Diatchyk A.G., Oglih V.V., Rieznik M.A.

*Oles Honchar Dnipropetrovsk National University, Ukraine*

**APPLICATION OF NEURAL NETWORKS TO DETECT ENTERPRISE BANKRUPTCY RISK**

The globalization processes in economy, finance and society have resulted in skyrocketing complexity of modern financial and social organizational systems and, consequently, the increase in instability and indefiniteness of these systems. Social and economic institutes are influenced more and more often by external and intersystem events which lead to disastrous losses. Therefore, implementation of mechanisms which regulate the sensitivity to the approach of risk and limitations caused by the risks of losses into the management process are becoming more and more urgent, that, in turn, assumes the use of modelling in economics.

The method of neural networks has become widespread in the field of modeling of financial stability of enterprises. The application of this technology, which was developed by U.Makkalok and U.Pitts in 1943, makes it possible to model the nonlinear processes, to work with noisy data, as well as to adapt to various spheres of research. The fact that artificial neural networks do not need programming distinguishes them from regular software systems. The most valuable characteristic of neural networks lies in their ability to be trained on a set of examples when mechanisms of situation development and relations of the input data with output data aren't known. Thus, neural networks are capable of solving problems, having leaning on incomplete, corrupted, noisy and internally inconsistent input data.

Neural network is applied primarily when the exact type of a relation between input data and the result is not certain. It is sufficient to be aware that there is a relation of the entrance and target data. Thus, dependence will be derived in the process of neural network training [3].

Application of neuronet technology is appropriate in cases when formalization of a process of problem-solving is complicated or even impossible. They are a very powerful tool of modeling because of their nonlinear nature [2].

There is a positive experience of the use of neural networks in the analysis of financial risks and other types of risks. In the conditions of a crisis the necessity for a qualitative tool of the analysis of the enterprise financial condition and of detection of the risk of bankruptcy has increased, and the application of traditional statistical methods cannot help prevent threats in proper time. It is very important to trace the dynamics of the enterprise development and detect the stage this enterprise is in.

The task complexity is determined by the necessity to analyze absolutely various data. Neural networks, however, don't require any certain type of distribution of the initial data or the linearity of target functions. They don't need a specially qualified user, as opposed to the statistical methods which require fundamental knowledge of the theory of probability and mathematical statistics. Neural networks are capable to work easily with a great volume of entrance data. They also accelerate the process of finding a result due to simultaneous processing of all data by neurons.

Artificial neural networks are a big group of systems whose architecture is similar to the structure of the nerve tissue. One of widespread architectures is that of a multilayered perceptron with the back propagation of an error. The multilayered perceptron simulates the work of neurons as a part of a hierarchical network where each neuron of a higher level is connected by its inputs with the outputs of neurons of a lower layer. The value of input parameters which are the basis to make some decisions, to predict the situation development, etc., are passed on to the neurons of the lowest layer. These values are considered as signals and transferred to the following layer where they can make relations of numerical values, which are attributed to interneuron communications, weaker or stronger. As a result, a certain value which is considered as the response (reaction) of all networks to the value of input parameters is made as an output. Before using the neural network it is necessary to "train" it on the basis of the data received earlier, when the input parameters and the right answers are known. The training consists in selecting the scales of interneuron communications which provide the greatest accuracy of the network answers to the known right answers.

In our opinion, in order to divide the enterprises in two categories (potential bankrupts and stable companies) it is sufficient to use one of simple neural networks, the three-layer perceptron with the algorithm of return distribution, as a training one.

Such a network can model the function of virtually any degree of complexity.

The development of entering parameters of a network makes a strong impact on the accuracy of a forecast. It is offered to use financial factors which are used for inconsistency prediction as entrance knots. The value of the only knot of an output target layer is an indicator of financial solvency of the enterprise. “One” means the state of bankruptcy, “zero” value means financial well-being. Sigmoid function is chosen to be the activation function.

A certain set of financial coefficients serve as entrance data.

The comparison of various variants of sets of financial coefficients has shown that the accuracy scatter makes 25%.

If we take the set of indicators which are typical for the discriminant analysis and the analysis by means of indistinct models, particularly, the precision in distributing enterprises as potential bankrupts and financially-stable companies has made 87.1 %. It is less, than while using of these indicators in indistinct model [3].

The use of indicators which display different kinds of profitability gives the accuracy of result which makes 70 % [3].

The following set of entrance financial factors has shown the highest level of accuracy of the forecast [2]:

1) |
Working capital/Total asset value |
4) |
Authorized capital/Amount of indebtedness |

2) |
Net profit/Total asset value |
||

3) |
Net sale proceeds/Total asset value |
5) |
Proceeds from sale/Total asset value |

The set of indicators has been changed further to increase the accuracy of classification [2]:

1) |
Net working capital/ Total asset value |
3) |
Coefficient of current assets |

2) |
Profit/ Owned capital |
4) |
Profit rate |

The accuracy of classification of enterprises at the given financial coefficients has made 95%.

In conclusion, we can say that currently the use of neural networks in modeling and forecasting financial stability of enterprises is one of the most exact methods. However, the use of indicators which are generally used in the statistical analysis gives the low accuracy of results. The right choice of the input financial factors will provide high effectiveness of bankruptcy forecasting.

**Literature:**

1. Теребух А. А. Порівняльний аналіз моделей оцінювання фінансових загроз суб’єктів господарювання/ А. А. Теребух, Н. О. Діброва// Збірник науково-технічних праць: Науковий вісник НЛТУ України. – 2010. – №20.11. – С.228-240.

2. Зайченко Ю.П. Сравнительный анализ методов оценки риска банкротства предприятий Украины/ Ю. П. Зайченко, С. Рогоза, В. Столбунов// Information Science and Computing. – 2008. – С.103-110.

3. Матвійчук А. Моделювання фінансової стійкості підприємств із застосуванням теорій нечіткої логіки, нейронних мереж і дискримінантного аналізу/ А. Мотвійчук// Вісник НАН України. – 2010. – №9. – С. 24-46.