基於改良式調和搜尋之支援向量分類模型於中國企業危機預測

AN IMPROVED HARMONY SEARCH-BASED SUPPORT VECTOR CLASSIFICATION MODEL FOR CHINA'S BUSINESS FAILURE PREDICTION

　　本研究使用支援向量機(Support Vector Machine, SVM)模型，來識別並預測企業危機(Business Failure, BF)問題。因進行SVM分類模型訓練時，存在會影響SVM分類正確率的待訂參數，傳統上大都使用格點搜尋或基因演算法，本研究使用另一新的最佳化搜尋法，稱改良式調和搜尋(Improved Harmony Search, IHS)演算法，來決定支援向量分類模型的參數，使支援向量機有較佳的分類正確率。並使用主成份分析(Principle Component Analysis, PCA)對自變數作因次縮減(Dimensionality Reduction, DR)，混合模型稱PCA-SVMIHS，並與傳統前饋多層感知器網路(Feedforward Multilayer Perceptron Networks, FMPN)模型做比較。本研究以1996至2005年間於中國上海與深圳掛牌上市公司，共884家，其中268家危機公司與616家正常公司，為本研究訓練與測試樣本，所得結果顯示，PCA-SVMIHS模型預測正確率優於FMPN模型，故本研究主要貢獻在提供一較佳之PCA-SVMIHS模型，可作為企業危機預測之另一可行模型。

　　In this paper, the support vector machine is combined with improved harmony search (IHS) and principle component analysis (PCA), by which the dimensionalities of independent variables are condensed, to construct a classification model for business failure prediction (BFP), so called PCA-SVMIHS model. The sample in this study includes 884 Chinese companies listed in Shanghai Stock Exchange or in Shenzhen Stock Exchange during 1996 to 2005, which contains 268 failure companies and 616 health companies. In comparison with the conventional feed-forward multilayer perceptron network (FMPN) model, the proposed PCA-SVMIHS model performs better not only results in a higher correct rate for predicting the failure of company, but also produce a higher AUC value. This empirical result shows that the PCA-SVMIHS model is superior to the FMPN model as an alternative to predict business failure.