Modification of Earning Manipulation Prediction Model with Emphasis on Environmental Variables and Hybrid Artificial Neural Network and Meta-heuristic Algorithms

Earning of companies is one of the important factors in economic growth and development and earning manipulation is one of the main challenges of market efficiency that researchers often use accounting data to predict earning manipulation, while non-accounting data also play an important role in predicting earning manipulation.Due to the fact of the conducted research in order to develop the Beneish model has been formed solely because of accounting data, so the effects and consequences of non-accounting variables in all models have been ignored. This study tries to examine the nonlinear relationships of accounting and non-accounting variables and examine the effect of both variables simultaneously. The purpose of this study is to measure the predictive power of Beneish model and the development of the Beneish model (DBM) by non-accounting variables and to compare the accuracy of earning manipulation prediction of the research models using a hybrid Artificial Neural Network trained by Particle Swarm Optimization (PSO) algorithm and Imperialist Competition Algorithm (ICA).Theoretical Framework: The development of the Beneish model (DBM) was done through emphasizing non-accounting variables, including the Information Asymmetry (IS) and Product Market Competition (PMC).(Asgari Alouj et al, 2020). Another study by (Pourali & Kouchaki Tajani, 2021) was conducted to compare the accuracy of companies' profit manipulation predictions using colonial competition algorithm and genetic algorithm. The results showed that colonial competition algorithm with 93% accuracy and 7% error and genetic algorithm with 76% accuracy And 24% error could have predicted the coefficients of the variables of the profit manipulation model. The results also showed that the ability to predict the accuracy of profit manipulation model coefficients by colonial and genetic competition algorithms is more accurate than the prediction of the initial model of Banish (1999) .

This research has been developed the Beneish model) BM) with non-accounting variables including information asymmetry (IS) and competition in the product market (PMC). The data of 184 companies listed on the Tehran Stock Exchange during 2006-2017 has been collected and the prediction accuracy of research models has been compared by two algorithms in training of Artificial Neural Network (ANN): Particle Swarm Optimization (PSO) and Imperialist Competition Algorithm (ICA) in detecting and identification of earning-manipulator companies. In this research, the auditor's report has been used as an alternative solution and the review process has been done such that the audit report of the sample companies has been fully reviewed and studied and if there were the cases as an index of earning manipulation (regardless of the type of report acceptable - adjusted - rejected and no comment), the sample companies would be selected as the earning-manipulator firm and the number 1 would be allocated. Also, if there were no clauses as an index of earning manipulation, for example, the report is adjusted for another reason, it would be selected as a non-earning manipulator and the number zero would be allocated.

After reviewing and auditing the audit reports of the sample companies of 1840 data-year, 900 data-year companies has been classified at the low level of earning manipulator companies and 940 data -year companies has been classified at the high level of earning manipulator companies. In this study, the prediction power of earning manipulation companies has been investigated by hybrid Artificial Neural Network method and Particle Swarm Optimization (PSO) algorithm and also by hybrid Artificial Neural Network method and Imperialist Competition Algorithm (ICA) and a comparison has been made between the accuracy of the research models. The areas under Receiver operating characteristic (ROC) curve of the Beneish model have been estimated up to 0.6001 and 0.5538 using the hybrid neural network trained by Imperialist competition algorithm and particle swarm optimization algorithm, respectively. The area under the ROC curve in the Beneish model has been estimated in the range of 0.5 - 0.6 and indicates the Beneish model test has been rejected in detecting and identifying earning manipulator companies. Therefore, it can be seen that the separation of the two groups of earning manipulator and non-manipulator companies is not significantly different from the separation of the chance model and it can be said that the Beneish model is a completely random model in the Tehran Stock Exchange and cannot be used to identify earning manipulator companies. Also, the best prediction accuracy of the Beneish model has been estimated up to 57.55 and 55.71 percentages using the hybrid neural network method trained by the Imperialist competition algorithm and the particle swarm optimization algorithm, respectively.

Findings indicate that the prediction accuracy of the proposed model has increased from 57.55 to 63.86 percentages and 55.71 to 59.84 percentages by the ANN-ICA and ANN-PSO, respectively. Development of the model, area under curve (AUC) of ROC has been increased and the prediction error has been reduced to 6.31 percentages by the ANN-ICA and to 4.13 percentages by the ANN-PSO, but the test result is still poor. In fact, the accuracy of model prediction by the ANN-PSO has been improved compared to the ANN-ICA.However, it can be seen that relying on these variables by itself could not easily identify earning manipulator and non-manipulator companies. Considering that the proposed model with the variables of Competition in the Product Market and Information Asymmetry has not significantly improved the accuracy of the prediction model, it can be seen that there is a not significant relationship between these variables and earning manipulation variable. In order to judgement whether or not the results of ANN-ICA and ANN-PSO of research models are significantly different, the Wilcoxon test has been performed at a significance level of 5% as the statistical method of non-parametric. The results of Wilcoxon test show that the normal statistic of Wilcoxon test is more than the critical value of 1.64 and the significance level is less than 0.05 in both methods .Also, the average rank has been calculated up to 548.5 before the development of the model and has been calculated up to 5549.7 after the development of the model, so the research hypothesis is confirmed.