Financial Reporting Fraud Scheme Prediction via Machine Learning Approach – Multiclass Classification

This paper attempts to evaluate the performance of machine learning models in fraudulent financial Reporting schemes prediction via a multi-classification approach and using an unbalanced dataset. Therefore, the financial statements of 134 companies listed on the Tehran Stock Exchange from 2009 to 2021 were investigated by Logistic Regression, Decision Tree, Boosting Algorithms, and Support Vector Machine. Models were programmed with Python and Performance indicators were calculated and compared. Furthermore, the machine learning model’s performance was investigated in binary classification with the balanced dataset to predict each fraud scheme exclusively. According to the results via a multi-classification approach, then the significant difference between machine learning models’ performance was approved. Support Vector Machin was preferred in multiclass problem space with the unbalanced data set. To predict fraud schemes via binary classification, a significant difference between machine learning models’ performance was not approved except to predict the “Overstatement assets and income” scheme. Support Vector Machin was preferred to Logistic Regression and Decision Tree model. The present research attempts to fill the research gap in the research area by developing machine learning models with a multi-classification approach.