Fighting Accounting Fraud Through Forensic Data Analytics (1805.02840v1)

Abstract: Accounting fraud is a global concern representing a significant threat to the financial system stability due to the resulting diminishing of the market confidence and trust of regulatory authorities. Several tricks can be used to commit accounting fraud, hence the need for non-static regulatory interventions that take into account different fraudulent patterns. Accordingly, this study aims to improve the detection of accounting fraud via the implementation of several machine learning methods to better differentiate between fraud and non-fraud companies, and to further assist the task of examination within the riskier firms by evaluating relevant financial indicators. Out-of-sample results suggest there is a great potential in detecting falsified financial statements through statistical modelling and analysis of publicly available accounting information. The proposed methodology can be of assistance to public auditors and regulatory agencies as it facilitates auditing processes, and supports more targeted and effective examinations of accounting reports.

• The paper presents a novel forensic data analytics framework leveraging machine learning to identify fraudulent financial activity.
• It employs stratified sampling and industry-specific financial ratios with Mann-Whitney tests to build robust predictive fraud models.
• Advanced ensemble methods like Boosted Trees and Random Forests significantly enhanced fraud detection metrics across diverse industries.

This paper "Fighting Accounting Fraud Through Forensic Data Analytics" (1805.02840) addresses the growing challenge of detecting accounting fraud, which poses a significant threat to financial system stability and public trust. Traditional auditing methods are often insufficient due to the dynamic and hidden nature of fraudulent schemes. The paper proposes a forensic data analytics approach leveraging machine learning techniques and publicly available financial data to improve fraud detection rates and provide actionable insights for auditors and regulatory agencies.

The core objective is to build predictive models that can effectively differentiate between fraudulent and non-fraudulent companies using financial statement information. Furthermore, the paper aims to identify specific financial indicators (red flags) that are particularly relevant for different industries, assisting examiners in focusing their investigations.

The methodology involves several key steps:

1. Data Collection: A comprehensive dataset of accounting fraud cases between 1990 and 2012 was compiled using SEC Accounting Series Releases (ASR) and Accounting and Auditing Enforcement Releases (AAER). This dataset includes 1,594 fraud-year observations across all SIC industries, notably including the financial services sector often excluded in prior studies. Non-fraudulent companies were selected to match the fraud cases based on industry and fiscal year.
2. Sample Selection: To mitigate the class imbalance problem inherent in fraud detection (where fraud cases are rare), a stratified sampling approach was used, pairing each fraud observation with a non-fraud observation from the same industry and fiscal year. This creates a more balanced dataset for training predictive models.
3. Variable Engineering & Selection: The paper focuses on 20 financial ratios derived from publicly available financial statements, categorized into Leverage, Profitability, Liquidity, and Efficiency. These ratios are widely used indicators of a firm's financial health and potential areas of manipulation. A univariate analysis using the non-parametric Mann-Whitney test was performed for each industry to identify ratios that showed statistically significant differences between fraud and non-fraudulent firms. This step is crucial for developing industry-specific models. Correlation analysis (Kendall correlation) was also conducted to understand relationships between variables and identify potential multicollinearity issues, although the primary variable selection relied on the Mann-Whitney test results per industry. Table \ref{summary_ratios} provides a summary of the selected significant financial ratios by industry.
4. Machine Learning Models: A range of machine learning models were implemented to classify companies as fraudulent or non-fraudulent. The models assessed include:
   • Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) as benchmarks.
   • Logistic Regression (LR), another common benchmark.
   • Ensemble methods based on decision trees: AdaBoost (AB), Boosted Trees (BT), and Random Forests (RF). These methods were chosen for their potential for higher accuracy and the interpretability offered by tree-based structures.
5. Model Assessment: Recognizing that misclassifying fraud (false negative) is typically more costly than misclassifying a non-fraud (false positive), the paper goes beyond overall accuracy. Key performance metrics included Specificity (correctly identifying non-fraud), Sensitivity (correctly identifying fraud), Precision, G-Mean (geometric mean of sensitivity and specificity), F-Measure (harmonic mean of precision and sensitivity), and Area Under the ROC Curve (AUC). These metrics are better suited for evaluating performance on imbalanced and cost-sensitive datasets. Model performance was evaluated using a stratified 10-fold cross-validation approach on out-of-sample data.

The empirical results show varied performance across industries and models (detailed in Table \ref{resultsbyind}). Generally, more advanced ensemble methods like Boosted Trees and Random Forests achieved better performance metrics (G-Mean, F-Measure, AUC) compared to benchmark models like LDA and Logistic Regression, particularly in capturing fraudulent cases (Sensitivity). Performance was best in industries like Agriculture, Mining/Construction, and Public Administration, and more challenging in Manufacturing and Services.

A significant practical contribution of the paper is the identification of industry-specific financial red flags extracted from the decision tree models. These provide concrete indicators that auditors and regulators can use during examinations:

• Mining and Construction: High values of Inventory to Total Assets (IVTA > 0.0118) or high values of Accounts Receivable to Total Sales (RVSA > 0.234) when IVTA is low.
• Manufacturing: High Retained Earnings to Total Assets (RETA > -0.292), low Current Assets to Total Assets (CATA < 0.347), and high Total Liabilities to Total Equity (TLTE > 1.132) occurring together are strong indicators.
• Transportation, Communications, Electric, Gas and Sanitary Service: Zero or negative Inventory to Total Sales (IVSA <= 0) or high Accounts Payable to Cost of Goods Sold (PYCOGS > 0.282) when IVSA is not zero.
• Wholesale Trade and Retail Trade: Moderate Retained Earnings to Total Assets (0 < RETA < 0.186) combined with high Inventory to Total Sales (IVSA > 0.189), or high RETA (> 0.186) with very high IVSA (> 0.335).
• Finance, Insurance and Real Estate: Accounts Payable to Cost of Goods Sold (PYCOGS) <= 0 coupled with Long-Term Debt to Total Assets (LTDTA) > 0, or PYCOGS > 22.82 combined with very high Total Liabilities to Total Equity (TLTE > 19.05).
• Services: Low Total Sales to Total Assets (SATA < 0.256) and high Inventory to Cost of Goods Sold (IVCOGS > 0.032).
• Public Administration: High Inventory to Total Sales (IVSA > 0.063).

These findings translate theoretical modeling into practical, interpretable rules that can support and prioritize auditing efforts. The proposed methodology provides a data-driven tool that can be combined with expert knowledge to enhance the effectiveness and efficiency of accounting fraud detection.

However, the paper acknowledges limitations, including the difficulty in obtaining a complete sample of all accounting fraud cases and the dynamic nature of fraudulent practices. Future work could explore the inclusion of additional predictive variables (qualitative data, time-series features), investigate different techniques for handling class imbalance, evaluate the impact of varying classification thresholds, and apply the methodology to more granular industry sub-sectors.