An Application of Data Mining Technique to Build the Model of Credit Rating for Banking Industry—Integration of Genetic Algorithm and Neural Networks

• Main Authors: SHENG, SHIH-CHUAN, 盛世娟
• Other Authors: GOO, YEONG-JIA
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/61369980163925486951

碩士 === 國立臺北大學 === 企業管理學系 === 92 === The bank is an important intermediary role in financial system, it gather public deposit to raise capital efficiency, for this reason the public are sensitive for bank credit. Hence, this research will build the model of credit rating for banking industry. It helps the public to choose their own banks based on their risk preference. The government authority can also refer this study to increase financial audit frequency with lower rating banks and make effective distribution resources. This research is application of SAS-IML program to design the model that integrated the genetic algorithm with neural networks. The empirical results are as the followings : 1.Through Genetic Algorithm: after 600 generations, the optimal independent variable combinations are “deposit market share ratio, loan/deposit ratio, return on assets, interest rate sensitive gap over net worth ratio, capital adequacy ratio, operating income per person, operating income ratio, operating profit ratio, unearned interest revenue ratio,”in which, there are two intelligence capital ratio, therefore the banking should be to pay much attention to intelligence capital concept. 2.In the model of credit rating internal and external validity, the average hit ratio to be up to 98.48% and 87.87% separately, hence they should be able to prove have superior explain and forecast ability. 3.According to research result, the performance excellence banking have government to regard as prop, external support ability is strong, or is the leadership in banking, to lie in economic scale, moreover the performance inferior banking is established after1991, profitability and constitution is not healthy, and market share is lower. 4.In accordance with actual rating grade, the research compare forecast value (network calculate rating) to actual value (taiwan rating), the error condition modulus to be smaller than two, the hit ratio to be up to 85.85%, it can be prove the model accuracy superior even didn’t blend the rating grade.