A STUDY OF ARTIFICIAL NEURAL NETWORK APPROACH TO THE CONSTRUCTION OF CLASSIFICATION MODELS--APPLICATION ON FINANCIAL DATA

• Main Authors: SHIESH, JUNE-HORNG, 謝俊宏
• Other Authors: Ching-Tang Hsieh
• Format: Others
• Language: zh-TW
• Published: 1998
• Online Access: http://ndltd.ncl.edu.tw/handle/79639005391821241893

博士 === 淡江大學 === 管理科學研究所 === 86 === The Classification problem is a ubiquitous problem in business application.Traditionally,researchers interested in business classification applications have applied parametric methodsuch as multiple discriminant analysis, probit, logit and regression. Most of these statisticalmethods assume the data being used to be distributed Gaussian. In addition to assumptions ofthe distributions involved, statistical methods are restricted by the omitted variables, multi-collinearity, auto-correlation. Finally, assumption of regression functions to be linear orquadratic might induce additional bias in estimating parameters. Artificial Neural Network(ANN) represents a field of study within the ArtificialIntelligence area where researchers are studying a "biologically inspired" way of processinginformation. To this point, ANN have proven to be good at solving classification decision problems. The purpose of this thesis is to construct classification models based on ANNtechniques and financial applications are presented as case studies to measure modelsperformance. The major contributions of this thesis are the following: 1. We present a modified algorithm, which conjugates gradient and Newton method, to improve the gradient descent minimization technique for artificial neural network learning. The result shows the modified learning algorithm has better performance in terms of classification accuracy and convergence speed than the traditional back-propagation algorithms. 2. We present the technique of SOM as a classification tool. The reason is due to its capability such as dimensionality reduction, topological ordering of data and simple computation of its training algorithms. The result of our experimental study reveals that the SOM model is undoubtedly a good tool to support decision making in classification. It not only clusters companies properly in different regions according to certain financial ratios, but also provides more information than other techniques on the financial state of the company. 3. We present a new approach to solving classification problems by combining the prediction of a similarity-measure tool with the prediction of the perceptron to create an integrated classifier. The study indicates that the integrated classifier performs significantly better than either of the individual techniques. 4. We develop three hybrid models for classification. These models are : an MDA-based model, an ID3-based model and a SOFM-instar model. Both the MDA-based model and the ID3- based model are the neural network models operating with the input variables selected by the MDA and ID3. The SOFM- instar model combines a SOFM model (unsupervised learning), a LVQ model and a Grossberg-instar model. The results of our experimental studies reveal that these hybrid models are very promising neural network models for classification.