Applying SVMs to Predict Earnings Management and Analyze its Determinant Factors

• Main Authors: Feng-You Wei, 魏逢佑
• Other Authors: 陳國泰
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/06384513049540363067

碩士 === 國立臺灣大學 === 會計學研究所 === 97 === Corporate earning is a very important information for financial statements users. However, due to various reasons and motivations, corporate management might attempt to manipulate earning through certain methods or processes. To address the concern of earnings management, many studies have attempted to investigate its determinant factors. It is hoped that by discerning these factors, earnings management can be prevented or detected. Based on the 20 determinant factors found by pervious studies, this research applies support vector machines to build classifiers for earning management prediction. In addition, we adopt a process of features selection to filter out the most important factors. To validate the prediction power of the SVM classifier, we compare its prediction accuracy against that of the logistic regression model. We use the grid search technique with 5-fold cross-validation and perform features selection by filter model and hybrid model to analyze the data of Taiwan’s listed firms during the period of years 2003 to 2007. The experiment results show that a SVM classifier with only 6 determinant factors possess the highest prediction accuracy rate of 78.05%. These six factors are: performance threshold, market value/book value, extreme earning, cash flow from operations, debt ratio, and corporate size. By comparison, the step-wise logistic model with 11 factors has a prediction accuracy rate of just 66.84%.