Research of Possibility of Bankruptcy Diagnostics Applying Neural Network
Research of Possibility of Bankruptcy Diagnostics Applying Neural Network

Summary/Abstract: This paper analyses the possibilities of prediction of the enterprise bankruptcy, applying neural network. The prediction results of the failed enterprises are compared to the prediction results of the profitable enterprises. In this way we can see the peculiarities and reliability of neural network usage for the bankruptcy diagnosis. Lithuanian enterprises work in different conditions than other foreign enterprises, but all introduced models were based only on foreign enterprises, so their applicability for the diagnostics of bankruptcy remains disputable. Estimations of enterprises for the bankruptcy in definite time, calculated from “Z” scores of different authors, are different so they have to be analysed taking into account of their changing tendencies. This article discuses the application of neural networks to analyse the possibility of enterprise bankruptcy. The classification of neural networks, estimation of the number of hidden layers and their size, the methods of training are described in special scientific literature. Perceptrone neural network was constructed of 3 layers. To train it the backpropagation method was used. The algorithms of training and the programmes to implement them require a lot of samples of enterprises – over ten times more than inputs of the enterprise state. To train the network following indicators were used: the indicator of net profitability of assets, coefficient of short-term solvency; debt ratio; ratio of short-term liquidity of the years 1998-2001. The authors had data of 13 enterprises, so they increased the number by including the same enterprises in the list several times. In this way 284 enterprises were obtained: 161 failed and 123 profitable. By training various networks with different inputs it was researched what indicators of the enterprise were the best to forecast the bankruptcy. Therefore the neural network was trained in the optimisation mode. The programme used different combinations of inputs and checked 408 different versions of the neural networks. As a result, all the used inputs could forecast bankruptcy, except the profitability of assets of the year 2000 and the short-term liquidity ratio of the year 1998. According to the small amount of enterprises (8 profitable and 5 insolvent) and their 4 financial ratios used to train the neural network, the percentage of the right diagnosis is 84. But when increased the number of enterprises to 284 (written the same enterprises a few times in the same list), the results of right diagnosis rose to 92 per cents. It is good result of the method of the neural network prognosis. The research is being continued.

• Details
• Contents