財務危機預測問題(Financial distressed prediction problem)一直以來是個重要且已被廣泛討論的問題，其中又以特徵挑選及學習演算法為兩大重心。本研究著重於找尋新的特徵以幫助預測，過往的研究大多使用財務比率(Financial Ratio)，部分使用公司治理指標(Corporate Government Indicator)進行財無危機預測，卻少有研究使用公司的負面新聞對台灣地區的公司進行未來的財務危機預測，在本研究中我使用TEJ的看門狗資料庫中所蒐集並定義的負面新聞事件分類，接著使用統計方法分析後挑選出了其中八個負面新聞事件，提取欲預測年份的前一年的發生次數作為特徵值去建模在透過DET Curve及cost ratios分析，並證實了在大部分的cost ratio 下使用ensemble Bagged Tree建模這些負面事件對預測表現是有幫助的。;The financial distressed prediction problem has always been an important and widely discussed issue, with feature selection and learning algorithms as the two main focuses. This study focuses on finding new features that can help improve the prediction. Most of the previous studies used the Financial Ratio, and some used the Corporate Government Indicator for financial crisis prediction. However, few studies used the company′s negative news to predict the financial crisis. In this study, we proposed eight negative news events to build the prediction model. For each event, calculate the number of occurrences of the year before predict year as the event feature value. After we analysis the result by DET Curve and different cost ratio analysis , we learned that these negative events are helpful for predicting performance over most of the cost ratios.
