Title: 模糊理論與類神經網路在企業信用評等之應用
Applying Fuzzy Theory and Neural Network in Business Rating
Authors: 李恩傑
En-Jie Li
Lee-Ing Tong
Keywords: 企業信用評等;類神經網路;模糊理論;business rating;neural network;fuzzy theory
Issue Date: 2001
Abstract: 企業在彼此貿易時以企業信用評等來降低交易的風險。然而要將傳統的數學模式應用在多等級判別上往往十分困難,僅有60%左右的正確率,這使得企業間的交易常處於不確定的高風險中。基於上述原因,本研究發展出一個具彈性且判別能力頗佳的模式,利用模糊理論與類神經網路結合企業信用評等與企業破產預測兩個領域。此模式在混合不同產業與企業規模資料中,分類的正確率亦可達到 80%以上。混合主觀判斷和客觀事實,此類神經網路使組織可以正確有效的評估一個企業的信用等級與可能破產的危機。
Enterprises trade with each other by carefully considering business ratings to reduce investment risks. However, conventional mathematic models have difficulty in discriminating between multiple ranks. A prediction accuracy of only 60% among departed models exposes traders to unnecessarily high risks. Based on the above, we should develop a flexible and accurate neural network structure that applies artificial intelligence in fuzzy theory to combine business ratings and bankruptcy prediction. The proposed model can discriminate multiple ranks with 80% accuracy in dissimilar industry and scales of enterprise. By incorporating subjective judgment and objective fact, the neural network structure proposed herein allows an organization to assesses the degree of risk and whether an enterprise will become insolvent. Most organizations evaluate business ratings subjectively, normally based on their professional knowledge and experience. Most enterprises also lack objective models capable of assessing trade partners’ credit actively. However, neural networks discriminate between multiple ranks inaccurately, and some samples that locate an ambiguous situation cannot be easily separated. For instance, while applying Back Propagation Network and Fuzzy Theory for multiple ranks of investment and two ranks of bankruptcy, prediction of bankruptcy can usefully reduce the risk in business rating alone. Combining different outputs in network structure can improve accuracy in multiple ranks and control the risk in limited area. Applying fictitious variables can enhance to discriminating multiple ranks effectively.
Appears in Collections:Thesis