國內外企業破產事件頻傳,財務危機預測與企業破產的相關研究,一直是國內外學者所關心的議題,近期學者運用了各種人工智慧方法建構危機預測模型。但為建構適合台灣地區企業財務危機預警模型和比較各種危機預測方法之準確性與穩定性,本研究以台灣上市櫃公司財務比率資料,分別以區別分析法(Multiple Discriminate Analysis,MDA)、羅吉斯法(Logistic regression model,Logit )、倒傳遞類神經法(Back-Propagation Network,BPN)、支向機(Support Vector Machine, SVM)、基因支向機(Genetic Algorithms-Support Vector Machine , GA-SVM)及人工免疫系統法(Artificial Immune Systems,AIS)六個方法來建構財務危機預警模型,期望找出最適合台灣上市櫃公司的財務危機預警模型。本研究將樣本區分為訓練與測試樣本。將88年至93年之樣本作為訓練樣本,以建立預測模型; 94年至96年之樣本作為測試樣本,以檢驗訓練樣本預測的準確度與模型的穩定度。 研究結果發現(1)人工免疫系統法的參數初始ARB值為 6、親和力門檻值為1.5、學習循環設定為 300、衰退率為1%時,人工免疫系統法所建立的模型能夠建構最佳的預測能力;(2)本研究將人工免疫系統法的變數區分為二類型,第一類型使用全部變數投入,第二類型經由區別分析逐步迴歸方法所萃取顯著變數再進行人工免疫系統模式的分類,驗證二種不同類型的變數投入孰優孰劣。實證結果顯示,使用第二類型變數投入的AIS方法預測準確性及穩定性皆優於第一類型;(3)Izan (1984)、Platt & Platt (1990)及LH Lin (2002)利用產業平均值來建構穩定的財務危機模型,AIS模型之穩定性與前述學者相當接近;(4)在比較六種研究方法所建立的財務危機預警模型中,人工免疫系統法(AIS)的預測準確性優於區別分析法(MDA)、羅吉斯模型(Logit)及倒傳遞類神經法(BPN);(5) 人工免疫系統(AIS)模型穩定度優於區別分析法(MDA)、羅吉斯模型(Logit)、倒傳遞類神經法(BPN)、支向機(SVM)及基因支向機(GA-SVM)。
A lot of bankruptcy incidents have happened. Therefore, financial distress prediction relevant research has been the popular subject which the scholar quite cares about both at home and abroad all the time. Recently, a lot of stud ties used artificial intelligence (AI) methods to set up the model. The aim of this study is try to use Multiple Discriminate Analysis, (MDA), Logistic regression model (Logit), Back-Propagation Network, (BPN), Support Service Machine (SVM), Genetic Algorithm-Support Service Machine (GA-SVM) and Artificial Immune Systems (AIS) to establish financial distress predictable model. For building up Taiwan bankruptcy prediction model, the data set is arbitrarily split into two subsets: about 2/3 of the data is used for a training set (1999 to 2004 ) and 1/3 for a validation set (from 2005 to 2007), comparing the prediction accuracy with the firms from 1999 to 2006 dataset. The empirical findings from the research shows that (1) As initial ARB value setting up for 6, affinity threshold value setting up for 1.5, and study circulation setting up for 300, the prediction accuracy of the artificial immune system model is the best. (2) When using extract parameters in the artificial immune system model, the predictive accuracy and the stability will be better than using all parameters. (3) Izan (1984), Platt & Platt (1990), and LH Lin (2002) used Industry Relative Ratio to building stable failure prediction models. The stability of the AIS model is as well as theirs. (4) The predictive accuracy of the AIS model is better than BPN, Logit and MDA models, but less than SVM and GA-SVM. (5)The stability of the AIS model is better than MDA, Logit, BPN, SVM, and GA-SVM models.