本研究以台灣證券交易所上市公司為研究對象,採取1:1配對方式,選取於1995-2005年間112家曾發生財務危機事件之公司及112家正常公司,並以危機發生時點為基準,往前選取十季的財務比率各58項及公司監理變數各6項為解釋變數,利用過去文獻中,學者所採用的t檢定法、單變量檢定法(ANOVA)、Mann-Whiteney-Wilcoxon法(MWW)及倒傳遞類神經網路法,篩選出顯著變數,再以常用於財務危機預測模型研究之統計方法:區別分析法、Logit及類神經網路法來建構財務危機預警模式,藉以分析上述變數篩選法是否會影響財務危機預警模式預測的正確率,並比較何種變數篩選法篩選出之變數所建構的模式可得到較佳正確率。本研究結果發現。1.以t檢定及單變量檢定法(ANOVA)篩選變數法結合區別分析法及Logit法所建構之財務危機預警模型具有相同的預測效果。2. 三種財務危機預警模型之預測力,以類神經網路法的預測效果最佳,而以區別分析法之預測效果最差,且具最高的型II誤差。3.以危機前ㄧ季的資料所建構的財務危機模式具有最高的預測正確率。
In this study, we have constructed several models to predict financial distress by using empirical data from Taiwan Stock Exchange Corporation. There are 112 distress companies and another 112 well-performed companies were selected ranging from years 1995 to 2005. Fifty-eight financial ration and six corporate governance variables are chosen as explanatory variable. For each company, ten quarter’s data are collected before the financial distress occurred. Several variable selection methods were used to identify significant variables among those sixty four explanatory variables. There are T-test, One-Way ANONA, Mann-Whiteney-Wilcoxon and Back-propagation Neural Networks. After identifying significant variables, three methods were used to construct the financial distress prediction model. There are discriminate analysis, Logit analysis and Neural networks. By analyzing whether different variable extraction methods would affect the accuracy of financial distress prediction models and comparing which one is more accurate for predicting financial distress, there are three conclusions drawn from this thesis: 1. If we select variables by T-test or One-Way ANOVA, the accuracy of prediction is the same no matter to model constructed by discriminant analysis approach or by Logit approach; 2. Neural Network has a superior ability to predict the financial distress firms than other models and variable selection methods done. In contrast, the Discriminant analysis has the worst predicting accuracy and the highest Type II error. 3. If we construct the financial distress prediction model with data in last quarter, the accuracy of model is the highest.