MACHINE LEARNING MEETS FINANCIAL FORENSICS: PREDICTING FINANCIAL STATEMENT FRAUD WITH DECISION TREE USING BENEISH M-SCORE RATIOS IN PAKISTAN

Abstract

This research developed a financial statement fraud prediction model in case Pakistani non-financial firms listed on the Pakistan Stock Exchange (PSX). The auditor opinion is taken as a class variable, while eight ratios used in Beneish M-score were taken as predictors to define fraudulent reporting. Six different variants of Decision tree models are also deployed to proposed final models. However, the imbalanced data problem is addressed using oversampling through the SMOTE algorithm before applying the decision tree models. Results showed that random forest provided the highest predictability, i.e., 83%, among other selected models. Random forest outperformed other evaluation matrices, including individual class true positive rate, f-score, ROC, and PRC. Detailed analysis also explored how inflating receivables and internal pressure contribute to the predictability of adverse and/or qualified opinions. The findings suggested that stakeholders use the proposed random forest model to identify the potential Financial statement fraud (FSF) in the case of Pakistani non-financial firms.

Keywords:Financial statement fraud; Fraudulent reporting; Decision tree models; Auditor opinion; Random forest model Pakistani non-financial firms