摘要 國內未上市公司由於財務報表透明性不高,在授信上銀行以擔保品為主,以致呆帳率居高不下影響銀行授信品質及投資大眾權益,本文主要目的在建構國內未上市公司財務危機預測模型,藉由模型預測以減少授信判斷錯誤。 本文以支向機(Support Vector Machines, SVM)與GMDH(Group Method of Data Handling)作為建構財務危機模式,以國內90年至93年360家未上市公司財務危機前一年之正常240個公司與危機120個公司分別以2:1作為研究樣本,並從模型中瞭解在不同訓練樣本數量下,驗證測試樣本之預測正確率及穩定性。 實證結果: 一、在建構企業財務危機預警模型,GMDH模型利用不同臨界值比較找出最佳預測能力;支向機利用不同參數C、γ值找出最佳預測正確率,在相同樣本及財務指標研究下,支向機與GMDH所建構財務危機預警模型比較,無論訓練樣本及測試樣本支向機預測能力及穩定性要比GMDH模式佳。 二、在驗證測試樣本預測正確率方面,支向機、GMDH模型,兩者模型均呈過度適配(Overfitting)現象。
ABSTRACT The bad debt rate is very high in Taiwan unlisted companies which will impact the bank loan quality and public investors’ right. The reason is that the financial information of the unlisted companies is not transparent enough and the bank loan mainly relies on the collateral only. The objective of this article is to establish a bankruptcy prediction model to reduce the judgment mistakes in the credit loan. This article is based on SVM (Support Vector Machines) and GMDH(Group Method of Data Handling)to establish the financial bankruptcy prediction model. The research sample is based on 360 unlisted companies in Taiwan during year 2001-2004 including 240 normal companies and 120 bankrupted companies. The analysis data uses the financial figures of one year before the bankruptcy. The research is also to understand the accuracy and the satiability of the holdout data in the prediction model for different number of the training data. The experiment results show that: 1. In building the bankruptcy prediction model, GMDH model get the best prediction rate by using cut-off values. SVM model gets the best prediction rate by different keneral function parameters C、γ. In the same sample and financial indicators, the prediction ability and the satiability of SVM is better than GMDH in training data and holdout data. 2. For the prediction accuracy of the holdout data, SVM and GMDH are both overfitting in the model. .