Evaluation of Solvency of potential customers of a company
The manufacturing sector is one of the main pillars of an advanced economy. It is the first sector to indicate potential national economic problems. In a similar way it is the first sector to show signs of recovery when an economy is coming out of recession or crisis. The aim of this article is to a...
Gespeichert in:
| Datum: | 2016 |
|---|---|
| Hauptverfasser: | , , |
| Format: | Artikel |
| Sprache: | English |
| Veröffentlicht: |
Інститут телекомунікацій і глобального інформаційного простору НАН України
2016
|
| Schriftenreihe: | Математичне моделювання в економіці |
| Schlagworte: | |
| Online Zugang: | https://nasplib.isofts.kiev.ua/handle/123456789/131837 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Назва журналу: | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| Zitieren: | Evaluation of Solvency of potential customers of a company / M. Vochozka, Z. Rowland, J. Vrbka // Математичне моделювання в економіці. — 2016. — № 1(5). — С. 5-18. — Бібліогр.: 15 назв. — англ. |
Institution
Digital Library of Periodicals of National Academy of Sciences of Ukraine| id |
nasplib_isofts_kiev_ua-123456789-131837 |
|---|---|
| record_format |
dspace |
| spelling |
nasplib_isofts_kiev_ua-123456789-1318372025-02-09T09:53:30Z Evaluation of Solvency of potential customers of a company Оценка платежеспособности потенциальных клиентов компании Оцінка платоспроможності потенційних клієнтів компанії Vochozka, M. Rowland, Z. Vrbka, J. Інформаційні технології в економіці The manufacturing sector is one of the main pillars of an advanced economy. It is the first sector to indicate potential national economic problems. In a similar way it is the first sector to show signs of recovery when an economy is coming out of recession or crisis. The aim of this article is to apply a neural network to be able to predict potential financial problems in manufacturing companies in the Czech Republic. Строительная отрасль является одной из ключевых отраслей всех развитых экономик мира. Раньше других отраслей она указывает на потенциальные проблемы национальной экономики. Раньше других отраслей она указывает и на потенциальное улучшение экономики, которая выходит из рецессии или даже из кризиса. Целью этой статьи является использование нейронных сетей для прогнозирования возможных финансовых трудностей строительных компаний в Чешской Республике. Будівельна галузь є однією з ключових галузей всіх розвинених економік світу. Раніше інших галузей вона вказує на потенційні проблеми національної економіки. Раніше інших галузей вона вказує і на потенційне поліпшення економіки, яка виходить з рецесії або навіть з кризи. Метою цієї статті є використання нейронних мереж для прогнозування можливих фінансових труднощів будівельних компаній в Чеській Республіці. 2016 Article Evaluation of Solvency of potential customers of a company / M. Vochozka, Z. Rowland, J. Vrbka // Математичне моделювання в економіці. — 2016. — № 1(5). — С. 5-18. — Бібліогр.: 15 назв. — англ. 2409-8876 https://nasplib.isofts.kiev.ua/handle/123456789/131837 004.77 en Математичне моделювання в економіці application/pdf Інститут телекомунікацій і глобального інформаційного простору НАН України |
| institution |
Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| collection |
DSpace DC |
| language |
English |
| topic |
Інформаційні технології в економіці Інформаційні технології в економіці |
| spellingShingle |
Інформаційні технології в економіці Інформаційні технології в економіці Vochozka, M. Rowland, Z. Vrbka, J. Evaluation of Solvency of potential customers of a company Математичне моделювання в економіці |
| description |
The manufacturing sector is one of the main pillars of an advanced economy. It is the first sector to indicate potential national economic problems. In a similar way it is the first sector to show signs of recovery when an economy is coming out of recession or crisis. The aim of this article is to apply a neural network to be able to predict potential financial problems in manufacturing companies in the Czech Republic. |
| format |
Article |
| author |
Vochozka, M. Rowland, Z. Vrbka, J. |
| author_facet |
Vochozka, M. Rowland, Z. Vrbka, J. |
| author_sort |
Vochozka, M. |
| title |
Evaluation of Solvency of potential customers of a company |
| title_short |
Evaluation of Solvency of potential customers of a company |
| title_full |
Evaluation of Solvency of potential customers of a company |
| title_fullStr |
Evaluation of Solvency of potential customers of a company |
| title_full_unstemmed |
Evaluation of Solvency of potential customers of a company |
| title_sort |
evaluation of solvency of potential customers of a company |
| publisher |
Інститут телекомунікацій і глобального інформаційного простору НАН України |
| publishDate |
2016 |
| topic_facet |
Інформаційні технології в економіці |
| url |
https://nasplib.isofts.kiev.ua/handle/123456789/131837 |
| citation_txt |
Evaluation of Solvency of potential customers of a company / M. Vochozka, Z. Rowland, J. Vrbka // Математичне моделювання в економіці. — 2016. — № 1(5). — С. 5-18. — Бібліогр.: 15 назв. — англ. |
| series |
Математичне моделювання в економіці |
| work_keys_str_mv |
AT vochozkam evaluationofsolvencyofpotentialcustomersofacompany AT rowlandz evaluationofsolvencyofpotentialcustomersofacompany AT vrbkaj evaluationofsolvencyofpotentialcustomersofacompany AT vochozkam ocenkaplatežesposobnostipotencialʹnyhklientovkompanii AT rowlandz ocenkaplatežesposobnostipotencialʹnyhklientovkompanii AT vrbkaj ocenkaplatežesposobnostipotencialʹnyhklientovkompanii AT vochozkam ocínkaplatospromožnostípotencíjnihklíêntívkompaníí AT rowlandz ocínkaplatospromožnostípotencíjnihklíêntívkompaníí AT vrbkaj ocínkaplatospromožnostípotencíjnihklíêntívkompaníí |
| first_indexed |
2025-11-25T12:33:43Z |
| last_indexed |
2025-11-25T12:33:43Z |
| _version_ |
1849765697501003776 |
| fulltext |
~ 5 ~
Математичне моделювання в економіці, №1, 2016
UDK 004.77
M. VOCHOZKA, Z. ROWLAND, J. VRBKA
EVALUATION OF SOLVENCY OF POTENTIAL CUSTOMERS
OF A COMPANY
Abstract. The manufacturing sector is one of the main pillars of
an advanced economy. It is the first sector to indicate potential national
economic problems. In a similar way it is the first sector to show signs of
recovery when an economy is coming out of recession or crisis.
The aim of this article is to apply a neural network to be able to predict
potential financial problems in manufacturing companies in the Czech
Republic.
Data on all manufacturing companies in the Czech Republic over the
period 2003-2013 were used for the modelling of the neural network. The
data file contained 67,000 records. These records included both financial
statements and non-accounting data (e.g. data on company employees).
The following networks were used for modelling the neural network:
a linear network, a probabilistic neural network (PNN), a generalised
regression neural network (GRNN), a radial basis function network
(RBF), a three-layer perceptron network (TLP) and a four-layer
perceptron network (FLP).
The analysis resulted in a concrete model of an artificial neural network.
The neural network is able to determine with more than ninety per cent
accuracy whether a company is able to overcome potential financial
problems, within how many years a company might go bankrupt, or
whether a company might go bankrupt within one calendar year. The text
also includes the basic statistical characteristics of the examined sample
and the achieved results (sensitivity analysis, confusion matrix, etc.).
The model can be exploited in practice by manufacturing company
managers, investors looking for a suitable company for capital
investment, competitors, etc.
Key words: manufacturing company, financial problems, prediction,
artificial neural network, model.
Introduction
Vomocil, Hajek and Olej [13] present that solvency in general means the ability of
an entity to duly fulfil its obligations. Evaluation of creditworthiness of customers
by business partners or banks is based on solvency, while high solvency indicates
a low credit risk and low solvency indicates a high credit risk. An analysis of
ІНФОРМАЦІЙНІ ТЕХНОЛОГІЇ В ЕКОНОМІЦІ
Ó M. Vochozka, Z. Rowland, J. Vrbka, 2016
~ 6 ~
Математичне моделювання в економіці, №1, 2016
customer solvency is a necessary condition for healthy business relation according
to [3]. According to [10] evaluation of solvency of potential customers belongs
among strategic decisions of a company management, where the so called credit
management, whose aim is to evaluate objectively the solvency of customers with
selection of the procedure to be applied to the seller, decides on a deal.
Classification of customers according to their solvency and importance for the
company is used for this purpose, while evaluation of solvency represents a process
of their multidimensional classification [10].
A solvent customer can only operate healthily if it perfectly manages not only
the business, but particularly the finance aspect of its business activity [15].
A thorough investigation into potential customer’s solvency is particularly used
in the bank sector in provision of loans. Mansouri and Dastoori [7] claim that a
credit risk is one of the biggest challenges banks and similar institutions face in all
economic systems. Solvent and insolvent customers thus have to be classified.
A company should be able to identify customers that represent a potential risk.
According to [10] this is mainly the payment risk, which might be the risk of
inability to pay, unwillingness to pay and the risk of payment delay. Unless
customers are solvent and not able to pay their debts, this might even lead to
serious threat to company health [7]. It is not easy to find the balance between the
ability to pay own bills, funding the operation capital and having satisfied solvent
customers. Solvency of customers differs in the individual countries and sectors,
while it depends not only on financial abilities of customers, but also on local
habits [3].
Availability of trade credits is a factor that particularly determines the existence
of small and medium sized companies according to [15]. Trade credit is a crucial
element of company credit policy [4]. It is necessary to introduce a suitable system
of evaluation of client’s solvency as selection of suitable business partners is the
key aspect of company success [15].
A so called Customer Profitability Analysis – CPA, which operates on the base
of historic data, particularly on the short-term base is often used for evaluation of
buyers’ solvency [4]. However numerous methods of assessment of customer’s
solvency exist. Bank information, solvency models (Kralicek Quicktest, Solvency
Index etc.), bankruptcy models (Altman Model, Index IN05 etc.), combined
models (Balanced Scorecards) or a financial analysis are among those most
frequently used.
The Kralicek Quicktest is always based on one representative of each of the
four basic spheres of financial analysis (finance, liquidation, profitability and
return), while the selection of these indicators is based on their economic
significance. Marks are always assigned to the indicators and their sum gives the
overall evaluation of the company [6].
The Solvency Index is based on multivariate discriminant analysis, it is mainly
used in Central European Countries and consists of six financial ratios.
The Altman Model [1] belongs among the best known general indices of
evaluation of a company business situation according to [11], while several
modifications have been created through the years (a variant for public limited
companies, a variant for private limited companies, a variant for non-
manufacturing companies or a variant for Czech companies). The discriminant
analysis, which reveals coefficients in a linear combination of the individual
financial ratios is applied to the calculation.
~ 7 ~
Математичне моделювання в економіці, №1, 2016
Index IN05, suitable for a less liquid capital market belongs to a series of
variants (IN95, IN99, IN01, IN05) designed by the Neumaiers [11]. They also
contain indicators of activity, return, indebtedness and liquidity [9].
The Balanced Scorecard method focuses on the strategic approach to
performance, including value processes. It consists of individual sections, which
are assigned to four perspectives (finance, customer, learning and growth and
internal processes), while fulfilment of one goal leads to fulfilment of the next one
(causal relation). These perspectives contain strategic goals, which then contain the
key indicators of performance [5].
According to [10] the cluster analysis is the most suitable applicable method. Its
aim is to classify a given number of objects into several relatively homogenous
clusters on the base of the requirement that the objects inside a cluster are as
similar as possible and the objects belonging to different clusters are as different as
possible. Modelling assisted by rating is the most frequently used solvency
modelling method according to [13]. Rating is an independent evaluation based on
quantitative and qualitative parameters and its goal is to reveal how a given entity
is able and willing to fulfil its obligations in time.
To keep pace with competitors companies must keep their customers satisfied,
which depends mainly on an analysis and evaluation of their solvency. Analyse
information [4] on customers (the information which cannot be handled by
mathematical models and optimization techniques) by means of artificial neural
networks.
The approach to evaluation of customer’s solvency has been changing a lot
recently with the growth of the progress in the IT sphere. The classical statistics,
expert systems or artificial intelligence are some of the methods used for evaluation
of solvency according to [2]. These systems have an advantage in the possibility to
evaluate the financial health of a customer from more points of view with regard to
interaction between the individual parameters. Neural networks are one of the
calculation models utilized in artificial intelligence, and they are based on the
example of biologic structures [8]. They are utilized for distributed parallel data
processing. Neutrons, which have an arbitrary number of inputs but just one output
are interconnected in a network, exchange signals between each other, which they
transform by means of transmission functions according to [12]. It is hard to
understand neural networks and to interpret them, they might moreover suffer from
overlearning [8]. Artificial neural networks undoubtedly represent a contribution in
the sense of objectivity and effectiveness of decision making on solvency, while
they also reduce the risk of the unfavourable influence of the human factor on
decision making [2]. Financial health of customers also depends on numerous
parameters that might change quickly. This is why it is also necessary to evaluate
the level of adaptability of artificial neural networks to new conditions, which is in
fact also excellent [4].
E.g. [14] deal with evaluation of solvency of energetic customers. They use
learning vector quantisation of a neural network to create an analytic credit model.
Vomocil, Hajek and Olej [13] deal with modelling of solvency of communities
by means of neural networks, which partially differs from modelling of solvency of
customers. The final model is based on a forward-propagation neural network with
several variants, which are verified, and their quality is evaluated according to the
criterion of mean square deviation of the data in the tested set. The best results are
achieved by the variants with the Quick Propagation learning function [13].
~ 8 ~
Математичне моделювання в економіці, №1, 2016
The aim of this article is to utilize neural networks for creation of a model for
determination of solvency of customers of a manufacturing company.
1. Material and methodology
The data file was created on the base of a randomly selected sample of 110
manufacturing companies. Specializations of the companies was not important for
the selection. However it was necessary to choose companies with longer than
20-year history, which means somehow established companies. The data relates to
2014.
As mentioned above the group contains information on 110 companies, namely:
– company name;
– place of business;
– the complete financial statement data for the examined year;
– selected financial ratios (return on assets, return on equity, return on capital
employed, 1-3 level liquidity, debt etc.).
MS Excel was used for the preparation of the data file. The data for each
company was always presented on one line. The file thus contains (apart from the
heading) 110 lines and 167 characteristics in each line. It was imported into
Statistica software by DELL. The data was subsequently processed by an
intelligent problem solver, actually this was a classification problem processed by
means of artificial intelligence (artificial neural networks).
An artificial neural structure was sought that would be able to classify each
company on the basis of the input data into one of the following two groups:
1. solvent company,
2. insolvent company.
First we determined the characteristics of the individual companies. We had to
define the output category quantity. In this case it was obviously the value
presented in the column “final status” of the MS Excel sheet. A solvent company
shows a positive difference between the working capital, trade liabilities and has no
overdue debts. Unless these conditions are met we talk about an insolvent
company. Such a company is considered to some extent risky to trade with, i.e.
required a higher guarantee for unpaid debts (cash payment in advance, insurance,
a letter of credit etc.). The other quantities are input ones and are continuous except
for the place of business. The place of business is an input – categorical quantity.
Once this exercise was completed, 1,000 artificial neural structures1 were
generated, of which the 10 most suitable were retained. For the model, linear neural
networks, probabilistic neural networks, radial basis function neural networks,
three-layer perceptron networks and four-layer perceptron networks, were utilised.
For the radial basis function neural network we used 1 up to 21 hidden neurons.
The 2nd layer of the three-layer perceptron network contained 1 up to 100
hidden neurons.
The 2nd and the 3rd layers of the four-layer perceptron network both contained
1 up to 100 hidden neurons.
1 Unless the improvement in the individual trained networks is significant the training of networks can be
shortened.
~ 9 ~
Математичне моделювання в економіці, №1, 2016
2. Results – production function
1,000 artificial neural networks were generated on the basis of the set parameters.
10 artificial neural networks showing the best characteristics were retained for
further assessment and subsequent processing. The results of the analysis are given
in Table 1.
Table 1 – Models of artificial neural networks showing the best characteristics2
Profile Train
Perf.
Select
Perf.
Test
Perf.
Train
Error
Select
Error
1 Linear 1:1-1:1 0.627907 0.714286 0.619048 0.481156 0.886400
2 Linear 2:2-1:1 0.674419 0.476190 0.428571 0.477751 0.883533
3 PNN 1:1-43-2-2:1 0.558140 0.666667 0.571429 0.513807 0.538365
4 PNN 2:2-43-2-2:1 0.604651 0.666667 0.476190 0.535191 0.530754
5 RBF 45:45-8-1:1 0.906977 0.857143 0.904762 0.287553 0.399138
6 RBF 45:45-12-1:1 0.953488 0.809524 0.904762 0.244266 0.361051
7 RBF 45:45-15-1:1 0.930233 0.857143 0.857143 0.215383 0.330831
8 MLP 91:91-99-59-1:1 0.953488 0.904762 0.904762 0.000000 0.100207
9 MLP 99:99-41-1:1 1.000000 0.904762 0.761905 0.001118 0.062911
10 MLP 99:99-55-1:1 0.953488 0.857143 0.952381 0.000256 0.012762
Profile Test
Error
Training/ Members Inputs Hid
(1)
Hid
(2)
1 Linear 1:1-1:1 0.53625 PI 1 0 0
2 Linear 2:2-1:1 0.53456 PI 2 0 0
3 PNN 1:1-43-2-2:1 0.51521 1 43 2
4 PNN 2:2-43-2-2:1 0.51369 2 43 2
5 RBF 45:45-8-1:1 0.35694 KM, KN, PI 45 8 0
6 RBF 45:45-12-1:1 0.33981 KM, KN ,PI 45 12 0
7 RBF 45:45-15-1:1 0.35566 KM, KN, PI 45 15 0
8 MLP 91:91-99-59-1:1 10.31337 BP100, CG20, CG1b 91 99 59
9 MLP 99:99-41-1:1 4.70291 BP100, CG20, CG58b 99 41 0
10 MLP 99:99-55-1:1 7.14166 BP100, CG20, CG0b 99 55 0
Source: Authors
Particularly linear neural networks were retained among the ten best networks,
then probabilistic neural networks, radial basic function neural networks, multiple
perceptron networks with two hidden layers and a multiple perceptron network
with four layers.
Figure 1 shows a schematic illustration of a linear neural network Linear 1:1-1:1.
2 A linear neural network is indicated as Linear, probabilistic neural network as PNN, generalized regression
neural network as GRNN, radial basis network as RBF and multi-layer perceptron network as MLP.
~ 10 ~
Математичне моделювання в економіці, №1, 2016
Source: Authors
Figure 1 – Graph of artificial neural network (Linear 1:1-1:1)
The first layer (from the left) represents the input to the model. The triangle
represents an input for the model, namely a continuous quantity. The model does
not have high success rate, it oscillates depending on the set (the training set, the
validation set and the data test set) from nearly 62% to nearly 71.5%. The model is
nevertheless interesting because of the fact that it is only based on a single quantity
that may be simply identified in the practice, even by a layperson – we namely
derive the correct result from the value of the tangible assets. These are assets with
the time of use longer than one year (buildings, land, machines, groups of movable
things etc.). The model is particularly suitable for quick identification of seller’s
solvency and for determination of the condition for further transactions between
a supplier and a customer.
Figure 2 shows a graphic illustration of a linear neural network in the Linear 2:2-1:1
configuration.
Source: Authors
Figure 2 – Graph of artificial neural network (Linear 2:2-1:1)
~ 11 ~
Математичне моделювання в економіці, №1, 2016
In this case the network did not reach the optimum values either. The prediction
or classification power oscillates across the sets of the group within nearly 43%
and nearly 67.5%. This model is characterized by high simplicity as well.
An evaluator is able to determine the resulting state on the base of two quantities.
Like in the first case these are the tangible assets. The material and energy
consumption is an additional indicator. The second indicator is very useful – it
presumes the level of company performance. Unfortunately it has caused lower
accuracy of the model. And as the model has not reached the prediction power
higher than 50%, it is useless in the practice (although it might be useful in the
theory).
Figure 3 shows a graphic illustration of a probabilistic neural network
PNN 1:1-43-2-2:1.
Source: Authors
Figure 3 – Graph of artificial neural network (PNN 1:1-43-2-2:1)
In the instance of the described probabilistic neural network the probability of
prediction oscillates from nearly 56% to more than 66%. The model is thus more
successful than the second linear network. Although there is one input indicator, it
does not use the same linear models as tangible assets. On the other hand it avoids
the use of financial ratios and uses the number of employees as the basic indicator.
It is applicable in the practice for its simplicity. The prediction power is also higher
than 50%. The application of this network is however less beneficial than that of
the first linear network.
Figure 4 shows a graphic illustration of a probabilistic neural network
PNN 2:2-43-2-2:1.
~ 12 ~
Математичне моделювання в економіці, №1, 2016
Source: Authors
Figure 4 – Graph of artificial neural network (PNN 2:2-43-2-2:1)
In the instance of the second probabilistic neural network the classification
accuracy is set across the sets of data from more than 47.5% to more than 66%.
With regard to the fact that the value is lower than 50% for one set, the file cannot
be used in practice. There is an interesting point that the model utilizes two inputs,
namely the number of employees and the type of company according to the
financial statement. The economic interpretation of the results is thus not fully
logical either.
Figure 5 shows a graphic illustration of a radial basis function neural network
RBF 45:45 -8 -1:1.
Source: Authors
Figure 5 – Graph of artificial neural network (RBF 45:45 -8 -1:1)
~ 13 ~
Математичне моделювання в економіці, №1, 2016
In the instance of this radial basis function neural network the classification
power oscillates at the level from 85.7% to 90.7%. The prediction accuracy is thus
very high. The networks uses 45 continuous quantities as the input values. These
are various data on employees, all the financial statement components as well as
selected financial ratios, namely from the number of employees via current assets
up to the proportion of bank loans in the total liabilities. We admit that a large
volume of data is necessary here. On the other hand it is a remarkable quality
proportion from the point of view of prediction. 85.7% represents relatively
reliable classification of a company and the whole model is thus applicable in the
practice (even despite the number of input data).
Figure 6 also shows a graphic illustration of a radial basis function neural
network RBF 45:45-12-1:1.
Source: Authors
Figure 6 – Graph of artificial neural network (RBF 45:45-12-1:1)
In the instance of the second neural network the probability of correct
classification of a company is also relatively high. It oscillates from 80.9% up to
95.3%. Although it is lower, the model is still applicable in the practice. This
network is also characterized by high demand for input quantities. It also needs 45.
The input data is identical to that of the network No. 15. We may then see the
difference between the two networks in the number of hidden neurons only.
Figure 7 shows a graphic illustration of a third radial basis function neural
network RBF 45:45-15-1:1.
~ 14 ~
Математичне моделювання в економіці, №1, 2016
Source: Authors
Figure 7 – Graph of artificial neural network (RBF 45:45-15-1:1)
The third RBF offers reliability at the level from 85.7% to 83%. This means that
it is characterized by the highest success in company solvency prediction. This
RBF also works with 45 input data identical with those of the previous two
networks. However it utilizes 15 neutrons in the hidden layer. This model is the
most successful one so far, and moreover, despite the need of significant volume of
data it is applicable in the practice.
Figure 8 shows a graphic illustration of a multi-layer perceptron network with
two hidden layers MLP 91:91-99-59-1:1.
Source: Authors
Figure 8 – Graph of artificial neural network (MLP 91:91-99-59-1:1)
~ 15 ~
Математичне моделювання в економіці, №1, 2016
The four-layer perceptron network MLP 91:91-99-59-1:1 shows the success rate
in classification of solvent and insolvent companies at the level from 90.5% to
95.3%. It is thus the most suitable model so far. Unfortunately it utilizes 91 various
company data as input values. It uses 99 neutrons in the first hidden layer and 59 in
the second one. The prediction power is however very high, more than 90%.
Thanks to this fact the neural network is applicable in the practice.
Figure 9 shows a multi-layer perceptron network with one hidden layer
MLP 99:99-41-1:1.
Source: Authors
Figure 9 – Graph of artificial neural network (MLP 99:99-41-1:1)
The three-layer perceptron network MLP 99:99-41-1:1 shows prediction
capability from 76% to 100%. Unfortunately the difference between the individual
data sets in terms of prediction power is too big, nearly 24%. This handicaps the
practical applicability of the whole network. It moreover needs 99 input data items.
Figure 10 also represents a three-layer perceptron network, namely
MLP 99:99-55-1:1.
~ 16 ~
Математичне моделювання в економіці, №1, 2016
Source: Authors
Figure 10 – Graph of artificial neural network (MLP 99:99-55-1:1)
The second perceptron network utilizes 99 input data items for the calculation
as well. However its success rate is remarkably better than in the previous case. Its
prediction ability oscillates from 85.7% to 95.2%. This means that this network is
applicable in the practice as well.
Only on the base of the graphic representation of the individual networks and
their basic characteristics we can say that the multi-layer perceptron network with
two hidden layers MLP 91:91-99-59-1:1 seems the best – throughout three data
sets (a training one, a validation one and a testing one).
To be able to confirm the results derived from the schemes of the individual
networks and the prediction power as the most important characteristic we will use
the ROC Curve (Receiver Operating Characteristic). The curve is shown in figure 10.
The ROC Curve examines the sensitivity and specificity of the individual
selected neural networks. We ideally seek for a point approximating the point [0,1].
Figure 10 demonstrates the course of 30 curves, i.e. 10 neural networks, each with
three data sets.
An ideal network would offer the result in the form of three 1s. If we examine
the curves and the keys to these curves, we choose the most successful networks
among numbers 18, 19 and 20. Two of their data sets always give 1 as the result.
We thus determine the most successful network according to the highest value of
the third set. This is 0.963 of network No. 18.
We can then say that we have found and confirmed the most successful model
for prediction of solvency and possible risk of potential clients of a manufacturing
company. It is the four-layer perceptron network MLP 91:91-99-59-1:1, which
defines the required result with at least 90 per cent accuracy.
~ 17 ~
Математичне моделювання в економіці, №1, 2016
Source: Authors
Figure 11 – Receiver Operating Characteristic Curve
Conclusion
The aim of the article was to utilize neural networks for creation of a model
for determination of solvency of potential customers of a manufacturing company.
On the base of obtained data on a group of 110 manufacturing companies 1,000
artificial neural networks were generated. Ten of these networks, which showed the
best results, were retained for further processing. After an analyses of the
individual networks and upon application of the ROC Curve a four-layer
perceptron neural network with two hidden layers MLP 91:91-99-59-1:1 was
determined as the best one.
The obtained neural network predicts solvency or possible risk of a company
(potential customer of a manufacturing company) with the probability higher than
90%. Such a prediction power predetermines the model to practical application.
The other networks are not so far as successful in company classification as the
one. The fact that the network needs 91 company characteristics to provide the
required result is a negative side of the model. If we wanted to use a neural
network for some kind of a quick test only the linear network Linear 1:1-1:1 would
be suitable, as it identifies the result with relatively high rate of reliability upon
a single variable – the tangible assets.
We should note that the model becomes applicable in the practice thanks to its
properties. It may be utilized by the company itself or by the customers of the
company or other entities.
~ 18 ~
Математичне моделювання в економіці, №1, 2016
REFERENCES
1. Altman, E. I. (1968). Financial ratios, discriminant analysis and the prediction of
corporate bankruptcy. The Journal of Finance, 23(4), pp. 589–609.
2. Buchtíková, A. (1998). Příspěvek k hodnocení finanční bonity bankovních klientů.
Praha: Česká národní banka, Institut ekonomie, 67 p.
3. Čermák, P. (2015). Customer Profitability Analysis and Customer Life Time Value
Models: Portfolio Analysis. Procedia Economics and Finance, (25), pp. 14–25.
4. Ersoz, S., Yaman, N., & Birgoren, B. (2008). Modeling and Analyzing Customer Data
in Customer Relationship Management with Artificial Neural Networks. Journal of the
Faculty of Engineering and Architecture of Gazi University, 23(4), pp. 759–767.
5. Kaplan, R. S., and Norton, D. P. (2005). Balanced scorecard: strategický systém měření
výkonnosti podniku. 4th ed., Praha: Management Press, 267 p.
6. Kralicek, P. (1991). Grundlagen der Finanzwirtschaft: Bilanzen, Gewinn – und
Verlustrechnung, Cashflow. Kalkulationsgrundlagen, Fruehwarnsysteme Finanzplannung.
Wien: Ueberrauter.
7. Mansouri, S., & Dastoori, M. (2013). Credit Scoring Model for Iranian Banking
Customers and Forecasting Creditworthiness of Borrowers. International Business
Research, 6(10), pp. 25–39.
8. Mohamad, H. H., Ibrahim, A. H., & Massoud, H. H. (2014). Modelling the financial
performance of construction companies using neural network via genetic algorithm.
Canadian Journal of Civil Engineering, 41(11), pp. 945–954.
9. NEUMAIER, I., & NEUMAIEROVÁ, I. (1995). Zkuste spočítat svůj INDEX IN 95. In:
Terno, (5), pp. 7–10.
10. Režňáková, M. (2009). Klasifikace obchodních partnerů s využitím metod shlukové
analýzy. 7. Mezinárodní konference Finanční řízení podniků a finančních institucí, 7, pp. 1–7.
11. Sedláček, J. (2007). Finanční analýza podniku. 1st ed., Brno: Computer Press, 154 p.
12. Šnorek, M. (2002). Neuronové sítě a neuropočítače. 1st ed., Praha: Vydavatelství
ČVUT, 156 p.
13. Vomočil, M., Hájek, P., & Olej, V. (2007). Modelování bonity obcí pomocí dopředných
neuronových sítí. Scientific Papers of the University of Pardubice, Series D, (11), 172–181.
14. Wang, J., & Wen, Y. (2008). Application of Genetic LVQ Neural Network in Credit
Analysis of Power Customer, Fourth International Conference on Natural Computation,
pp. 305–309.
15. Wodyńska, A. (2009). Evaluation of Customer’s Creditworthiness as the Instrument of
Corporate Trade Credit Policy. Contemporary Economics, 3(1), pp. 89–103.
Стаття надійшла до редакції 02.02.2016
1. Material and methodology
1. Material and methodology
2. Results – production function
Conclusion
|