基於資訊熵的粒化屬性於企業破產
預測之分類器研究

企業破產預測的研究顯示，其研究樣本資料通常是範圍很大的高維度數據集，本論文就此議題提出一解決方案。先使用獨立樣本T檢定進行條件屬性刪減，再以基於資訊熵的粒化演算法，將研究樣本具連續數值的條件屬性加以粒化，進一步試圖找尋最佳粒化數目，使得企業破產預測分類器仍能保持甚至提升預測正確率。預測模型之分類器選用倒傳遞神經網路(BPNN)與支援向量機(SVM)。研究樣本取自1996年至2005年間於中國上海與深圳掛牌共884上市企業，其中268家破產企業，另616家為正常企業。本研究將各條件屬性粒化為兩級(2Rank)或三級(3Rank)，分別進行模擬實驗測試。結果顯示BPNN分類器搭配2Rank及3Rank，分別可得出幾近相等之預測正確率84%、87%；而SVM分類器也分別可得出幾近相等之預測正確率89%、90%。再則，搭配粒化條件屬性之分類器所得預測正確率，對於分類器訓練資料集選取之隨機性的敏感度較小，此結果表示本研究所提出之搭配離散粒化條件屬性之分類器的強健性較佳。另外研究結果也顯示，先以獨立樣本T檢定做屬性刪減處理後，BPNN分類器預測正確率稍有提升，而SVM分類器則沒有改變，此結果表示先使用獨立樣本T檢定進行屬性刪減處理，除可消除冗餘的條件屬性，還可提升分類器預測正確率。

關鍵字

企業破產預測；資訊熵；粒化計算；支援向量機；強健性

並列摘要

Prior research of corporate failure prediction shows that the large scope with high dimensions characterized by the research sample makes the process of the performance in model prediction time-consuming. This study addresses this issue by proposing a method, which is processed as follow. First, use independent sample T-test to original conditional attributes for dimension reduction. Secondly, the information entropy-based granulation method is used to discretize conditional attributes with continuous numerical value of the research sample to maintain and even improve the prediction accuracy. The classifier of the prediction models includes back-propagation neural networks (BPNN) and support vector machine (SVM). The research sample contains 884 Chinese firms listing in Shanghai Stock Exchange and Shenzhen Stock Exchange during 1996 to 2005, and it includes 268 financial crisis firms and 616 normal firms. In this study, the conditional attributes are discretized into two (2Rank) and three (3Rank) level for test simulation. The results showed that model with BPNN classifier is able to draw the classification accuracy near to 84% and 87% for 2Rank and 3Rank respectively; while SVM classifier can be drawn nearly equal prediction accuracy of 89% and 90%. Other interesting results showed that the sensitivity of the classifier performance with the discrete granulation conditions to the random selection of the classifier training data set is relatively small, implying that the classifier with discrete granulating conditional attributes proposed in this study is quite robust. Furthermore, the simulation results indicate that BPNN classifier can reach as high as 90% performance which is slightly higher than the classifier without reduction of the attributes. This suggests that conditional attribute reduction prior to constructing the models has its efficacy on the failure prediction of the classifier models.