The Use Of Hybrid Manifold Learning In The Prediction Of Credit Assessment Forsamplingchinese Listed Companies

With the rapid development of science and technology combined with the extension of economic globalization, how toconduct an effective credit risk assessment is an essential issue in the modern financial field. Accurate risk assessment is especially important in bank loan industry, and even a small improvement for predicting default probability will bring to bank more additional profit. However, facing a huge client database, bank employees can hardly exert effective analysis and take advantage of it while data mining technology could provide a strong support of seeking for the rule of extant business data and developing the bank decision support system. In order to insure the high effect of data mining, a special preprocessing needs to be conduct before importing the original data to guarantee the good performance of data mining algorithm in front of huge data and high dimensionality. Surprisingly, manifold learning, as a new method of machine learning, satisfies the requirement for dimension reducing. For that reason, this paper proposes a hybrid manifold learning and data mining model to conduct the credit assessment research.The proposed credit assessment model based on manifold learning is shown as follows:(1) Exert the z-score standardized data preprocessing on thenonlinear financial data of 250 sampling A-share listed companies.(2) Conduct a feature extraction of financial data through Isometric Feature Mapping(ISOMAP) of classical manifold learning.(3) Import the extracted feature into Support Vector Machine(SVM) to classify data and predict credit risk for enterprises. In order to validate the effectiveness of proposed model, the performance of ‘PCA+SVM’, ‘LLE+SVM’, ‘SVM’ are compared with the proposed hybrid model of ‘ISOMAP+SVM’.(4) Make the clustering based on classification and do the credit ranking for company clusters to make corresponding loan strategy.This paper combines the qualitative and quantitative analysis, process the financial data by Matlab R2012 a and gets the conclusiosions as follows :(1) The result of data preprocessed by z-score normalized method is superior to the result without preprocessing.(2) Compared with ‘PCA+SVM’ and ‘LLE+SVM’, our hybrid manifold learning model for credit assessment not only has the best classi?cation rate, but also produces the lowest incidence ofType II errors, and is capable of achieving an improved predictive accuracy and of providing guidancefor decision makers to detect and prevent potential ?nancial crises in the early stages.(3) After the clustering based on classification,250 listed companies are divided into 7 clusters while we develop the credit ranking based on clustering results which contribute to credit risk assessment and make the corresponding credit strategy.(4) Prediction accuracy and clustering precision could be imporved by using manifold learning and PCA to reduce the dimensionality of nonlinear data while credit classification cost could be keeping lower.However, the performance of dimensionality reduction by ISOMAP and LLE is a little bit superior to PCA for nonlinear data.