تبیین الگوی پیش بینی مدیریت سود با استفاده از ترکیب روش های یادگیری ماشین (مقاله علمی وزارت علوم)
درجه علمی: نشریه علمی (وزارت علوم)
آرشیو
چکیده
شناخت مدیریت سود برای استفاده کنندگان از اطلاعات حسابداری به دلیل ارزیابی عملکرد، پیش بینی سودآوری و تعیین ارزش واقعی شرکت بسیار حائز اهمیت است. هدف از این پژوهش ارائه مدلی جهت تشخیص مدیریت سود اقلام تعهدی و مدیریت سود واقعی از طریق ارزیابی عملکرد با استفاده از روش های یادگیری ماشین از جمله درخت تصمیم، ماشین بردار پشتیبان، k- نزدیک ترین همسایه، یادگیری عمیق و ترکیب آنها با روش انتخاب ویژگی ریلیف و تحلیل مؤلفه های اصلی است. برای دستیابی به این هدف، تعداد 180 شرکت پذیرفته شده در بورس تهران به عنوان نمونه آماری برای سال های 1389 تا 1400 انتخاب گردید. همچنین برای آزمون فرضیه ها از معیارهای میانگین صحت پیش بینی، خطاهای نوع اول و دوم استفاده گردید یافته های پژوهش بیانگر آن است عملکرد روش های پیش بینی مدیریت سود اقلام تعهدی بر اساس الگوی انتخاب ویژگی مبتنی بر ریلیف نسبت به الگوی انتخاب ویژگی مبتنی بر تحلیل مؤلفه های اصلی از توانای بهتری برخوردار است. این نتیجه در کلیه روش های پیش بینی مورد تایید قرار گرفت. اما نتایج برتری الگوی انتخاب ویژگی مبتنی بر ریلیف نسبت به الگوی انتخاب ویژگی مبتنی بر تحلیل مؤلفه های اصلی را در پیش بینی مدیریت سود واقعی نشان نداد. همچنین، یافته ها نشان دادند مدیریت سود اقلام تعهدی را می توان بادقت بالاتری نسبت به مدیریت سود واقعی پیش بینی کرد. نتایج پژوهش می تواند موردتوجه سرمایه گذاران، اعتباردهندگان، تحلیلگران مالی و حسابرسان قرار گیرد. استفاده از ترکیب روش های یادگیری ماشین، می تواند به شناسایی فعالیت های بالقوه مدیریت سود کمک کند.Explaining the Earnings Management Prediction Model Using the Hybrid of Machine Learning Methods
Knowledge of earnings management is essential for users of accounting information due to performance evaluation, profitability forecasting, and determining the true value of the company. The purpose of this research is to provide a model to diagnose accrual-based earnings management and real earnings management through performance evaluation of machine learning methods including decision tree, support vector machine, k-nearest neighbor, deep learning, and combining them with feature selection methods based on relief and principal component analysis. To achieve this goal, 180 companies admitted to the Tehran Stock Exchange were selected as a statistical sample from 2010 to 2021. Also, to test the hypotheses, the criteria of average accuracy and type I and type ΙΙ errors were used. The results show that the performance of accrual-based earnings management forecasting methods based on the relief-based feature selection model is better than the feature selection model based on principal component analysis. This result was confirmed in all prediction methods. However, the results did not show the superiority of the relief-based feature selection model over the principal component analysis-based feature selection model in predicting real earnings management. Also, the findings showed that accrual earnings management can be more accurately predicted than real earnings management. The research results can be of interest to investors, creditors, financial analysts, and auditors. Incorporating machine learning methods can help identify potential earnings management activities. IntroductionEarnings management can be described as the discretion utilized by managers to provide generally accepted accounting principles (GAAP)-based financial reports that can affect the relevance and reliability of the presented accounting information. EM can be performed either (1) through deviations from normal business practices to purposefully manipulate earnings; this is called real earnings management (Roychowdhury, 2006), and it affects cash flow from operating activities; or (b) by manipulating reported earnings through accruals, that is accrual-based earnings management, to achieve a suitable earnings figure. As a corporation's earnings are used by different financial statement users (such as shareholders, creditors, and financial analysts) to gauge its performance, detection of earnings management can be interesting and crucial for them. In this context, this study attempts to present prediction tools that aid in detecting earnings management activities. For this purpose, six machine learning methods have been discussed to predict earnings management.Methods & MaterialA sample of 180 companies listed on the Tehran Stock Exchange during the period 2010-2021 was selected for testing hypotheses. The performance of each machine learning method at predicting accrual-based earnings management and real earnings management was evaluated based on three criteria: type Ι error, type ΙΙ error, and average accurac.FindingThe results show that the performance of accrual-based earnings management forecasting methods based on the relief-based feature selection model is better than the feature selection model based on principal component analysis. This result was confirmed in all prediction methods. However, the results did not show the superiority of the relief-based feature selection model over the principal component analysis-based feature selection model in predicting real earnings management. Also, the findings showed that accrual earnings management can be more accurately predicted than real earnings management.Conclusion & ResultsEarnings management would affect accounting data, in particular, the earnings reported in accounting other than the actual earnings of an enterprise. Therefore, the prediction of earnings management is still an issue of great importance. The purpose of this research is to use machine learning methods such as decision tree, support vector machine, k-nearest neighbor, and deep learning to predict earnings management. Also, this research relies on feature selection to identify the most optimal features for use in the prediction model. Even though the determinants of earnings management have been studied for a long time, the ability of these factors to predict earnings management has received less attention.The results show that the combination of feature selection based on relief with each of the forecasting methods provides a more accurate performance for predicting accruals earnings management than the feature selection based on principal components analysis. However, the results did not show the superiority of the relief-based feature selection model over the principal component analysis-based feature selection model in predicting real earnings management. Also, the findings showed that accrual earnings management can be more accurately predicted than real earnings management. In addition, the results indicated that the most important features for prediction are related to the auditor's features in the first place and then to the features of the company's ownership structure. In other words, investors should pay a lot of attention to the features of the auditor and the ownership structure of companies in predicting earnings management. Based on the obtained results, the hybrid method based on deep learning and relief feature selection has the highest prediction accuracy (89/62) among other hybrid methods for forecasting accruals earnings management, and the hybrid method based on deep learning and principal component analysis feature selection has the highest prediction accuracy (82/65) among other hybrid methods for forecasting real earnings management.The findings of this study can expedite earnings management detection for financial statement users by improving earnings management prediction accuracy. These results may be applied to reduce investment risks and losses and increase investment benefits for investors and creditors if they are better able to predict misleading financial reports due to earnings management. In summary, the results of this study suggest tools to decision makers that help in predicting earnings management with relatively high accuracy.