在企業經營管理或個人投資風險評估中，若能有效利用財務報表分析，結合資料探勘(Data Mining)與機器學習(Machine Learning)的方式，預測台灣上市櫃公司是否發生財務危機問題，將極具參考價值。 運用資料探勘(Data Mining)技術與多分類器(Multiple Classifiers)組合機制建立財務預測(Financial Prediction)模型，針對台灣上市櫃公司是否發生財務危機做探討。培養在實務應用中，善用專業背景知識與實作技巧，以融合跨領域的研究內容。觀察到證交所定義的財務危機類型(紓困、跳票或經營疑慮等)各有其特殊意義，可依危機定義將危機類別切割成不同的子類別作探討，每個子類別對應於一種危機型態，其中紓困類與經營疑慮類為發生率最高者。初步以SVM (Support Vector Machine，支持向量機)演算法建構此兩專精型分類器(紓困型、繼續經營疑慮型)，並給予兩分類器不同的訓練資料集與特徵集，目的為建構各專精型分類器並有效預測該類財務危機。由於專精型分類器用於處理特定危機類型，預測結果方透過本研究提出的OR組合機制對危機與非危機類別做判別。 本論文貢獻在於提出新的多分類器OR組合機制用於財務預測領域，相較於傳統單一分類器，我們的多分類器架構除了能有效提升整體危機公司與非危機公司樣本的預測準確率，更能大幅降低危機公司誤判率，富有研究價值。 How to effectively predict financial distress is an important issue in corporate financial management. We use data mining and machine learning methodology to analysis financial statement or financial ratio. Traditional approaches usually formalize financial prediction problem as two-class problem, attempting to differentiate the financially distressed companies (the distressed class) from the normal companies (the non-distressed class). However, there are many factors contributing to a company’s financial crisis. Taiwan Stock Exchange Corporation (TWSE) defines several kinds of financial crisis which show distinct reason. This observation motivates us to further segment the distressed class into a few subclasses. Each subclass corresponds to one crisis type. We propose new methods to design multiple classifier system. Each classifier for a subclass gives a meaningful training set and feature set. It makes that each classifier is professional for each sub-problem. This model is different from the existing approaches that each classifier is not designed for the same pattern recognition problem. The prediction accuracy is superior to traditional approaches by using our prediction model.
