台灣地區上市公司信用風險衡量與績效評估

The Evaluation and Examination of Credit Risk for the Companies Listed on TSE

本研究以2001年至2005年發生財務危機的國內上市公司為研究對象，並依據Beaver（1966）與Altman（1968）的準則，每一家危機公司配對兩家正常公司，選出危機樣本51家，配對樣本102家，而為了驗證財務預警模型之預測能力，將所有樣本分為訓練樣本與測試樣本，首先以因素分析法進行財務變數的篩選，模型所使用的變數除財務變數及非財務變數，並嘗試納入信用風險變數於財務預警模型中，探討納入違約機率後，是否能提升財務預警模型之預測能力，並比較何種財務預警模型加入違約機率後預測正確率較佳。依據本文實證結果分析，可得到以下三點結論：（1）類神經網路模型的預測正確率較羅吉斯迴歸模型差；（2）根據績效評估模型CAP曲線及ROC曲線的面積指標顯示，類神經網路模型的預測能力優於羅吉斯迴歸模型；（3）在類神經網路模型建構的兩種模式中，加入信用風險變數後，危機發生前一年的訓練樣本及測試樣本預測危機發生正確率皆會提升。因此，就企業採用之財務預警模型，建議企業可選擇類神經網路模型並加入信用風險變數做為財務危機之預測模型。

The purpose of this study is to evaluate if the credit risk could affect the accuracy of financial distress model. The objectives of this investigation are the listed companies on TSE which have financial distress between 2001 and 2005. According to Beaver (1966) and Altman (1968), each financial distress company allots two financial normal companies, and select 51 financial distress companies and 102 financial normal companies as samples. Then we separate the samples of training samples from test samples to examine the accuracy of financial distress model. First of all, we use the factor analysis to extract financial variables, and evaluate whether the accuracy of financial distress model may advance by adding the credit risk variables. Then the financial distress models are compared to investigation which is more accurate when adding probability of default. This investigation draws three conclusions: (1) The accuracy of ANN model is worse than Logit model; (2) According to the evaluation model, the area of CAP and ROC curve points that the accuracy of ANN model is better than Logit model; (3) By adding the credit risk variables to these two models, which are applied of ANN model, before the risk happens for previous one year, the accuracy of the training sample and testing sample are both increased. Therefore, to a financial distress model, the suggestion is that the companies may adopt ANN model adding a credit risk variables.