類神經網路在銀行授信風險之應用

Abstract

自政府開放新銀行設立後，從此銀行界進入激烈競爭的時代。銀行槓 桿不若從前，利息收入直線下滑。大多數的銀行都開始擴展其貸款業務， 然而可能因此忽略了貸款的品質。此外，銀行在不景氣時期所給予的貸款 ，可能會面臨貸款者無力償還或要求展期的情況。如此將使得銀行的經營 環境變得更具風險與不確定性。過去的研究者嘗試運用統計方法如多變量 區別分析及 logistic 分析來預測銀行的授信風險，但是使用線性方程式 並不能有效解決分類問題。因此本研究的目的為（ 1 ）運用原始及經產 業調整過後的財務比率發展一套跨產業的模式來衡量貸款的違約風險。 （ 2 ）建構一最適類神經網路模式並與區別分析與logistic 分析就 分類及預測能力進行比較。 研究 樣本包含 1992 至 1995 年的 159 家違約企業， 依其產業別及資本額配 對 159 家正常企業。樣本分割為兩部份，訓練樣本含 232 家企業，測試 樣本包括 86 家企業。選用 23 個財務比率做為輸入變數，分別以統計及 類神經網路方法建構模式，並比較其區別和預測效果。 研究結果顯示，類神經網路無論在分類及測試效果上皆優於統計方法，並 且以相對產業比率作為模式輸入變數更適於建構跨產業模式。三層類神經 網路的分類及預測正確率分別是 95.45% 及 81.75%；兩層類神經網路是 86.67%、77.55%，logistic 分析是 84.76%、 76.03%，區別分析是 82.76%、72.33%。 The government has relaxed the restriction about the establishmentof new banks since 1992. There is a heated competition among thebanking. Then the banking can't use leverage as ever, and both therevenues of interest and the profits are decline. Most banks startincreasing the loan business, they may neglect the quality of loan. Inaddition, the debtors are involved with recession, then they may facecash insolvency and cash inadequacy. So the environment of bankoperation become more uncertain and risk y. The past researches usetraditional statistic methods like multiple discriminateanalysis(MDA), logistic analysis to predict the risk of bank credit.But using specific function can' t solve classification problems verywell. The purposes of the thesis are: (1)to use original andindustry-relative financial ratios to develop a cross industry modelfor measuring the debt violation risk; (2) to make a proper artificialneural network model to compare with MDA and logistic analysis aboutthe classified and p redictive ability. The sample of firms consisted of a total 318 firms, 159 of whichwent debt violation from 1992 to 1995 and 159 nonviol- ative frimsmatched on industry and capital. This sample was divided into- twogroups. The first subsameple of 232 firms data was used as thetraining set. The second subsameple consisted of 86 firms and was usedas the predictive set. The twenty-three financial ratios were selectedas input variables. Statistic and neural network methods are used tobuild model. We can compare the performanc es of term. The results of this research reveal the neural network performedbetter on both the correct classification and prediction, andindustry- relative ratios might be applicable to building a crossindustry model. The correct classification and prediction ofthree-layers neural network are separately 95.45% and 81.75%; oftwo-layers neural network are 86.67% and 77.55%; of logistic analysisare 84.76% and 76.03%; of MDA are 82.76% and 72.33%.