公開及非公開發行公司財務危機之比較 -使用MDA、Logit、BPN、SVM及GA-SVM

To Compare the Publicly Traded and Non-Publicly Traded Distress Firms Using MDA、Logit、BPN、SVM and GA-SVM

洪信豪(Sin-Hao Hong)

指導教授 ： 林利萱

元智大學/管理學院/會計學系/碩士(2007年)

摘要

近年來在國內外發生了很多企業破產的事件，例如:美國的安隆案和台灣的力霸案等…這些大企業的倒閉，不僅僅只是影響了公司員工的生計問題，更影響了廣大的投資大眾。因此，財務危機預測或是企業的破產的相關研究，一直是國國內外學者所關心的議題。
本研究是藉由實數型基因遺傳演算法(Rea-valued Genetic Algorithm)結合支持向量機的模型(GA-SVM)，使能最佳化SVM二個參數值(C和δ2)，籍此提高SVM的預測準確度，並且與支持向量機(Support Vector Machine, SVM )，倒傳遞類神經網路(Back-Propagation Network, BPN)及傳統的統計方法(Logit and MDA)模型的預測準確性進行比較。實證結果顯示如下:
1.在公開發行公司與非公開發行公司這一部份中，不管在哪一個模型(GA-SVM, SVM, BPN, Logit及MDA )公開發行公司的預測凖測性皆大於非公開發行公司，原因可能是公開發行公司的財務結構相對於非公開發行公司要來得健全，而且非公司發行公司的財務報表也沒有經過會計師的查核。
2.在比較各模型的預測凖確性部份，GA-SVM模型不管在訓練組或是驗證組，都具有最高的預測準確性。
3.在比較各模型的預測穩定性部份，GA-SVM模型仍是這幾個模型(SVM, BPN, Logit及MDA)中最佳的模型。
4.由於本研究是採用財務比率做為研究的變數，而財務變數大多不符合常態分配，這有違MDA的基本假設，所以MDA模型的預測準確性是最差的。

關鍵字

倒傳遞類神經網路 ； 支持向量機 ； 基因演算法

並列摘要

In recent years, a lot of bankruptcy incidents have happened like Enron event and Taiwan has similar situation that takes place, for example: CHINA REBAR CO,LTD，not only influence on the companies’ employees but also thousands of investors, therefor, financial distress prediction relevant research, has been the popular subject which the scholar quite cares about both at home and abroad all the time.
The aim of this study is try to use Multiple Discriminant Analysis, (MDA), Logit, Back-Propagation Network, (BPN), Support Service Machine (SVM) and Genetic Algorithm (GA-SVM) to establish financial distress predictable model. After that, applying Training set and Validation set to compare the models prediction accuracy. The GA-SVM model was tested on the prediction of financial distress of Taiwan to compare its accuracy with those of other models on traditional statistics methods (such as MDA and Logit) and other Artificial Intelligence (AI) methods (BPN, SVM and GA-SVM). After empirical result indicated that GA-SVM model is performed highest accuracy than the prediction accuracy of other 4 models. The empirical findings from the research shows the in the following:
1.To compare with the publicly traded firms and non-publicly traded firms that the empirical results indicated the publicly traded firm prediction accuracy outperform non-publicly traded firms in all 5 different models (GA-SVM, SVM, BPN, Logit and MDA) from 2000 to 2004. The main reason may be the financial structure of non-publicly traded firms does not performed well better than publicly traded firms. Furthermore, the financial statement of the non-publicly traded firms has not been audited completely by the CPA.
2.To Evaluate the GA-SVM model using the real financial ratio data and compare it with other 4 models (SVM, BPN, Logit and MDA).The results shows that the GA-SVM model was effective in finding the optimal parameters of SVM, and that it improved the prediction of bankruptcy. The GA-SVM is more reliable than previously reported traditional models (MDA and Logit) and AI models (such as SVM and BPN) models. Restated the GA-SVM model, the most accurate predictive ability of financial distress whether applied to the training set or validation set. However, the results of this work demonstrate that the predictive accuracy of the GA-SVM in forecasting the financial distress of firms is significantly increased by optimizing its parameters and the experimental results also shows that the GA-SVM model performs the best among 5 models.
3.This study analyzes the stability on traditional models (MDA and Logit) and AI models (GA-SVM, SVM and BPN). Experimental results show that the GA-SVM model outperforms SVM, BPN, Logit and MDA.
4.Most financial rations did not satisfy the normality assumption for multivariate statistical models such as the MDA. Thus, MDA exhibited the worst predictive accuracy and the largest errors.

並列關鍵字

Back-Propagation Network(BPN) ； Support Vector Machine ； Genetic Algorithm

參考文獻

1.Altman, E. I. “Financial Ratios, Discriminant Analysis and Prediction of Corporate Bankruptcy,” Journal of Finance, Vol. 23, 1968, pp.589-609.

2.Beaver, W, “Financial Ratios as Predictors of Failure,” Journal of Accounting Research, Vol. 4, 1966, pp.77-102.

3.Beneish, M. D. and E Press, "Interrelation Among Events of Default,” Contemporary Accounting Research, 1995l, pp.57-84"

4.Blum M. “Failing Company Discriminated Analysis,” Journal of Accounting Research, Vol.12, 1974, pp.1-25.

5.Buxton, B., Trotter, M., Burbidge, R., Holden, S. “Drug design by machine learning: support vector machines for pharmaceutical data analysis,” Computer and Chemistry, Vol.26, 2001, pp.5-14.