本文對於財務危機模型的建構,有別於傳統危機預測模型,以往變數資料多以連續變數表達,本研究將變數的型態分為連續變數和經過K-means分群法之離散化變數,探討在不同資料型態類神經網路和決策樹之最適變數呈現方式;為了尋求危機模型之最佳變數,使用相關性特徵選擇和支援向量機遞迴特徵消去方法進行重要變數的選取,提供給社會大眾從最適變數提早了解公司的財務狀況;最後將變數精簡下最好的結果以連續變數和離散化變數分別進行兩組類神經網路模型訓練,利用不同的人工神經元做運算處理,再合併整合演算,以組合式類神經網路模型建構財務預警模型,並與類神經網路、決策樹和TEJ信用評等相互比較,發展出最適模型。 研究結果顯示: (1)決策樹模型都能處理連續變數或K-means分群法之離散化變數;類神經網路資料呈現方式連續變數優於離散化變數。(2)變數選取下支援向量機遞迴特徵消去方法比相關性特徵選擇較佳。(3)組合式類神經網路的整體預測率比傳統類神經網路和決策樹更能提高其模型的預測準確性,預測準確率高於96.97%;並將組合式類神經網路和TEJ信用評等相比較,信用評等為7整體預測率為83.33%,誤判成本率隨著成本率的增加而高低起伏,而組合式類神經網路之誤判成本變化幅度較小,故組合式類神經網路為預測能力最佳之模型。
This research work employs an efficient neural network composition framework combined with two different feature selection methodologies to construct a financial crisis prediction model and evaluate its performance. The selection of important variables for the composition framework includes Correlation Based Feature Selection and Support Vector Machine Recursive Feature Elimination. This model separates input variables of neural network into two categories comprising continuous variables and discrete variables. Continuous variables are then divided into the required k groups by means of K-means analysis. Each category of data has been trained by an independent neural network respectively. The outcomes of both trained networks are combined and integrated by another specific neural network. The complexity of the composition framework is subject to the categories of the input data. The experiment results are shown as follows: (1) The process of variable discretization or continuous can improve the performance and construction of the decision tree model. (2) The support vector machine recursive feature elimination is superior to the correlation-based feature selection. (3) The overall prediction accuracy for the composition framework of neural network outperforms the decision tree and other neural network approaches.