A Novel Computational Approach to Predict Financial Statements Fraud using Clustering and Classification Techniques: Evidence from Listed Companies in Tehran Stock Exchange

Recent scandals and corporate failures have shaken the confidence of investors as accounts which were purported to reflect a “true and fair view” of businesses have been misleading. In an era with evolutionary financial frauds, computer assisted automated fraud detection mechanisms will be more effective and efficient with specialized domain knowledge. Statistics and machine learning based technologies have been shown to be an effective way to deter and defect fraud. Therefore, this study deals with the identification of factors related to fraudulent financial statements and investigating the effectiveness of data mining techniques in detecting firms that issue fraudulent financial statements.

 Research Questions

In order to achieve the objectives of this research, the following questions are developed:1. Can the fraudulent and non-fraudulent financial statements be detected from the text of annual corporate audit report fillings?
2. Can a quantitative fraud detection model be developed that will provide an analytical procedure for automating detection of potential fraud?

Data mining is used in many domains, including finance, engineering and biomedicine. There are two categories of data mining methods: unsupervised and supervised. Identifying financial statements fraud can be regarded as a typical classification problem. Data mining proposes several classification methods derived from the fields of statistics and artificial intelligence. Three methods, which enjoy a good reputation for their classification capabilities, are employed in this research study. These methods are Decision Trees, Support Vector Machines and Boosting algorithm. In this research, nineteen red flags of fraud extracted from auditing standard No. 240 along with data mining techniques such as principal components analysis and modified K-Means clustering were used to discriminate financial statements fraud cases. We consider data mining based financial fraud detection techniques such as Decision Trees, Support Vector Machines and Boosting approach in order to identify fraud. Based upon the fraudulent financial reporting related literature, 40 predictor variables are selected as input variables. These variables represent measurement proxies for a firm’s attributes of financial leverage, profitability, asset composition, liquidity, efficiency, size, growth, overall financial position, audit firm size, audit firm tenure and auditor change. These predictor variables are collected from the financial statements of listed companies in Tehran Stock Exchange (TSE) for the period of 2003-2015.

This research uses a quantitative approach on textual data to discriminate fraud cases using text mining techniques. The first research question was: can the fraudulent and non-fraudulent financial statements be detected from the text of annual corporate audit report fillings? Modified k-Means clustering demonstrated a good ability to discriminate fraud cases from non-fraud cases and thus it answered the first research question. the successful implementation of the Decision Trees, Support Vector Machines and Boosting algorithms also answers the second research question: can a quantitative fraud detection model be developed that will provide an analytical procedure for automating detection of potential fraud? In comparative assessment of the models’ performance, Support Vector Machines achieved the best performance to correctly classify validation sample in a 5-fold cross validation procedure. Decision trees and boosting algorithms also achieve a satisfactorily high performance.

The present study contributes to auditing and accounting research by examining the suggested variables that can best discriminate cases of financial statements fraud. These results suggest that there is validity for detecting financial statement fraud using data mining techniques such as Decision Trees, Support Vector Machines and Boosting approach. Our analysis provides insight for auditors, taxation authorities, investors, the stock exchange and banking system.