標題: 運用類神經網路建構企業債信分類模式
The Research of Using Nrueal Netwrok in Corporate Credit Classification Model
作者: 王葆真
Wang, Po-Chen
巫永森
Yung-Sun Wu
管理科學系所
關鍵字: 企業債信分類模式;類神經網路;企業債信評等;倒傳遞網路;財務比率;區別分析;Corporate credit classification model;Neural network;Corporate credit rating;Back-propagation network;Financial ratio;Discriminant analysis
公開日期: 1996
摘要: 本研究主要在探討運用類神經網路於企業債信分類的情形。隨著國內 金融市場逐漸開放,直接融資趨勢的來臨,近年來公司債發行量大增,也 因此帶動企業債信評等的需求。然而,企業債信評等所使用的主觀評等程 序遭人質疑,因此有需多學者嘗試以較科學化的方法來進行企業債信評等 分類,包括統計方法及類神經網路。 本研 究以台灣經濟新報社之上市上櫃公司評等為基礎,選取二十個財務比率做 為模式的輸入變數,以類神經網路中的倒傳遞網路模式及統計中的區別分 析來建構企業債信分類模式。輸入變數分別採用二十個原始財務比率及七 個經因素分析過後之共同因素。網路架構方面則分別採用單一隱藏層及雙 隱藏層架構來建構類神經網路之企業債信分類模式。 研究結果發現 ,運用類神經網路於企業債信分類,其分類效果及穩定度皆比區別分析為 佳。 類神經網路模式之分類正確率達 75.01%,而區別分析模式只 有48.61%。其次,在類神經網路模式中,使用原始財務比率作為模式輸入 變數會比使用共同因素有較好的分類效果。最後,若以原始變數為模式的 輸入變數,則以雙隱藏層架構較佳;若以共同因素為模式的輸入變數,則 以單一隱藏層架構為最佳。 This research primarily discuss how to use neural network in corporatecredit classification. In recent years, Taiwan's financial market is graduallyopening. With the trend of direct financing, the volume of corporation bondsbeen released have greatly increased. There are increasing demand of creditrating to evaluate the risk of default. However, many people doubt about creditrating agencies' rating sequences, and many scholars try to use more scientificway to classify corporate credit using both stati stic and neural networkmethods. This research is based on the Taiwan Corporate Credit Risk Index. The 20financial ratios were selected as input variables of this model. Usingback-propagation network of neural network and discriminant analysis ofstatistic build the corporate credit classification model. This model separatelyuses original 20 financial ratios and 7 common factors as the input variables.About the structure of neural network, this research separately usesone-hidden- layer structure and two-hidden-layers structure. We can find that both the classification effect and reliability of neuralnetwork's corporate credit classification model are better than discriminantanalysis's. In this research, the classification rate of neural network is75.01%, but of discriminant analysis is only 48.61%. Second, in neural networkmodel, using original financial ratios as input variables have a betterclassification effect than using common factors after the factor analysis.Finally, the appreciate structure of this model is separatel y two-hidden- layerson original financial ratios and one-hidden-layer on common factors.
URI: http://140.113.39.130/cdrfb3/record/nctu/#NT850457036
http://hdl.handle.net/11536/62197
Appears in Collections:Thesis