Predicting Earnings Management level By Using Artificial Neural Networks

Document Type : Research Paper

Authors

Abstract

Journal of Accounting Advances (J.A.A)
Vol. 5, No. 1, 2013, Ser. 64/3
 
 
Extended Abstract
 
Predicting Earnings Management level By Using Artificial Neural Networks
 




Dr.   G. Kordestani
Imam Khomeini International University


J.   Masomi
Hazratee   Masoumeh University (HMU)


V.   Baghaee
 




 
Introduction
Earning management is one of the most fascinating and controversial issues in accounting and finance that has attracted the attention of many researchers due to the existence of contradictions, deceptions, secrecy and a sense of uncertainty. Interest rates in financial statements have been the subject of interest to investors, creditors, employers, financial analysts, customers and suppliers of materials who make decisions based on such data. Thus presentation of timely and reliable financial information is to the benefit of the users. However, this provides an opportunity for managers to mislead users by manipulating earnings in order to achieve their goals. For example, when managers' bonuses are tied to their financial performance, they have plenty of incentives to manage earnings to meet analysts' expectations financially. Most studies of the earnings management literature have examined the factors that significantly affect the level of earning management, but these variables are not used directly to predict the level of earnings management. There is one only possible correlation between these variables and the earning management. Therefore, a model designed to predict the level of earnings management to reduce risks of financial crisis management and the profits to help investors, creditors and other users of financial information, to avoid suffering major losses in capital market seems to be necessary.
One of the most practical methods in modelling and forecasting is neural networks which have many applications in the field of finance. Given the benefits of the neural network to predict the level of earnings management, capital markets will help users of financial statements. In this study, using a neural network multilayer perception (MLP) model for a level of earnings management is offered in Tehran Stock Exchange.
 
Design, test and neural network results
Classification of earnings management
In order to predict the level of earnings management by different levels of artificial neural network, first we must define it. Two observations follow from beginning to end to manage the surge in profits and profit sharp reductions in our classified data. The next classification is large increase profits and decrease of high-profit management, and finally the last category, the lowest level of earnings management, is earnings minor management.
 
The artificial neural network structure     
20 networks with different parameters examined in this study are to determine the optimal neural network to predict the level of earnings management. Table 2 shows these 20 networks with the desired parameters.
 
Table 2: Parameters of neural network





The number of hidden layer neurons


The   number of epochs




1000


1500


2000


2500




8


M1


M6


M11


M16




12


M2


M7


M12


M17




16


M3


M8


M13


M18




24


M4


M9


M14


M19




32


M5


M10


M15


M20





 
Also according to studies carried out in the hidden layer tangent sigmoid transfer function and the output layer of linear transfer function are used. The network training was done using error back propagation algorithm and Markwardt-Levenbery.
 
Neural network training and test results
M14 network with accuracy of 0.94 percent in the training phase and 0.69 percent in the test phase have the best performance. As a result, the network with 24 neurons in the hidden layer with 2000 epoch of the process is chosen as the best-trained network to predict the level of earnings management in Tehran Stock Exchange.
 
Conclusion 
This paper is a preliminary study to predict the level of earning management that uses neural network models. These predictive models are useful for users of financial statements who make decisions depending on the earnings amounts to avoid suffering a great loss. In addition, predicting level of earnings management in advance is a new application for neural networks. In this research we used three layers perceptron (MLP) with error back propagation algorithm. Network parameters are determined with continued testing. As result, a neural network with accuracy of 0.94 percent in the training phase and 0.69 percent in the test phase has the best performance for predicting of earning management level.
  
 
 

Keywords