運用分類演算法檢測財務報表詐欺－印尼上市公司之實證

Financial Statement Fraud Detection Using Classification Algorithms: Evidence from Indonesian Listed Companies

莫爾凡(IRFANDY AZIS MUH.)

指導教授 ： 李正文

中原大學/商學院/商學博士學位學程/博士(2024年)

https://doi.org/10.6840/cycu202400270

摘要

摘要
這項研究旨在辨識幾個關鍵變數，這些變數可能作為金融報表欺詐的潛在警示信號，並評估各種分類算法在印尼上市公司中檢測此類欺詐的效果。使用多元線性回歸和邏輯回歸分析，以及包括邏輯回歸（LR）、k-最近鄰（KNN）、支持向量機（SVM）、決策樹（DT）和隨機森林（RF）在內的分類算法。該研究基於精確度、召回率、準確度和F1分數評估了每個模型的性能。多元線性回歸分析顯示，金融報表欺詐的顯著潛在警示信號包括應收帳款周轉率（ART）、應收帳款未收天數（DOAR）、應付帳款未付天數（DPO）、毛利的對數（Log(GP)）、毛利率（GPM）、庫存銷售比（ITS）和總資產周轉率（TAT）。ART、DOAR、GPM、ITS 和 TAT 與欺詐呈正相關，表明這些指標的較高數值可能與增加的欺詐風險相關。相反，DPO 和 Log(GP) 的負相關顯著，表明較長的付款期限和較高的盈利能力減少了欺詐可能性。邏輯回歸分析支持這些發現，DOAR、Log(GP) 和 GPM 為顯著的負向預測因子，而淨銷售變化（ΔNS）、淨收入比率（NIR）和存貨與流動負債比率（ICL）則是欺詐的正向指標。麥克法登R平方值為0.0997，顯示了適度的解釋力但具有統計顯著性。在分類算法中，隨機森林被證明是最有效的金融報表欺詐檢測模型，展示了在訓練和測試數據集上更好的性能。邏輯回歸和支持向量機也表現良好，而KNN和決策樹顯示出過度擬合問題，限制了它們的實際應用性。理論上的影響強調了「壓力」作為推動印尼上市公司欺詐活動的關鍵動機，反映了達到財務目標的激烈文化和經濟期望。實踐上的影響強調了強大的內部控制、有針對性的欺詐檢測和法規的完善。通過關注ART、DOAR、DPO、Log(GP)、GPM、ICL、ITS、ΔNS、NIR和TAT等關鍵財務指標，稽核師、財務分析師和監管機構可以提升欺詐檢測的效率和準確性，從而改善財務報告的誠信。

關鍵字

財務報表 ； 詐欺檢測 ； 迴歸 ； 分類演算法

並列摘要

Abstract
This study aimed to identify several key variables that serve as potential red flags for financial statement fraud and evaluate the effectiveness of various classification algorithms in detecting such fraud among Indonesian listed companies. Using multiple linear regression and logistic regression analyses, alongside classification algorithms including Logistic Regression (LR), k-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF). This study assessed each model's performance based on precision, recall, accuracy, and F1-Score. The multiple linear regression analysis revealed that significant potential red flags for financial statement fraud include Accounts Receivable Turnover (ART), Days Outstanding Accounts Receivable (DOAR), Days Payables Outstanding (DPO), Logarithm of Gross Profit (Log(GP)), Gross Profit Margin (GPM), Inventory to Sales (ITS), and Total Asset Turnover (TAT). ART, DOAR, GPM, ITS, and TAT were positively correlated with fraud, suggesting that higher values in these metrics might be associated with increased fraud risk. Conversely, DPO and Log(GP) were negatively significant, indicating that longer payment periods to suppliers and higher profitability reduce fraud likelihood. Logistic regression analysis supported these findings, with DOAR, Log(GP), and GPM as significant negative predictors, and Change in Net Sales (ΔNS), Net Income Ratio (NIR), and Inventory to Current Liabilities (ICL) as positive indicators of fraud. The McFadden R-squared value of 0.0997 demonstrated modest explanatory power but statistical significance. Among the classification algorithms, Random Forest emerged as the most effective model for detecting financial statement fraud, exhibiting better performance across both training and testing datasets. Logistic Regression and SVM also performed well, while KNN and Decision Tree showed overfitting issues, limiting their practical applicability. Theoretical implications highlight "pressure" as the key motivation driving fraudulent activities in Indonesian listed companies, reflecting intense cultural and economic expectations to meet financial targets. Practical implications emphasize the importance of robust internal controls, targeted fraud detection, and regulatory refinements. By focusing on critical financial metrics such as ART, DOAR, DPO, Log(GP), GPM, ICL, ITS, ΔNS, NIR, and TAT, auditors, financial analysts, and regulatory bodies can enhance the efficiency and accuracy of fraud detection, thereby improving financial reporting integrity.

並列關鍵字

Financial Statement ； Fraud Detection ； Regression ； Classification Algorithms

參考文獻

REFERENCES

ACFE. (2014). Report to the Nations on Occupational Fraud and Abuse. ACFE.

ACFE. (2022). Occupational Fraud 2022: A Report to the Nations. ACFE.

Achakzai, M. A. K., & Juan, P. (2022). Using machine learning Meta-Classifiers to detect financial frauds. Finance Research Letters, 48, 102915. https://doi.org/https://doi.org/10.1016/j.frl.2022.102915

Aghghaleh, S. F., Mohamed, Z. M., & Rahmat, M. M. (2016). Detecting Financial Statement Frauds in Malaysia: Comparing the Abilities of Beneish and Dechow Models. Asian Journal of Accounting & Governance, 7.