While selecting a model, accuracy provides a general measure of predictive performance, precision focuses on the correctness of the positive class, and AUC measures a model’s ability to distinguish between classes across thresholds . Higher precision reduces false positives, crucial for identifying manipulators with confidence. A high AUC indicates a model's competence in discrimination over a range of thresholds. The trade-off involves balancing these metrics to ensure a model does not overly favor one without compromising on others, like choosing Logistic Regression despite slightly lower N/A than an alternative with the best balance across metrics .

AdaBoost yielded lower accuracy (63.33%) compared to other models, suggesting it is less reliable for predicting earnings manipulations in this context . While its precision is moderate (83.33%), which indicates decent performance in identifying true manipulators, the overall low accuracy implies a higher rate of false negatives. AdaBoost might not capture the complexity of financial data manipulations as effectively as other models like Logistic Regression.

Undersampling adjusts the dataset by reducing the majority class examples, thereby balancing the dataset for training models . This prevents the model from being biased towards the majority class and enhances its ability to detect the minority class accurately. While effective, the limitation of undersampling is potential loss of valuable information from the majority class, which could impact model performance. It may also lead to models that do not generalize well where real-world class distributions differ from the balanced dataset created by undersampling.

Sales Growth Index (SGI) is identified as a key predictor in both Logistic Regression and Random Forest models, highlighting its importance in detecting earnings manipulation . In Logistic Regression, it is the second most significant predictor, suggesting a direct correlation between sales growth and manipulation probability. Random Forest also ranks SGI highly, emphasizing its role in distinguishing manipulators, potentially due to rapid or unexpected sales growth patterns that may involve manipulation. SGI's prominence across models underscores its predictive value in signaling potential risky behavior in financial reporting.

The Logistic Regression model outperformed the Decision Tree model with a higher accuracy (86.67% vs 82%), precision (95.24% vs 80%), and a greater area under the curve (0.863 vs 0.721). These metrics suggest that Logistic Regression is more effective in correctly identifying true positives and minimizing false positives when predicting earnings manipulation. The higher AUC indicates better discrimination ability across various thresholds, making it a superior choice for this application.

The precision metric varies significantly across the models, with Logistic Regression achieving the highest precision (95.24%), followed by Random Forest (89.74%), Decision Tree (80%), and AdaBoost (83.33%). Higher precision in Logistic Regression indicates its superior ability to correctly identify true manipulators while reducing false positives. This variance in precision reflects each model's effectiveness in accurately predicting true positives within the dataset, with Logistic Regression being the most effective at minimizing false alarms in identifying earnings manipulation.

The Logistic Regression Model identifies ACCR, SGI, DSRI, GMI, and AQI as the most influential predictors, ranked by their importance. ACCR (Accruals) is the most significant predictor, highlighting earnings management through discretionary accruals . The model's equation indicates positive contributions of each predictor to the likelihood of earnings manipulation, with ACCR having the largest coefficient (14.0498), followed by SGI (Sales Growth Index) and DSRI (Days Sales Receivable Index). These contribute to effective identification of manipulators by capturing the nuances of financial statement manipulation strategies.

The study addresses the challenge of unbalanced data by using undersampling techniques, which involve reducing the number of instances of the majority class (non-manipulators) to ensure that the data set is balanced . This technique allows the machine learning models to more accurately and fairly evaluate the minority class (manipulators) without being biased towards the majority class, enhancing performance and precision.

In the Random Forest model, the key predictors identified are SGI, ACCR, LEVI, SGAI, and DSRI, with SGI (Sales Growth Index) being the most significant . This differs from the Logistic Regression model where ACCR is the most impactful predictor. The inclusion of LEVI (Leverage Index) and SGAI (Sales General and Administrative Expenses Index) in the Random Forest model suggests that changes in sales-related expenses and leverage are crucial when determining manipulative practices, highlighting differences in how each model processes financial indicators to predict manipulation.

The Beneish Model indices are suitable for detecting earnings manipulation as they quantify financial statement anomalies across various activities such as sales, receivables, and accruals, which are common manipulation targets . Each index captures specific deviations that could indicate manipulation, such as inflated sales or unusual accruals, making them robust for detection. Their utility in this study highlights their relevance in providing a structured approach to identifying red flags indicative of earnings manipulation in financial data.