تجسم و مدل ارزیابی پویا از ساختار مالی شرکت ها با نقشه خود سازماندهی و رگرسیون بردار پشتیبانی

Prediction of financial bankruptcy has been a focus of considerable attention among both practitioners and researchers. However, most research in this area has ignored the non-stationary nature of corporate financial structures. Specifically, financial structures do not always present consistent statistical tests at each point of time, resulting in dynamic relationships between financial structures and their predictors. This characteristic of financial bankruptcy presents a significant challenge for any single artificial prediction technique. Therefore, this paper will propose a multi-phased and dynamic evaluation model of the corporate financial structure integrating both the self-organizing map (SOM) and support vector regression (SVR) techniques. In the 1st phase, the inputs to the SOM are financial indicators derived from listed companies’ public financial statements adopting the principle component analysis (PCA) to extract useful indicators with a strong influence that each year determines the company's position on the SOM. In addition, we used the SOM to visualize and cluster each corporate in the 2D map. We also investigated each cluster and classified them into healthy and bankrupt-prone ones based on their regions in visualizing the 2D map. In the 2nd phase, we drew the trajectory for the healthy and the bankrupt-prone companies for consecutive years in a 2D map. Therefore, several visualized and dynamic patterns of corporate behavior could be recognized. In the 3rd phase, we used the SVR method to forecast the future trend for corporate financial structure. In addition, this research also compared the hybrid SOM–SVR architecture with single SOM, SVR, and Learning Vector Quantization (LVQ) algorithms. The results showed that the proposed methodology outperformed the other methods in both prediction accuracy and ease of use.

Leading economists called the financial crisis of 2007–2010 the worst financial crisis since the Great Depression of the 1930s [6]. There were many causes leading up to this major economic shock in the U.S., such as the housing bubble, credit boom, sub-prime lending, predatory lending, and incorrect pricing of risk and collapse of the shadow banking system. All of these contributed to the failure of key businesses, a decline in consumer wealth estimated to be in the trillions of U.S. dollars, substantial financial commitments incurred by the U.S. governments, and a significant decline in economic activity [3]. Initially the companies affected were those directly involved in home construction and mortgage lending such as Northern Rock and Countrywide Financial, as they could no longer obtain financing through the credit markets. Over 100 mortgage lenders went bankrupt during 2007 and 2008. The crisis hit its peak in September and October 2008. Several major institutions either failed, were acquired under duress, or were taken over by the government, including included Lehman Brothers, Merrill Lynch, Fannie Mae, Freddie Mac, Washington Mutual, Wachovia, and AIG [2]. As a result, the U.S. economic crisis affected the global economy and put pressure on all the major sources of external revenue for developing countries, including exports, remittances, foreign direct investment, portfolio equity flows, and aid. In Taiwan, the bankruptcy of the Procomp Corp. and the Cdbank Corp. also caused a major upheaval in Taiwan's financial market, and any related investors incurred heavy losses. Today, financial distress and bankruptcy forecasting has become an increasingly important function with a significant impact on the lending decisions being made and the profitability of financial institutions. The prediction of failure of financial firms has been an extensively researched area since the late 1960s. A variety of statistical methods have been applied to solve the bankruptcy prediction problem for listed companies. Beaver introduced a univariate technique for the classification of firms into two groups using financial ratios [4]. These financial ratios were classified into six categories, including cash flow, net-income, debt to total-asset, liquid-asset to total-asset, liquid-asset to current debt, and turnover ratios. The data was collected from financial statements. Altman was the first researcher to use the multivariate discriminant analysis (MDA) and Z-score to predict the failures of firms in different industries [1]. As the result, the accuracy rate for determining a healthy company is 79% and 93.5% for one that will go bankrupt. Blum used the factor analysis to determine the critical financial ratios to construct the Failing Company Model [5]. Ohlson proposed a logistic regression (logit) model to determine the impact factors for 2058 healthy and 105 bankruptcy-prone companies [23]. The result showed that company size, financial structure, management performance, and asset liquidity influenced bankruptcy probability. More recently, Jones and Hensher compared the mixed logit model with the multinomial logit models for predicting firm distress [16]. Canbas et al. combined discriminant analysis, logistic regression, probit and principal component analysis (PCA) to implement an early warning system for bankruptcy prediction [7]. Doganay et al. integrated multiple regression, discriminant analysis, logit and probit to construct a financial distress prediction model [11]. Recently, many studies have demonstrated that artificial intelligence (AI) such as neural networks can be an alternative method for financial distress prediction [17], [24] and [25]. Current statistical and AI methods used in this field present some drawbacks. Real world applications of statistical methods are limited by strict assumptions such as linearity, normality, independence among predictor variables and pre-existing functional forms relating to the criterion variable and the predictor variable [14]. Researchers using AI fail to explain why they adopt particular financial ratios as inputs for neural networks, and fail to investigate the trajectory of healthy and bankrupt companies, resulting in a lack of visualized or dynamic patterns of corporate behavior for current and future trends. In addition, improving accuracy prediction is still the leading concern in the field [15]. Especially for stock market predictions, even a slight improvement on prediction accuracy could have a positive impact on investment profitability. Finally, a hybrid system for prediction and classification outperforms the traditional system [9], [13] and [22]. Tay and Cao found combining a self-organizing map (SOM) with support vector regression (SVR) can successfully predict financial time series and stock index activity [27]. Therefore, this paper proposes a model of financial distress prediction integrating the SOM and SVR techniques. The proposed model uses statistical methods to process feature selection and draw trajectories for healthy/bankrupt companies. Proper feature selection can reduced data complexity and improve prediction accuracy. The main objectives of this paper are to (1) apply the SOM technique to construct a visualization and dynamic evaluation model for corporate financial behavior, (2) use the SVR technique to improve the accuracy of financial distress predictions, and (3) provide investors and investment monitoring organizations with financial information and investment suggestions. The rest of this paper is organized as follows. A literature review of related studies is provided in Section 2. In Section 3, we discuss our research methodologies and pre-processing of our experiment materials. In Section 4, we use SOM to construct the static financial bankruptcy prediction model. Then in Section 5 we establish a trajectory using these experimental companies to construct a dynamic model. To prove the performance of the future trend prediction of our approach, we carried out a SVR with several experiments as described in Section 6. In Section 7, we draw our conclusions regarding financial distress forecasting and discuss some future work.

After the experiments we summarized four critical contributions. First, we found that when we several times applied all of the 13 financial ratios and 9 macroeconomic indexes with PCA into the SOM clustering models, we obtained a better total explained variance for the proposed model. This means that the model has a better prediction performance. In addition we found that the non-financial ratios and macroeconomic indexes did not have a direct impact on the occurrence of a bankruptcy. In this research we found that the primary reasons for financial distress were always the related financial ratios that went from bad to worse. Second, the SOM offers valuable new insights into the analysis of financial statements. The static SOM model makes it possible to distinguish different clusters of corporate financial structures in a 2D map. In addition, we also found that the more imminent the bankruptcy the more different clusters can be recognized. This seems to indicate that our proposed method is an easy-to-use visualization tool for analyzing financial statements. Third, the dynamic SOM model makes it possible to recognize different patterns of corporate behavior and to find the attributes associated with those patterns. A point worth noting is that in this research it is important to use information from several consecutive years. It seems that the information used in this study is suitable for the construction of trajectory maps. Also, trajectory patterns for corporate financial structure can be observed with the dynamic SOM model. Finally, applying the SVR for the prediction of improvement seems to be successful. The hybrid SOM–SVR architecture can better capture the financial characteristics of a company by decomposing the whole financial series into smaller homogenous regions. After decomposing the data, the SVR can better predict the financial indices. The results of this study show that the hybrid SOM–SVR architecture is better than the single SOM, SVR, and LVQ algorithms. Future research should further test the idea of the Growing Hierarchical SOM (GHSOM) architecture to improve the traditional SOM capability of handling huge datasets. In addition, our proposed methodology can also be applied to examine different financial statements or time series analysis.