توسعه مدل پيش‌بيني دستكاري سود با روش تركيبي شبكه عصبي و الگوريتم هاي كيهان شناسي

Development of Earning Manipulation Prediction Model Applying Hybrid Neural Network and Cosmology Based Algorithms

افزايش ميزان تقلب در گزارشگري مالي در سال‌هاي اخير باعث عدم اعتماد سرمايه‌گذاران به بازار‌هاي سرمايه شده است. به همين علت پيش‌بيني دقيق مديريت سود يكي از نيازهاي اساسي براي تصميم‌گيري استفاده‌كنندگان صورت‌هاي مالي است. اين پژوهش با ارائه يك مدل توسعه‌يافته بر مبناي مدل بنيش با تأكيد بر متغير‌هاي نظام راهبري شركتي مشتمل بر ساختار كميته حسابرسي، بازرس قانوني و حسابرس مستقل، ساختار هيئت‌مديره و ساختار مالكيت شركتي سعي درافزايش دقت پيش‌بيني مديريت سود دارد. داده‌هاي 81 شركت پذيرفته‌شده در بورس اوراق بهادار تهران طي سال‌هاي 1397-1391 با روش تركيبي شبكه عصبي پرسپترون چندلايه و الگوريتم‌هاي كيهان‌شناسي سياه­‌چاله، مه­‌بانگ – مه‌رمب و ازدحام ذرات كهكشاني مورد تحليل قرارگرفته است. دقت پيش‌بيني مدل با روش تركيبي شبكه عصبي مصنوعي و الگوريتم‌­هاي كيهان‌­شناسي ازدحام ذرات كهكشاني، مه­‌بانگ – مه­‌رمب و سياه­‌چاله به ترتيب از 59/08 ، 63/49 و 57/5 درصد به 87/3، 79/72 و 74/25 درصد افزايش پيداكرده است كه حاكي از بهبود قدرت پيش­‌بيني مدل در كشف شركت‌­هاي دستكاري كننده سود است. در مدل پيشنهادي سطح زير منحني مربوط به دو الگوريتم سياه‌­چاله ومه­‌بانگ-مه‌رمب، توسط الگوريتم ازدحام ذرات كهكشاني احاطه‌شده است ونشان دهنده اين است كه خطاي شبكه عصبي پرسپترون چند­لايه تركيبي با الگوريتم ازدحام ذرات كهكشاني به 12/7 درصد كاهش‌يافته است.

1- INTRODUCTION Accurately predicting earning manipulation in order to detect and identify manipulation of financial statements has always been one of the most fundamental challenges ahead of financial reports users. Because of increasing financial reporting fraud, this fact resulted in investor distrust of capital markets in recent years. The purpose of this study is to answer the questions whether it is possible to detect earning manipulation in financial statements based on the Beneish model? is it possible to detect earning manipulation in financial statements based on the proposed model? and does the proposed model predict better than the Beneish model in detecting earning manipulation? The findings of this research can be concidered by investors, creditors, auditors, regulators and other users to help them in making decisions and offering appropriate solutions. In order to detect earning manipulation and enhance the predictive accuracy of the earning manipulation model, it has been designed and presented a developed model based on Beneish model (1999). this study uses corporate governance variables i.e., audit committee structure, legal inspector and independent auditor, board of director's structure and corporate ownership structure requirements. 2- THEORETICAL FRAMEWORK  Benish (1999) investigated 74 earning-manipulator companies applying probit analysis during 1982-1992. He assigned the number 1 to the manipulative companies and the number zero to the non-manipulative companies and calculated the coefficients of the independent variables. The cut-off point of this model was -1.78. Therefore, if the M-score is greater than -1.78, it is likely that the company is earning manipulator. The overall accuracy of the model was confirmed at 76%. By using eight accounting variables in its model, Beneish showed that the probability of earning manipulation increases with unusual increase in receivables, decrease in gross profit margin, decrease in asset quality, sales growth and increase in accruals. But what is hidden from view in this model, is the attention to the control of mechanism to reduce transaction and agency costs. Studies conducted to develop the Beneish model have also been based solely on accounting data and have ignored the implications of control mechanisms in model development. Therefore, in order to improve the predictive power of the beneish model, corporate governance system can be considered as a deterrent factor from earning manipulation. 3- METHODOLOGY  The data of this study are drawn from the annual financial statements and reports of a sample of 81 non-financial listed companies on TSE over the period 2012-2018 i.e., 567 firm-year observations and analyzed by the hybrid multi-layer perceptron (MLP) neural network and cosmology based algorithms i.e., black-hole based optimization (BHBO), big bang-big crunch (BBBC) and galactic swarm optimization (GSO). It has been applied feed forward net to design the initial and final neural network model by structure of 8-17-1-1 for Beneish model (1999) and of 25-17-1-1 for the proposed model.This study also compares models based on hybrid neural networks and cosmological algorithms and the best and weakest cosmological algorithms are determined in neural network training to detect earning manipulation. 4- RESULTS & DISCUSSION The numerical value of the area under the curve (AUC) of the receiver operating characteristic gives an idea about the detection power which it has been in the rejected range of 0.74 to 0.55 for the Beneish model (1999) and in the accepted range of 0.97 to 0.75 for the proposed model. The best cut-off points and the best accuracy of the Beneish model (1999) have been estimated up to 0.4014, 63.49%, respectively, by the hybrid multi-layer perceptron neural network and big bang-big crunch algorithm (MLP-BBBC). The best cut-off point and the best accuracy of the proposed model have been estimated up to 0.4023 and 87.30 percent, respectively, by the hybrid multi-layer perceptron neural network and galactic swarm optimization algorithm (MLP-GSO). The estimated accuracy of the model by the hybrid methods of multi-layer perceptron neural network and galactic swarm optimization algorithm (MLP-GSO), hybrid multi-layer perceptron neural network and big bang-big crunch algorithm (MLP-BBBC) and hybrid multi-layer perceptron neural network and black-hole based optimization algorithm (MLP-BHBO) has been increased from 59.08, 63.49 and 57.5 percentages to 87.3, 79.72 and 74.25 percentages, respectively. 5- CONCLUSIONS & SUGGESTIONS This evidence indicates that predictive power of the model has been enhanced in detecting earning-manipulator companies and training error of the network has been decreased up to 12.7 percentages by hybrid method of multi-layer perceptron neural network and galactic swarm optimization algorithm (MLP-GSO). Therefore, it can be concluded that the integration of corporate governance variables as non-accounting variables to the Beneish model (1999) has been more effective in detecting and identifying earning manipulation. This evidence is consistent with this fact that a significant reduction in mean square error (MSE) is up to 23.81% and as a result the predictive power of proposed model has been significantly improved. The area under the curve of the black-hole based optimization (BHBO) and big bang-big crunch (BBBC) algorithms is covered by the galactic swarm optimization (GSO) algorithm in the proposed model. Also the area under the curve of the black-hole based optimization (BHBO) and galactic swarm optimization (GSO) algorithms is covered by the big bang-big crunch (BBBC) algorithm in Beneish model(1999). Therefore, the best algorithms for training the multi-layer perceptron neural network belong to the galactic swarm optimization (GSO) algorithm in the proposed model by the 12.7% error and to the Big Bang- Big Crunch (BB-BC) algorithm in Beneish model (1999) by the 36.51% error compared to the other cosmological algorithms in this study to detect and identify of manipulator companies.