Page 233

AD ALTA

JOURNAL OF INTERDISCIPLINARY RESEARCH

per share, main enterprise costs, total costs of annual wages,
main entrepreneurship income, net profit after tax, return on total
assets, or profitability on equity (Zhang and Zhong, 2015, p.
179). The output is represented by a value copying the course of
economic indicator by which the network was determined – a
whole range of economic indicators may be used, while the most
suitable are difficult to be set and complex or testifying
(Galushkin, 2012). Different types of artificial neural networks
may create the network architecture – according to what the
neuron transmission function is, how neurons are interconnected
mutually, how many input neurons, hidden layers there are, etc.
(García, Giménez and Guijarro, 2013). Similarly, the amount of
layers depends on the given model´s author´s consideration. If
there are not so many in the neural networks, they learn quicker.
If there are bigger numbers of layers, they are able to generalize
better (Elsawy, Hosny and Razek, 2011).

The disadvantage is that we never know how the network makes
its decisions, and why it has decided the way. It is thus
impossible to know the inner structure of this system. That is
why neural networks are also termed as ´black boxes´ (Tzeng
and Ma, 2005). The networks are very comfortable and practical,
but the way they evaluate the enterprise exactly, is not always
very clear (Shi, Bian and Zhang, 2010, p. 640). We also always
operate with the probability that the response will be, in certain
percentage, wrong. This fact considerably limits its use in areas
with one-hundred-percent-flawlessness (Slavici, Mnerie and
Kosutic, 2012). While evaluating an enterprise, networks are
also sensitive to organization and preparation of data, but also to
the whole configuration. To apply them, a high computing
power is needed, and their processing takes a long time (Kim,
An and Kang, 2004).

Advantages in using artificial neural networks during a complex
enterprise evaluation are the following (Ciobanu and Vasilescu,
2013):



Simple implementation,



Possibility of parallel processing,



Learning and generalization ability,



Adaptation ability,



Distributed representation and calculation.

Nevertheless, there is a range of disadvantages while using
artificial neural networks while evaluating an enterprise in a
complex manner (Knez-Riedl and Mulej, 2014):

They work with a so-called ´black-box approach´ - their inner
functionality is not directly known,
At the training stage they are computing-power consuming,
Their processing often takes a long time,
Networks are unable to solve other, similar problems other than
those they are trained to solve,
Networks are an approximation of the required solution – it is
necessary to always count on certain error rate,
Networks are prone to be over-trained.

Only a few authors dedicate their work to applying neural
networks in order to evaluate an enterprise in a complex manner.
The reasons are probably reasonably significant disadvantages of
artificial neural networks, and the existence of many complex,
and often simpler models for enterprise evaluation. Zhang and
Zhong (2015) have suggested a model based on artificial neural
networks, which has a high prediction accuracy and its results
are objective and exact. A similar model is presented by
Makeeva and Bakurova (2012). The background for its creation
is profitability, liquidity, indebtedness and return indicators. Al-
Shayea and El-Refae (2012) have created a model for insolvency
prediction based on less used types of neural networks –
GMDH

Counter Propagation and fuzzy ARTMAP

networks.

The most influencing factors when evaluating, are, according to
them, net profit, total equity, costs on sale, sales, cash flow, and
credits. Net profit, annual volume of work and work capital are

Group Method of Data Handling – Networks with inductive modelling.

Adaptive Resonance Theory MAP – neural network hybrid architecture.

the main indicators of financial performance of any building
company (Mohamad et al., 2014). It was Mohamad et al. (2014)
who have developed a hybrid model (an artificial neural network
technique + genetic algorithm) with the aim to predict, based on
the previously published data on financial statements, the
amount of the three main given indicators of building
companies´ financial performance. Complex enterprise
evaluation methods created via artificial neural networks are
often used by banks when considering credit requests – credit
risk evaluation in a given enterprise (Mansouri and Dastoori,
2013). A complex enterprise evaluation´s aim in this case is to
minimize credit risk and improve decision-making process while
establishing business relationships in economic, legal and social
sphere (Yongli et al., 2013). A model based on GRNN

neural

network, designed by Zhu et al. (2015) may serve as an example.
It may evaluate credit risk efficiently.

Complex enterprise evaluation methods are nowadays created by
modern analytical models using computers and sophisticated
mathematical models (Gholizadeh et al., 2011). Neural network
imperfections, however, point to the fact that this technology
still undergoes the process of development and improvement.
Even so, they may be used as a complex enterprise evaluation
indicator, while complemented and combined with other models
very often. Many authors have proven that complementation by
other models improves the calculation, and raises the efficiency
and accuracy of the result (Ciobanu and Vasilescu, 2013, p.
448). The obtained model outputs may be further compared to
the other enterprises´ results or to the results of best enterprises
working in the same branch (Rosillon and Alejandra, 2009).
The aim of this contribution is to create a model, using neural
networks, which will be able to predict the enterprise´s ability to
survive possible financial distress.

2 Data and Methods

Basic data about enterprises, which is going to be analysed and
examined comes from the Albertina database. These are
enterprises classified among building enterprises by the Czech
Statistical Office. These enterprises fall among the classification
F-section in CZ-NACE (economic activity classification). The
resulting file includes exactly 65 536 data lines. Each line
consists of a hundred characteristics. Specifically, they are
financial parameters and non-financial indicators.

Financial parameters include all data from financial statements i.
e. balance sheets, profit and loss statements, cash flow
statements. Further, earnings before interest and tax (EBIT) is
included. Non-financial indicators include enterprise
identification (name and identification number), enterprise
business district, number of employees and the enterprise
auditor´s statement.

It is common to start the paper with input data analysis from the
perspective of their objective interpretation. Data analysis has
been carried out, but only on the level of variable classification,
not from the perspective of ´economic fundaments´. In case
some available data is excluded already at the stage of data file
preparation, we could reach a situation of excluding a variable,
which, although refused by current economic theory, may
significantly influence the result. Thus, we are facing a dilemma
whether to include a greater amount of variables (some even
against the sense of current knowledge) and obtain a result,
which may be economically difficult to interpret, or whether the
amount of variables should be decreased to values possible to be
relatively easy interpreted today. I have chosen the first option.
The economic environment has changed so much as we can not
describe it using the same variables as we had done several
decades ago.

To prepare a data file MS Excel will be utilized. The data file
will be imported into the DELL Statistica software in version No
12 and version No 7 (result visualization). Subsequently it will

Generalized Regression Neural Network.

- 233 -