銀行信用評等模型之研究

近年來銀行不斷發生超貸、掏空、或資產管理不當，面對這些問題，銀行監理單位無法有效監督，造成需要付出巨額的社會成本，因此銀行信用評等顯得額外重要。本研究蒐集中華信用評等的資料共29家銀行，共計103個樣本，期間為民國91年到95年，84個訓練樣本，期間為民國91年到94年，19個測試樣本，期間為民國95年，以約略集合、支援向量機、類神經網路、決策樹和判別分析，分別建立銀行信用評等模型，讓決策者可以自由選擇適用的評等模型。實驗結果發現約略集合、支援向量機和類神經網路分類的正確率最高，達到89.47%，但約略集合可以產生決策規則找出銀行信用評等分類錯誤的原因並提早發現那些銀行財務比率是有問題；類神經網路可以提供每個變數的重要性；決策樹可以產生最精簡的規則；判別分析可以看出判別函數中的每個變數的顯著水準。

關鍵字

本國銀行信用評等；約略集合；支援向量機；類神經網路；決策樹；判別分析；決策規則

並列摘要

In recent years many banks often occur over loan,embezzlement or inappropriate asset management. Facing these problems, banking supervision cannot supervise these banks efficiently, so it takes a huge of social cost. Hence, the importance of bank credit rating is extraordinarily significant. This study collects 29 banks from Taiwain ratings corporation. There are 103 samples from 2002 to 2006. There are 84 training samples from 2002 to 2005. There are 19 testing samples from 2006. This study constructs bank credit rating model by using rough sets,support vector machine,neural network,decision tree and discriminant analysis. The empirical evidence shows that rough sets,support vector machine and neural network have the highest hit rate which is 89.47%. Using rough sets not only can derive decision rules but also can identify the reason why the banks are misclassified. Moreover, by using rough sets, it can find out the banks which have suspicious financial ratio in advance. Neural network can identify which financial ratio is important. Decision tree can derive most simplify rules. Discriminant analysis can observe each significant level of variable in discriminant function.