Please use this identifier to cite or link to this item: http://hdl.handle.net/11455/28861
標題: 台灣地區農會信用部金融預警評等系統之研究
A Study on the Financial Early Warning System by Rating for the Credit Department of Farmers'''''''' Associations in Taiwan
作者: 蔡碩倉
Tsai, Shuo-Tsang
關鍵字: financial early warning system;金融預警系統;business distress theory;portfolio theory;back-propagation neural network;multiple discriminant analysis;ordered-logit model;order-probit model;企業經營危機理論;投資組合理論;倒傳遞類神經網路;多變量區別分析;ordered-logit 模型;order-probit 模型
出版社: 農業經濟學系
摘要: 
台灣地區農會信用部係於「非股金制」的制度設計下,利用法令規範並搭配寬鬆優惠政策共築支撐架構,以發揮多元化組織功能,因此二者間之支撐平衡將直接影響農會信用部之經營良窳。一旦農會信用部之支撐體質一方失去平衡,其經營體質脆弱一面極易遭受打擊。另方面存款戶在面臨農會信用部金融危機時,由於營運資訊係處於不對稱情況下,並無法理性正確判讀營運狀況,連帶影響對農會信用部之金融信心。因此,建構一套可及早洞悉營運狀況之農會信用部金融預警系統,對解決目前普遍對農會信用部金融信心不足問題,具正面積極意義。
本文旨在結合企業經營危機理論與投資組合理論,並搭配倒傳遞類神經網路預警模型,量身裁製符合台灣地區農會信用部經營特性之金融預警系統,經由驗證三項待驗假說釐清幾點具體結論彙整如下:
1.藉由CAMEL指標描述台灣地區農會信用部經營輪廓,由於農會信用部非以獲利為唯一目標,故法令限制其業務經營範圍,促使低風險穩健經營,然因市場空間遭致壓縮,導致資產品質(A)普遍惡化。另基於「非股金制」資本累積不易,故政策寬鬆自有資本淨值要求,致資本適足性(C)嚴重不足。至於轉存農業行庫及免稅等優惠措施,配合農會信用部區域性經營優勢,使其在管理績效(M)、獲利性(E)及流動性(L)之優異表現可與金融同業相互抗衡。
2.農會信用部經營良窳係屬不同投資組合下之槓桿操作結果,而複雜的投資組合間存在程度上不同之抵換關係,須藉由金融預警系統綜合判定其營運等級。至於擠兌應以單純引發事件視之,並無法衍生為經營不善關係,否則將嚴重產生統計上型I與型II誤差,且任何農會信用部皆有引發擠兌的可能性。
3.本文結合企業經營危機理論與投資組合理論建構農會信用部金融預警系統理論模型,內容涵蓋農會信用部生存因素、經營投資組合、經營指標屬性、預警模型及農會信用部營運等級五大部份,其運作方式係農會信用部在面臨經營特性因素、競爭特性因素與經濟環境因素等生存因素觀點下發揮金融仲介功能,透過農會經營者的投資組合策略決定營運內容,而營運績效將藉由CAMELS經營指標予以揭露,最後再經預警模型綜合評等完成營運等級分類,俾發出營運等級訊號反饋至農會經營者,以決定是否重新調整投資組合策略,而此一診斷→評等→反饋之迴圈途徑,即表資料庫之擷取與累積過程,故足以構成「金融預警系統」之基本雛形。.
4.本文利用階層式集群分析將663家農會信用部樣本資料,依所選定之二十九項經營績效指標觀察值,「物以聚類」為五類評等等級,依其優劣排序分別為A、B、C、D、E。另以農會信用部經營類型為區分營運等級之區隔變數,則優劣排序分別為:「鄉村型」、「混合II型」、「都市型」、「混合I型」。
5.農會信用部於經營失敗過程中具有明顯的危機警訊可供金融預警系統事前偵測,蓋農會信用部經營失敗過程具有連續軌跡可供搜尋,而此連續過程亦代表不同營運評等等級之差異展現。若農會信用部經營投資組合具全方位表現,則常歸屬於A、B評等等級,而管理績效不彰為C評等等級之主要徵兆,接續資產品質開始惡化,連帶造成極易遭受市場風險影響,此一階段已進入D評等等級,一旦高風險槓桿操作仍無法獲利,則注定邁向失敗一途,此最後階段為E評等等級。
6.經比較四種預警模型之估計樣本預測準確率,以使用原始變數BPN中之模型I(Delta-Rule為學習法則,TanH為轉換函數)預測效率(100%)為最佳模型,其他依序為MDA(92.18%)、Ordered Logit(87.59%)、Ordered Probit(86.90%)。再利用事先預留之保留樣本鑑別預警模型實際預測能力,其結果與估計樣本雷同,以使用原始變數BPN中之模型I預測效率(93.86%)為最佳模型,其他依序為MDA(91.67%)、Ordered Logit(83.77%)、Ordered Probit(83.33%)。
7.為展現BPN於應用時無須任何先驗假設限制優勢,本文採用經因素分析處理後之十一項變數與原始二十九項變數兩組對照樣本進行實驗,結果未經因素分析處理之原始二十九項變數預測效率明顯優於另一組樣本,意謂應用BPN應多給予網路愈多訊息,以達學習訓練目的,反觀因素分析係利用萃取後之因素變數取代原始變數,以減少變數間之共線性問題,但連帶遺漏許多變數訊息,而影響BPN的學習訓練,進而降低預測效果。

Under the designed non-profit organizational structure , the credit department operation of farmers'''''''' associations in Taiwan has been supported by both special regulations and beneficiary re-depositing as to accomplish multi-dimensional organization functions. On one hand, the weak financial asset accumulation is easily under attack of financial crisis. On the other hand, the limited or asymmetry of information on internal operation would also loose confidence easily, As a result, the construction of an early warning system by rating operational performance for the credit departments of farmers'''''''' associations may reserve positive evaluation.
This study links the business distress and portfolio theories together with the back-propagation neural network on the financial early warning system by rating for the credit department of farmer'''''''' associations in Taiwan. The estimated results may be concluded as followings:
1. From the CAMEL indicators, most credit departments of farmers'''''''' associations gain the competitiveness on management , earning and liquidity under the special allowance of re-depositing savings and loose quality of assets and sufficiency of capital due to limited business by regulations and the non-profit restrictions.
2. The management performance of the credit departments are highly related to the trade-off the leverage operations on trade-off among different portfolios. The aggregated rating results on management performance may be derived by the early warning system. Cases of bank runs may be viewed as exceptional due to the weak linkage to operational distress. Other wise, Type-I and Type II errors may occur. More specifically, any credit department may fall into the situation of bank run.
3.Using business distress and portfolio theories, five main vectors including the survival vector , the portfolio vector, the management indicator feature, the early warning model and the management rating vector are constructed and performed in order. The management performance will be explored by the operational indicators CAMELS. Finally, the aggregated rating for performance classes are accomplished by the early warning model. The signal of rated classes are then feedback to business managers for the decision of portfolio adjustments. Such loop of diagnosis to rating to feedback represents data drawing and accentuation process which in turn become the basis of the financial early warring system.
4.Totally, 29 management performance indicators for 633 samples of credit departments during the 1995~1997 period are ranked from the best A through the worst E by using hierarchical cluster analysis. Accordingly, the operational classes from the best to the worst are also classified and ranked as "rural type", "mixture II", "city type", and "mixture I" in order.
5.It is argued that there are continuous tracks for detecting the operational failure of the credit departments through the surveillance system for financial institution . The continuous rating results represent signals of performance difference. For those will better and stable performance, the rating outcomes by year will locus on A or B levels. Those ranked C provides single of worse management performance and continuous asset quality deterioration risk become highly sensitive, the performance may be ranked D. Finally, once market the high-risk portfolio leverage operation may be failed the institution will be ranked E.
6.The predictability comparison provides the highest accuracy for BPN(93.86%) in the surveillance system, followed by MDA(91.67%), Ordered Logit(83.77%), and Ordered Probit (83.33%).
7.Comparing the predictability for the 29 original variables in BPN and the 11 variables determined by factor analysis, the originally selected 29 variables show better prediction efficiency. It implies that the application of BPN will need more learning information to gain training purpose. The selected 11 variables leaving out of collinearity provide less information for learning process in BPN which in turn lowering the prediction efficiency.
URI: http://hdl.handle.net/11455/28861
Appears in Collections:應用經濟學系

Show full item record
 

Google ScholarTM

Check


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.