نوع مقاله : مقاله پژوهشی

نویسندگان

• مرضیه پورساعدی ¹
• محمود همت فر ²
• سید عنایت‌اله علوی ³
• رویا نصیرزاده ⁴

¹ دانشجوی دکتری حسابداری، گروه حسابداری، واحد بروجرد، دانشگاه آزاد اسلامی، بروجرد، ایران

² دانشیار، گروه حسابداری، واحد بروجرد، دانشگاه آزاد اسلامی، بروجرد، ایران

³ استادیار گروه کامپیوتر، دانشکده مهندسی، دانشگاه شهید چمران اهواز، اهواز، ایران

⁴ استادیار، گروه آمار، دانشکده علوم، دانشگاه فسا، فسا، ایران

چکیده

هدف این مطالعه، مدل‌سازیِ تشخیصِ تقلب مالی شرکت‌ها تحتِ اجرای ارزیابی الگوریتم‌های مطلوبیتِ شبکه عصبی مصنوعی می‌باشد. در این مطالعه با استفاده از فرآیندهای برنامه‌ریزی مجذوری « » در الگوریتم‌های شبکه عصبی مصنوعی، تلاش شده است تا طی چندین مرحله بر اساس داده‌های زمانی 1392 تا 1401، اقدام به تعیین الگوریتم پایه در وهله‌ی اول و انتخاب پارامترهای تکنیکال شبکه عصبی مصنوعی در وهله‌ی دوم گردد. سپس با توسعه یک مدل تشخیصی بر اساس دو مقیاس آزمون و کنترل، الگوریتم‌هایی فر ابتکاری که بالاترین ضرایب دقت در پیش‌بینی صحت تقلب مالی را دارند، در سطح شرکت‌های بازار سرمایه موردبررسی قرار گیرند. لذا بر اساس فرآیند نمونه‌گیری سیستماتیک، تعداد 95 شرکت بورس اوراق بهادار انتخاب شدند تا بر اساس 950 مشاهده (سال-شرکت)، حد فاصل شرکت‌های دارای سلامت مالی با شرکت‌های دارای احتمال تقلب مالی از طریق دهک‌بندی تعیین گردد و شرکت‌های قرارگرفته در دهک‌های دارای تقلب مالی، از طریق پارامترهای مطلوبیت شبکه عصبی مصنوعی موردبررسی قرار گیرند. نتایج مطالعه نشان داد، الگوریتم یادگیری بدون نظارت که شامل مجموعه پارامترهای ارزیابی مبتنی بر الگوریتم فرا ابتکاری است، از صحت پیش‌بینی‌های مبتنی بر داده‌های برآورده شده بالاتری برخوردار می‌باشد. همچنین نتایج ناشی از صحت پیش‌بینی تقلب‌های مالی شرکت‌های دهک‌بندی‌شده بر اساس دو الگوریتم انتخابی ژنتیک و کلونی زنبورعسل نشان می‌دهد، الگوریتم کلونی زنبورعسل از ضریب دقت بالاتری در پیش‌بینی صحت احتمال تقلب شرکت‌های موردبررسی برخوردار است. همچنین مشخص شد، نسبت سود خالص به فروش مهم‌ترین معیار ارزیابی دقت پیش‌بینی احتمال تقلب در شرکت‌های موردبررسی می‌باشد.

کلیدواژه‌ها

• تقلب مالی
• شبکه عصبی مصنوعی
• برنامه‌ریزی مجذوری

موضوعات

• حسابداری مالی

عنوان مقاله [English]

Modeling the Detection of Firms Financial Fraud under the Implementation of Artificial Neural Network Evaluation Algorithms

نویسندگان [English]

• Marzieh Poursaedi ¹
• Mahmood Hematfar ²
• Seyed Enayatallah Alavi ³
• Roya Nasirzadeh ⁴

¹ Phd Student, Department of Accounting, Borujerd Branch, Islamic Azad University, Borujerd, Iran

² Associate Professor, Department of Accounting, Borujerd Branch, Islamic Azad University, Borujerd, Iran

³ Assistant professor of computer Department, faculty of Engineering Shahid Chamran university of Ahvaz. Ahvaz, Iran

⁴ Assistant Professor, Department of Statistics, Faculty of Science, Fasa University, Fasa, Iran

چکیده [English]

The aim of this study is to model the detection of firms’ financial fraud using artificial neural network evaluation algorithms. Quadratic Programming (QP) processes were applied in artificial neural network algorithms to, first, determine the basic algorithm and, second, choose the technical parameters of the artificial neural network, based on time-series data from 2013 to 2022. A diagnostic model was then developed using test and control scales to examine innovative algorithms with the highest accuracy coefficients in predicting financial fraud at the level of capital market companies. Based on systematic sampling, 95 stock exchange companies were selected, providing 950 firm-year observations. The distinction between financially healthy companies and those with the potential for financial fraud was determined through decimalization, and companies placed in the fraud-prone deciles were examined using the parameters of the artificial neural network's effectiveness.

Introduction

With the advent of technology and digital business exchanges, today's era is far more affected by the negative consequences of fraud in financial statements than in the past, making the development of fraud detection methods a technical necessity in the field of financial knowledge. By providing appropriate assessment opportunities for optimal decision-making, highly efficient financial markets facilitate a balanced flow of information among firms, thereby maintaining the attractiveness of entering this market compared to other markets, such as money markets. In particular, implementing such processes can be considered a strategic financial necessity in developing economies that face serious challenges due to resource constraints. However, reference to scientific evidence and the operational reality of financial markets in these countries indicates the existence of information rents for certain groups of financial decision-makers, which create information asymmetry without requiring expertise, technical evaluation methods, or fundamental analysis.
 

2. Literature Review

Disclosure of transparent financial information is considered an important resource in stakeholders’ economic decision-making and can contribute significantly to balancing investment markets. However, with theoretical and structural changes in the field of financial transparency, fraud as a behavioral and functional issue has become a challenge to the financial transparency of companies. To understand the official definition of financial fraud, the best reference is Section 240, Clause 4 of the Iranian Auditing Standards, which defines this phenomenon as the intentional actions of companies’ executive managers, governing bodies, employees, or third parties that result in the acquisition of illegitimate benefits and cause widespread damage to other stakeholders. An important point noted in Clause 9 of the same standard is the distinction between fraud and error, where intent is considered the only distinguishing feature. In practice, however, drawing a clear boundary between these two in order to protect stakeholders’ rights is not an easy task.
 

Methodology

In terms of data type, this study is classified as a semi-experimental and post-event study in the field of positive financial research, implemented using the neural network analysis method and related techniques. In terms of results, this study can be classified as applied research, and in terms of implementation, it is correlational and algorithmic. First, based on a set of analytical procedures using software such as MATLAB and WEKA, the evaluation processes of artificial neural network algorithms are examined. Then, by comparing the selected algorithms, the most effective type of analysis is determined from the perspective of evaluating financial ratios to predict the probability of fraud in companies. It should be noted that the analytical implementation in this process is based on three steps: extracting financial ratios, evaluating extracted ratios, and finally applying neural networking to the evaluated algorithms.
 

Result

In this study, a systematic review of the literature related to the research field over recent years was first conducted to select financial ratios appropriate for evaluating fraud in capital market companies. Then, based on quadratic programming (QP), the basic algorithm for evaluating the accuracy of corporate fraud was determined using firm-year observations by minimizing the gap between predicted and actual data obtained from the identified financial ratios. Through the adaptive neural fuzzy inference system (ANFIS) test and the cross-validation process (k-fold), it was determined that the unsupervised learning algorithm, which incorporates evaluation parameters based on a meta-heuristic approach and provides higher prediction accuracy, was selected as the foundation for the algorithms in this study. To create a reference index for assessing fraud probability based on financial ratios, the decile method was used to identify which companies with a coefficient of D>1 could be distinguished between financially healthy firms and those with fraud probability, under the constant return to scale (CRS) and variable return to scale (VRS) evaluation scales. The results indicate that companies with fraud probability were concentrated in four deciles, suggesting that two algorithms, genetic and bee colony selection, were used to further evaluate prediction accuracy. Finally, it was found that the bee colony algorithm had a higher accuracy coefficient in predicting fraud accuracy probability compared to the genetic algorithm. It was also found that the ratio of net profit to sales is the most important criterion for evaluating the accuracy of fraud prediction in the companies under study.
 

Discussion

In interpreting the results, it should first be noted that the bee colony algorithm performs better in solving complex problems due to its multi-process optimization through collective intelligence. This algorithm can be more effective in financial decision-making at the capital market level because it shows higher convergence power and accuracy in predicting the probability of fraud in the companies under study compared to the genetic algorithm. In addition, the coefficients obtained from the bee colony algorithm indicate more effective optimization of financial ratios in predicting the probability of fraud. Financial decision-makers can therefore use this algorithm for more accurate evaluations based on financial ratios, enabling them to identify companies with a probability of fraud and avoid purchasing their shares when forming a portfolio.
 

Conclusion

Given the importance of fraud in financial decision-making, investors are advised to reduce the risk of predicting financial fraud in companies by improving their level of technical analysis training when forming a stock portfolio. The most significant analytical techniques are related to prediction methods based on unsupervised learning algorithms. Focusing on this set of algorithms provides decision-makers with deeper insights by enabling them to recognize the true nature of the data and, without manipulating or remapping it, identify predictable patterns, structures, and relationships of financial fraud in companies.

کلیدواژه‌ها [English]

• Artificial Neural Network
• Financial Fraud
• Quadratic Programming