Using Data Mining Techniques for Going Concern Prediction－Sentiment Analysis of Media

• Main Authors: Yi-Hsin Liao, 廖宜心
• Other Authors: 林嬋娟
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/6thfbm

碩士 === 國立臺灣大學 === 會計學研究所 === 107 === This study applies both data mining techniques（Random Forest and Support Vector Machine）and the traditional statistical method（i.e., Logistic Regression）to construct a going concern diagnostic model. This study also tries to assess the impact of media coverage on auditors’ going concern opinion by using sentiment word analysis. The empirical results show that all methods above achieve the best predictive performance after selecting key features from 72 variables. Among the three methods, Random Forest has the highest Recall value, about 0.92, indicating among all companies receiving going concern opinion, the prediction accuracy of Random Forest is over 90%. However, the empirical results from Random Forest show that financial and operational variables remain the most important factors considered by auditors while assessing the likelihood of going concern. The importance of variables related to corporate governance, audit quality, and media sentiment, however, is relatively low. Specifically, this study finds that the media information has no significant effect on the issuance of going concern audit opinion, whether samples are confined to the companies receiving going concern opinion in the first year or not. In sum, these results may suggest that media coverage adds little incremental value beyond the operating and financial information already considered by auditors in making going concern opinion decision. One possible reason might be that most of media content has been incorporated in the operating and financial information disclosed by clients.