یک روش داده کاوی برای پیش بینی شکست شرکت های بزرگ

This paper uses a data mining approach to the prediction of corporate failure. Initially, we use four single classifiers — discriminant analysis, logistic regression, neural networks and C5.0 — each based on two feature selection methods for predicting corporate failure. Of the two feature selection methods — human judgement based on financial theory and ANOVA statistical method — we found the ANOVA method performs better than the human judgement method in all classifiers except discriminant analysis. Among the individual classifiers, decision trees and neural networks were found to provide better results. Finally, a hybrid method that combines the best features of several classification models is developed to increase the prediction performance. The empirical tests show that such a hybrid method produces higher prediction accuracy than individual classifiers.

Due to recent changes in the world economy and as more firms, large or small, seem to fail now more than ever corporate failure prediction is of increasing importance. Corporate failure prediction is not only an interesting but also a challenging problem that has led to several studies over the past four decades. Numerous statistical classifiers have been constructed for the prediction of corporate failure. The main techniques used include discriminant analysis (DA) [1] and [28], logistic regression (LG) [19], [23] and [32], probit analysis [9] and [27], mathematical programming [7], expert systems [15], artificial neural networks (NN) [2] and [30], decision tree method [14], rough sets [11] and multicriteria decision aid [18]. In this paper, we will use the data mining approach, which is a systematic approach to find hidden patterns, trends and relationships in data and sometimes we refer to it as knowledge discovery. Altman [1], Argenti [4] and Lincoln [22] argue that corporate failure is not an instantaneous occurrence but that it is a process which evolves over a considerable period of time. Companies do not fail overnight. Since corporate failure evolves over a considerable period of time, this gives us the foundation for predicting corporate failure. According to UK 1985 Act (Section 228): ‘every balance sheet and profit and loss account shall give a true and fair view’, so the financial statement data are a set of facts which should include extensive warning signals pointing toward failures. Using these facts, we can extract the discovery of patterns or rules describing a firm's financial status; such patterns or rules potentially lead to decision support using methods such as classification and forecasting.

In this study, we use two feature selection method to choose independent input variables. The first one is human judgement based on financial theory, which has been used by many researchers before, the second one is the ANOVA statistical method. For all the models, DA, LG, NN, C5.0 and hybrid classifiers, we found the ANOVA feature selection is better than human judgement feature selection except for DA. The reason may be up to now there is no complete integrated theory of corporate failure that exists. So using the ANOVA computing method to select features gives a better performance. Also, we analysed the performance of several models developed for the problem of corporate failure prediction. The machine learning methods (NN and decision trees) show better performance than the statistical approach. We have suggested the hybrid algorithm that combines statistical and machine learning methods. In this algorithm, DA, LG, NN and decision tree C5.0 can be used as the base classifier to produce a hybrid classifier. The hybrid 1 classifier is the combination of four classifiers: DA, LG, NN and C5.0. The hybrid 2 classifier is the combination of three classifiers: DA, NN and C5.0 and the hybrid 3 classifier is the combination of two classifiers: LG and C5.0. We used real world data as training samples plus these three different hybrid classifiers to show its effectiveness. The main finding is that the hybrid methods produce better results in the prediction of corporate failure at 1 year before failure.