulatory capital. The
new regulatory framework of Basel 3 also seems to con-
firm the sensibility of the Committee about the increas-
ing sophisticated models for calculating capital charges
Although the research has been conducted jointly by the two authors,
p
aragraphs 1, 2, 3 and 5 can be attributed to Vincenzo Pacelli, pa
r
a-
graphs 3.1 and 3.2 can be attributed to Michele Azzollini, while para-
graph 4 is a collaborative effort of the two authors.

An Artificial Neural Network Approach for Credit Risk Management 104
and managing credit risk.
The objective of this paper is to analyze the ability of
the artificial neural network model developed to forecast
the credit risk of a panel of Italian manufacturing com-
panies. In a theoretical point of view, this paper intro-
duces a detail literature review on the application of arti-
ficial intelligence systems for credit risk management. In
an empirical point of view, this research compares the
architecture of the artificial neural network model de-
veloped in this research to another one, built for a re-
search conducted in 2004 with a similar panel of compa-
nies, showing the differences between the two neural net-
work models.
2. A Literature Review
In the last years, the literature has produced several stud-
ies about the application of artificial intelligence systems
for credit risk management. Among the studies on the
application of artificial intelligence systems within the
classification and discrimination of economic phenomena,
with particular attention to the management of the credit
risk, we can mention Tam and Kiang [6], Lee, Chiu, Lu
and Chen [7], Altman, Marco and Varetto [8], Zhang,
Cao and Schniederjans [9], Huang, Chen, Hsu, Chen and
Wu [10], Ravi Kumar and Ravi [11], Angelini, Tollo and
Roli [12], Chauhan, Ravi and Chandra [13], Hsieh and
Hung [14].
The paper of K. Y. Tam and M. Y. Kiang [6] intro-
duces a neural network approach to perform discriminant
analysis in business research. Using bank default data,
the neural approach is compared with linear classifier.
Empirical results show that neural model is a promising
method of evaluating bank conditions in terms of predic-
tive accuracy, adaptability and robustness.
The objective of the paper of T. S. Lee, C. C. Chiu, C.
J. Lu and I. F. Chen [7] is to explore the performance of
credit scoring by integrating the back propagation neural
networks with traditional discriminant analysis approach.
To demonstrate the inclusion of the credit scoring result
from discriminant analysis would simplify the network
structure and improve the credit scoring accuracy of the
designed neural network model, credit scoring tasks are
performed on one bank credit card data set. As the results
reveal, the proposed hybrid approach converges much
faster than the conventional neural networks model.
Moreover, the credit scoring accuracies increase in terms
of the proposed methodology and outperform traditional
discriminant analysis and logistic regression approaches.
E. I. Altman, G. Marco and F. Varetto [8] analyze the
comparison between traditional statistical methodologies
for distress classification and prediction, i.e., linear dis-
criminant (LDA) or logit analyses, with an artificial neu-
ral networks (ANN). Analyzing over 1000 healthy, vul-
nerable and unsound industrial Italian firms from 1982 –
1992, this study was carried out at the Centrale dei Bi-
lanci in Turin (Italy) and is now being tested in actual
diagnostic situations. The results are part of a larger ef-
fort involving separate models for industrial, retailing/
trading and construction firms. The results indicate a
balanced degree of accuracy and other beneficial charac-
teristics between LDA and ANN. The authors are par-
ticularly careful to point out the problems of the
“black-box” ANN systems, including illogical weight-
ings of the indicators and over fitting in the training stage
both of which negatively impacts predictive accuracy.
Both types of diagnostic techniques displayed acceptable,
over 90%, classification and holdout sample accuracy and
the study concludes that there certainly should be further
studies and tests using the two techniques and suggests a
combined approach for predictive reinforcement.
W. Zhang, Q. Cao and M. J. Schniederjans [9] present
a comparative analysis of the forecasting accuracy of
univariate and multivariate linear models that incorporate
fundamental accounting variables (i.e., inventory, ac-
counts receivable, and so on) with the forecast accuracy
of neural network models. Unique to this study is the
focus of their comparison on the multivariate models to
examine whether the neural network models incorporat-
ing the fundamental accounting variables can generate
more accurate forecasts of future earnings than the mod-
els assuming a linear combination of these same variables.
They investigate four types of models: univariate-linear,
multivariate-linear, univariate-neural network, and multi-
variate-neural network using a sample of 283 firms. This
study shows that the application of the neural network
approach incorporating fundamental accounting variables
results in forecasts that are more accurate than linear fo-
recasting models. The results also reveal limitations of
the forecasting capacity of investors in the security mar-
ket when compared to neural network models.
Z. Huang, H. Chen, C. J. Hsu, W. H. Chen and S. Wu
[10] introduce a relatively new machine learning tech-
nique, support vector machines (SVM), in attempt to
provide a model with better explanatory power. They use
back propagation neural network (BNN) as a benchmark
and obtain prediction accuracy around 80% for both
BNN and SVM methods for the United States and Tai-
wan markets. However, only slight improvement of SVM
is observed. Another direction of the research is to im-
prove the interpretability of the AI-based models. They
applied the research results in neural network model in-
terpretation and obtain relative importance of the input fi-
nancial variables from the neural network models. Based
on these results, they conduct a market comparative ana-
lysis on the differences of determining factors in the

An Artificial Neural Network Approach for Credit Risk Management105
United States and Taiwan markets.
P. Ravi Kumar and V. Ravi [11] present a comprehen-
sive review of the work done, during the 1968-2005, in
the application of statistical and intelligent techniques to
solve the bankruptcy prediction problem faced by banks
and firms. The review is categorized by taking the type
of technique applied to solve this problem as an impor-
tant dimension. Accordingly, the papers are grouped in the
following families of techniques: 1) statistical techniques,
2) neural networks, 3) case-based reasoning, 4) decision
trees, 5) operational research, 6) evolutionary approaches,
7) rough set based techniques, 8) other techniques sub-
suming fuzzy logic, support vector machine and isotonic
separation and 9) soft computing subsuming seamless
hybridization of all the above-mentioned techniques.
What particular significance is that in each paper, the re-
view highlights the source of data sets, financial ratios
used, country of origin, time line of study and the compa-
rative performance of techniques in terms of prediction
accuracy wherever available. The review also lists some
important directions for future research.
E. Angelini, G. Tollo and A. Roli [12] describe the
case of a successful application of neural networks to
credit risk assessment. They develop two neural network
systems, one with a standard feed forward network, while
the other with a special purpose architecture. The appli-
cation is tested on real-world data, related to Italian small
businesses. They show that neural networks can be very
successful in learning and estimating the in bonis/default
tendency of a borrower, provided that careful data analy-
sis, data pre-processing and training are performed.
In the study of N. Chauhan, V. Ravi and D. K. Chandra
[13], differential evolution algorithm (DE) is proposed to
train a wavelet neural network (WNN). The resulting
network is named as differential evolution trained wavelet
neural network (DEWNN). The efficacy of DEWNN is
tested on bankruptcy prediction datasets of US banks,
Turkish banks and Spanish banks. Moreover, Garson’s
algorithm for feature selection in multi layer perceptron
is adapted in the case of DEWNN. The performance of
DEWNN is compared with that of threshold accepting
trained wavelet neural network (TAWNN) and the origi-
nal wavelet neural network (WNN) in the case of all data
sets without feature selection and also in the case of four
data sets where feature selection was performed. The
whole experimentation is conducted using 10-fold cross
validation method. Results show that soft computing
hybrids outperform the original WNN in terms of accu-
racy and sensitivity across all problems. Furthermore,
DEWNN outscore TAWNN in terms of accuracy and
sensitivity across all problems except Turkish banks da-
taset.
The paper of N. C. Hsieh, L. P. Hung [10] focuses on
predicting whether a credit applicant can be categorized
as good, bad or borderline from information initially
supplied. This is essentially a classification task for credit
scoring. They introduce the concept of class- wise classi-
fication as a pre-processing step in order to obtain an
efficient ensemble classifier. This strategy would work
better than a direct ensemble of classifiers without the
pre-processing step. The proposed ensemble classifier is
constructed by incorporating several data mining tech-
niques.
3. The Methodology: A Neural Network
Approach
Generally two essential linear statistical tools, discrimi-
nant analysis and logistic regression, were most com-
monly applied to develop credit scoring models. Dis-
criminant analysis is the first tool to be used in building
credit scoring models. However, the utilization of linear
discriminant analysis (LDA) has often been criticized
because of its assumptions of the categorical nature of
the credit data and the fact that the covariance matrices
of the good and bad credit classes are unlikely to be
equal [15].
Logistic regression is an alternative to develop credit
scoring models. Basically the logistic regression model
was emerged as the technique of choice in predicting
dichotomous outcomes.
In addition to these linear methodologies, non-linear
methods, as the artificial neural networks, are applied to
develop credit scoring models. Neural networks provide
a new alternative to LDA and logistic regression, par-
ticularly in situations where the dependent and inde-
pendent variables exhibit complex non-linear relation-
ships. Even though neural networks have shown to have
better credit scoring capability than LDA and logistic
regression, they are, however, also criticized for its long
training process in designing the optimal network’s to-
pology.
Artificial neural networks rise from the desire to artifi-
cially simulate the physiological structure and function-
ing of human brain structures.
Artificial neural networks consist of elementary com-
putational units, known as Processing Elements (PE),
proposed by Mc Cullock and Pitts in 1943 [16].
As illustrated in Figure 1, the Input Layers are the in-
put neurons, which receive the incoming stimuli. Input
neurons process, according to a particular function called
Transfer function, the inputs selected and distribute the
result to the next level of neurons.
Then the input neurons forward the information to all
neurons of the layer 2 (Middle Layers).
Information is not simply sent to the intermediate neu-
rons, but is weighed. It means that the result obtained

An Artificial Neural Network Approach for Credit Risk Management 106
from each neuron is sized according to the weight of the
connection between the two neurons. Specifically, as
shown in Figure 2, the weight of connection is repre-
sented by Wj,i.
Each neuron is characterized by a transition function
and a threshold value.
The threshold is the minimum value that input must
have to activate the neuron.
The Middle Layers are the neurons that constitute the
middle layer.
Each neuron of this layer sum the inputs that are pre-
sented to its incoming connections. In mathematical terms,
each neuron performs the summation of inputs, which are
the product of output neurons of the first layer and weight
of the connection. The result of this sum is again drawn
on the basis of the transfer function of each neuron. The
result obtained is in turn forwarded to the next layer of
neurons, multiplied by the weight between neurons.
Before the neural network can be applied to the prob-
lem at hand, a specific tuning of its weights has to be
done. This task is accomplished by the learning algorithm
which trains the network and iteratively modifies the
weights until a specific condition is verified. In most
applications, the learning algorithm stops when the dis-
crepancy (error) between desired output and the output
produced by the network falls below a predefined thresh-
old. There are three typologies of learning mechanisms
for neural networks [12]:
 supervised learning;
 unsupervised learning;
 reinforced learning.
Supervised learning is characterized by a training set
which is a set of correct examples used to train the net-
work. The training set is composed of pairs of inputs and
corresponding desired outputs. The error produced by the
network then is used to change the weights. This kind of
learning is applied in cases in which the network has to
learn to generalize the given examples.
A typical application is classification. A given input
has to be inserted in one of the defined categories.
In unsupervised learning algorithms, the network is only
provided with a set of inputs and no desired output is
given. The algorithm guides the network to self-organize
and adapt its weights. This kind of learning is used for
tasks such as data mining and clustering, where some
regularities in a large amount of data have to be found.
Finally, reinforced learning trains the network by in
troducing prizes and penalties as a function of the net-
work response. Prizes and penalties are then used to mo-
dify the weights. Reinforced learning algorithms are ap-
plied, for instance, to train adaptive systems which per-
form a task composed of a sequence of actions. The final
outcome is the result of this sequence, therefore the con-
Figure 1. Artificial neural network.
Figure 2. Processing element.
tribution of each action has to be evaluated in the context
of the action chain produced.
The learning algorithm is one of the most significant
among the factors which help to define the specific con-
figuration of a neural network and thus determine the con-
dition and capacity of the network itself to provide cor-
rect answers to the specific problem.
In general, learning algorithms have some common
features, such [17]:
 the values of synaptic weights of the network are
assigned randomly within a small range of varia-
tion;
 the modification of synaptic values (weights) (Δwij)
of the neural network is calculated after each pres-
entation of a single pattern (learning or online
courses) or at the end of the presentation of all
patterns of training (learning epochs). The new
configuration values is synaptic calculated by add-
ing the change obtained [Δwij (t)] to the previous
configuration synaptic [Wij (t – 1) + Δwij (t)].
Learning therefore concerns the overlap of new know-
ledge on an already-established prior knowledge. To en-
sure that this eraser and distort what has been learned,
learning proceeds recursively and gradual. The learning
speed is regulated by a constant ή, called learning rate,

An Artificial Neural Network Approach for Credit Risk Management107
which defines the portion of change that is applied to the
values of synaptic.
The architecture of a neural network is usually classi-
fied according to two characteristics: the dynamics and
topology.
With reference to the distinction of architectures of neu-
ral networks based on their dynamic, we can classify
static and dynamic networks. This distinction concerns
the way in which the flow of information travels from
input nodes to output ones. In static architectures the flow
of information travels in one direction (from the nodes of
the layers below those of the upper layers).
In dynamic architectures, instead, the flow of informa-
tion does not travel in one direction, because there are
feedback connections. Dynamic architecture allows re-
ception of signals of neurons of the same layer or upper
layers of neurons.
Specially the primary task of a single artificial neuron
is to perform a weighted sum of input signals and apply
activation function output.
The activation function is intended to limit the output
of the neuron, usually between the values [0, 1] or [–1,
+1].
Typically it is used the same activation function for all
neurons in the network, even if it is not necessary. The
activation functions that are most commonly used are:
 identity function:
f(x) = x, for each x.
In this case, the output of a neuron is simply equal to
the weighted sum of inputs signals.
 step function with threshold θ:
the output y of this transfer function is binary, depending
on whether the input meets a specified threshold, θ.
This function is used in perceptron model and often
shows up in many other models. It performs a division of
the space of inputs by a hyper plane. It is especially use-
ful in the last layer of a network intended to perform bi-
nary classification of the inputs. It can be approximated
from other sigmoid functions by assigning large values to
the weights.
Activation functions of this type are necessary if you
want a neural network to convert the input signal in a
binary signal (1 or 0) or bipolar (–1 or 1).
 Sigmoid function:
It is the most used function.
This function produces an output value between 0 and
1 and, for this reason this pattern is also called logistic
sigmoid.
Furthermore, the sigmoid function is continuous and
differentiable. For this is used in neural network models
in which the algorithm learning requires the involvement
of formulas in which appear derived.
 hyperbolic tangent function:
This function produces an output value between –1
and 1 and has a similar pattern to the sigmoid function.
The topological distinction, however, refers to the num-
ber of layers of neural network and, therefore, we have
networks with single layer and multilayer networks, as
described below.
3.1. The Perceptron
The simplest network consists of a single neuron with
“n” inputs and one output. The basic learning algorithm
of the perceptron analyzes the configuration (pattern) in-
put and weighting variables through synapses, deciding
which category of output is associated with the configu-
ration.
However, this architecture presents the major limita-
tion to solve only linearly separable problems (for each
neuron output), the output values that activate the neuron
must be clearly separate from the disabled through a hy-
per-plane separation size 1 - n.
3.2. The Multi Layer Perceptron—MLP
The neural network with one input layer, one or more
layers of intermediate neurons and an output layer is
called the Multi Layer Perceptron. In a network-type
feed-forward signals propagate from input to output only
through intermediate neurons, failing to tie lines, or in
feedback.
These networks use, in most cases, the Back Propaga-
tion learning algorithm. It calculates the appropriate syn-
tactic weights between inputs and neurons of intermedi-
ate layers and between them and outputs, starting from
random weights to them and making small changes, gra-
dual and progressive, determined by estimating the error
between the result produced by the network and the de-
sired one.
The learning phase is based, then, on a sequence of pre-
sentations of a finite number of configurations in which
the learning algorithm converges to the desired solution.
Networks learn through a series of attempts, sometimes
prolonged, that allow to model the weights that link the
input with output, through the hidden layers of neurons.
There are many other types of more complex architec-
ture, but architecture supervised Back Propagation is the
most widely used and disseminated for the capabilities
that this set of models has to generalize the results for a
large number of financial problems.
The following diagram illustrates a perceptron net-
work with three layers (Figure 3).
This network has an input layer (on the left) with three
neurons, one hidden layer (in the middle) with three neu-
rons and an output layer (on the right) with three neurons.
There is one neuron in the input layer for each predictor
variable.

An Artificial Neural Network Approach for Credit Risk Management 108
A vector of predictor variable values (x1xp) is pre-
sented to the input layer. The input layer (or processing
before the input layer) standardizes these values so that
the range of each variable is –1 to 1. The input layer dis-
tributes the values to each of the neurons in the hidden
layer. In addition to the predictor variables, there is a
constant input of 1.0, called the bias that is fed to each of
the hidden layers; the bias is multiplied by a weight and
added to the sum going into the neuron. Arriving at a neu-
ron in the hidden layer, the value from each input neuron
is multiplied by a weight (wji), and the resulting weighted
values are added together producing a combined value uj.
The weighted sum (uj) is fed into a transfer function, σ,
which outputs a value hj. The outputs from the hidden
layer are distributed to the output layer.
Arriving at a neuron in the output layer, the value from
each hidden layer neuron is multiplied by a weight (wkj),
and the resulting weighted values are added together
producing a combined value vj. The weighted sum (vj) is
fed into a transfer function, σ, which outputs a value yk.
The y values are the outputs of the network.
The network diagram shown above is a full-connected,
three layer, feed-forward, perceptron neural network.
“Fully connected” means that the output from each input
and hidden neuron is distributed to all of the neurons in
the following layer. “Feed forward” means that the val-
ues only move from input to hidden to output layers; no
values are fed back to earlier layers.
All neural networks have an input layer and an output
layer, but the number of hidden layers may vary.
The goal of the training process is to find the set of
weight values that will cause the output from the neural
network to match the actual target values as closely as
possible. There are several issues involved in designing
and training a multilayer perceptron network (Figure 3):
 Selecting how many hidden layers to use in the net-
work: for nearly all problems, one hidden layer is
sufficient. Two hidden layers are required for mod-
eling data with discontinuities such as a saw tooth
wave pattern. Using two hidden layers rarely im-
proves the model, and it may introduce a greater
risk of converging to a local minima. There is no
theoretical reason for using more than two hidden
layers.
 Deciding how many neurons to use in each hidden
layer: one of the most important characteristics of a
perceptron network is the number of neurons in the
hidden layer(s). If an inadequate number of neu-
rons are used, the network will be unable to model
complex data, and the resulting fit will be poor. If
too many neurons are used, the training time may
become excessively long, and worse, the network
may over fit the data. When over fitting occurs, the
network will begin to model random noise in the
data. The result is that the model fits the training
data extremely well, but it generalizes poorly to
new, unseen data. Validation must be used to test
for this.
 Finding a globally optimal solution that avoids lo-
cal minima: a typical neural network might have a
couple of hundred weighs whose values must be
found to produce an optimal solution. If neural
networks were linear models like linear regression,
it would be a breeze to find the optimal set of
weights. But the output of a neural network as a
function of the inputs is often highly nonlinear; this
makes the optimization process complex.
 Converging to an optimal solution in a reasonable
Figure 3. Perceptron network.

An Artificial Neural Network Approach for Credit Risk Management109
period of time: Most training algorithms follow this
cycle to refine the weight values:1) run a set of
predictor variable values through the network using
a tentative set of weights, 2) compute the difference
between the predicted target value and the actual
target value for this case, 3) average the error in-
formation over the entire set of training cases, 4)
propagate the error backward through the network
and compute the gradient (vector of derivatives) of
the change in error with respect to changes in
weight values, 5) make adjustments to the weights
to reduce the error. Each cycle is called an epoch.
Because the error information is propagated back-
ward through the network, this type of training me-
thod is called backward propagation. The back pro-
pagation training algorithm was the first practical
method for training neural networks. The original
procedure used the gradient descent algorithm to
adjust the weights toward convergence using the
gradient. Because of this history, the term “back-
propagation” or “back-prop” often is used to denote
a neural network training algorithm using gradient
descent as the core algorithm. Back-propagation
using gradient descent often converges very slowly
or not at all. On large-scale problems its success
depends on user-specified learning rate and mo-
mentum parameters. There is no automatic way to
select these parameters, and if incorrect values are
specified the convergence may be exceedingly slow,
or it may not converge at all. While back-propaga-
tion with gradient descent is still used in many
neural network programs, it is no longer considered
to be the best or fastest algorithm. A newer algo-
rithm, Scaled Conjugate Gradient was developed in
1993 by Martin Fodslette Moller. The scaled con-
jugate gradient algorithm uses a numerical ap-
proximation for the second derivatives (Hessian
matrix), but it avoids instability by combining the
model-trust region approach from the Leven-
berg-Marquardt algorithm with the conjugate gra-
dient approach. This allows scaled conjugate gra-
dient to compute the optimal step size in the search
direction without having to perform the computa-
tionally expensive line search used by the traditional
conjugate gradient algorithm. Of course, there is a
cost involved in estimating the second derivatives.
Tests performed show the scaled conjugate gradi-
ent algorithm converging up to twice as fast as tra-
ditional conjugate gradient and up to 20 times as
fast as back-propagation using gradient descent.
 Validating the neural network to test for over fitting:
this step is to choose the estimator to prevent the
over fitting. Using the low biased Cross-Validation
estimator is a good way to choose between differ-
ent models. A simple way to solve this significant
problem is to rate the different complexity models
with their Cross-Validation error estimator, and to
choose the better one. Cross-Validation is used both
to choose between different kinds of Supervised
Learning algorithms, and then to determine their
hyper parameters.
4. The Artificial Neural Network Model
Developed
This research compares a feed-forward multi-layers neu-
ral network developed for this study to another feed-
forward neural network, built for a research conducted in
2004. The two neural networks are similar, but they dif-
fer for the activation function adopted.
The neural network built in 2004 [18], using an algo-
rithm produced by Easy NN, has an architecture com-
posed of three hidden layers, composed respectively by 8,
4 and 4 neurons, and one output layer, composed by the
rating.
The feed-forward network architecture built for this
research, instead, is composed of an input layer, composed
of 24 neurons, two hidden layers, composed respectively
of 10 and 3 neurons, and one output layer, composed by
the score. The error function used is a linear one.
This network has been built taking advantage of the
library Fann (F i gure 4).
Both network models have been trained by means of a
supervised algorithm, namely the back propagation algo-
rithm. This algorithm performs an optimization of the
network weights, trying to minimize the error between
desired and actual output.
For training, an incremental algorithm, the standard
back propagation algorithm, was used, where the weights
are updated after each training pattern. This mean that the
weights are only updated once during an epoch function
Inputlayer
middlelayer
outputlayer
Figure 4. The neural networ k model developed.

An Artificial Neural Network Approach for Credit Risk Management 110
error. So for training two vectors were used: one contain-
ing the input patterns and another containing the pattern
desired responses (targets). Each training pattern is then
composed of a pair of input vectors/target. This is com-
pared to the outputs provided by the network.
As activation function, for the first network, a logistic
function was chosen, as in previous studies it was that
which provided more reliable results. This function is
presented by the following formula:
1
1i
i
P
ye 

Generally the response of the function determines
values between minimum and maximum in the case of
neural networks “yi” lies between 0 and 1.
In fact, the logistic function reaches the extreme values,
0 and 1, only with very high weights, and then in indi-
vidual cases we are content to approximate values.
The weight or intensity is determined by the process to
maximize the degree of matching of model predictions to
reality examined.
The neural network makes it possible to change the
learning rate set equal to 0.7, the momentum set equal to
0.8.
The activation function used for the network built for
this research is the sigmoid symmetric stepwise function.
Stepwise linear approximation to symmetric sigmoid is
faster than symmetric sigmoid but a bit less precise. This
activation function gives output that is between –1 and 1.
In this case the learning rate is equal to 0.8 and the mo-
mentum is equal to 0.5.
The data base on which were built the two networks
consists of a set of Italian manufacturing companies hav-
ing the legal form of limited liability companies. Only
companies with a workforce of under 500 units and class
turnover below 50 millions of euro were considered. We
have excluded from our study companies that showed
outliers, i.e. variables were much higher or lower than
the generality of cases.
The companies considered in the study of 2004
amounted to 273 units. The database was divided into 3
subgroups: training set, validation test and test set.
The data used were provided by “Central Balance
Sheet”, specializing in credit counseling for credit lines,
which maintains a continuous system of monitoring the
Italian companies.
Here are the ten indicators that the network of 2004
considered the most discriminating of the 80 indicators
provided by the Central Balance Sheet:
 total fixed capital to total fixed assets;
 revenues;
 Ebitda/revenues;
 Cash flow/assets;
 Number of employees;
 Working capital/assets;
 Equity/assets;
 Cash/assts;
 Functional operating working capital/net sales;
 Tangible equity + debt + group members and con-
vertible bonds /Total debt – cash.
The results obtained by network, after 101.470 cycles
of learning are shown in Table 1.
It is possible immediately to notice the increase in the
percentage of correct classification for companies con-
sidered safe and vulnerable by the Central Financial,
while there is a clear condition of misclassification for
companies at risk.
So, according to the results obtained by the neural net-
work, lots of mistakes in the model of Central Balance
Sheet were found in the classification of companies, ini-
tially considered at high risk.
The companies considered for this study amounted to
507 units, of which 359 were used for the training phase
and 148 for the stage of validation test.
Specifically, the sample was divided into three classes:
the first class includes the safe companies, the second
class includes the vulnerable companies and the third
class includes the risky ones. The 70% of companies
belonging to each of the rating classes mentioned above
was used for the training, while the remaining 30% of
each class was used for validation testing. The objective
of this choice lies in the wish to have uniform data in
terms of classes for submission to the stage of training.
The variables used as input to our network are:
 % turnover;
 % EBITDA;
 % capital;
 % equity;
 ROE;
 ROI;
 ROA;
 operating turnover;
 cash flow/assets;
 dividends/net income;
 tangible assets/operational value added;
 overall depreciation rate;
 degree of depreciation;
 operational value added/intangible assets;
Table 1. Neural network’s results.
Rating Safe Vulnerable Risk
Safe 84.2% 15.8% 0%
Vulnerable 23.1% 73.9% 3.0%
Risk 15.2% 50.0% 34.8%

An Artificial Neural Network Approach for Credit Risk Management111
 liquidity;
 short term debt;
 day average escort;
 credits vs customers;
 credits vs suppliers;
 equity/total debts;
 debts vs banks + ics/financial debts;
 financial burden./EBIDTA;
 self-financing/intangible assets;
 Equity/assets.
All inputs were normalized between the values from –1
to +1.
This step serves to ensure that data are processed, so
that they are more easily readable by the network. The
data are included in a given range; in our case the inter-
val is equal to [–1, 1]1.
The variable used as output is only one, and it is the
score.
The purpose of the network is to minimize the differ-
ence between the desired response and the one provided
by the network.
The aim of this network is to correctly classify the
companies of our sample, to create classes more homo-
geneous internally and more heterogeneous among them-
selves.
The network has undergone a phase of training; spe-
cifically, 10.000 iterations were performed on 359 units
(Figure 5).
As we can see from the graph chart above, the value of
error, in this phase, provided by the network as a result is
equal to 0.3308.
Performing the validation with the 148 units used for
the stage of validation set, we obtain the results shown in
Figure 6.
The green line represents the expected results. The
graph shows that a subdivision of the companies into the
three classes described above was expected from the net-
work. The results obtained from the network do not match
those expected. In fact, as we can see from the graph, the
network has not been able to classify companies, putting
all the 148 units used for the validation test in the first
class, as the red line of the graph indicates.
The value of error, in this phase, provided by the net-
work as a result is equal to 0.3311.
5. Concluding Remarks
The objective pursued by this paper is to describe the
artificial neural network model developed to forecast the
credit risk of a panel of Italian manufacturing companies.
In an empirical point of view, this research compares the
structure of the artificial neural network model developed
in this research to another one, built for a research con-
ducted in 2004 with a similar panel of companies. The
results on the use of neural networks recognize these
undoubted advantages. Neural networks represent an
alternative to traditional methods of classification because
they are adaptable to complex situations. As highlighted
in other papers [19], in fact, the artificial neural networks
are particularly suited to analyze and interpret—revea-
ling hidden relationships that govern the data—complex
and often obscure phenomena and processes, which are,
for example, those governing the dynamics of the vari-
ables in financial markets. The neural network models
certainly present some limits as the risk of inability to
exit from local minima, the need a lot of examples to
extract the prototype cases to be included in the training
set and the lack of transparency in the identification of
parameters most discriminatory.
We can conclude that the flexibility and objectivity of
Figure 5. Training phase’s results.
1Normalization, which consists of a linear scaling of data, is done using
the following formula:
I = Imin + (Imax – Imin)*(D – Dmin)/(Dmax – Dmin)
where Dmin and Dmax are the endpoints of the input ranges of vari-
ables, D is the value actually observed (or on) and lmax and lminare
the new ends of the range in which we report the standard variable
whose value will become. Figure 6. Validation set’s results.

An Artificial Neural Network Approach for Credit Risk Management 112
neural networks models can provide strong support in
combination with linear methods of analysis for the effi-
ciency of the processes of credit risk management of a
bank. It is not possible in fact to state if traditional me-
thods are better than non-linear one in forecasting credit
defaults, but only that the traditional methods and neural
networks have different strengths and weaknesses, which
must be carefully evaluated by the analyst during the
elaboration of the credit risk forecasting model.
6. Acknowledgement
The authors acknowledge to the anonymous referees for
their thoughtful and constructive suggestions and Maria
Rosaria Di Muro for the valuable support.
REFERENCES
[1] T. W. Anderson, “An Introduction to Multivariate Statis-
tical Analysis,” Wiley, New York, 1984.
[2] W. R. Dillon and M. Goldstein, “Multivariate Analysis
Methods and Applications,” Wiley, New York, 1984.
[3] D. J. Hand, “Dis- crimination and Classification,” Wiley,
New York, 1981.
[4] D. F. Morrison, “Multivariate Statistical Methods,”
McGraw-Hill, New York, 1990.
[5] R. A. Johnson and D. W. Wichern, “Applied Multivariate
Statistical Analysis,” 4th Edition, Prentice-Hall, Upper
Saddle River, 1998.
[6] K. Y. Tam and M. Kiang, “Managerial Applications of
Neural Networks: The Case of Bank Failure Predictions,”
Management Science, Vol. 38, No. 7, 1992, pp. 926-947.
[7] T. S. Lee, C. C. Chiu, C. J. Lu and I. F. Chen, “Credit
Scoring Using the Hybrid Neural Discriminant Tech-
nique”, Expert System with Applications, Vol. 23, No. 3,
2002, pp. 245-254.
[8] E. I. Altman, G. Marco and F. Varetto, “Corporate Dis-
tress Diagnosis: Comparisons Using Linear Discriminant
Analysis and Neural Networks (the Italian Experience),”
Journal of Banking and Finance, Vol. 18, No. 3, 1994, pp.
505-529.
[9] W. Zhang, Q. Cao and M. J. Schniederjans, “Neural
Network Earnings per Share Forecasting: A Comparative
Analysis of Alternative Methods,” Decision Science, Vol.
35, No. 2, 2004, pp. 205-237.
[10] Z. Huang, H. Chen, C. J. Hsu, W. H. Chen and S. Wu,
“Credit Rating Analysis with Support Vector Machines
and Neural Networks: A Market Comparative Study,”
Decision Support System, Vol. 37, No. 4, 2004, pp.
543-558.
[11] P. Ravi Kumar and V. Ravi, “Bankruptcy Prediction in
Banks and Firms via Statistical and Intelligent Tech-
niques-A Review,” European Journal of Operational
Research, Vol. 180, No. 1, 2007, pp. 1-28.
[12] E. Angelini, G. Tollo and A. Roli, “A Neural Network
Approach for Credit Risk Evaluation,” The Quarterly Re-
view of Economics and Finance, Vol. 48, No. 4, 2008, pp.
733-755.
[13] N. Chauhan, V. Ravi and K. Chandra, “Differential Evo-
lution Trained Wavelet Neural Networks: Application to
Bankruptcy Prediction in Banks,” Expert System with Ap-
plications, Vol. 36, No. 4, 2009, pp. 7659-7665.
[14] N. C. Hsieh and L. P. Hung, “A Data Driven Ensemble
Classifier for Credit Scoring Analysis,” Expert Systems
with Applications, Vol. 37, No. 1, January 2010, pp.
534-545. doi:10.1016/j.eswa.2009.05.059
[15] J. A. Anderson and E. Rosenfeld, “Neurocomputing:
Foundations of Research,” MIT Press, Cambridge, 1988.
[16] W. McCulloch and W. Pitts, “A Logical Calculus of Ideas
Immanent in Nervous Activity,” Bulletin of Mathematical
Biophysics, Vol. 5, No. 1-2, 1943, pp. 99-115.
[17] V. Pacelli, “Un Modello Interpretativo Delle Dinamiche
dei Corsi Azionari delle Banche: Elaborazione Teorica e
Applicazione Empirica,” Banche e Banchieri, No. 6,
2007.
[18] D. Summo and M. Azzollini, “L’Analisi Discriminante e
la Rete Neurale per la Previsione delle Insolvenze Azi-
endali: Analisi Empirica e Confronti,” Quaderni di Dis-
cussione, No. 24, Istituto di Statistica e Matematica
Università degli studi di Napoli “Parthenope,” 2004.
[19] V. Pacelli, “An Intelligent Computing Algorithm to Ana-
lyze Bank Stock Returns,” In: Huang et al. Eds. Emerg-
ing Intelligent Computing Technology and Applications,
Lectures Notes on Computer Sciences, No. 5754,
Springer Verlag, New York, 2009, pp. 1093-1104.