Developing an algorithmic framework for detecting accounting fraud

Today, accounting fraud has emerged as a serious issue that undermines corporate financial credibility, shakes investor confidence, and threatens the stability of economic systems. Major global accounting scandals such as those involving Enron, WorldCom, and Xerox have not only led to the collapse of the respective companies but also disrupted national economic balances, resulting in a widespread loss of confidence in capital markets. These scandals have necessitated a comprehensive restructuring of financial reporting standards and independent audit practices. In this context, the early detection of accounting fraud has become critically important not only for ensuring corporate sustainability but also for maintaining transparency in capital markets and preserving public trust. This study proposes a dynamic and algorithmic model for detecting accounting fraud, tailored to the structural characteristics and sectoral differences of companies operating in Turkey. The model is based on the Beneish M-Score and has been enhanced by incorporating two additional financial ratios proposed by Cemal Küçüksözen—namely, the ratio of gross sales to total sales and the ratio of financial expenses to total sales. The performance of the model was further improved using the SMOTE algorithm to address class imbalance. In addition, Benford's Law was integrated into the study, thereby establishing a multi-layered control mechanism. The dataset consists of financial statements from manufacturing firms listed on Borsa Istanbul (BIST). Firms were classified as manipulators or non-manipulators based on the type of audit opinions stated in their independent audit reports and their inclusion in the Corporate Governance Index. During the analysis process, sector specific models were developed, and unique threshold values were defined for each sector to build a dynamic risk assessment framework. In this way, the study aims to go beyond traditional models that rely on fixed thresholds by offering an approach that considers sectoral variations. The developed model was integrated into Python and Django-based software, enabling users to upload financial data easily and receive instant visual reports for analysis. The software was tested using both real company data and artificially constructed manipulation scenarios, and it successfully identified high-risk firms. It is believed that the proposed approach may provide a framework compatible with auditing processes in Turkey and contribute meaningfully to the field of accounting fraud detection.