| Summary: | 碩士 === 華梵大學 === 資訊管理學系碩士班 === 95 === Financial Early warning system can not only help the management of the financial institutions but also diagnose their common operations. Since the early 1970s, many related researches have already made. However, most of them use traditional statistic ways to build the early warning system until recent years. Because of the vigorous development of the data mining techniques, many researches begin to apply those techniques to various fields also including early warning system. Data mining doesn’t need to satisfy many statistical antecedent assumptions and can transform enormous original data into meaningful and useful information.
To build the early warning system model, the related financial laws, data, and operation management rules need to be taken into consideration. However, the number of features is too large and not all of them are helpful to prediction. Data sets with unimportant, noisy or high correlated features will significant decrease the classification accuracy rate. By removing these features, the efficiency and accuracy rate can obtain a better result.
Back-propagation neural network (BPN), support vector machine (SVM) and decision tree (DT) are well-known data mining techniques, which can be applied to various fields and have higher classification ability. However, data mining techniques may suffer the problem of parameters settings. Bad parameter setting of data mining techniques will result worse accuracy rate. Therefore, this paper utilize one meta-heuristic, particle swarm optimization (PSO), to obtain suitable parameter optimization and select a subset of feature without degrade the classification accuracy rate. By the meta-heuristic global search characteristic, the parameters of BPN, SVM and DT can be optimized and the feature selection can be done at the same time to obtain the minimum set of features which can result in higher accuracy effectively.
In order to evaluate the proposed approach, this research taken the report of the Taiwan Ratings to be the authority. The “Condition and Performance of Domestic Banks” from the Central bank of China, Republic of China (Taiwan) and the “Statistics of Financial Institutions” from the Financial Supervisory commission, Executive Yuan are planed to be the source data. Banks will be classified as one of three categories ( ”well”, ”average”, and ”risky”).
In the experiment, although BPN and SVM have the high accuracy of forecast, the processes among them are black-box testing. Professionals can’t take these results into their future judgments. By the tree structure which was obtained from the proposed PSO+DT architecture, experts can obtain the best decision rules and thus make further evaluation and correction of our early warning system model. The experiment results shown that our proposed approaches can reduce unnecessarily features and improve classification accuracy significantly.
|
|---|
