A Neural Network Approach for Credit Risk Evaluation

Eliana Angelini
Author’s Affiliation

Giacomo di Tollo
Author’s Affiliation

Andrea Roli
Author’s Affiliation

Abstract. The Basel Committee on Banking Supervision proposes a capital ade-

quacy framework that allows banks to calculate capital requirement for their banking
books using internal assessments of key risk drivers. Hence the need for systems to
assess credit risk. Among the new methods, artificial neural networks have shown
promising results. In this work, we describe the case of a successful application of
neural networks to credit risk assessment. We developed two neural network systems,
one with a standard feedforward network, while the other with a special purpose
architecture. The application is tested on real-world data, related to Italian small
businesses. We show that neural networks can be very successful in learning and
estimating the bonis/default tendency of a borrower, provided that careful data
analysis, data pre-processing and training are performed.

Keywords: credit risk, Basel II, neural networks

1. Introduction

The Basel Committee on Banking Supervision, with its revised capital

adequacy framework “International Convergence of Capital Measure-
ment and Capital Standards” (Basel Committee on Banking Supervi-
sion, 2005) - commonly known as Basel II - proposes a more flexible
capital adequacy framework to encourage banks to make ongoing im-
provements in their risk assessment capabilities . The text reflects
the results of extensive consultations with supervisors and bankers
worldwide1 . Regulators allow banks the discretion to calculate capital
requirement for their banking books using “internal assessments” of
key risk drivers, rather than the alternative regulatory standardized
model: the risk weights and thus capital charge are determined through
the combination of quantitative inputs provided by bank and formulas
1
The Committee subsequently released the first round of proposals for revising
the capital adequacy framework in June 1999; than in January 2001 and April 2003
and furthermore conducted three quantitative impact studies related to its proposals.
As a results of these labours, many valuable improvements have been made to the
original proposals.

c 2006 Kluwer Academic Publishers. Printed in the Netherlands.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.1

2

specified by the Committee. For the first time, banks will be permitted
to rely on their own assessments of a borrower’s credit risk.
Credit risk has long been an important and widely studied topic in
bank lending decisions and profitability. For all banks, credit remains
the single largest risk, difficult to offset, despite advances in credit
measurement techniques and the diversification of portfolio. Contin-
uing increases in the scale and complexity of financial institutions and
in pace of their transactions demand that they employ sophisticated
risk management techniques and monitor rapidly changing credit risk
exposures. At the same time, fortunately, advances in information tech-
nology have lowered the cost of acquiring, managing and analysing
data, in an effort to build more robust and sound financial systems.
In recent years, a number of the largest banks have developed so-
phisticated systems in an attempt to asses credit risk arising from
important aspects of their business lines. What have been the benefits of
the new model-based approach to risk measurement and management?
The most important is that better risk measurement and management
contribute to a more efficient capital allocation2 . When risk is better
evaluated, it can be more accurately priced and it can be more easily
spread among a larger number of market participants. The improve-
ment in credit risk modelling has led to the development of new markets
for credit risk transfer, such as credit derivatives and collateralised debt
obligations (CDOs). These new markets have expanded the ways that
market participants can share credit risk and have led to more efficient
pricing of that risk (R.W.Ferguson, 2001).
The aim of this paper is to apply the neural network approach in the
small-business lending analysis to asses credit risk of listed companies.
The focus of this article is on the empirical approach, by using two
different architectures of neural networks trained on real-world data.
The paper is organized as follows. The paper begins (section 2) by
stating the overall objectives of the Basel Committee in addressing
the topic of sound practices to the credit risk management process.
Moreover, the paper presents important elements of regulatory func-
tion. Section 3 presents an analysis of the conceptual methodologies to
credit risk modelling and focuses on the various techniques used for pa-
rameter estimation. We have tried to simplify the technical details and
analytics surrounding these models. Finally, we emphasize on neural

2
“. . .More accurate risk measurement and better management do not mean the
absence of loss. Those outcomes in the tails of distributions that with some small
probability can occur, on occasion do occur; but improved risk management has
meant that lender and investors can more thoroughly understand their risk exposure
and intentionally match it to their risk appetite, and they can more easily hedge
unwanted risk. . .” (R.W.Ferguson, 2001)

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.2

 3

network models, extremely useful for analysing small-business lending

problems. In section 4 we first briefly overview the main principles
and characteristics of neural networks; then we describe the models we
developed and tested. In section 5 the date set used is illustrated: We
select 30 companies as samples and assign each of them to one of the
two groups: a “good” one, which means that the economic and financial
situation is good, and a “bad” one, which means that the company is
close to default. Section 6 discusses the experimental settings and the
results we obtained, leading to considerable accuracy in prediction. The
paper concludes with a discussion of advantages and limitations of the
solution achieved and the future work for improvement.

2. Key Elements of the New Basel Capital Accord

2.1. Why new capital requirements?

In the last analysis, the Basel Supervisor’s Committee has published its
document describing the proposal in its near-final form. The industry
demonstrated an actual support for this survey and 265 banks from
nearly 50 countries concrete assessment of how the Committee’s propos-
als will functioning as we had hoped. Committee efforts are geared for
improving risk management, measuring risk more accurately; commu-
nicating those measurements to the management, to supervisors, and
to the public; and, of course, relating risk both to capital requirement
and to supervisory focus (R.W.Ferguson, 2003).
The existing rules, based on relatively simple 1988 Basel Accord,
represented an important step in answering the age-old question of how
much capital is enough for banks to hedge economic downturns. Specifi-
cally, under the current regulatory structure, virtually all private- sector
loans are subject to the same 8 percent capital ratio with no account of
the size of the loan, its maturity and, most importantly, the credit qual-
ity of the borrower. Thus, loans to a firm near bankruptcy are treated
(in capital requirement terms) in the same fashion as loans to a AAA
borrowers. Moreover, the current capital requirement is additive across
all loans; there is no allowance for lower capital requirements because
of a grater degree of diversification in the loan portfolio (A.Saunders,
1999).
By the late 1990s, it became clear that the original Accord was
becoming outdated. Its nature has had a tendency to discourage certain
types of bank lending. It has also tended to encourage transactions
whose sole benefits is regulatory capital relief (W.J.McDonough, 2003).
Further, the international competition, the globalisation and the im-
provements in the risk management tools, changed the way that banks

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.3

4

monitor and measure risk in manner that the 1988 Accord could not
anticipate. To response to these challenges, the Committee began a
few years ago to develop a more flexible capital adequacy framework.
The New Accord consists of three pillars: minimum capital require-
ments, supervisory review of capital adequacy and public disclosure.
The Committee believes that all banks should be subject to a cap-
ital adequacy framework comprising minimum capital requirements,
supervisory review, and market discipline.
The objective is reached by giving banks a range of increasingly
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) approach. The standardized
approach is similar to the current Accord: banks will be expected to
allocate capital to their assets based on the risk weights assigned to
various exposures. It improved on the original Accord by weighting
those exposures based on each borrower’s external credit risk rating.
Clearly, the IRB approach is a major innovation of the New Accord:
bank internal assessments of key risk drivers are primary inputs to
the capital requirements. For the first time, banks will be permitted
to rely on their own assessments of a borrower’s credit risk. The close
relationship between the inputs to the regulatory capital calculations
and banks’ internal risk assessments will facilitate a more risk sensitive
approach to minimum capital. Changes in a client’s credit quality will
be directly reflected in the amount of capital held by banks.
How will the New Basel Accord promote better corporate gover-
nance and improve risk management techniques?
First, the Basel Committee has expressly designed the New Accord
to provide tangible economic incentives for banks to adopt increasingly
sophisticated risk management practices. Banks with better measure
of their economic risks will be able to allocate capital more efficiently
and more closely in line with their actual sensitivity to the underlying
risks. Second, to achieve those capital benefits, the more advanced
approaches to credit and operational risk require banks to meet strong
process control requirements. Again, the increasing focus on a bank’s
control environment gives greater weight to the management disciplines
of measuring, monitoring and controlling risk (W.J.McDonough, 2003).

2.2. “Core components” in the IRB Approach

In the IRB Approach to credit risk there are two variants: a foun-
dation version and an advanced version. In the first version, banks

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.4

 5

must provide internal estimates of probability of default (PD)- which

measures the likelihood that the borrower will default over a given
time horizon. In addition, in the advanced approach, banks, subject
to certain minimum conditions and disclosure requirements, can deter-
mine other elements needed to calculate their own capital requirements.
They are: (a) Loss given default (LGD), which measures the proportion
of the exposure that will be lost if a default occurs; (b) Exposure at
default (EAD), which for loan commitments measures the amount of
the facility that is likely to be drawn if a default occurs; (c) Maturity
(M), which measures the remaining economic maturity of the expo-
sure (Basel Committee on Banking Supervision, 2005).In others words,
the two approaches differ primarily in terms of the inputs that are
provided by the bank based on its own estimates and those that have
been specified by the Committee.The risk weights and thus capital
charges are determined through the combination of quantitative inputs
provided by banks and formulas specified by the supervisor. The risk
weight functions have been developed for separate asset classes. For
corporate, sovereign and banks formulas are:

−50×PD 1−e−50×PD
Correlation(R)= 0.12 × 1−e 1−e−50
+ 0.24 × 1 − 1−e−50
;

Maturity adjustment(b)= (0.11852 − 0.05478 timesln(PD))2 ;

Capital requirement(K)=
   0.5  
−0.5 R
LGD × N (1 − R) × G(PD) + 1−R × G(0.999) − PD × LGD ×
×(1 − 1.5 × b)−1 × (1 + (M − 2.5) × b);

Risk-weighted assets (RWA)= K × 12.5 × EAD.

Where: N (x) denotes the cumulative distribution function for a

standard normal random variable (i.e. the probability that a normal
random variable with mean zero and variance of one is less than or
equal to x). G(z) denotes the inverse cumulative distribution function
for a standard normal random variable (i.e. the value of x such that
N (x) = z). PD and LGD are measured as decimals, and EAD is mea-
sured as currency, except where explicitly noted otherwise.This risk
weight function, based on modern risk management techniques, trans-
late a bank’s inputs into a specific capital requirement. Under the IRB
approach for corporate credits, banks will be permitted to separately
distinguish exposures to small and medium sized entities (SME). They
are defined as corporate exposures where the reported sales for the

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.5

6

consolidates group of which the firm is part is less than e50 million from
those to large firms. A firm size adjustment [0.04 × 1 − (S − 5)/45]
is made to the corporate risk weight formula for exposures to SME
borrowers (Basel Committee on Banking Supervision, 2005).
−50×PD


1−e−50×PD
Correlation(R)= 0.12 × 1−e1−e−50
+ 0.24 × 1 − 1−e−50
− 0.04×
×(1 − (S − 5)/45).

S is expressed as total annual sales in millions of euros with value of

S falling in the range of equal to or less than e50 million or greater than
or equal to e5 million. Another major element of the IRB approach
is the treatment of credit risk mitigants, namely, collateral, guarantees
and credit derivatives; in particularly the LGD parameter provides a
great deal of flexibility to assess the potential value of credit risk mitiga-
tion techniques.In these formulas we can see that the most important
risk components is the probability of default. PD estimates must be
a long-run average of one year realised default rates for borrowers in
the grade. For corporate and bank exposures, the PD is the greater
of the one year PD associated with the internal borrower grade to
which that exposure is assigned, or 0.03%. Banks may use one or more
of the three specific methods (internal default experience, mapping to
external data and statistical default models) as well as other infor-
mation and techniques as appropriate to estimate the average PD for
each rating grade.Improvements in the rigour and consistency of credit
risk measurement, the flexibility of models in responding to changes in
the economic environment and innovations in financial products may
produce estimates of credit risk that better reflect the credit risk of
exposure. However, before a modelling approach could be used in the
formal process of setting regulatory capital requirements for credit risk,
regulators would have to be confident not only that models are being
used to actively manage risk, but also that they are conceptually sound
and empirically validated. Additionally, problems concerning data lim-
itations and model validation must be cleared before models may be
implemented in the process of setting regulatory capital requirement.
At present, there is no commonly accepted framework for periodically
verifying the accuracy of credit risk models; it is important to note
that the internal environment in which a model operate – including
the amount of management oversight, the quality of internal controls,
the rigour of stress testing, the reporting process and others traditional
features of credit culture – will also continue to play a key part in the
evaluation of bank’s risk management framework (Basel Committee on
Banking Supervision, 1999).

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.6

 7

3. Conceptual approaches to credit risk modelling

Over the last decade, enormous strides have been made in the art and
science of credit risk measurement. Banks have devoted increased at-
tention to measuring credit risk and have made important gains, both
by employing innovative and sophisticated risk modelling techniques
and also by strengthening their more traditional practices (L.H.Meyer,
2000). Measuring credit risk accurately allows banks to engineer future
lending transactions, so as to achieve targeted return/risk characteris-
tics. However, credit risk models are not a simple extension of their
market risk counterparts for two key reasons (Basel Committee on
Banking Supervision, 1999).

− The specification of the process of default and rating migration is

severely constrained by a lack of data on the historical performance
of loans and other modelled variables; most credit operations are
not market to market and the predictive nature of a credit risk
model does not derive from a statistical projection of future prices
based on a comprehensive record of historical prices. The diffi-
culties in specification are exacerbated by the longer term time
horizons used in measuring credit risk, which suggest that many
years of data, spanning multiple credit cycles, may be needed to
estimate the process of default. Even if individual default probabil-
ities could be modelled accurately, the process of combining these
for a portfolio might still be hampered by the scarcity of data with
which to estimate reliably the correlations between numerous vari-
ables. Hence, in specifying models parameters, credit risk models
require the use of simplifying assumptions and proxy data.
− The validation of credit risk models is fundamentally more diffi-
cult than the back testing of market risk models. Where market
risk models typically employ a horizon of a few days, credit risk
models generally rely on a time frame of one year or more; the
longer holding period, coupled with the higher confidence intervals
used in credit risk models, presents problems to model- builders
in assessing the accuracy of their models; the effect of modelling
assumptions on estimates of the extreme tails of the distributions
is not well understood.

The new models – some publicly available and some partially propri-
etary – try to offer “internal model” approaches to measure the credit
risk of a loan or a portfolio of loans. In this section, we do not propose to
make a taxonomy of these approaches, but aim to discuss key elements
of the different methodologies.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.7

8

First, within the current generation of credit risk models, banks

employ either of two conceptual definitions of credit loss, the default
mode (DM) paradigm or the mark to market (MTM) paradigm. In the
first paradigm a credit loss arises only if a borrower defaults within
the planning horizon3 . In the absence of a default event, no credit
loss would be incurred. In the case that a client defaults, the credit
loss would reflect the difference between the bank’s credit exposure
and the present value of future net recoveries. In contrast to the DM
paradigm, in the MTM models a credit loss can arise in response to
the deterioration in an asset’s credit quality. Given the rating transition
matrix associated with each client, Monte Carlo methods are generally
used to simulate migration paths for each credit position in the port-
folio4 .Second, there are different methodologies for the unconditional
and conditional models. Unconditional approaches typically reflect cus-
tomer or facility-specific information5 . Such models are currently not
designed to capture business cycle effects, such as the tendency for
internal ratings to improve (deteriorate) more during cyclical upturns
(downturns). Instead, conditional models incorporate information on
the state of the economy, such as levels and trends in indicators of eco-
nomic and financial health, in domestic and international employment,
in stock prices and interest rates, ecc. In these models, rating transition
matrices are increased likelihood of an upgrade during an upswing in
a credit cycle and vice versa6 .
Finally, there are different techniques for measuring the interdepen-
dence of factors that contribute to credit losses. In measuring credit
risk, the calculation of a measure of the dispersion of credit risk re-
quires consideration of the dependencies between the factors determin-
ing credit- related losses, such as correlations among defaults or rating
migrations, LGDs and exposures, both for the same borrower and
among different borrowers (Basel Committee on Banking Supervision,
2000).

3
The choice of a modelling horizon of one year reflects the typical interval over
which: (a) new capital could be raised; (b) new customer information could be
revealed; (c) loss mitigation actions could be undertaken; (d) internal budgeting,
capital planning and accounting statements are prepared; (e) credits are normally
reviewed for renewal.
4
See J.P.Morgan’s Credit MetricsTM framwork (1997) for MTM approach;
Credit Risk PlusTM of Credit Suisse Financial Product (1997) for DM approach.
5
Examples are CreditMetricsTM and Credit Risk +TM; these modelling frame-
works derive correlation effects on relationship between historical defaults and
borrower-specific information such as internal risk ratings. The data is estimated
over (ideally) many credit cycles.
6
One example is McKinsey and Company’s Credit Portfolio ViewTM.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.8

 9

In closing, the fundamental elements of credit risk measurement are

easy to describe in the abstract but are far more difficult to apply case
by case. Each situation is unique, built around the roles and capabilities
of individuals, the technology system, activities and objectives of the
institutions. Anyway, to remain competitive, institutions must adapt
and constantly improve their credit risk measurement techniques.

4. The neural network approach

Neural networks have recently emerged as an effective method for credit

scoring (Wu and Wang, 2000; Rong-Zhou et al., 2002; Pang et al.,
2002; Piramuthu, 1999; Atiya, 2001). They differ from classical credit
scoring systems, such as the Z–score model (Altman, 1968), mainly
in their blackbox nature and because of they assume a non-linear re-
lation among variables. Neural networks are learning systems which
can model the relation between a set of inputs and a set of outputs,
under the assumption that the relation is nonlinear. They are consid-
ered blackboxes, since, in general, it is not possible to extract symbolic
information from their internal configuration.
In this section, we introduce neural networks with the aim of defin-
ing the background needed to discuss the model we propose and the
experimental results.

4.1. Neural networks: an introduction

Neural networks are machine learning systems based on a simplified

model of the biological neuron (S.Hykin, 1999). In the same way as
the biological neural network changes itself in order to perform some
cognitive task (such as recognizing faces or learning a concept), artificial
neural networks modify their internal parameters in order to perform
a given computational task.
Typical tasks neural networks perform efficiently and effectively are:
Classification (i.e., deciding which category a given example belongs
to), recognizing patterns in data, prediction (such as the identification
of a disease from some symptoms, or the identification of causes, once
effects are given).
The two main issues to be defined in a neural network application
are the network typology and structure and the learning algorithm (i.e.,
the procedure used to adapt the network so as to make it able to solve
the computational task at hand).

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.9

10

inputs
Σin i
ai

output

Figure 1. Artificial neuron basic structure.

An artificial neural network7 is composed of a bunch of neurons,

connected in a predefined topology. There are some possible topologies,
usually depending on the task the network has to learn. Usually, the
network topology is kept constant, but in some applications (for in-
stance in robotics) the topology itself can be considered as a parameter
and can dynamically change. The connections (links) among neurons
have associated a weight which determines type and intensity of the
information exchanged. The set of weights represents, in essence, the
information the network uses to perform the task, i.e., given a topology,
the weights represent the functions that defines the network behavior.
The artificial neuron is an extremely simplified model of the bio-
logical neuron and it is depicted in Fig.1. Neurons are the elementary
computational units of the network. A neuron receives inputs from
other neurons and produces an output which is transmitted to other
destination neurons.
The generation of the output is divided in two steps. In the first
step, the weighted sum of inputs is evaluated, i.e., every single input is
multiplied by the weight on the corresponding link and all these values
are summed up. Then, the activation is evaluated by applying a par-
ticular activation function to the weighted sum of inputs. In formulas,
for neuron i receiving inputs from neurons in the set I,

yi = j∈I Wj,i aj input evaluation
ai = g(y) the activation function g is applied
where Wj,i is the weight of the link connecting neuron j with neuron
i, and aj is the activation of neuron j.
Some kinds of activation function are used, for instance: linear, step
function and sigmoid (see Fig.2).
The most used network topologies are the following:
7
In the following, we will skip the adjective “artificial”, since we will not deal
with biological neural networks.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.10

 11
ai ai ai

ini ini ini

(a) linear function (b) step function (c) sigmoidal function

Figure 2. Examples of activation functions.

(a) completely connected (b) feedforward (c) recurrent

Figure 3. Examples of network topologies.

− Layered

− Completely connected

Networks of the first category have neurons subdivided in layers.

If the connections are only in one direction (i.e., each neuron receives
inputs from the previous layer and sends output to the following layer),
they are called feedforward networks. Otherwise, if also ’loops’ are al-
lowed, the network is called recurrent network . Completely connected
networks, on the other hand, have neurons which are all connected with
each other8 . In Fig.3, examples of the three main kinds of topologies
are depicted.
Before the neural network can be applied to the problem 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 appli-
cations, the learning algorithm stops as soon as the discrepancy (error )
between desired output and the output produced by the network falls
8
In principle, a completely connected network can be seen as a special case of a
recurrent network with one level of n neurons and n2 connections.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.11

12

below a predefined threshold. There are three typologies of learning

mechanisms for neural networks:

− 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 network. 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 introducing prizes
and penalties as a function of the network response. Prizes and penalties
are then used to modify the weights. Reinforced learning algorithms are
applied, for instance, to train adaptive systems which perform a task
composed of a sequence of actions. The final outcome is the result of this
sequence, therefore the contribution of each action has to be evaluated
in the context of the action chain produced.9
Diverse algorithms to train neural networks have been presented in
the literature. There are algorithms specifically designed for a particular
kind of neural networks, such as the backpropagation algorithm (Wer-
bos, 1988) or general purpose algorithms, such as genetic algorithms (Mitchell,
1998) and simulated annealing (Kirkpatrick et al., 1983).
At the end of this introductory section, we would like to remark
advantages and limits of systems based on neural networks. The main
advantages have to be found in their learning capabilities and in the fact
the derived model does not make any assumpion on the relations among
input variables. Conversely, a theoretical limit of neural networks is
that they are black-box systems and the extraction of symbolic knowl-
edge is awkward. Moreover, design and optimization neural network
9
In these cases it is difficult, if not impossible, to produce an effective training
set, hence the need of the use of prizes and penalties, instead of the definition of an
explicit error as in the case of supervised learning.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.12

 13

Figure 4. A four layers feedforward network.

methodologies are almost all empirical, thus the experience and sensi-
biloty of the designer have a strong contirbution in the final success.
Nevertheless, with this work we show that some useful general design
and parameters optimization guidelines exist.
In the next section we will describe in more detail the neural network
models we adopted. We used a classical feedforward network and a vari-
ation of feedforward network with ad hoc connections. Both networks
are trained with the backpropagation algorithm.

4.2. Our models

In our experiments, we used a feedforward neural network –in the

classical topology– and with a feedforward neural network with ad hoc
connections.
The feedforward network architecture is composed of an input layer,
two hidden layers and an output layer (composed of only one neuron).
In Fig.4 an example of a classical two hidden layers feedforward network
is represented. In the following, we will indicate with Wk,j the weights
between input and hidden neurons and with Wj,i the weights between
hidden and output neurons. Moreover, the activation will be Ik , Hj and
Oi for input, hidden and output neurons respectively.
The feedforward network with ad hoc connections (thereinafter re-
ferred to as ad hoc network) is a four layers feedforward network with
the input neurons grouped by three. Each group is connected to one

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.13

14

Figure 5. The ad hoc network used in our experiments.

neuron of the following layer. The reasons for choosing this topology
are to be found in the actual date we used and will be described in the
next section. The ad hoc network topology is depicted in Fig.5.
Both the networks have been trained by means of a supervised algo-
rithm, namely the backpropagation algorithm. This algorithm performs
an optimisation of the network weights trying to minimize the error
between desired and actual output. For each output neuron i, the error
is: Erri = Ti − Oi , where Ti is the desired output. The weight update
formula used to change the weights Wj,i is the following:
Wj,i ← Wj,i + ηHj Erri g (ini ),
where η is a coefficient which controls the amount of change (the learn-
 is the first derivative of the activation function g and in
ing rate), g i
is equal to j Wj,i Hj .
The formula for updating weights Wk,j is similar:
Wk,j ← Wk,j + ηIk ∆j ,

where Ik is the k-th input10 and ∆j = g (inj ) i Wj,i Erri g (ini ).
The algorithm is iterated until a stopping criterion is satisfied11 .

The definition of both the network structure and the algorithm needs
the careful choice of parameters, such as the number of neurons at
each layer and the value of the learning rate. Moreover, both the data
10
The activation of input neurons is simply the input value.
11
E.g., the overall network error is below a predefined threshold or the maximum
amount of allotted time is reached.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.14

 15

used to train the network (the training set) and the data on which
the performance of the network is validated (the test set) need to be
carefully chosen. In Sec.6, we will describe our choices along with the
benchmark used. In the next section we will describe the data set used
in the experiments.

5. Data set

For our experiments, we used data of 76 small businesses from a bank

in Italy. For each business we have data across three years (2001-2003).
For each business we have 11 fields: 8 of them are financial ratios drawn
from the balance sheet of the firms: The ratios (fields) on the sample
set are the following:
Cash flow
1. Total debt ,

Sales
2. Stock value

Short-term liability
3. Sales

Equity
4. Total assets ,

Financial costs
5. Total debts ,

Circulating capital
6. Total assets ,

Consumer credits
7. Sales ,

Value added
8. Total assets ,

The remaining 7 are calculated analysing the credit positions with

the supplying (“Andamentale” ratios) bank
Utilized credit line
9. Accorded credit line ;

Outstanding debt (quantity)

10. S.B.F. effects

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.15

16
Outstanding debt (value)
11. S.B.F. effects (value) .

and with the overall Italian Banking System (“Centrale dei Rischi”12
ratios).
Transpassing short-term
12. Accorded credit line short-term ;
Transpassing medium-long term
13. Accorded credit line M-L term ;
Utilized credit line short-term
14. Accorded credit line short-term ;
Utilized credit line M-L term
15. Accorded credit line M-L term .

The sample businesses are categorized in two groups: the “bonis”

group (composed of firms repaying the loan obligation at the end of
the analysing period) and the “default” group (composed of firms not
repaying the loan obligation at the end of the analysing period). Before
feeding the neural net with our data we decided to operate some pre-
processing operations. Here below we describe the issues about data-
pre-processing.

5.0.0.1. Missing and wrong values Some values were missing from the
data about firms. This occurrence can be due to two different reasons:

1. the value is missing because in that year there is no value in the

data-base for that firm in that particular field, or because the value
is not in the theoretical allowed range13 ;

2. the value is missing because of a computation-error. This happens

because the variables we are working on are mathematical ratios,
and in our sample some ratios can derive by a division by zero.

The usual way of overcoming this problem is to discard from the data
set the all the entries of the corresponding firm (this is the approach
pursued on several works about neural nets and credit risk), but oper-
ating in that way we would loss a significant amount of information. In
order to preserve this information we will perform the replacement of
missing values with other values. We decide to handle differently these
two situations.
12
This data are drawn from the “Centrale Dei Rischi”, a “Banca D’Italia” kept
data-base.
13
Such a kind of errors may occur because of typos and transcription errors
accidentally introduced by bank employees. This situation is quite common and
constitutes one of the sources of noise in real-world data.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.16

 17

1. In the first case (missing because of none or wrong value) we decide

to substitute the empty space with the arithmetical mean of the
field, being the mean calculated as the mean of the existing values
belonging to that field for all the businesses in the overall period of
collecting (column substitution);
2. In the latter case we decide to replace the missing value with the up-
per limit of the Normalization interval (see below). This choice can
easily understood: as this occurrence is given by a division by zero,
we can imagine the result of this ratio being ∞. If we normalize the
belonging field, this occurrence will be replaced by the maximum
value in the range. We mentioned earlier in our experiment this
upper limit being 1.

5.0.0.2. Erasing useless fields We discussed about replacing missing

and wrong values with suitable ones in order to allow the net using
those data and preserving useful information. This operation can be
harmful if the processed field show too many missing and wrong values:
replacing them would convey the net to draw wrong conclusions about
the variables-dynamic. Furthermore, as we will see, the overall-sample
splitting into training and test set is made using some percentage ratios.
For this purpose we will use the ratios [70% 30%]. In the worst case, if
for a specific field there are more than 30% of missing and wrong values
(on the overall observations) they can be all included in the test set.
This will result in an incredible loss of generalization capability of the
net. For this reason we decided not to use the fields containing more
than 30% of missing and wrong values and precisely the following fields
belonging to “Andamentale” and “Centrale dei Rischi” ratios.

5.0.0.3. Data Normalization Data normalization must be performed

in order to feed the net with data ranging in the same interval for each
input node. We choose to use the interval [0, 1] for each input node.
The most common way for normalizing data is the Min-Max linear
transformation to [0, 1]14 , but we cannot use this formula because there
are “outliers”15 in the data-base we are using. Using the Min-Max
formula we would lose a lot of useful information and several fields
would have almost all data close to one of the limit of normalization.
For this reason we decided to use the logarithmic formula to normalize
data. This formula is the most flexible because can be defined by the
14 xia −mina
The general formula of this kind of normalization is xia = max a −mina
, where
xia is the after-normalizing value and xia is the actual value.
15
Outlier means that the value is too different from the other values in the field
we are analysing.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.17

18

user, provided that the argument of the formula being < 1. For our
ratios we used the following formula:

x = logm (x + 1)
Where m is near to the actual maximum of the field we are analysing.
We add 1 to the argument to avoid the value being > 1.

5.0.0.4. Correlation analysis In order to decide which ratios to use as

input variables we performed a correlation analysis: with this operation
we want to find out the most strongly correlated variables and remove
them from our further experiments. The results of this analysis show
that there is no strong correlation between couples of variables, so we
use the 11 fields examined so far.

5.0.0.5. Training and test set We select 53 firms to consist of training

set and 23 firms to consist of test set. We want to reserve, defining both
sets, the ratio between bonis and default firms existing in the overall
sample set. For this reason we use 33 bonis examples and 20 default
examples to define the training set. For the same purpose we use 15
bonis examples and 8 default examples to define the test set.

6. Experimental Results

The two network architectures introduced in Sec.4.2 have been trained

with backpropagation using training and test sets as described in the
previous section.
The inputs of the networks are the eleven (normalized) attributes. In
the classical feedforward network, they are simply given as an ordered
array, while in the ad hoc network they are first grouped by three, each
group corresponding to the values of an attribute over three years.
The output y of the network, a real value in the range [0, 1] is
interpreted as follows:

− If y < 0.5 then the input is classified as bonis;

− otherwise, the input is classified as default

A large number of tests has been performed trying to optimally

tune parameters of the network and the algorithm. Despite many years
of research in neural networks, parameter tuning is usually still lack-
ing a theoretical framework, therefore this design phase is still per-
formed empirically. Nevertheless, some general methodological guide-
lines are possible and useful. For instance, our approach is based on

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.18

 19

a simple –but effective– systematic optimisation procedure. More for-

mally, we applied a procedure similar to a gradient ascent algorithm
in the space of parameters (Blum and Roli, 2003). First, we selected
for each parameter ph (e.g., number of hidden neurons or backpropaga-
tion parameters such as learning rate and momentum) a set of values
Vh (h = 1, . . . , Nparam ). Thus, the problem is to find an assignment
{(p1 , v1 ), . . . , (pNparam , vNparam )}, vh ∈ Vh , that leads to a network with
optimal performance. We start from a heuristically generated assign-
ment {(p1 , v10 ), . . . , (pNparam , vN 0
param
)} and we iterate the following pro-
cedure: at step h the current assignment is {(p1 , v1∗ ), . . . , (ph−1 , vh−1 ∗ ),

. . . , (ph , vh ), . . . , (pNparam , vNparam )} (only parameters p1 , . . . , ph−1 have

0 0

been assigned) and the network is trained on each instance for every
possible value of parameter ph , while the other parameters are kept
constant. Then, the optimal value vh∗ is chosen. At the end, we obtain
the assignment {(p1 , v1∗ ), . . . , (ph , vh∗ ), . . . , (pNparam , vN ∗
param
)} that is, at
least, not worse than any other partial assignment {(p1 , v1∗ ), . . .
∗ ), . . . , (p , v 0 ), . . . , (p
. . . , (ph−1 , vh−1 Nparam , vNparam )}.
0
h h
We performed a long series of experiments, of which we report only
the ones corresponding to the networks achieving the best results in
terms of classification. We obtained a very effective tuning for the clas-
sical feedforward network by applying a slight variation of the standard
backpropagation algorithm that does not propagate errors if they are
below a given threshold. With this technique we were able to produce
a network with null error on the training set and an error of 8.6%,
corresponding to only wrong bonis classifications (i.e., an input that
should be classified as bonis is classified as default). We remark the fact
that this network was able to correctly classifying all the default cases,
that are considerably riskier than bonis ones. As in machine learning
and data mining techniques is often discussed (Han and Kamber, 2000),
false negatives are usually much more dramatically important in real
world cases (e.g., in diagnosis). As reported in table I, the network per-
formance is very robust with respect to the number of hidden neurons.
In the table, we report the number of neurons in the hidden layer, the
wrong bonis (misbo) and default (misdef ) classifications and the global
error (i.e., the overall error on the training/test set), in the training set
and the test set respectively. All errors are reported in percentage. For
completeness, we report also the values of the algorithm parameters:
η = 0.2, β = 0, δ = 0.1; initial weights are randomly chosen in the
range [−1, 1].
With the ad hoc network we could also achieve very good results,
even with a lower error on the test set that was equal to 4.3% in our best
configuration. Nevertheless, the errors are all related to false negative

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.19

20
Table I. Best results achieved with a classical feedforward network. we report the number of neurons in the hidden layer, t

# of hidden tr. set tr. set tr. set test set test set test set
neurons misbo misdef error misbo misdef error

25 0% 0% 0% 13.3% 0% 8.6%

26 0% 0% 0% 13.3% 0% 8.6%

27 0% 0% 0% 13.3% 0% 8.6%

28 0% 0% 0% 13.3% 0% 8.6%

29 0% 0% 0% 13.3% 0% 8.6%

33 0% 0% 0% 13.3% 0% 8.6%

cases, i.e., the network classifies as bonis inputs that should be classified
as default. Table II summarizes the performance of the best found
configuration. For this reason, we may consider the two networks as
complementary: The one returns safe answers, while having a – rather
very small – error; the second has a very high performance in terms
of overall error on the test set, but it wrongly classifies positive cases.
The parameters used for training the ad hoc network are the following:
η = 0.8, β = 0.2, δ = 0; initial weights are randomly chosen in the
range [−1, 1].

Table II. Best results achieved with the ad hoc feedforward network. We report the wrong bonis (misbo) and default (mis

# of hidden tr. set tr. set tr. set test set test set test set
neurons misbo misdef error misbo misdef error

11 + 11 (2 layers) 0% 5% 1.8% 0% 12.5% 4.3%

7. Conclusion and future work

In this paper, we have presented an application of artificial neural

network to credit risk assessment. We have discussed two neural ar-
chitecture for the classification of borrowers into two distinct classes:
bonis and default. The system has been trained and tested on data

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.20

 21

related to Italian small businesses. One of the system we developed

is based on a classical feedforward neural network, while the other
one has a special purpose feedforward architecture. Results in both
cases show that the approach is very effective and leads to a system
able to correctly classify the inputs with a very low error. The overall
performance of the networks we developed can be considered as a state-
of-the-art result. One of the reasons for this performance is the careful
analysis of the data at hand. In fact, real-world data are often noisy
and incomplete; therefore, our analysis was aimed at eliminating wrong
values and replace empty entries with meaningful values. Moreover,
since also data normalization plays an important role in the final per-
formance, we investigated some normalization procedures in order to
keep as much information as possible in the inputs used to feed the
network.
This empirical work proves, on the one hand, evidence for the actual
applicability of neural networks in credit risk applications, especially as
black-box non-linear systems to be used in conjunction with classical
rating and classification systems. On the other hand, this research also
shows that the critical issues for developing such a kind of systems are
data analysing and processing.
Future work is focused on both methodological and application is-
sues. As to methodology, we are currently working on the design of
procedural techniques for data analysis and processing and for param-
eter optimisation. On the side of applications, we plan to assess the
generalization capabilities of the networks we obtained by testing them
on wider data bases.

References

Altman, E.: 1968, ‘Financial ratios, discriminant analysis and the prediction of
corporate bankruptcy’. J. Finance 13.
A.Saunders: 1999, Credit Risk Measurement. John Wiley & Sons.
Atiya, A.: 2001, ‘Bankruptcy Prediction for Credit Risk Using Neural Networks: A
survey and New Results’. IEEE Transactions on Neural Networks 12(4).
Basel Committee on Banking Supervision: 1999, ‘Credit risk modelling: Current
Practice and applications’. mimeo.
Basel Committee on Banking Supervision: 2000, ‘Overview of Conceptual Ap-
proaches to Credit Risk Modelling, Part II’. mimeo.
Basel Committee on Banking Supervision: 2005, ‘International Convergence of Cap-
ital Measurement and Capital Standards. A revised framework’. Bank for
International Settlements, Basel.
Blum, C. and A. Roli: 2003, ‘Metaheuristics in Combinatorial Optimization:
Overview and Conceptual Comparison’. ACM Computing Surveys 35(3),
268–308.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.21

22

Han, J. and M. Kamber: 2000, Data Mining: Concepts and Techniques, The Morgan
Kaufmann Series in Data Management Systems. Morgan Kaufmann Publishers.
Kirkpatrick, S., C. Gelatt, and M. Vecchi: 1983, ‘Optimization by simulated
annealing’. Science, 13 May 1983 220(4598), 671–680.
L.H.Meyer: 2000, ‘Why risk management is important for global financial institu-
tions?’. BIS Review (68).
Mitchell, M.: 1998, An introduction to genetic algorithms. Cambridge, MA: MIT
press.
Pang, S., Y. Wang, and Y. Bai: 4-5 Novembre 2002, ‘Credit scoring model based
on neural network’. In: Proc. of the First International Conference on Machine
Learning and Cybernetics. Beijing.
Piramuthu, S.: 1999, ‘Financial credit-risk evaluation with neural and neurofuzzy
systems’. European Journal of Operational Research 112.
Rong-Zhou, L., P. Su-Lin, and X. Jian-Min: 4-5 Novembre 2002, ‘Neural Net-
work Credit-Risk Evaluation Model based on Back-Propagation Algorithm’.
In: Proceedings of the First International Conference on Machine Learning and
Cybernetics. Beijing.
R.W.Ferguson: 2001, ‘Credit risk management: models and judgment’. BIS Review
(85).
R.W.Ferguson: 2003, ‘Basel II: A case study in Risk Management’. BIS Review (20).
S.Hykin: 1999, Neural Networks: A Comprehensive Foundation. Prentice Hall
International, Inc., second edition.
Werbos, P.: 1988, ‘BackPropagation, Past and future’. In: Proceedings of the IEEE
International Conference on Neural Networks.
W.J.McDonough: 2003, ‘Risk Management, supervision and the New Basel Accord’.
BIS Review (5).
Wu, C. and X. Wang: 2000, ‘A Neural Network approach for analyzing small business
lending decisions’. Review of Quantitative Finance and Accounting 15.

angelini_ditollo_roli.tex; 10/02/2006; 18:34; p.22