產業別財務危機預警模型建構之研究

本研究為建構財務危機預警模型，採用三業別（含半導體業、金融業、航運業）及三種研究模型（含Logistic迴歸分析法、Probit迴歸分析法、多元區別分析分析法），以觀察產業別、不同研究方法所產生之差異及實證結果之優劣，以供進行應用財務危機預警模型之參考。本研究資料，除財務比率變數（計25個）外，亦選用非財務變數，包括公司治理變數（計5個）、總體經濟變數（計4個），合計34個研究變數，致其考慮構面較為廣泛周全，以提高財務危機預警模型之估計確度。其中選取公司樣本數，採取「1：2 配對抽樣」，半導體業有48家（含危機公司16家、正常公司32家），金融業有24家（含危機公司8家、正常公司16家），航運業有9家（含危機公司3家、正常公司6家）。關於逐步迴歸分析，由於本研究變數高達34個，經共線性檢定及逐步迴歸分析，剔除具有共線性變數，並採用依逐步迴歸分析後所產生具有影響力之變數，才放進各模型進行實證結果分析。其中三業別被選用四次者，計有二個變數（含保留盈餘占資產總額比率、固定資產週轉率），選用三次者，計有三個變數（含稅後淨利率、董監事質押比率、應收帳款週轉率），另選用二次及一次者各有7個及13個變數。首先採用Logistic迴歸模型進行實證結果，半導體業前一年、前二年、前三年之正確率為91.61%、74.23%、71.49%；金融業前一年、前二年、前三年之正確率為85.07%、84.88%、89.50%；至於航運業前一年、前二年、前三年之正確率均為100%，顯示各業別估計結果為佳，大致以接近危機發生前一年之準確度更為突出。其次採用Probit迴歸模型估計結果，半導體業前一年至前三年之正確率為93.75%、87.50%、85.42%；金融業前一年至前三年之正確率為91.67%、87.50%、91.67%；至航運業前一年至前三年之正確率均為100%。顯示各業別之正確率頗高，尤其接近危機發生前一年，其正確率更高。再次採用多元區別分析模型進行實證結果，由於Logistic迴歸與Probit迴歸其因變數有二類（如危機公司、正常公司），而多元區別分析其因變數則有三類（危機公司、最優公司、次優公司）。經觀察多元區別分析估計結果之正確率，其中半導體業前一年至前三年之正確率各為79.2%、75.0%、62.5%；金融業前一年至前三年之正確率各為58.3%、41.7%、70.8%；而航運業同期各為70.8%、62.5%、79.2%。顯示各業別之正確率偏低，致採用此方法似不如預期。再觀察各業別三個模型正確度之比較，就半導體業而言，以Probit迴歸模型前一、二、三年之正確率最高（為93.8%、87.5%、85.4%）；其次大體上為Logistic迴歸模型（為91.6%、74.2%、71.5%），而多元區別分析第三（為79.2%、75.0%、62.5%）。若觀察金融業與航運業亦復如此，綜結此三個模型估計企業財務危機之正確率，係以Probit迴歸模型最佳，其次為Logistic迴歸模型，而多元區別分析模型居末，此亦與過去多數學者所做之研究結果相脗合。

關鍵字

財務危機；預警模型； Logistic迴歸； Probit迴歸；多元區別分析

並列摘要

In order to construct a financial crisis early warning model, this study adopts three industry categories (including semiconductor industry, financial industry, and shipping industry) and three research models (including Logistic regression analysis method, Probit regression analysis method, and multivariate difference analysis analysis method) to observe the industry. The differences between different research methods and the pros and cons of the empirical results can be used as a reference for the application of the financial crisis early warning model. In addition to financial ratio variables (25 in total), non-financial variables, including corporate governance variables (5 in total) and overall economic variables (4 in total), were selected for the research data, totaling 34 research variables. The scope is more extensive and comprehensive to improve the estimation accuracy of the financial crisis early warning model. Among them, the number of company samples is selected, and "1:2 pairing sampling" is adopted. There are 48 companies in the semiconductor industry (including 16 crisis companies and 32 normal companies), and 24 companies in the financial industry (including 8 crisis companies and 16 normal companies). , there are 9 shipping companies (including 3 crisis companies and 6 normal companies). Regarding the stepwise regression analysis, since there are as many as 34 variables in this study, after the collinearity test and stepwise regression analysis, the variables with collinearity are eliminated, and the influential variables generated by the stepwise regression analysis are used, and then they are put into each model for demonstration. Result analysis. Among them, if the three industries are selected four times, there are two variables (including the ratio of retained earnings to total assets and the turnover rate of fixed assets), and those selected three times, there are three variables (including after-tax net interest rate, directors and supervisors pledge ratio, Accounts Receivable Turnover Ratio), and there are 7 and 13 variables for secondary and primary respectively. First, the logistic regression model is used to carry out the empirical results. The correct rates of the semiconductor industry in the previous year, the first two years, and the first three years are 91.61%, 74.23%, and 71.49%; the financial industry is correct in the previous year, the first two years, and the first three years. The accuracy rates were 85.07%, 84.88%, and 89.50%; as for the shipping industry, the accuracy rates for the previous year, the first two years, and the first three years were all 100%, indicating that the estimated results of various industries were better, and were roughly close to the year before the crisis occurred. The accuracy is more prominent. Secondly, the Probit regression model was used to estimate the results. The accuracy rates of the semiconductor industry from the previous year to the first three years were 93.75%, 87.50%, and 85.42%; the accuracy rates of the financial industry from the previous year to the first three years were 91.67%, 87.50%, and 91.67%. %; the correct rate is 100% from the previous year to the previous three years in the shipping industry.Shows a high rate of correctness across industries, especially close to the year before the crisis. Again, the multivariate discriminant analysis model is used to carry out the empirical results. Because Logistic regression and Probit regression have two types of dependent variables (such as crisis companies and normal companies), and multivariate discriminant analysis has three types of dependent variables (crisis companies, optimal companies, second-class companies) excellent company). After observing the correct rate of the estimated results of the multivariate difference analysis, the correct rate of the semiconductor industry from the previous year to the previous three years was 79.2%, 75.0%, and 62.5% respectively; the correct rate of the financial industry from the previous year to the previous three years was 58.3%, 41.7%, 70.8%; while the shipping industry was 70.8%, 62.5%, and 79.2% in the same period. It shows that the accuracy rate of various industries is low, so the use of this method does not seem to be as expected. Looking at the comparison of the accuracy of the three models in each industry, as far as the semiconductor industry is concerned, the Probit regression model has the highest accuracy rates (93.8%, 87.5%, 85.4%) in the first one, two and three years; Logistic regression model (91.6%, 74.2%, 71.5%), and multivariate discriminant analysis third (79.2%, 75.0%, 62.5%). If we look at the financial industry and the shipping industry, the same is true. Summarizing the accuracy of the three models for estimating corporate financial crisis, the Probit regression model is the best, followed by the Logistic regression model, and the multivariate discriminant analysis model is the last. The results of the studies done by most scholars in the past are consistent with each other.