本研究主要以發生財務危機之上市櫃公司為研究對象,將會計資訊、盈餘管理、公司治理及景氣循環等影響公司財務危機的四個不同構面變數投入,萃取公司財務危機前五年之數據為樣本資料,建構出11組不同信用評等早期預警模型。本研究樣本以2000年至2005年為訓練組樣本以建構預測模型之正確性,並將預測結果以檢驗2006年至2007年測試組樣本之穩定性,本研究預測方法則比較不同之條件隨機域(Conditional Random Field;CRF)、基因演算法(Genetic Algorithm;GA)、支撐向量機(Support Vector Machine;SVM)、倒傳遞網路法(Back Propagation Networks;BPN)、邏吉斯迴歸分析(Logistic Regression;Logit)、多變量區別分析(Multiple Discriminant Analysis;MDA)等並檢驗模型之預測正確率及穩定度。 研究結果發現:(1)、最佳信用評等預測模型是以包括會計資訊、盈餘管理、公司治理及景氣循環等構面變數為最優之預測能力模型;(2)、比較上述六種不同財務危機預測方法應用於信用評等預測模型加以檢驗,發現以條件隨機域(CRF)之預測正確率與穩定度最佳,其次依序為基因演算法、支撐向量機、倒傳遞網路、邏吉斯迴歸分析以及多變量區別分析;(3)、採用公司財務危機前五年之數據為樣本資料,在建構最佳信用評等預警系統模型方面,以財務危機發生前一年之財報資訊最能有效預測公司信評等級。
The study of this paper is to develop 11 Credit-Rating Forecasting Models based on the financial ratios、earnings management、corporate government and business cycle variables using the data 5 years prior to company bankruptcy to develop the early credit rating warning system . We collected the data from year 2000 to 2005 to build up the model and then to develop the prediction accuracy ability and using the data from year 2006 to 2007 to examine the stability of each various models. Through employing the 6 different methods of Conditional Random Field (CRF)、Genetic Algorithm (GA)、Support Vector Machine (SVM)、Back Propagation Networks (BPN)、Logistic Regression (Logit)、Multiple Discriminant Analysis (MDA), to forecasting the prediction accounting and test stability of our research model. The empirical results shows that the best model is the one including finance ratios、earnings management、corporate government and business cycles variables. We employed the 6 prediction methods mentioned above to test the research model. Among all the methods, the Conditional Random Field (CRF) indicates the best result and then the Genetic Algorithm (GA)、Support Vector Machines (SVM)、Back Propagation Networks (BPN)、Logistic Regression (Logit)、Multiple Discriminant Analysis (MDA). Regarding the best credit-rating system, we found that using the data one year prior to the company bankruptcy shows the best prediction accuracy and stability.