Detecting financial misstatements in emerging markets: a machine learning approach

Detecting financial misstatements in emerging markets: a machine learning approach | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Detecting financial misstatements in emerging markets: a machine learning approach Hoa Thi Thanh Tieu, Thanh Hien Hoang, Hung Ngoc Tran This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7360630/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This study develops a machine learning–based framework for detecting material misstatements in the financial statements of Vietnamese listed companies. Using 10,286 firm-year observations from 2016–2023, the research applies two ensemble algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), to a binary classification task based on audit-adjusted profit discrepancies. To address data imbalance and improve prediction reliability, the Synthetic Minority Over-sampling Technique (SMOTE) is applied within a stratified cross-validation procedure, while Bayesian optimization tunes hyperparameters to enhance generalization performance. Both RF and XGBoost achieved high predictive accuracy (~ 0.839) and strong discriminative power (AUC-ROC ~ 0.91), outperforming logistic regression. Model interpretability was improved through the Least Absolute Shrinkage and Selection Operator (LASSO), which selected key financial and non-financial predictors from over 50 variables. RF’s feature importance analysis further highlighted the influence of listing exchange characteristics, prior misstatement history, and forward-looking performance indicators. The proposed framework offers auditors and regulators a scalable, data-driven tool for risk-based audit planning and regulatory oversight—particularly valuable in emerging markets with limited confirmed fraud data. Physical sciences/Engineering Physical sciences/Mathematics and computing Ensemble machine learning Financial statement misstatements Fraud detection LASSO Random Forest XGBoost SMOTE Bayesian optimization Stratified cross-validation Figures Figure 1 Figure 2 Figure 3 1. Introduction Accurate financial statements are essential for ensuring transparency, enabling informed decision-making, and maintaining stakeholder trust. However, material misstatements, whether caused by unintentional errors or deliberate fraud, can distort the true financial condition of an entity and result in significant harm to investors, regulators, and the broader economy. Auditing standards such as SAS 99 and ISA 240 emphasize the heightened risks associated with fraudulent misstatements due to their intentional and often concealed nature. High-profile corporate failures like Enron and Lehman Brothers underscore the devastating consequences that undetected financial reporting fraud can have. While fraud-related misstatements account for a smaller portion of total occupational fraud cases globally, they are the most financially damaging (ACFE, 2018–2022). A critical issue is the time lag in detection, averaging two years, which allows manipulative behaviors to continue unchecked (Kaminski et al., 2004 ). Early detection, therefore, is not only desirable but essential for protecting financial systems. Despite global advancements in fraud analytics, Vietnam faces unique challenges in this domain. Specifically, the country lacks a centralized, publicly accessible database of confirmed financial reporting fraud cases. This data limitation prevents direct fraud analysis and restricts the applicability of models developed in other contexts. To address this gap, the current study focuses on material misstatements as a practical and theoretically grounded proxy for fraudulent behavior. These misstatements, which are defined in alignment with Vietnam Auditing Standard No. 320, offer a reliable signal of potential fraud risk and are more consistently observable within Vietnam's regulatory environment. This study addresses a critical research problem: the absence of effective, data-driven tools for detecting financial misstatements in Vietnamese listed companies. The primary objective is to develop and validate a machine learning-based predictive model to identify companies likely to produce materially misstated financial reports. By integrating financial indicators with non-financial contextual variables, the model aims to provide early warning signals to auditors, regulators, and investors. This research offers three key contributions. First, it is among the first empirical studies to apply ensemble machine learning algorithms, specifically Random Forest and XGBoost, to predict financial misstatements in the context of Vietnam. Second, it employs LASSO regression to enhance model interpretability and relevance by selecting the most influential features from a large dataset. Third, it responds to a pressing need in the Vietnamese auditing landscape by offering a practical, scalable solution that can be used by auditing firms and regulators for risk-based planning and regulatory oversight. By aligning with both international auditing standards and Vietnam’s current data landscape, this study contributes to strengthening financial integrity and audit quality in an emerging market context. It supports the broader objective of enhancing transparency, accountability, and investor confidence in Vietnam’s capital markets. 2. Literature review Detecting material misstatements in financial statements—often symptomatic of fraudulent activity—has long been a critical concern in accounting and auditing research. Early studies predominantly relied on traditional statistical models, with one of the most widely recognized tools being the Beneish M-score (Beneish, 1999 ). This model uses a set of financial ratios to detect earnings manipulation and has been applied in various national contexts (Arshad et al., 2015 ; Herawati, 2015 ; Hołda, 2020 ). However, its predictive accuracy remains modest, typically ranging from 70–75%, and recent evaluations have called for caution when applying it as a standalone tool (Bhavani & Amponsah, 2017 ; Marais et al., 2023 ). To improve fraud detection capabilities, researchers have expanded the scope of predictive inputs. For example, Dechow et al. ( 2011 ) developed the F-score by integrating off-balance sheet indicators, market performance data, and governance-related variables. Similarly, studies by Kanapickienė and Grundienė ( 2015 ) and Zainudin and Hashim ( 2016 ) demonstrate that combining financial and non-financial variables—such as ownership concentration, audit opinions, or listing status—can substantially increase model accuracy. Despite these improvements, traditional statistical models face limitations in handling nonlinear relationships, high-dimensional data, and multicollinearity. These challenges have prompted a methodological shift toward machine learning (ML) techniques. Groundbreaking work by Green and Choi ( 1997 ) introduced neural networks to fraud detection, and subsequent studies have adopted algorithms such as decision trees, support vector machines, and ensemble methods like Random Forest and XGBoost (Gaganis, 2009 ; Hajek & Henriques, 2017 ; Jan, 2018 ), 2018 ). These models consistently outperform logistic regression, particularly when applied to datasets that incorporate both financial and contextual non-financial data (Kirkos et al., 2007 ; Lokanan et al., 2019 ; Ravisankar et al., 2011 ). Recent reviews (Gupta & Mehta, 2024 ; Shoetan et al., 2024 ) affirm that ensemble methods deliver superior accuracy and resilience in fraud detection, addressing both nonlinearity and feature selection challenges. These techniques are especially relevant in complex datasets where conventional assumptions do not hold. While the existing literature offers valuable insights, it remains heavily concentrated in developed markets with strong institutional frameworks and accessible fraud data. In contrast, emerging markets like Vietnam face unique challenges, particularly the absence of public datasets on confirmed fraud cases. This significantly limits the applicability of fraud detection frameworks developed in Western or mature economies. Instead, material misstatements—detected through audit adjustments—must serve as proxy indicators for fraud risk in data-constrained environments. In Southeast Asia, studies have begun to explore the utility of machine learning in fraud detection. Lokanan Lokanan et al. ( 2019 ) applied ML to Vietnamese firms and found improved accuracy over statistical models. Similarly, Omar et al. ( 2017 ) demonstrated the predictive power of neural networks in the Malaysian context. Research from Indonesia and Thailand supports the adaptability of ML models to regional data environments with similar regulatory and reporting constraints. Despite this growing interest, Vietnam remains underrepresented in empirical research. Most prior studies either focus on global datasets or lack contextual adaptation for Vietnam’s regulatory structure, audit practices, and data access limitations. This study seeks to fill that gap by developing a Vietnam-specific machine learning model that uses both financial and non-financial variables to predict material misstatements, thereby offering practical applications for auditors, regulators, and corporate governance professionals operating in transitional economies. 3. Research methodology 1.1 Machine learning approach According to (Bennett et al., 2022 ), data scientists and statisticians are often inconsistent when determining the best approach: machine learning or statistical modeling to solve an analytical challenge. However, machine learning and statistical modeling are complementary because they are all based on mathematical principles. The difference is only in the use of instruments for the overall analysis process. The choice of both approaches or just one depends on the problem to be solved and the results to be achieved as well as the characteristics of the data and the context of the analysis. The determination of the approach is mainly based on empirical evidence, such as the size and completeness of the data, the number of variables, whether the assumptions are warranted, and whether the expected results are predictive or causal (Bennett et al., 2022 ). Statistical modeling typically makes assumptions that predictors or characteristics are known, models are parametric, and that testing research hypotheses and uncertainty are paramount (Breiman, 2001 ), meanwhile machine learning does not make these assumptions (Bennett et al., 2022 ). In machine learning, many models are based on non-parametric approaches, in which the structure of the model is unspecified or unknown, and assumptions about normal distribution, linearity, or residuals are not required for modeling (Carmichael & Marron, 2018 ). The purpose of machine learning is to focus on predictive performance by using learning algorithms to find little-known, unrelated, and complex patterns in data without prior insight into the underlying structures (Carmichael & Marron, 2018 ). Whereas, in statistical modeling, the motivation is to consider inferences, correlations, and effects of a small number of variables (Breiman, 2001 ). On the same page, (Bzdok et al., 2018 ) observed that statistical modeling draws inferences about the population from a sample of studies while machine learning finds generalizable predictive patterns from data. A literature review shows that predictive models are built on a wide range of features (variables) including both financial indicators and non-financial information. In the early stages, researchers mainly applied statistical models. Logistic regression was widely used; however, the predictive performance was relatively low because of the tight constraints of assumptions that may not be achieved for the data characteristics and the relatively large number of features of this type of research. Although statistical models can detect simple fraud cases effectively, they have difficulty with the increasing complexity and sophistication of modern fraud schemes (Shoetan et al., 2024 ). Machine learning approaches overcome this problem by analyzing large amounts of financial data to detect patterns and anomalies that may indicate fraud (Ashtiani & Raahemi, 2021 ; Lakhan et al., 2022 ). Therefore, many studies have approached machine learning methods in identifying fraud in financial statements and have achieved better results than traditional statistical methods. 1.2 Data 1.2.1 Labelling The dependent variable in this study is the presence of material misstatements in financial statements, serving as a proxy for fraudulent reporting. Following auditing best practices and standards (Vietnam Auditing Standard No. 320; VACPA sample guidelines), a material misstatement is defined as an absolute change of 10% or more in profit after tax (PAT) between pre-audited and post-audited financial reports. Profit was selected as the basis for measuring misstatements because it comprehensively reflects fluctuations in revenue, expenses, assets, and liabilities. The categorical variable "Misstatements" was constructed using data from the FinnPro platform, which aggregates annual pre- and post-audit PAT for all companies listed on Vietnamese stock exchanges. Leveraging FinnPro enables cost-effective data collection and preprocessing while ensuring a high degree of reliability and transparency. Observations where the absolute value of the profit variance ratio is greater than or equal to 10% are labeled 1 (material misstatement), and those below this threshold are labeled 0 . This computation method is straightforward, consistent with both international and local auditing guidance, and therefore valid as a proxy indicator for financial manipulation (see Table 1 ). Table 1 Measurement of categorical variable Categorical variable Variable name Measurement Material misstatement in financial statements Misstatement Binary variable Misstatement = 1: The absolute value of the Profit variance Ratio is greater than or equal to 10%; Otherwise = 0 In there, Profit variance ratio = (Profit after tax after auditing - Profit after tax before auditing)/Profit after tax after auditing Source: Authors’ calculation The final dataset includes 10,286 firm-year observations from non-financial companies listed on Vietnamese stock exchanges, covering the period from 2016 to 2023. Firms in the banking and finance sector were excluded due to their distinct financial reporting structure. The class distribution comprises 4,314 firms with material misstatements and 5,972 without, resulting in a 42:58 ratio. 1.2.2 Feature Selection and Data Sources This study employs a combination of financial and non-financial features to train the predictive model, consistent with the approach used in prior fraud detection research. Financial statement analysis techniques, especially analytical procedures such as trend and ratio analysis, are widely recommended by auditing standards like SAS No. 56 and SAS No. 99 to identify anomalies and assess the risk of fraud (Albrecht et al., 2008 ; Kaminski et al., 2004 ). Although financial ratios remain a foundational tool, their predictive power alone is often insufficient due to changes in the nature of fraud over time and evolving business practices (Bhavani & Amponsah, 2017 ; Somayyeh, 2015 ). As such, integrating non-financial information, such as audit firm characteristics and ownership structure, has proven useful in improving detection accuracy and contextualizing risk. To ensure a robust and representative feature set, this study uses 50 input variables, including 47 financial indicators and 3 non-financial attributes. These were drawn from two high-quality sources: FiinPro Platform and Kreston Vietnam (details of features are shown in Table 5 in Appendix 1 ). The FiinPro Platform, developed by StoxPlus (a joint venture with Nikkei), is one of the most reputable financial databases in Vietnam. It provides standardized, audited, and time-consistent financial data covering over 1,000 listed firms. Its widespread use by analysts, fund managers, and researchers enhances the credibility and consistency of the information. In addition, the data was cross-checked and cleaned to minimize errors, missing values, or inconsistencies. Kreston Vietnam, a member firm of Kreston Global (a top 13 global accounting network), was selected for supplementary data because of its access to audited reports, market insights, and classification expertise. The inclusion of these two data sources ensures data reliability, regulatory alignment, and domain relevance, particularly in the Vietnamese context where publicly available fraud data is limited. Given the relatively large number of financial variables (over 80 available), the study adopts LASSO (Least Absolute Shrinkage and Selection Operator) regression as a feature selection method. LASSO is an embedded technique widely used in machine learning to reduce dimensionality, eliminate multicollinearity, and improve model generalizability (Muthukrishnan & Rohini, 2016 ; Tibshirani, 1996 ). It applies a penalty to the absolute size of regression coefficients, forcing less significant ones to shrink to zero, thereby identifying the most predictive variables. This improves model interpretability and avoids overfitting. LASSO is particularly suited to financial applications and has shown strong performance in fraud prediction and bankruptcy forecasting (Liu et al., 2021 ; Paraschiv et al., 2023 ; Wang & Liu, 2020 ). In summary, the combination of verified data from FiinPro and Kreston Vietnam, careful labeling aligned with auditing standards, and feature selection via LASSO ensures a reliable, structured, and interpretable dataset for predicting material misstatements in Vietnamese financial statement. 1.2.3 Data processing and Model development To ensure the development of a reliable and generalizable predictive model, data preprocessing and model construction were carried out in several stages, as illustrated in Fig. 1 . First, the dataset was partitioned into independent training and testing sets in an 80:20 ratio to prevent data leakage during evaluation. The training set was then used for model fitting and hyperparameter tuning. A 10-fold stratified cross-validation procedure was employed to estimate model performance. Within each fold, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to address the class imbalance, where the positive class (material misstatements) accounted for approximately 16% of all observations. Stratification ensured that both training and validation folds preserved the original class distribution, thereby providing cross-validation results that more accurately reflect the generalization error. Bayesian optimization was used during cross-validation to identify the optimal hyperparameters. The final model was retrained on the entire training set with SMOTE applied to the training data, and its performance was evaluated on the independent test set using metrics derived from the confusion matrix (precision, recall, F1-score) and the area under the receiver operating characteristic curve (AUC). 1.2.4 Choosing the machine learning algorithm Detecting material misstatements in financial statements is a binary classification problem based on labeled historical data, making supervised machine learning a suitable methodological approach (Agrawal & Chatterjee, 2015 ). This study selects Random Forest (RF) and Extreme Gradient Boosting (XGBoost) due to their proven effectiveness in structured, high-dimensional financial data, especially in fraud detection contexts (Bao et al., 2020 ; Cheng et al., 2021 ; Gupta & Mehta, 2021 ). Both are ensemble learning algorithms that aggregate decision trees to improve model stability and accuracy. RF uses bagging to reduce variance, while XGBoost applies boosting with regularization to minimize bias and overfitting (Pham et al., 2021 ). These methods are particularly effective at capturing complex nonlinear interactions between financial and non-financial variables, which are common in financial misstatement data (Lokanan et al., 2019 ; Ravisankar et al., 2011 ). Additionally, both algorithms provide built-in feature importance metrics, allowing researchers and auditors to interpret the relative influence of input variables—a critical factor in auditing where model transparency is required (Gilpin et al., 2018 ). Alternative machine learning models were considered but not selected for this study. Logistic regression, although widely used, underperformed in our benchmark analysis and is constrained by assumptions of linearity and independence that may not hold in complex financial environments (Hajek & Henriques, 2017 ). Support Vector Machines (SVM) and Artificial Neural Networks (ANN) offer high accuracy but lack interpretability and can be computationally demanding for large datasets, limiting their usefulness in audit contexts where explainability is essential (Dickinson & Meyer, 2022 ; Rane et al., 2023a ). Simpler classifiers such as k-Nearest Neighbors (k-NN), Naïve Bayes, or single decision trees often suffer from poor generalization and sensitivity to outliers (Huang & Huang, 2023 ). Although LightGBM is a strong competitor to XGBoost, the latter was chosen due to its broader empirical validation in financial fraud literature and more extensive integration with interpretability tools (Xu et al., 2022 ). Therefore, Random Forest and XGBoost offer the most appropriate combination of accuracy, efficiency, and transparency for detecting material misstatements in the Vietnamese financial reporting landscape. 1.2.5 Model performance evaluation Confusion matrix metrics Each machine learning model is described by a set of model parameters. The job of a machine learning algorithm is to find the optimal model parameters, and this is closely related to evaluation metrics. It is the process of finding model parameters so that the evaluation metrics achieve the best results. There are many methods to evaluate classifier model performance from different perspectives depending on the research objectives (Hand, 2012 ). In which, good prediction results can be understood as having few misclassified data points (Hand, 2012 ; Tharwat, 2020 ). This evaluation aspect can be done through the Confusion matrix. It is a matrix that describes the classification results. Table 2 shows that confusion matrix is commonly used to calculate classification model evaluation indicators such as Accurary_score; Precision_score, Recall_score, and F1_score (Hand, 2012 ; Tharwat, 2020 ). Table 2 Confusion matrix Actual value Actual value Negative Positive Predicted value Negative True negative (TN) False negative (FN) Predicted value Positive False Positive (FP) True Positive (TP) Source: Authors’ calculation Note: The contents of the Confusion Matrix are interpreted as follows: "Positive": There is a material misstatement in the financial statements "Negative": There are no material misstatements in the financial statements. True Positive (TP): predicted and actual values are both positive False Positive (FP): the predicted value is positive while the actual value is negative True negative (TN): predicted and actual values are both negative False negative (FN): the predicted value is negative while the actual value is positive Accuracy score = (TP + TN)/Sample Precision score = TP/(TP + FP) Recall score = TP/(TP + FN) F1 score = (2 x Precision score x Recall score)/(Precision score + Recall score). Area Under the Receiver Operating Characteristic Curve (AUC) The receiver operating characteristic (ROC) curve is a graphical tool for visualizing the performance of a binary classifier by plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) at various classification thresholds. It illustrates the trade-off between detecting positive instances and avoiding false positives across different thresholds. The area under the ROC curve (AUC) provides a single scalar value summarizing the classifier’s ability to discriminate between the positive class (material misstatements) and the negative class (non-material misstatements). A higher AUC indicates better discriminative performance, making it a particularly valuable metric in imbalanced data contexts where accuracy alone can be misleading. The TPR (sensitivity or recall) measures the proportion of actual positives correctly identified: \(\:TPR=\raisebox{1ex}{$TP$}\!\left/\:\!\raisebox{-1ex}{$(TP+FN)$}\right.\) The FPR measures the proportion of actual negatives incorrectly classified as positives: $$\:FPR=\raisebox{1ex}{$FP$}\!\left/\:\!\raisebox{-1ex}{$(FP+TN)$}\right.$$ In this study, ROC–AUC is used to evaluate the performance of Random Forest (RF), XGBoost, and Logistic Regression, all of which produce predicted probabilities rather than just binary labels. For RF, the probability of the positive class is the average proportion of trees voting for that class; for XGBoost, it is computed from the additive ensemble of decision trees trained sequentially via gradient boosting. Logistic Regression produces probabilities through the logistic function, making it inherently well-suited for ROC analysis. By comparing AUC values across models, we can objectively assess their ability to distinguish between firms with and without material misstatements in financial statements. 4. Classification Results and Discussion 1.3 Classification results Model training and optimization were conducted on the training set using stratified cross-validation, SMOTE, and Bayesian optimization. The final predictive performance of the three best-performing classifiers was evaluated on an independent hold-out test set, with results summarized in Table 3 and illustrated by their ROC curves. XGBoost achieved the highest overall accuracy (0.8391) and precision (0.8262), as well as the highest AUC–ROC value (0.91). These results indicate its strong ability to correctly identify financial statements containing material misstatements while minimizing false positives—an important advantage in audit settings where unnecessary investigations can be costly. Random Forest achieved the highest recall (0.8142) and the second-highest F1-score (0.8031), along with an AUC–ROC of 0.90. This suggests that it is particularly effective at detecting the majority of misstatements, which is critical in minimizing the risk of overlooking materially misstated reports. Although Logistic Regression recorded the lowest performance across all metrics (accuracy = 0.7871, precision = 0.7523, recall = 0.7419, F1-score = 0.7471, AUC–ROC = 0.85), it remains valuable due to its high interpretability and ease of explaining predictions to stakeholders. As shown in Fig. 2 , the ROC curve for the Logistic Regression model demonstrates its moderate discriminative ability compared to the other classifiers. This transparency can be crucial in regulatory and audit contexts where justifying model decisions is as important as predictive accuracy. Overall, these results reinforce the evidence that ensemble-based machine learning algorithms—particularly XGBoost and Random Forest—can deliver superior predictive performance compared to traditional statistical methods, making them promising tools for risk-based audit planning and fraud detection. Figure 3 compares the ROC curves of the Random Forest and XGBoost models, highlighting their higher discriminative performance relative to Logistic Regression. Table 3 Prediction performance evaluation score Classifier Accuracy score Precision score Recall score F1- score Logistic regression 0.7871 0.7523 0.7419 0.7471 Random Forest 0.8309 0.7924 0.8142 0.8031 XGBoost 0.8391 0.8262 0.7855 0.8054 Source: Authors’ calculation Source: Authors’ calculation To enhance both model efficiency and interpretability, this study employed two distinct feature selection and explanation techniques. First, LASSO regression was utilized in the pre-modeling phase for input attribute selection. This embedded method effectively reduced dimensionality by shrinking the coefficients of less informative variables to zero, thereby identifying a core subset of predictors with the highest relevance to material misstatement detection. The LASSO procedure improved generalizability, minimized overfitting, and strengthened the theoretical validity of the model by retaining variables with meaningful associations to fraud-related behavior. Subsequently, feature importance values from the trained Random Forest model were extracted to interpret which attributes contributed most significantly to the classification decisions. This post-modeling analysis enables explainability in the output dimension, revealing the variables that most influenced the model's performance across key evaluation metrics. Table 4 Important features (20 features) Ordinal number Features Important score Note 1 UPCoM 0.139723 New features with LASSO regression 2 LAST MISSTATE 0.096867 3 SIZE (ln(TA)) 0.087433 4 Ln(TL) 0.044385 5 Planed_NeP 0.041723 New features with LASSO regression 6 HOSE 0.039551 New features with LASSO regression 7 Planned_EBIT 0.035057 New features with LASSO regression 8 ln(CASH) 0.027775 9 ROA 0.023334 10 EBTtTA 0.023301 11 P/S 0.018986 12 ROE 0.018181 13 EtA 0.016182 14 REtA 0.015744 15 EPS_G 0.015028 New features with LASSO regression Source: Authors’ calculation As shown in Table 4 , the most influential variable was UPCoM, a dummy variable indicating listing on Vietnam’s Unlisted Public Company Market, which generally has less stringent disclosure requirements. This finding underscores the elevated risk profile of firms operating in less regulated environments and highlights the need for enhanced regulatory oversight. Interestingly, another dummy variable, HOSE—the largest stock exchange in Vietnam—also ranked sixth in importance, suggesting that listing venue characteristics can influence the detection of financial reporting errors. The second most important feature was LAST MISSTATE, reflecting a firm’s prior history of material misstatements. This result is consistent with previous studies (Lou & Wang, 2009 ) and supports the ACFE’s (2022) observation that financial fraud often persists over extended periods before detection. Firm size (SIZE), measured as the natural logarithm of total assets, ranked third, aligning with prior (Beneish, 1999 ; Cheng et al., 2021 ; Hajek, 2019 ), which suggests that large firms may have more complex reporting structures and greater incentives or opportunities for earnings manipulation. Among the LASSO-selected features, several forward-looking performance indicators, such as Planned EBIT, Planned Net Profit (Planned_NeP), and EPS Growth (EPS_G), emerged as significant predictors. These metrics capture pressures to meet projected performance targets and may serve as early warning indicators of earnings management. Their inclusion highlights the value of incorporating strategic financial planning indicators, beyond static ratios, into predictive fraud detection models. Traditional financial ratios, including Ln(Total liabilities), Ln(CASH), Return on Assets (ROA), Earnings Before Tax to Total Assets (EBTtTA), Return on Equity (ROE), Equity to Total Assets (EtA), and Trade Receivables to Total Assets (REtA), also remained relevant predictors. These variables reflect structural composition and balance sheet integrity, consistent with earlier findings in the fraud detection literature (Beneish, 1999 ; Dechow et al., 2011 ; Ravisankar et al., 2011 ). Overall, the classification results reflect a balanced integration of historical indicators, governance-related variables, forward-looking performance metrics, and conventional financial ratios. The dual application of LASSO and Random Forest feature importance provides complementary interpretability: LASSO strengthens feature selection and model efficiency, while Random Forest clarifies output-level feature impact. Together, these methods enhance the transparency, reliability, and practical applicability of the predictive models in financial statement auditing and regulatory risk assessment. 1.4 Discussing machine learning transparency Machine learning has become a powerful tool in financial analytics, enabled by the rapid growth of data availability and computing capacity. These algorithms excel at uncovering complex patterns from structured and unstructured data without requiring explicitly programmed rules (de Laat, 2018 ; Lepri et al., 2017 ). However, this power often comes at the expense of transparency. Many machine learning models, particularly ensemble methods like Random Forest and XGBoost, are viewed as “black boxes” because it can be challenging to understand how they arrive at their predictions (Mitchell et al., 2022 ; Rane et al., 2023a ). This lack of interpretability can limit stakeholder trust and hinder adoption in highly regulated domains like financial auditing. A key way to address this issue is through explainability, which refers to techniques that clarify how a machine learning model makes decisions. Explainability helps stakeholders—auditors, regulators, and investors—understand not only what a model predicts but why it does so (Gilpin et al., 2018 ). Models such as linear regression or decision trees are inherently interpretable, offering a clear mapping between input features and outcomes (Peng et al., 2021 ). However, their simplicity often makes them inadequate for capturing the non-linear relationships that are common in financial data. Ensemble models like Random Forest and XGBoost offer superior predictive accuracy because they can model such complexity (Huang & Huang, 2023 ; Xu et al., 2022 ). Nonetheless, their complexity necessitates interpretation tools like feature importance scores to enhance transparency. Machine learning algorithms generate predictions from sample data without explicit instructions from the user (Huang & Huang, 2023 ). This has led to them being considered “black boxes” due to the challenge of understanding how machine learning makes its predictions (Rane et al., 2023b ). As a result, the interpretability, reliability, and effectiveness of machine learning models are often difficult to assess (Mitchell et al., 2022 ). This issue is discussed by researchers in terms of machine learning transparency in decision-making, including algorithmic transparency. In this study, feature importance values derived from the Random Forest algorithm were used to assess which variables most significantly influenced the model's predictions of material misstatements. This approach contributes directly to machine learning transparency by identifying and communicating the relative weight of each input in the model’s decision-making process (Munkhdalai et al., 2019 ; Sonkavde et al., 2023 ). Among the top-ranked features, UpCoM, LAST_MISSTATE, and SIZE emerged as dominant predictors. Their practical implications are significant: firms listed on the less-regulated UpCoM exchange, companies with a history of misstatements, and larger enterprises with complex structures are more likely to engage in or obscure material misstatements. These findings align with audit theory and support targeted, risk-based audit planning. A further contribution to model transparency comes from understanding and communicating the predictive performance metrics used. The comparative analysis of the three classifiers reveals important trade-offs in model selection for detecting material misstatements in financial statements. Random Forest, with its highest recall (0.8142), offers a clear advantage when the primary objective is to minimize false negatives—cases where financial statements containing material misstatements are incorrectly classified as not misstated. From an auditing and risk management perspective, failing to detect a material misstatement can have severe regulatory and financial consequences. The ensemble characteristic of Random Forest, leveraging multiple decision trees with bootstrap aggregation, likely contributed to its strong sensitivity by capturing non-linear relationships between predictor features and the occurrence of material misstatements. Conversely, XGBoost exhibited the strongest precision (0.8262) and the highest discriminative ability (AUC-ROC = 0.91), indicating its capacity to effectively identify true positive cases while minimizing false positives. This model enables auditors to focus resources on cases that are indeed materially misstated. The gradient boosting mechanism of XGBoost, combined with optimized hyperparameters via Bayesian search, likely enhanced its ability to capture complex feature interactions and prioritize hard-to-classify cases. Although logistic regression yielded comparatively lower predictive performance, it retains value due to its interpretability, stemming from its linear nature. In situations where the ability to provide clear, transparent justification for predictions is as critical as accuracy, logistic regression remains an attractive supplementary tool. Many of the most influential variables were selected through LASSO regression, which enhances both predictive performance and interpretability. Features such as Planned EBIT, Planned Net Profit, and EPS Growth (EPS_G) reflect strategic financial planning behavior, suggesting that attempts to manage or project strong performance may be early signals of misstatement risk. These indicators, derived from forward-looking or growth-oriented metrics, suggest that firms under performance pressure may be more prone to earnings manipulation. Their inclusion highlights the importance of not only reviewing historical results but also understanding firms’ projected targets and capital efficiency when assessing audit risk. In summary, this study demonstrates the effectiveness of machine learning in predicting material misstatements and contributes to the field of explainable AI by clearly identifying which features drive model decisions and by interpreting the practical implications of key performance metrics. Such transparency supports trust, fosters informed stakeholder engagement, and facilitates the adoption of machine learning tools in financial statement auditing in Vietnam and beyond. 1.5 Implications for the auditing context While corporate management bears primary responsibility for the accuracy of financial reporting, stakeholders, including investors and regulators, are increasingly holding auditors accountable for failing to detect material misstatements (Hunt et al., 2022 ). In this context, integrating machine learning (ML) into audit practices offers a promising pathway to enhance audit quality, efficiency, and risk detection. Traditional auditing, particularly analytical procedures, has relied heavily on manual reviews and statistical comparisons. However, with the rise of Big Data and advanced analytics, auditors now have access to tools capable of processing large volumes of financial and non-financial data, uncovering subtle anomalies that might otherwise go undetected (Appelbaum et al., 2017 ; Warren et al., 2015 ). This study demonstrates that ML-based analytical procedures can predict material misstatements with higher accuracy than traditional statistical methods. Both Random Forest (RF) and XGBoost achieved strong accuracy (0.8309–0.8391), correctly classifying most firms regardless of misstatement status. XGBoost’s precision (0.8262) indicates that over four out of five flagged cases were indeed misstated, enabling auditors to focus resources on the most probable risks. RF’s recall (0.8142) shows it effectively identifies most actual misstatements, thereby reducing the risk of undetected fraud. The choice between these models should be based on the auditing firm’s tolerance for Type I and Type II errors: RF is preferable when avoiding undetected misstatements is the priority, while XGBoost is advantageous for optimizing resource allocation. A hybrid approach, using XGBoost for primary screening and RF for secondary verification, could combine their strengths, reducing false alarms while maintaining high detection capability. Feature importance analysis further highlights key audit risk indicators. UPCoM listing status ranked first, indicating that firms on Vietnam’s less-regulated secondary exchange present higher risks. LAST MISSTATE, a firm’s history of misreporting, ranked second, reinforcing the importance of incorporating prior audit outcomes into current risk assessments. Firm size (SIZE) ranked third, aligning with literature linking larger, more complex firms to greater misstatement risk. LASSO regression also identified forward-looking indicators such as Planned EBIT, Planned Net Profit (Planned_NeP), and EPS Growth (EPS_G), suggesting that performance pressures can trigger earnings management. Traditional financial ratios, including Ln(Total liabilities), Return on Assets (ROA), Return on Equity (ROE), and Earnings Before Tax to Total Assets (EBTtTA), also showed strong predictive value, reflecting both financial structure and leverage-related risks. These findings imply that auditors should balance their focus between historical results, strategic performance plans, and financial position when assessing misstatement risk. In the Vietnamese auditing context, where direct fraud data are scarce, this study shows that ML can bridge data gaps by using proxy indicators like material misstatements. Feature selection via LASSO ensures attention to the most relevant variables, supporting risk-based audit planning. As audit firms invest in AI-powered systems, integrating these insights can enhance audit quality, regulatory compliance, and stakeholder trust. In an increasingly complex reporting environment, adopting data-driven, transparent, and adaptive tools will be essential for maintaining integrity, accountability, and investor confidence in capital markets. Declarations Ethical Approval : “This article does not contain any studies with human participants performed by any of the authors.” Informed Consent : “This article does not contain any studies with human participants performed by any of the authors.” Funding: This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Author Contribution Thi Thanh Hoa Tieu: Conceptualization, methodology, formal analysis, writing, original Draft.Ngoc Hung Tran: Data curation, investigation, writing – original draft.Hien Thanh Hoang: Supervision, validation, project administration, writing – review & editing. Data Availability All relevant data are within the manuscript and its supporting information files. The full computational code for attribute selection (LASSO regression) and model implementation, as well as the research results, are openly available via the following Google Colaboratory links:LASSO regression and attribute selection:https://colab.research.google.com/drive/1EYgRtMdbQPTJM77fQcdkzNH5Ri-seAnp?usp=sharingResearch resultshttps://colab.research.google.com/drive/1K36ciNvp8KsbQpG0JMMQkAplNmvldtGr?usp=sharingThese resources can also be found in the Supporting Information (S1). Additional data or code are available from the corresponding author upon reasonable request. References Agrawal, K., & Chatterjee, C. (2015). Earnings management and financial distress: Evidence from India. Global Business Review , 16 (5_suppl), 140S-154S. Albrecht, W. S., Albrecht, C., & Albrecht, C. C. (2008). Current Trends in Fraud and its Detection. Information Security Journal: A Global Perspective , 17 (1), 2-12. https://doi.org/10.1080/19393550801934331 Appelbaum, D., Kogan, A., & Vasarhelyi, M. A. (2017). Big Data and Analytics in the Modern Audit Engagement: Research Needs. Auditing: A Journal of Practice & Theory , 36 (4), 1-27. https://doi.org/10.2308/ajpt-51684 Arshad, R., Iqbal, S. M., & Omar, N. (2015). Prediction of Business Failure and Fraudulent Financial Reporting: Evidence from Malaysia. Indian Journal of Corporate Governance , 8 (1), 34-53. https://doi.org/10.1177/0974686215574424 Ashtiani, M. N., & Raahemi, B. (2021). Intelligent fraud detection in financial statements using machine learning and data mining: a systematic literature review. Ieee Access , 10 , 72504-72525. Bao, Y., Ke, B., Li, B., Yu, Y. J., & Zhang, J. (2020). Detecting accounting fraud in publicly traded US firms using a machine learning approach. Journal of Accounting Research , 58 (1), 199-235. Beneish, M. D. (1999). The detection of earnings manipulation. Financial Analysts Journal , 55 (5), 24-36. Bennett, M., Hayes, K., Kleczyk, E. J., & Mehta, R. (2022). Similarities and differences between machine learning and traditional advanced statistical modeling in healthcare analytics. arXiv preprint arXiv:2201.02469 . Bhavani, G., & Amponsah, C. T. (2017). M-Score and Z-Score for detection of accounting fraud. Accountancy Business and the Public Interest , 1 (1), 68-86. Breiman, L. (2001). Random forests. Machine learning , 45 , 5-32. Bzdok, D., Altman, N., & Krzywinski, M. (2018). Statistics versus machine learning. Nat Methods , 15 (4), 233-234. https://doi.org/10.1038/nmeth.4642 Carmichael, I., & Marron, J. (2018). Data science vs. statistics: two cultures? Japanese Journal of Statistics and Data Science , 1 (1), 117-138. Cheng, C.-H., Kao, Y.-F., & Lin, H.-P. (2021). A financial statement fraud model based on synthesized attribute selection and a dataset with missing values and imbalanced classes. Applied Soft Computing , 108 , 107487. de Laat, P. B. (2018). Algorithmic Decision-Making Based on Machine Learning from Big Data: Can Transparency Restore Accountability? Philos Technol , 31 (4), 525-541. https://doi.org/10.1007/s13347-017-0293-z Dechow, P. M., Ge, W., Larson, C. R., & Sloan, R. G. (2011). Predicting Material Accounting Misstatements*. Contemporary Accounting Research , 28 (1), 17-82. https://doi.org/10.1111/j.1911-3846.2010.01041.x Dickinson, Q., & Meyer, J. G. (2022). Positional SHAP (PoSHAP) for Interpretation of machine learning models trained from biological sequences. PLOS Computational Biology , 18 (1), e1009736. Gaganis, C. (2009). Classification techniques for the identification of falsified financial statements: a comparative analysis. Intelligent Systems in Accounting, Finance and Management , 16 (3), 207-229. https://doi.org/10.1002/isaf.303 Gilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., Specter, M., & Kagal, L. (2018). Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th International Conference on data science and advanced analytics (DSAA), Green, B. P., & Choi, J. H. (1997). Assessing the risk of management fraud through neural network technology. Auditing , 16 , 14-28. Gupta, S., & Mehta, S. K. (2021). Data Mining-based Financial Statement Fraud Detection: Systematic Literature Review and Meta-analysis to Estimate Data Sample Mapping of Fraudulent Companies Against Non-fraudulent Companies. Global Business Review , 25 (5), 1290-1313. https://doi.org/10.1177/0972150920984857 Gupta, S., & Mehta, S. K. (2024). Data mining-based financial statement fraud detection: Systematic literature review and meta-analysis to estimate data sample mapping of fraudulent companies against non-fraudulent companies. Global Business Review , 25 (5), 1290-1313. Hajek, P. (2019). Interpretable Fuzzy Rule-Based Systems for Detecting Financial Statement Fraud. 559 , 425-436. https://doi.org/10.1007/978-3-030-19823-7_36 Hajek, P., & Henriques, R. (2017). Mining corporate annual reports for intelligent detection of financial statement fraud–A comparative study of machine learning methods. Knowledge-Based Systems , 128 , 139-152. Hand, D. J. (2012). Assessing the Performance of Classification Methods. International Statistical Review , 80 (3), 400-414. https://doi.org/10.1111/j.1751-5823.2012.00183.x Herawati, N. (2015). Application of Beneish M-Score models and data mining to detect financial fraud. Procedia-Social and Behavioral Sciences , 211 , 924-930. Hołda, A. (2020). Using the Beneish M-score model: Evidence from non-financial companies listed on the Warsaw Stock Exchange. Investment Management and Financial Innovations , 17 (4), 389-401. https://doi.org/10.21511/imfi.17(4).2020.33 Huang, A. A., & Huang, S. Y. (2023). Increasing transparency in machine learning through bootstrap simulation and shapely additive explanations. PLoS One , 18 (2), e0281922. https://doi.org/10.1371/journal.pone.0281922 Hunt, E., Hunt, J., Richardson, V. J., & Rosser, D. (2022). Auditor response to estimated misstatement risk: A machine learning approach. Accounting Horizons , 36 (1), 111-130. Jan, C.-l. (2018). An Effective Financial Statements Fraud Detection Model for the Sustainable Development of Financial Markets: Evidence from Taiwan. Sustainability , 10 (2), 513. https://doi.org/10.3390/su10020513 Kaminski, K. A., Sterling Wetzel, T., & Guan, L. (2004). Can financial ratios detect fraudulent financial reporting? Managerial Auditing Journal , 19 (1), 15-28. Kanapickienė, R., & Grundienė, Ž. (2015). The Model of Fraud Detection in Financial Statements by Means of Financial Ratios. Procedia - Social and Behavioral Sciences , 213 , 321-327. https://doi.org/10.1016/j.sbspro.2015.11.545 Kirkos, E., Spathis, C., & Manolopoulos, Y. (2007). Data Mining techniques for the detection of fraudulent financial statements. Expert Systems with Applications , 32 (4), 995-1003. https://doi.org/10.1016/j.eswa.2006.02.016 Lakhan, A., Mohammed, M. A., Ibrahim, D. A., Kadry, S., & Abdulkareem, K. H. (2022). ITS based on deep graph convolutional fraud detection network blockchain-enabled fog-cloud. IEEE Transactions on Intelligent Transportation Systems , 24 (8), 8399-8408. Lepri, B., Oliver, N., Letouzé, E., Pentland, A., & Vinck, P. (2017). Fair, Transparent, and Accountable Algorithmic Decision-making Processes. Philosophy & Technology , 31 (4), 611-627. https://doi.org/10.1007/s13347-017-0279-x Liu, Z., Ye, R., & Ye, R. (2021). https://doi.org/10.21203/rs.3.rs-640038/v1 Lokanan, M., Tran, V., & Vuong, N. H. (2019). Detecting anomalies in financial statements using machine learning algorithm: The case of Vietnamese listed firms. Asian Journal of Accounting Research , 4 (2), 181-201. Lou, Y.-I., & Wang, M.-L. (2009). Fraud risk factor of the fraud triangle assessing the likelihood of fraudulent financial reporting. Journal of Business & Economics Research (JBER) , 7 (2). Marais, A., Vermaak, C., & Shewell, P. (2023). Predicting financial statement manipulation in South Africa: A comparison of the Beneish and Dechow models. Cogent Economics & Finance , 11 (1). https://doi.org/10.1080/23322039.2023.2190215 Mitchell, R., Frank, E., & Holmes, G. (2022). GPUTreeShap: massively parallel exact calculation of SHAP scores for tree ensembles. PeerJ Comput Sci , 8 , e880. https://doi.org/10.7717/peerj-cs.880 Munkhdalai, L., Munkhdalai, T., Namsrai, O.-E., Lee, J. Y., & Ryu, K. H. (2019). An empirical comparison of machine-learning methods on bank client credit assessments. Sustainability , 11 (3), 699. Muthukrishnan, R., & Rohini, R. (2016). LASSO: A feature selection technique in predictive modeling for machine learning. 2016 IEEE international conference on advances in computer applications (ICACA), Omar, N., Johari, Z. A., & Smith, M. (2017). Predicting fraudulent financial reporting using artificial neural network. Journal of Financial Crime , 24 (2), 362-387. Paraschiv, F., Schmid, M., & Wahlstrøm, R. R. (2023). Bankruptcy prediction of privately held SMEs using feature selection methods. Available at SSRN 4588767 . Peng, C.-L., Chen, W.-K., & Wei, A.-P. (2021). Teaching CAPM for a Pre-Finance Graduate Program at the STEM Undergraduate Level: Linear Algebra Perspective. Mathematics , 9 (14). https://doi.org/10.3390/math9141668 Pham, K., Kim, D., Park, S., & Choi, H. (2021). Ensemble learning-based classification models for slope stability analysis. Catena , 196 . https://doi.org/10.1016/j.catena.2020.104886 Rane, N., Choudhary, S., & Rane, J. (2023a). Explainable Artificial Intelligence (XAI) approaches for transparency and accountability in financial decision-making. Available at SSRN 4640316 . Rane, N., Choudhary, S., & Rane, J. (2023b). Explainable Artificial Intelligence (XAI) Approaches for Transparency and Accountability in Financial Decision-Making. SSRN Electronic Journal . https://doi.org/10.2139/ssrn.4640316 Ravisankar, P., Ravi, V., Rao, G. R., & Bose, I. (2011). Detection of financial statement fraud and feature selection using data mining techniques. Decision support systems , 50 (2), 491-500. Shoetan, P. O., Oyewole, A. T., Okoye, C. C., & Ofodile, O. C. (2024). Reviewing the role of big data analytics in financial fraud detection. Finance & Accounting Research Journal , 6 (3), 384-394. Somayyeh, H. N. (2015). Financial ratios between fraudulent and non-fraudulent firms: Evidence from Tehran Stock Exchange. Journal of Accounting and Taxation , 7 (3), 38-44. https://doi.org/10.5897/jat2014.0166 Sonkavde, G., Dharrao, D. S., Bongale, A. M., Deokate, S. T., Doreswamy, D., & Bhat, S. K. (2023). Forecasting Stock Market Prices Using Machine Learning and Deep Learning Models: A Systematic Review, Performance Analysis and Discussion of Implications. International Journal of Financial Studies , 11 (3). https://doi.org/10.3390/ijfs11030094 Tharwat, A. (2020). Classification assessment methods. Applied Computing and Informatics , 17 (1), 168-192. https://doi.org/10.1016/j.aci.2018.08.003 Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B: Statistical Methodology , 58 (1), 267-288. Wang, W., & Liu, F. (2020). Research on Identification of Fraud in Financial Reporting Based on Adaptive Group Lasso-Logistic Regression 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), Warren, J. D., Moffitt, K. C., & Byrnes, P. (2015). How big data will change accounting. Accounting Horizons , 29 (2), 397-407. Xu, Y., Ye, W., Song, Q., Shen, L., Liu, Y., Guo, Y.,…Sun, X. (2022). Using machine learning models to predict the duration of the recovery of COVID-19 patients hospitalized in Fangcang shelter hospital during the Omicron BA. 2.2 pandemic. Frontiers in Medicine , 9 , 1001801. Zainudin, E. F., & Hashim, H. A. (2016). Detecting fraudulent financial reporting using financial ratio. Journal of Financial Reporting and Accounting , 14 (2), 266-278. https://doi.org/10.1108/jfra-05-2015-0053 Additional Declarations No competing interests reported. Supplementary Files APPENDIX1.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7360630","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":518495690,"identity":"d274374b-5294-4a4c-851f-a41c96ceb78a","order_by":0,"name":"Hoa Thi Thanh Tieu","email":"","orcid":"","institution":"Industrial University of Ho Chi Minh City","correspondingAuthor":false,"prefix":"","firstName":"Hoa","middleName":"Thi Thanh","lastName":"Tieu","suffix":""},{"id":518495692,"identity":"bcba7ec2-be9f-4fa6-aab4-14f0d7a671c7","order_by":1,"name":"Thanh Hien Hoang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAy0lEQVRIie3OMQrCMBTG8ZQHz+VJ14rgGR4UKqKHyWRv0FVB6FRwrbfwCJVARj2AgorQuZM4aDEFFwdN3RzyhwyB94NPCJfrD2MQArp1PVtCYb7YlhAX3iqVbYl4kbVuS4Yd0BfiA4TaLKsSJXg//05GC4zDO5cYaZRevjXkUFiGKYr6xEDRMWPopkr0cmkj/rUhQZj6FTzaEUJDFDOSAM8QP7ASbIZNZaCRN9k2JjvZqdKQifRTOJ9uyXiANvJWYR79cO9yuVyuTz0BZ8c8TW5M7EQAAAAASUVORK5CYII=","orcid":"","institution":"FPT University","correspondingAuthor":true,"prefix":"","firstName":"Thanh","middleName":"Hien","lastName":"Hoang","suffix":""},{"id":518495694,"identity":"c1013b91-cbb6-44c6-a0bc-83ccfd4bbce5","order_by":2,"name":"Hung Ngoc Tran","email":"","orcid":"","institution":"Industrial University of Ho Chi Minh City","correspondingAuthor":false,"prefix":"","firstName":"Hung","middleName":"Ngoc","lastName":"Tran","suffix":""}],"badges":[],"createdAt":"2025-08-13 04:23:10","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7360630/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7360630/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":92610262,"identity":"c71622da-b071-4e59-9177-1a596d52b69b","added_by":"auto","created_at":"2025-10-01 16:09:51","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":210354,"visible":true,"origin":"","legend":"","description":"","filename":"ManuscriptV21.docx","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/39e4b366327ddbebab255921.docx"},{"id":92610259,"identity":"1da68ef2-c9f0-4659-a4dd-47806b40e576","added_by":"auto","created_at":"2025-10-01 16:09:50","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5538,"visible":true,"origin":"","legend":"","description":"","filename":"022ec58ac7b04df8a64a70b1015c6baa.json","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/c67500f9da88fcbd5bd0d27e.json"},{"id":92610544,"identity":"a4887830-aa19-4e5c-9ced-35c682274783","added_by":"auto","created_at":"2025-10-01 16:17:51","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":152094,"visible":true,"origin":"","legend":"","description":"","filename":"022ec58ac7b04df8a64a70b1015c6baa1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/a41f2673433df822d4a0b357.xml"},{"id":92610543,"identity":"0157b047-4ce7-49cf-93ae-9bc4d7634a8e","added_by":"auto","created_at":"2025-10-01 16:17:51","extension":"jpeg","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1074,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/d01ecb6b2e17e53279daac9c.jpeg"},{"id":92610261,"identity":"90c76bec-6986-4544-8278-0dd48ce3b8cd","added_by":"auto","created_at":"2025-10-01 16:09:51","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119971,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/a545210565a63c5cce59ac21.jpeg"},{"id":92610548,"identity":"6757de5e-6284-43ca-8a31-d5465deebd98","added_by":"auto","created_at":"2025-10-01 16:17:51","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":237903,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/b4af6aae128851010ee1d0d6.jpeg"},{"id":92610264,"identity":"67a82a8e-9118-4513-b0f9-06bc0850bd9b","added_by":"auto","created_at":"2025-10-01 16:09:51","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":28873,"visible":true,"origin":"","legend":"","description":"","filename":"groupimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/afa6a0f1774b7224ccd0b38b.jpeg"},{"id":92611254,"identity":"ae8c18f9-a9f2-42ce-a01e-f273a98e1d9a","added_by":"auto","created_at":"2025-10-01 16:25:51","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":935,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/95ed919fb78554984148a54b.png"},{"id":92610268,"identity":"6424d32c-cb03-4910-b665-003458a4c9a2","added_by":"auto","created_at":"2025-10-01 16:09:51","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":31196,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/93ed78b91144585ab2eb3ff5.png"},{"id":92611513,"identity":"9455e759-b7e6-4114-a68e-4b2b3afdf85e","added_by":"auto","created_at":"2025-10-01 16:33:51","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":59227,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/eaee711b3bd3a9bcc86d65a1.png"},{"id":92611512,"identity":"e02bb707-ba38-46a4-9ba6-77fa0fb834de","added_by":"auto","created_at":"2025-10-01 16:33:51","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10697,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinegroupimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/2fab63d72c360d6c7e644027.png"},{"id":92610271,"identity":"6dade4d9-69a6-42da-b7f9-0d394d525347","added_by":"auto","created_at":"2025-10-01 16:09:51","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":151665,"visible":true,"origin":"","legend":"","description":"","filename":"022ec58ac7b04df8a64a70b1015c6baa1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/781a5d0e91778dd8d1f686c9.xml"},{"id":92610270,"identity":"d36444bf-f501-47ba-ba9d-b767b1d10e71","added_by":"auto","created_at":"2025-10-01 16:09:51","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":160677,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/e08b5d64a73cb476b98b3cec.html"},{"id":92610255,"identity":"f9d93ace-f112-4e85-b2c1-70e5e4ab4363","added_by":"auto","created_at":"2025-10-01 16:09:50","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":26918,"visible":true,"origin":"","legend":"\u003cp\u003eMODEL DEVELOPMENT\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/bca1e2f5ed97e5b3d9e92102.png"},{"id":92610542,"identity":"9265e155-47ed-4c85-922b-a0121e9a1fc0","added_by":"auto","created_at":"2025-10-01 16:17:50","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":39557,"visible":true,"origin":"","legend":"\u003cp\u003eRECEIVER OPERATING CHARACTERISTIC (LOGISTC REGRESSION)\u003c/p\u003e\n\u003cp\u003eSource: Authors’ calculation\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/e1d412a330f25d03d5b5c132.png"},{"id":92610258,"identity":"dc4b4eba-4840-4fb9-94db-404f613dba7d","added_by":"auto","created_at":"2025-10-01 16:09:50","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":78031,"visible":true,"origin":"","legend":"\u003cp\u003eRECEIVER OPERATING CHARACTERISTIC (RANDOM FOREST AND XGBOOST)\u003c/p\u003e\n\u003cp\u003eSource: Authors’ calculation\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/9eca2be2699b9b02cbff671c.png"},{"id":97421881,"identity":"e8775443-b66b-4c9c-a311-44442a628648","added_by":"auto","created_at":"2025-12-04 08:38:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":945922,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/77a4df65-0d72-4e78-b00f-4c52a24f61d5.pdf"},{"id":92610256,"identity":"405065f0-099c-4b46-97ce-678a0f3f55ea","added_by":"auto","created_at":"2025-10-01 16:09:50","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":18671,"visible":true,"origin":"","legend":"","description":"","filename":"APPENDIX1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7360630/v1/1f1fda3431b0833bce6524fd.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Detecting financial misstatements in emerging markets: a machine learning approach","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eAccurate financial statements are essential for ensuring transparency, enabling informed decision-making, and maintaining stakeholder trust. However, material misstatements, whether caused by unintentional errors or deliberate fraud, can distort the true financial condition of an entity and result in significant harm to investors, regulators, and the broader economy. Auditing standards such as SAS 99 and ISA 240 emphasize the heightened risks associated with fraudulent misstatements due to their intentional and often concealed nature. High-profile corporate failures like Enron and Lehman Brothers underscore the devastating consequences that undetected financial reporting fraud can have.\u003c/p\u003e\u003cp\u003eWhile fraud-related misstatements account for a smaller portion of total occupational fraud cases globally, they are the most financially damaging (ACFE, 2018\u0026ndash;2022). A critical issue is the time lag in detection, averaging two years, which allows manipulative behaviors to continue unchecked (Kaminski et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). Early detection, therefore, is not only desirable but essential for protecting financial systems.\u003c/p\u003e\u003cp\u003eDespite global advancements in fraud analytics, Vietnam faces unique challenges in this domain. Specifically, the country lacks a centralized, publicly accessible database of confirmed financial reporting fraud cases. This data limitation prevents direct fraud analysis and restricts the applicability of models developed in other contexts. To address this gap, the current study focuses on material misstatements as a practical and theoretically grounded proxy for fraudulent behavior. These misstatements, which are defined in alignment with Vietnam Auditing Standard No. 320, offer a reliable signal of potential fraud risk and are more consistently observable within Vietnam's regulatory environment.\u003c/p\u003e\u003cp\u003eThis study addresses a critical research problem: the absence of effective, data-driven tools for detecting financial misstatements in Vietnamese listed companies. The primary objective is to develop and validate a machine learning-based predictive model to identify companies likely to produce materially misstated financial reports. By integrating financial indicators with non-financial contextual variables, the model aims to provide early warning signals to auditors, regulators, and investors.\u003c/p\u003e\u003cp\u003eThis research offers three key contributions. First, it is among the first empirical studies to apply ensemble machine learning algorithms, specifically Random Forest and XGBoost, to predict financial misstatements in the context of Vietnam. Second, it employs LASSO regression to enhance model interpretability and relevance by selecting the most influential features from a large dataset. Third, it responds to a pressing need in the Vietnamese auditing landscape by offering a practical, scalable solution that can be used by auditing firms and regulators for risk-based planning and regulatory oversight.\u003c/p\u003e\u003cp\u003eBy aligning with both international auditing standards and Vietnam\u0026rsquo;s current data landscape, this study contributes to strengthening financial integrity and audit quality in an emerging market context. It supports the broader objective of enhancing transparency, accountability, and investor confidence in Vietnam\u0026rsquo;s capital markets.\u003c/p\u003e"},{"header":"2. Literature review","content":"\u003cp\u003eDetecting material misstatements in financial statements\u0026mdash;often symptomatic of fraudulent activity\u0026mdash;has long been a critical concern in accounting and auditing research. Early studies predominantly relied on traditional statistical models, with one of the most widely recognized tools being the Beneish M-score (Beneish, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e1999\u003c/span\u003e). This model uses a set of financial ratios to detect earnings manipulation and has been applied in various national contexts (Arshad et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Herawati, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Hołda, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). However, its predictive accuracy remains modest, typically ranging from 70\u0026ndash;75%, and recent evaluations have called for caution when applying it as a standalone tool (Bhavani \u0026amp; Amponsah, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Marais et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eTo improve fraud detection capabilities, researchers have expanded the scope of predictive inputs. For example, Dechow et al. (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) developed the F-score by integrating off-balance sheet indicators, market performance data, and governance-related variables. Similarly, studies by Kanapickienė and Grundienė (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) and Zainudin and Hashim (\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) demonstrate that combining financial and non-financial variables\u0026mdash;such as ownership concentration, audit opinions, or listing status\u0026mdash;can substantially increase model accuracy.\u003c/p\u003e\u003cp\u003eDespite these improvements, traditional statistical models face limitations in handling nonlinear relationships, high-dimensional data, and multicollinearity. These challenges have prompted a methodological shift toward machine learning (ML) techniques. Groundbreaking work by Green and Choi (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e1997\u003c/span\u003e) introduced neural networks to fraud detection, and subsequent studies have adopted algorithms such as decision trees, support vector machines, and ensemble methods like Random Forest and XGBoost (Gaganis, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Hajek \u0026amp; Henriques, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Jan, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). These models consistently outperform logistic regression, particularly when applied to datasets that incorporate both financial and contextual non-financial data (Kirkos et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Lokanan et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Ravisankar et al., \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2011\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eRecent reviews (Gupta \u0026amp; Mehta, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Shoetan et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e) affirm that ensemble methods deliver superior accuracy and resilience in fraud detection, addressing both nonlinearity and feature selection challenges. These techniques are especially relevant in complex datasets where conventional assumptions do not hold.\u003c/p\u003e\u003cp\u003eWhile the existing literature offers valuable insights, it remains heavily concentrated in developed markets with strong institutional frameworks and accessible fraud data. In contrast, emerging markets like Vietnam face unique challenges, particularly the absence of public datasets on confirmed fraud cases. This significantly limits the applicability of fraud detection frameworks developed in Western or mature economies. Instead, material misstatements\u0026mdash;detected through audit adjustments\u0026mdash;must serve as proxy indicators for fraud risk in data-constrained environments.\u003c/p\u003e\u003cp\u003eIn Southeast Asia, studies have begun to explore the utility of machine learning in fraud detection. Lokanan Lokanan et al. (\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) applied ML to Vietnamese firms and found improved accuracy over statistical models. Similarly, Omar et al. (\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) demonstrated the predictive power of neural networks in the Malaysian context. Research from Indonesia and Thailand supports the adaptability of ML models to regional data environments with similar regulatory and reporting constraints.\u003c/p\u003e\u003cp\u003eDespite this growing interest, Vietnam remains underrepresented in empirical research. Most prior studies either focus on global datasets or lack contextual adaptation for Vietnam\u0026rsquo;s regulatory structure, audit practices, and data access limitations. This study seeks to fill that gap by developing a Vietnam-specific machine learning model that uses both financial and non-financial variables to predict material misstatements, thereby offering practical applications for auditors, regulators, and corporate governance professionals operating in transitional economies.\u003c/p\u003e"},{"header":"3. Research methodology","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e1.1 Machine learning approach\u003c/h2\u003e\u003cp\u003eAccording to (Bennett et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e), data scientists and statisticians are often inconsistent when determining the best approach: machine learning or statistical modeling to solve an analytical challenge. However, machine learning and statistical modeling are complementary because they are all based on mathematical principles. The difference is only in the use of instruments for the overall analysis process. The choice of both approaches or just one depends on the problem to be solved and the results to be achieved as well as the characteristics of the data and the context of the analysis. The determination of the approach is mainly based on empirical evidence, such as the size and completeness of the data, the number of variables, whether the assumptions are warranted, and whether the expected results are predictive or causal (Bennett et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eStatistical modeling typically makes assumptions that predictors or characteristics are known, models are parametric, and that testing research hypotheses and uncertainty are paramount (Breiman, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2001\u003c/span\u003e), meanwhile machine learning does not make these assumptions (Bennett et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). In machine learning, many models are based on non-parametric approaches, in which the structure of the model is unspecified or unknown, and assumptions about normal distribution, linearity, or residuals are not required for modeling (Carmichael \u0026amp; Marron, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe purpose of machine learning is to focus on predictive performance by using learning algorithms to find little-known, unrelated, and complex patterns in data without prior insight into the underlying structures (Carmichael \u0026amp; Marron, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Whereas, in statistical modeling, the motivation is to consider inferences, correlations, and effects of a small number of variables (Breiman, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2001\u003c/span\u003e). On the same page, (Bzdok et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) observed that statistical modeling draws inferences about the population from a sample of studies while machine learning finds generalizable predictive patterns from data.\u003c/p\u003e\u003cp\u003eA literature review shows that predictive models are built on a wide range of features (variables) including both financial indicators and non-financial information. In the early stages, researchers mainly applied statistical models. Logistic regression was widely used; however, the predictive performance was relatively low because of the tight constraints of assumptions that may not be achieved for the data characteristics and the relatively large number of features of this type of research. Although statistical models can detect simple fraud cases effectively, they have difficulty with the increasing complexity and sophistication of modern fraud schemes (Shoetan et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Machine learning approaches overcome this problem by analyzing large amounts of financial data to detect patterns and anomalies that may indicate fraud (Ashtiani \u0026amp; Raahemi, \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Lakhan et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Therefore, many studies have approached machine learning methods in identifying fraud in financial statements and have achieved better results than traditional statistical methods.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e1.2 Data\u003c/h2\u003e\u003cdiv id=\"Sec6\" class=\"Section3\"\u003e\u003ch2\u003e1.2.1 Labelling\u003c/h2\u003e\u003cp\u003eThe dependent variable in this study is the presence of material misstatements in financial statements, serving as a proxy for fraudulent reporting. Following auditing best practices and standards (Vietnam Auditing Standard No. 320; VACPA sample guidelines), a material misstatement is defined as an absolute change of 10% or more in profit after tax (PAT) between pre-audited and post-audited financial reports. Profit was selected as the basis for measuring misstatements because it comprehensively reflects fluctuations in revenue, expenses, assets, and liabilities.\u003c/p\u003e\u003cp\u003eThe categorical variable \u003cb\u003e\"Misstatements\"\u003c/b\u003e was constructed using data from the FinnPro platform, which aggregates annual pre- and post-audit PAT for all companies listed on Vietnamese stock exchanges. Leveraging FinnPro enables cost-effective data collection and preprocessing while ensuring a high degree of reliability and transparency. Observations where the absolute value of the profit variance ratio is greater than or equal to 10% are labeled \u003cb\u003e1\u003c/b\u003e (material misstatement), and those below this threshold are labeled \u003cb\u003e0\u003c/b\u003e. This computation method is straightforward, consistent with both international and local auditing guidance, and therefore valid as a proxy indicator for financial manipulation (see Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eMeasurement of categorical variable\u003c/span\u003e\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCategorical variable\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eVariable name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eMeasurement\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMaterial misstatement in financial statements\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMisstatement\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eBinary variable\u003c/p\u003e\u003cp\u003eMisstatement\u0026thinsp;=\u0026thinsp;1: The absolute value of the Profit variance Ratio is greater than or equal to 10%;\u003c/p\u003e\u003cp\u003eOtherwise\u0026thinsp;=\u0026thinsp;0\u003c/p\u003e\u003cp\u003eIn there,\u003c/p\u003e\u003cp\u003eProfit variance ratio = (Profit after tax after auditing - Profit after tax before auditing)/Profit after tax after auditing\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"3\"\u003eSource: Authors\u0026rsquo; calculation\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe final dataset includes 10,286 firm-year observations from non-financial companies listed on Vietnamese stock exchanges, covering the period from 2016 to 2023. Firms in the banking and finance sector were excluded due to their distinct financial reporting structure. The class distribution comprises 4,314 firms with material misstatements and 5,972 without, resulting in a 42:58 ratio.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section3\"\u003e\u003ch2\u003e1.2.2 Feature Selection and Data Sources\u003c/h2\u003e\u003cp\u003eThis study employs a combination of financial and non-financial features to train the predictive model, consistent with the approach used in prior fraud detection research. Financial statement analysis techniques, especially analytical procedures such as trend and ratio analysis, are widely recommended by auditing standards like SAS No. 56 and SAS No. 99 to identify anomalies and assess the risk of fraud (Albrecht et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Kaminski et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). Although financial ratios remain a foundational tool, their predictive power alone is often insufficient due to changes in the nature of fraud over time and evolving business practices (Bhavani \u0026amp; Amponsah, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Somayyeh, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). As such, integrating non-financial information, such as audit firm characteristics and ownership structure, has proven useful in improving detection accuracy and contextualizing risk.\u003c/p\u003e\u003cp\u003eTo ensure a robust and representative feature set, this study uses 50 input variables, including 47 financial indicators and 3 non-financial attributes. These were drawn from two high-quality sources: FiinPro Platform and Kreston Vietnam (details of features are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e5\u003c/span\u003e in \u003cspan refid=\"Sec15\" class=\"InternalRef\"\u003eAppendix 1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThe FiinPro Platform, developed by StoxPlus (a joint venture with Nikkei), is one of the most reputable financial databases in Vietnam. It provides standardized, audited, and time-consistent financial data covering over 1,000 listed firms. Its widespread use by analysts, fund managers, and researchers enhances the credibility and consistency of the information. In addition, the data was cross-checked and cleaned to minimize errors, missing values, or inconsistencies. Kreston Vietnam, a member firm of Kreston Global (a top 13 global accounting network), was selected for supplementary data because of its access to audited reports, market insights, and classification expertise. The inclusion of these two data sources ensures data reliability, regulatory alignment, and domain relevance, particularly in the Vietnamese context where publicly available fraud data is limited.\u003c/p\u003e\u003cp\u003eGiven the relatively large number of financial variables (over 80 available), the study adopts LASSO (Least Absolute Shrinkage and Selection Operator) regression as a feature selection method. LASSO is an embedded technique widely used in machine learning to reduce dimensionality, eliminate multicollinearity, and improve model generalizability (Muthukrishnan \u0026amp; Rohini, \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Tibshirani, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e1996\u003c/span\u003e). It applies a penalty to the absolute size of regression coefficients, forcing less significant ones to shrink to zero, thereby identifying the most predictive variables. This improves model interpretability and avoids overfitting. LASSO is particularly suited to financial applications and has shown strong performance in fraud prediction and bankruptcy forecasting (Liu et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Paraschiv et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Wang \u0026amp; Liu, \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eIn summary, the combination of verified data from FiinPro and Kreston Vietnam, careful labeling aligned with auditing standards, and feature selection via LASSO ensures a reliable, structured, and interpretable dataset for predicting material misstatements in Vietnamese financial statement.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section3\"\u003e\u003ch2\u003e1.2.3 Data processing and Model development\u003c/h2\u003e\u003cp\u003eTo ensure the development of a reliable and generalizable predictive model, data preprocessing and model construction were carried out in several stages, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. First, the dataset was partitioned into independent training and testing sets in an 80:20 ratio to prevent data leakage during evaluation. The training set was then used for model fitting and hyperparameter tuning.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eA 10-fold stratified cross-validation procedure was employed to estimate model performance. Within each fold, the Synthetic Minority Over-sampling Technique (SMOTE) was applied to address the class imbalance, where the positive class (material misstatements) accounted for approximately 16% of all observations. Stratification ensured that both training and validation folds preserved the original class distribution, thereby providing cross-validation results that more accurately reflect the generalization error. Bayesian optimization was used during cross-validation to identify the optimal hyperparameters.\u003c/p\u003e\u003cp\u003eThe final model was retrained on the entire training set with SMOTE applied to the training data, and its performance was evaluated on the independent test set using metrics derived from the confusion matrix (precision, recall, F1-score) and the area under the receiver operating characteristic curve (AUC).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section3\"\u003e\u003ch2\u003e1.2.4 Choosing the machine learning algorithm\u003c/h2\u003e\u003cp\u003eDetecting material misstatements in financial statements is a binary classification problem based on labeled historical data, making supervised machine learning a suitable methodological approach (Agrawal \u0026amp; Chatterjee, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). This study selects Random Forest (RF) and Extreme Gradient Boosting (XGBoost) due to their proven effectiveness in structured, high-dimensional financial data, especially in fraud detection contexts (Bao et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Cheng et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Gupta \u0026amp; Mehta, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Both are ensemble learning algorithms that aggregate decision trees to improve model stability and accuracy. RF uses bagging to reduce variance, while XGBoost applies boosting with regularization to minimize bias and overfitting (Pham et al., \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). These methods are particularly effective at capturing complex nonlinear interactions between financial and non-financial variables, which are common in financial misstatement data (Lokanan et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Ravisankar et al., \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). Additionally, both algorithms provide built-in feature importance metrics, allowing researchers and auditors to interpret the relative influence of input variables\u0026mdash;a critical factor in auditing where model transparency is required (Gilpin et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAlternative machine learning models were considered but not selected for this study. Logistic regression, although widely used, underperformed in our benchmark analysis and is constrained by assumptions of linearity and independence that may not hold in complex financial environments (Hajek \u0026amp; Henriques, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Support Vector Machines (SVM) and Artificial Neural Networks (ANN) offer high accuracy but lack interpretability and can be computationally demanding for large datasets, limiting their usefulness in audit contexts where explainability is essential (Dickinson \u0026amp; Meyer, \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Rane et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2023a\u003c/span\u003e). Simpler classifiers such as k-Nearest Neighbors (k-NN), Na\u0026iuml;ve Bayes, or single decision trees often suffer from poor generalization and sensitivity to outliers (Huang \u0026amp; Huang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Although LightGBM is a strong competitor to XGBoost, the latter was chosen due to its broader empirical validation in financial fraud literature and more extensive integration with interpretability tools (Xu et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Therefore, Random Forest and XGBoost offer the most appropriate combination of accuracy, efficiency, and transparency for detecting material misstatements in the Vietnamese financial reporting landscape.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\u003ch2\u003e1.2.5 Model performance evaluation\u003c/h2\u003e\u003cp\u003e\u003cb\u003eConfusion matrix metrics\u003c/b\u003e\u003c/p\u003e\u003cp\u003eEach machine learning model is described by a set of model parameters. The job of a machine learning algorithm is to find the optimal model parameters, and this is closely related to evaluation metrics. It is the process of finding model parameters so that the evaluation metrics achieve the best results. There are many methods to evaluate classifier model performance from different perspectives depending on the research objectives (Hand, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). In which, good prediction results can be understood as having few misclassified data points (Hand, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Tharwat, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). This evaluation aspect can be done through the Confusion matrix. It is a matrix that describes the classification results. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows that confusion matrix is commonly used to calculate classification model evaluation indicators such as Accurary_score; Precision_score, Recall_score, and F1_score (Hand, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2012\u003c/span\u003e; Tharwat, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eConfusion matrix\u003c/span\u003e\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colspan=\"2\" morerows=\"1\" nameend=\"c2\" namest=\"c1\" rowspan=\"2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eActual value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eActual value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePredicted value\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNegative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTrue negative (TN)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eFalse negative (FN)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePredicted value\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePositive\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eFalse Positive (FP)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eTrue Positive (TP)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"4\"\u003eSource: Authors\u0026rsquo; calculation\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd colspan=\"4\"\u003eNote: The contents of the Confusion Matrix are interpreted as follows:\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\"Positive\": There is a material misstatement in the financial statements\u003c/p\u003e\u003cp\u003e\"Negative\": There are no material misstatements in the financial statements.\u003c/p\u003e\u003cp\u003eTrue Positive (TP): predicted and actual values are both positive\u003c/p\u003e\u003cp\u003eFalse Positive (FP): the predicted value is positive while the actual value is negative\u003c/p\u003e\u003cp\u003eTrue negative (TN): predicted and actual values are both negative\u003c/p\u003e\u003cp\u003eFalse negative (FN): the predicted value is negative while the actual value is positive\u003c/p\u003e\u003cp\u003eAccuracy score = (TP\u0026thinsp;+\u0026thinsp;TN)/Sample\u003c/p\u003e\u003cp\u003ePrecision score\u0026thinsp;=\u0026thinsp;TP/(TP\u0026thinsp;+\u0026thinsp;FP)\u003c/p\u003e\u003cp\u003eRecall score\u0026thinsp;=\u0026thinsp;TP/(TP\u0026thinsp;+\u0026thinsp;FN)\u003c/p\u003e\u003cp\u003eF1 score = (2 x Precision score x Recall score)/(Precision score\u0026thinsp;+\u0026thinsp;Recall score).\u003c/p\u003e\u003cp\u003eArea Under the Receiver Operating Characteristic Curve (AUC)\u003c/p\u003e\u003cp\u003eThe receiver operating characteristic (ROC) curve is a graphical tool for visualizing the performance of a binary classifier by plotting the True Positive Rate (TPR) against the False Positive Rate (FPR) at various classification thresholds. It illustrates the trade-off between detecting positive instances and avoiding false positives across different thresholds.\u003c/p\u003e\u003cp\u003eThe area under the ROC curve (AUC) provides a single scalar value summarizing the classifier\u0026rsquo;s ability to discriminate between the positive class (material misstatements) and the negative class (non-material misstatements). A higher AUC indicates better discriminative performance, making it a particularly valuable metric in imbalanced data contexts where accuracy alone can be misleading.\u003c/p\u003e\u003cp\u003eThe TPR (sensitivity or recall) measures the proportion of actual positives correctly identified:\u003c/p\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:TPR=\\raisebox{1ex}{$TP$}\\!\\left/\\:\\!\\raisebox{-1ex}{$(TP+FN)$}\\right.\\)\u003c/span\u003e\u003c/span\u003eThe FPR measures the proportion of actual negatives incorrectly classified as positives:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:FPR=\\raisebox{1ex}{$FP$}\\!\\left/\\:\\!\\raisebox{-1ex}{$(FP+TN)$}\\right.$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn this study, ROC\u0026ndash;AUC is used to evaluate the performance of Random Forest (RF), XGBoost, and Logistic Regression, all of which produce predicted probabilities rather than just binary labels. For RF, the probability of the positive class is the average proportion of trees voting for that class; for XGBoost, it is computed from the additive ensemble of decision trees trained sequentially via gradient boosting. Logistic Regression produces probabilities through the logistic function, making it inherently well-suited for ROC analysis. By comparing AUC values across models, we can objectively assess their ability to distinguish between firms with and without material misstatements in financial statements.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"4. Classification Results and Discussion","content":"\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e1.3 Classification results\u003c/h2\u003e\u003cp\u003eModel training and optimization were conducted on the training set using stratified cross-validation, SMOTE, and Bayesian optimization. The final predictive performance of the three best-performing classifiers was evaluated on an independent hold-out test set, with results summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e3\u003c/span\u003e and illustrated by their ROC curves. XGBoost achieved the highest overall accuracy (0.8391) and precision (0.8262), as well as the highest AUC\u0026ndash;ROC value (0.91). These results indicate its strong ability to correctly identify financial statements containing material misstatements while minimizing false positives\u0026mdash;an important advantage in audit settings where unnecessary investigations can be costly. Random Forest achieved the highest recall (0.8142) and the second-highest F1-score (0.8031), along with an AUC\u0026ndash;ROC of 0.90. This suggests that it is particularly effective at detecting the majority of misstatements, which is critical in minimizing the risk of overlooking materially misstated reports.\u003c/p\u003e\u003cp\u003eAlthough Logistic Regression recorded the lowest performance across all metrics (accuracy\u0026thinsp;=\u0026thinsp;0.7871, precision\u0026thinsp;=\u0026thinsp;0.7523, recall\u0026thinsp;=\u0026thinsp;0.7419, F1-score\u0026thinsp;=\u0026thinsp;0.7471, AUC\u0026ndash;ROC\u0026thinsp;=\u0026thinsp;0.85), it remains valuable due to its high interpretability and ease of explaining predictions to stakeholders. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the ROC curve for the Logistic Regression model demonstrates its moderate discriminative ability compared to the other classifiers. This transparency can be crucial in regulatory and audit contexts where justifying model decisions is as important as predictive accuracy.\u003c/p\u003e\u003cp\u003eOverall, these results reinforce the evidence that ensemble-based machine learning algorithms\u0026mdash;particularly XGBoost and Random Forest\u0026mdash;can deliver superior predictive performance compared to traditional statistical methods, making them promising tools for risk-based audit planning and fraud detection. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e compares the ROC curves of the Random Forest and XGBoost models, highlighting their higher discriminative performance relative to Logistic Regression.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003ePrediction performance evaluation score\u003c/span\u003e\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eClassifier\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eAccuracy score\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003ePrecision score\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eRecall score\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eF1- score\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eLogistic regression\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e0.7871\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7523\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.7419\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.7471\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eRandom Forest\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.8309\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.7924\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.8142\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.8031\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eXGBoost\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.8391\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.8262\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.7855\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.8054\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"5\"\u003eSource: Authors\u0026rsquo; calculation\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eSource: Authors\u0026rsquo; calculation\u003c/p\u003e\u003cp\u003eTo enhance both model efficiency and interpretability, this study employed two distinct feature selection and explanation techniques. First, LASSO regression was utilized in the pre-modeling phase for input attribute selection. This embedded method effectively reduced dimensionality by shrinking the coefficients of less informative variables to zero, thereby identifying a core subset of predictors with the highest relevance to material misstatement detection. The LASSO procedure improved generalizability, minimized overfitting, and strengthened the theoretical validity of the model by retaining variables with meaningful associations to fraud-related behavior.\u003c/p\u003e\u003cp\u003eSubsequently, feature importance values from the trained Random Forest model were extracted to interpret which attributes contributed most significantly to the classification decisions. This post-modeling analysis enables explainability in the output dimension, revealing the variables that most influenced the model's performance across key evaluation metrics.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eImportant features (20 features)\u003c/span\u003e\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eOrdinal number\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eFeatures\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eImportant score\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eNote\u003c/span\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e1\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eUPCoM\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.139723\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eNew features with LASSO regression\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e2\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eLAST MISSTATE\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.096867\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e3\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSIZE (ln(TA))\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.087433\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e4\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLn(TL)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.044385\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e5\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003ePlaned_NeP\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.041723\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eNew features with LASSO regression\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e6\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHOSE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.039551\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eNew features with LASSO regression\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e7\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003ePlanned_EBIT\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.035057\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eNew features with LASSO regression\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e8\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eln(CASH)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.027775\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e9\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eROA\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.023334\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e10\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eEBTtTA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.023301\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e11\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eP/S\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.018986\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e12\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eROE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.018181\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e13\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eEtA\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.016182\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e14\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eREtA\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.015744\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003e15\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eEPS_G\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.015028\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e\u003cspan type=\"SmallCaps\" class=\"SmallCaps\" name=\"Emphasis\"\u003eNew features with LASSO regression\u003c/span\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"4\"\u003eSource: Authors\u0026rsquo; calculation\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the most influential variable was UPCoM, a dummy variable indicating listing on Vietnam\u0026rsquo;s Unlisted Public Company Market, which generally has less stringent disclosure requirements. This finding underscores the elevated risk profile of firms operating in less regulated environments and highlights the need for enhanced regulatory oversight. Interestingly, another dummy variable, HOSE\u0026mdash;the largest stock exchange in Vietnam\u0026mdash;also ranked sixth in importance, suggesting that listing venue characteristics can influence the detection of financial reporting errors.\u003c/p\u003e\u003cp\u003eThe second most important feature was LAST MISSTATE, reflecting a firm\u0026rsquo;s prior history of material misstatements. This result is consistent with previous studies (Lou \u0026amp; Wang, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2009\u003c/span\u003e) and supports the ACFE\u0026rsquo;s (2022) observation that financial fraud often persists over extended periods before detection.\u003c/p\u003e\u003cp\u003eFirm size (SIZE), measured as the natural logarithm of total assets, ranked third, aligning with prior (Beneish, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e1999\u003c/span\u003e; Cheng et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Hajek, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2019\u003c/span\u003e), which suggests that large firms may have more complex reporting structures and greater incentives or opportunities for earnings manipulation.\u003c/p\u003e\u003cp\u003eAmong the LASSO-selected features, several forward-looking performance indicators, such as Planned EBIT, Planned Net Profit (Planned_NeP), and EPS Growth (EPS_G), emerged as significant predictors. These metrics capture pressures to meet projected performance targets and may serve as early warning indicators of earnings management. Their inclusion highlights the value of incorporating strategic financial planning indicators, beyond static ratios, into predictive fraud detection models.\u003c/p\u003e\u003cp\u003eTraditional financial ratios, including Ln(Total liabilities), Ln(CASH), Return on Assets (ROA), Earnings Before Tax to Total Assets (EBTtTA), Return on Equity (ROE), Equity to Total Assets (EtA), and Trade Receivables to Total Assets (REtA), also remained relevant predictors. These variables reflect structural composition and balance sheet integrity, consistent with earlier findings in the fraud detection literature (Beneish, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e1999\u003c/span\u003e; Dechow et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Ravisankar et al., \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2011\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eOverall, the classification results reflect a balanced integration of historical indicators, governance-related variables, forward-looking performance metrics, and conventional financial ratios. The dual application of LASSO and Random Forest feature importance provides complementary interpretability: LASSO strengthens feature selection and model efficiency, while Random Forest clarifies output-level feature impact. Together, these methods enhance the transparency, reliability, and practical applicability of the predictive models in financial statement auditing and regulatory risk assessment.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e1.4 Discussing machine learning transparency\u003c/h2\u003e\u003cp\u003eMachine learning has become a powerful tool in financial analytics, enabled by the rapid growth of data availability and computing capacity. These algorithms excel at uncovering complex patterns from structured and unstructured data without requiring explicitly programmed rules (de Laat, \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Lepri et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). However, this power often comes at the expense of transparency. Many machine learning models, particularly ensemble methods like Random Forest and XGBoost, are viewed as \u0026ldquo;black boxes\u0026rdquo; because it can be challenging to understand how they arrive at their predictions (Mitchell et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Rane et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2023a\u003c/span\u003e). This lack of interpretability can limit stakeholder trust and hinder adoption in highly regulated domains like financial auditing.\u003c/p\u003e\u003cp\u003eA key way to address this issue is through explainability, which refers to techniques that clarify how a machine learning model makes decisions. Explainability helps stakeholders\u0026mdash;auditors, regulators, and investors\u0026mdash;understand not only what a model predicts but why it does so (Gilpin et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Models such as linear regression or decision trees are inherently interpretable, offering a clear mapping between input features and outcomes (Peng et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). However, their simplicity often makes them inadequate for capturing the non-linear relationships that are common in financial data. Ensemble models like Random Forest and XGBoost offer superior predictive accuracy because they can model such complexity (Huang \u0026amp; Huang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Xu et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Nonetheless, their complexity necessitates interpretation tools like feature importance scores to enhance transparency.\u003c/p\u003e\u003cp\u003eMachine learning algorithms generate predictions from sample data without explicit instructions from the user (Huang \u0026amp; Huang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). This has led to them being considered \u0026ldquo;black boxes\u0026rdquo; due to the challenge of understanding how machine learning makes its predictions (Rane et al., \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2023b\u003c/span\u003e). As a result, the interpretability, reliability, and effectiveness of machine learning models are often difficult to assess (Mitchell et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). This issue is discussed by researchers in terms of machine learning transparency in decision-making, including algorithmic transparency.\u003c/p\u003e\u003cp\u003eIn this study, feature importance values derived from the Random Forest algorithm were used to assess which variables most significantly influenced the model's predictions of material misstatements. This approach contributes directly to machine learning transparency by identifying and communicating the relative weight of each input in the model\u0026rsquo;s decision-making process (Munkhdalai et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Sonkavde et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Among the top-ranked features, UpCoM, LAST_MISSTATE, and SIZE emerged as dominant predictors. Their practical implications are significant: firms listed on the less-regulated UpCoM exchange, companies with a history of misstatements, and larger enterprises with complex structures are more likely to engage in or obscure material misstatements. These findings align with audit theory and support targeted, risk-based audit planning.\u003c/p\u003e\u003cp\u003eA further contribution to model transparency comes from understanding and communicating the predictive performance metrics used. The comparative analysis of the three classifiers reveals important trade-offs in model selection for detecting material misstatements in financial statements. Random Forest, with its highest recall (0.8142), offers a clear advantage when the primary objective is to minimize false negatives\u0026mdash;cases where financial statements containing material misstatements are incorrectly classified as not misstated. From an auditing and risk management perspective, failing to detect a material misstatement can have severe regulatory and financial consequences. The ensemble characteristic of Random Forest, leveraging multiple decision trees with bootstrap aggregation, likely contributed to its strong sensitivity by capturing non-linear relationships between predictor features and the occurrence of material misstatements.\u003c/p\u003e\u003cp\u003eConversely, XGBoost exhibited the strongest precision (0.8262) and the highest discriminative ability (AUC-ROC\u0026thinsp;=\u0026thinsp;0.91), indicating its capacity to effectively identify true positive cases while minimizing false positives. This model enables auditors to focus resources on cases that are indeed materially misstated. The gradient boosting mechanism of XGBoost, combined with optimized hyperparameters via Bayesian search, likely enhanced its ability to capture complex feature interactions and prioritize hard-to-classify cases.\u003c/p\u003e\u003cp\u003eAlthough logistic regression yielded comparatively lower predictive performance, it retains value due to its interpretability, stemming from its linear nature. In situations where the ability to provide clear, transparent justification for predictions is as critical as accuracy, logistic regression remains an attractive supplementary tool.\u003c/p\u003e\u003cp\u003eMany of the most influential variables were selected through LASSO regression, which enhances both predictive performance and interpretability. Features such as Planned EBIT, Planned Net Profit, and EPS Growth (EPS_G) reflect strategic financial planning behavior, suggesting that attempts to manage or project strong performance may be early signals of misstatement risk. These indicators, derived from forward-looking or growth-oriented metrics, suggest that firms under performance pressure may be more prone to earnings manipulation. Their inclusion highlights the importance of not only reviewing historical results but also understanding firms\u0026rsquo; projected targets and capital efficiency when assessing audit risk.\u003c/p\u003e\u003cp\u003eIn summary, this study demonstrates the effectiveness of machine learning in predicting material misstatements and contributes to the field of explainable AI by clearly identifying which features drive model decisions and by interpreting the practical implications of key performance metrics. Such transparency supports trust, fosters informed stakeholder engagement, and facilitates the adoption of machine learning tools in financial statement auditing in Vietnam and beyond.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e1.5 Implications for the auditing context\u003c/h2\u003e\u003cp\u003eWhile corporate management bears primary responsibility for the accuracy of financial reporting, stakeholders, including investors and regulators, are increasingly holding auditors accountable for failing to detect material misstatements (Hunt et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). In this context, integrating machine learning (ML) into audit practices offers a promising pathway to enhance audit quality, efficiency, and risk detection. Traditional auditing, particularly analytical procedures, has relied heavily on manual reviews and statistical comparisons. However, with the rise of Big Data and advanced analytics, auditors now have access to tools capable of processing large volumes of financial and non-financial data, uncovering subtle anomalies that might otherwise go undetected (Appelbaum et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Warren et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThis study demonstrates that ML-based analytical procedures can predict material misstatements with higher accuracy than traditional statistical methods. Both Random Forest (RF) and XGBoost achieved strong accuracy (0.8309\u0026ndash;0.8391), correctly classifying most firms regardless of misstatement status. XGBoost\u0026rsquo;s precision (0.8262) indicates that over four out of five flagged cases were indeed misstated, enabling auditors to focus resources on the most probable risks. RF\u0026rsquo;s recall (0.8142) shows it effectively identifies most actual misstatements, thereby reducing the risk of undetected fraud. The choice between these models should be based on the auditing firm\u0026rsquo;s tolerance for Type I and Type II errors: RF is preferable when avoiding undetected misstatements is the priority, while XGBoost is advantageous for optimizing resource allocation. A hybrid approach, using XGBoost for primary screening and RF for secondary verification, could combine their strengths, reducing false alarms while maintaining high detection capability.\u003c/p\u003e\u003cp\u003eFeature importance analysis further highlights key audit risk indicators. UPCoM listing status ranked first, indicating that firms on Vietnam\u0026rsquo;s less-regulated secondary exchange present higher risks. LAST MISSTATE, a firm\u0026rsquo;s history of misreporting, ranked second, reinforcing the importance of incorporating prior audit outcomes into current risk assessments. Firm size (SIZE) ranked third, aligning with literature linking larger, more complex firms to greater misstatement risk. LASSO regression also identified forward-looking indicators such as Planned EBIT, Planned Net Profit (Planned_NeP), and EPS Growth (EPS_G), suggesting that performance pressures can trigger earnings management.\u003c/p\u003e\u003cp\u003eTraditional financial ratios, including Ln(Total liabilities), Return on Assets (ROA), Return on Equity (ROE), and Earnings Before Tax to Total Assets (EBTtTA), also showed strong predictive value, reflecting both financial structure and leverage-related risks. These findings imply that auditors should balance their focus between historical results, strategic performance plans, and financial position when assessing misstatement risk.\u003c/p\u003e\u003cp\u003eIn the Vietnamese auditing context, where direct fraud data are scarce, this study shows that ML can bridge data gaps by using proxy indicators like material misstatements. Feature selection via LASSO ensures attention to the most relevant variables, supporting risk-based audit planning. As audit firms invest in AI-powered systems, integrating these insights can enhance audit quality, regulatory compliance, and stakeholder trust. In an increasingly complex reporting environment, adopting data-driven, transparent, and adaptive tools will be essential for maintaining integrity, accountability, and investor confidence in capital markets.\u003c/p\u003e\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003e\u003cb\u003eEthical Approval\u003c/b\u003e:\u003c/h2\u003e\u003cp\u003e\u003cem\u003e\u0026ldquo;This article does not contain any studies with human participants performed by any of the authors.\u0026rdquo;\u003c/em\u003e\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003e\u003cb\u003eInformed Consent\u003c/b\u003e:\u003c/strong\u003e\u003cp\u003e\u003cem\u003e\u0026ldquo;This article does not contain any studies with human participants performed by any of the authors.\u0026rdquo;\u003c/em\u003e\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eThi Thanh Hoa Tieu: Conceptualization, methodology, formal analysis, writing, original Draft.Ngoc Hung Tran: Data curation, investigation, writing \u0026ndash; original draft.Hien Thanh Hoang: Supervision, validation, project administration, writing \u0026ndash; review \u0026amp; editing.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eAll relevant data are within the manuscript and its supporting information files. The full computational code for attribute selection (LASSO regression) and model implementation, as well as the research results, are openly available via the following Google Colaboratory links:LASSO regression and attribute selection:https://colab.research.google.com/drive/1EYgRtMdbQPTJM77fQcdkzNH5Ri-seAnp?usp=sharingResearch resultshttps://colab.research.google.com/drive/1K36ciNvp8KsbQpG0JMMQkAplNmvldtGr?usp=sharingThese resources can also be found in the Supporting Information (S1). Additional data or code are available from the corresponding author upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAgrawal, K., \u0026amp; Chatterjee, C. (2015). Earnings management and financial distress: Evidence from India. \u003cem\u003eGlobal Business Review\u003c/em\u003e,\u003cem\u003e 16\u003c/em\u003e(5_suppl), 140S-154S. \u003c/li\u003e\n\u003cli\u003eAlbrecht, W. S., Albrecht, C., \u0026amp; Albrecht, C. C. (2008). Current Trends in Fraud and its Detection. \u003cem\u003eInformation Security Journal: A Global Perspective\u003c/em\u003e,\u003cem\u003e 17\u003c/em\u003e(1), 2-12. https://doi.org/10.1080/19393550801934331 \u003c/li\u003e\n\u003cli\u003eAppelbaum, D., Kogan, A., \u0026amp; Vasarhelyi, M. A. (2017). Big Data and Analytics in the Modern Audit Engagement: Research Needs. \u003cem\u003eAuditing: A Journal of Practice \u0026amp; Theory\u003c/em\u003e,\u003cem\u003e 36\u003c/em\u003e(4), 1-27. https://doi.org/10.2308/ajpt-51684 \u003c/li\u003e\n\u003cli\u003eArshad, R., Iqbal, S. M., \u0026amp; Omar, N. (2015). Prediction of Business Failure and Fraudulent Financial Reporting: Evidence from Malaysia. \u003cem\u003eIndian Journal of Corporate Governance\u003c/em\u003e,\u003cem\u003e 8\u003c/em\u003e(1), 34-53. https://doi.org/10.1177/0974686215574424 \u003c/li\u003e\n\u003cli\u003eAshtiani, M. N., \u0026amp; Raahemi, B. (2021). Intelligent fraud detection in financial statements using machine learning and data mining: a systematic literature review. \u003cem\u003eIeee Access\u003c/em\u003e,\u003cem\u003e 10\u003c/em\u003e, 72504-72525. \u003c/li\u003e\n\u003cli\u003eBao, Y., Ke, B., Li, B., Yu, Y. J., \u0026amp; Zhang, J. (2020). Detecting accounting fraud in publicly traded US firms using a machine learning approach. \u003cem\u003eJournal of Accounting Research\u003c/em\u003e,\u003cem\u003e 58\u003c/em\u003e(1), 199-235. \u003c/li\u003e\n\u003cli\u003eBeneish, M. D. (1999). The detection of earnings manipulation. \u003cem\u003eFinancial Analysts Journal\u003c/em\u003e,\u003cem\u003e 55\u003c/em\u003e(5), 24-36. \u003c/li\u003e\n\u003cli\u003eBennett, M., Hayes, K., Kleczyk, E. J., \u0026amp; Mehta, R. (2022). Similarities and differences between machine learning and traditional advanced statistical modeling in healthcare analytics. \u003cem\u003earXiv preprint arXiv:2201.02469\u003c/em\u003e. \u003c/li\u003e\n\u003cli\u003eBhavani, G., \u0026amp; Amponsah, C. T. (2017). M-Score and Z-Score for detection of accounting fraud. \u003cem\u003eAccountancy Business and the Public Interest\u003c/em\u003e,\u003cem\u003e 1\u003c/em\u003e(1), 68-86. \u003c/li\u003e\n\u003cli\u003eBreiman, L. (2001). Random forests. \u003cem\u003eMachine learning\u003c/em\u003e,\u003cem\u003e 45\u003c/em\u003e, 5-32. \u003c/li\u003e\n\u003cli\u003eBzdok, D., Altman, N., \u0026amp; Krzywinski, M. (2018). Statistics versus machine learning. \u003cem\u003eNat Methods\u003c/em\u003e,\u003cem\u003e 15\u003c/em\u003e(4), 233-234. https://doi.org/10.1038/nmeth.4642 \u003c/li\u003e\n\u003cli\u003eCarmichael, I., \u0026amp; Marron, J. (2018). Data science vs. statistics: two cultures? \u003cem\u003eJapanese Journal of Statistics and Data Science\u003c/em\u003e,\u003cem\u003e 1\u003c/em\u003e(1), 117-138. \u003c/li\u003e\n\u003cli\u003eCheng, C.-H., Kao, Y.-F., \u0026amp; Lin, H.-P. (2021). A financial statement fraud model based on synthesized attribute selection and a dataset with missing values and imbalanced classes. \u003cem\u003eApplied Soft Computing\u003c/em\u003e,\u003cem\u003e 108\u003c/em\u003e, 107487. \u003c/li\u003e\n\u003cli\u003ede Laat, P. B. (2018). Algorithmic Decision-Making Based on Machine Learning from Big Data: Can Transparency Restore Accountability? \u003cem\u003ePhilos Technol\u003c/em\u003e,\u003cem\u003e 31\u003c/em\u003e(4), 525-541. https://doi.org/10.1007/s13347-017-0293-z \u003c/li\u003e\n\u003cli\u003eDechow, P. M., Ge, W., Larson, C. R., \u0026amp; Sloan, R. G. (2011). Predicting Material Accounting Misstatements*. \u003cem\u003eContemporary Accounting Research\u003c/em\u003e,\u003cem\u003e 28\u003c/em\u003e(1), 17-82. https://doi.org/10.1111/j.1911-3846.2010.01041.x \u003c/li\u003e\n\u003cli\u003eDickinson, Q., \u0026amp; Meyer, J. G. (2022). Positional SHAP (PoSHAP) for Interpretation of machine learning models trained from biological sequences. \u003cem\u003ePLOS Computational Biology\u003c/em\u003e,\u003cem\u003e 18\u003c/em\u003e(1), e1009736. \u003c/li\u003e\n\u003cli\u003eGaganis, C. (2009). Classification techniques for the identification of falsified financial statements: a comparative analysis. \u003cem\u003eIntelligent Systems in Accounting, Finance and Management\u003c/em\u003e,\u003cem\u003e 16\u003c/em\u003e(3), 207-229. https://doi.org/10.1002/isaf.303 \u003c/li\u003e\n\u003cli\u003eGilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., Specter, M., \u0026amp; Kagal, L. (2018). Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th International Conference on data science and advanced analytics (DSAA), \u003c/li\u003e\n\u003cli\u003eGreen, B. P., \u0026amp; Choi, J. H. (1997). Assessing the risk of management fraud through neural network technology. \u003cem\u003eAuditing\u003c/em\u003e,\u003cem\u003e 16\u003c/em\u003e, 14-28. \u003c/li\u003e\n\u003cli\u003eGupta, S., \u0026amp; Mehta, S. K. (2021). Data Mining-based Financial Statement Fraud Detection: Systematic Literature Review and Meta-analysis to Estimate Data Sample Mapping of Fraudulent Companies Against Non-fraudulent Companies. \u003cem\u003eGlobal Business Review\u003c/em\u003e,\u003cem\u003e 25\u003c/em\u003e(5), 1290-1313. https://doi.org/10.1177/0972150920984857 \u003c/li\u003e\n\u003cli\u003eGupta, S., \u0026amp; Mehta, S. K. (2024). Data mining-based financial statement fraud detection: Systematic literature review and meta-analysis to estimate data sample mapping of fraudulent companies against non-fraudulent companies. \u003cem\u003eGlobal Business Review\u003c/em\u003e,\u003cem\u003e 25\u003c/em\u003e(5), 1290-1313. \u003c/li\u003e\n\u003cli\u003eHajek, P. (2019). Interpretable Fuzzy Rule-Based Systems for Detecting Financial Statement Fraud.\u003cem\u003e 559\u003c/em\u003e, 425-436. https://doi.org/10.1007/978-3-030-19823-7_36 \u003c/li\u003e\n\u003cli\u003eHajek, P., \u0026amp; Henriques, R. (2017). Mining corporate annual reports for intelligent detection of financial statement fraud\u0026ndash;A comparative study of machine learning methods. \u003cem\u003eKnowledge-Based Systems\u003c/em\u003e,\u003cem\u003e 128\u003c/em\u003e, 139-152. \u003c/li\u003e\n\u003cli\u003eHand, D. J. (2012). Assessing the Performance of Classification Methods. \u003cem\u003eInternational Statistical Review\u003c/em\u003e,\u003cem\u003e 80\u003c/em\u003e(3), 400-414. https://doi.org/10.1111/j.1751-5823.2012.00183.x \u003c/li\u003e\n\u003cli\u003eHerawati, N. (2015). Application of Beneish M-Score models and data mining to detect financial fraud. \u003cem\u003eProcedia-Social and Behavioral Sciences\u003c/em\u003e,\u003cem\u003e 211\u003c/em\u003e, 924-930. \u003c/li\u003e\n\u003cli\u003eHołda, A. (2020). Using the Beneish M-score model: Evidence from non-financial companies listed on the Warsaw Stock Exchange. \u003cem\u003eInvestment Management and Financial Innovations\u003c/em\u003e,\u003cem\u003e 17\u003c/em\u003e(4), 389-401. https://doi.org/10.21511/imfi.17(4).2020.33 \u003c/li\u003e\n\u003cli\u003eHuang, A. A., \u0026amp; Huang, S. Y. (2023). Increasing transparency in machine learning through bootstrap simulation and shapely additive explanations. \u003cem\u003ePLoS One\u003c/em\u003e,\u003cem\u003e 18\u003c/em\u003e(2), e0281922. https://doi.org/10.1371/journal.pone.0281922 \u003c/li\u003e\n\u003cli\u003eHunt, E., Hunt, J., Richardson, V. J., \u0026amp; Rosser, D. (2022). Auditor response to estimated misstatement risk: A machine learning approach. \u003cem\u003eAccounting Horizons\u003c/em\u003e,\u003cem\u003e 36\u003c/em\u003e(1), 111-130. \u003c/li\u003e\n\u003cli\u003eJan, C.-l. (2018). An Effective Financial Statements Fraud Detection Model for the Sustainable Development of Financial Markets: Evidence from Taiwan. \u003cem\u003eSustainability\u003c/em\u003e,\u003cem\u003e 10\u003c/em\u003e(2), 513. https://doi.org/10.3390/su10020513 \u003c/li\u003e\n\u003cli\u003eKaminski, K. A., Sterling Wetzel, T., \u0026amp; Guan, L. (2004). Can financial ratios detect fraudulent financial reporting? \u003cem\u003eManagerial Auditing Journal\u003c/em\u003e,\u003cem\u003e 19\u003c/em\u003e(1), 15-28. \u003c/li\u003e\n\u003cli\u003eKanapickienė, R., \u0026amp; Grundienė, Ž. (2015). The Model of Fraud Detection in Financial Statements by Means of Financial Ratios. \u003cem\u003eProcedia - Social and Behavioral Sciences\u003c/em\u003e,\u003cem\u003e 213\u003c/em\u003e, 321-327. https://doi.org/10.1016/j.sbspro.2015.11.545 \u003c/li\u003e\n\u003cli\u003eKirkos, E., Spathis, C., \u0026amp; Manolopoulos, Y. (2007). Data Mining techniques for the detection of fraudulent financial statements. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e,\u003cem\u003e 32\u003c/em\u003e(4), 995-1003. https://doi.org/10.1016/j.eswa.2006.02.016 \u003c/li\u003e\n\u003cli\u003eLakhan, A., Mohammed, M. A., Ibrahim, D. A., Kadry, S., \u0026amp; Abdulkareem, K. H. (2022). ITS based on deep graph convolutional fraud detection network blockchain-enabled fog-cloud. \u003cem\u003eIEEE Transactions on Intelligent Transportation Systems\u003c/em\u003e,\u003cem\u003e 24\u003c/em\u003e(8), 8399-8408. \u003c/li\u003e\n\u003cli\u003eLepri, B., Oliver, N., Letouz\u0026eacute;, E., Pentland, A., \u0026amp; Vinck, P. (2017). Fair, Transparent, and Accountable Algorithmic Decision-making Processes. \u003cem\u003ePhilosophy \u0026amp; Technology\u003c/em\u003e,\u003cem\u003e 31\u003c/em\u003e(4), 611-627. https://doi.org/10.1007/s13347-017-0279-x \u003c/li\u003e\n\u003cli\u003eLiu, Z., Ye, R., \u0026amp; Ye, R. (2021). https://doi.org/10.21203/rs.3.rs-640038/v1 \u003c/li\u003e\n\u003cli\u003eLokanan, M., Tran, V., \u0026amp; Vuong, N. H. (2019). Detecting anomalies in financial statements using machine learning algorithm: The case of Vietnamese listed firms. \u003cem\u003eAsian Journal of Accounting Research\u003c/em\u003e,\u003cem\u003e 4\u003c/em\u003e(2), 181-201. \u003c/li\u003e\n\u003cli\u003eLou, Y.-I., \u0026amp; Wang, M.-L. (2009). Fraud risk factor of the fraud triangle assessing the likelihood of fraudulent financial reporting. \u003cem\u003eJournal of Business \u0026amp; Economics Research (JBER)\u003c/em\u003e,\u003cem\u003e 7\u003c/em\u003e(2). \u003c/li\u003e\n\u003cli\u003eMarais, A., Vermaak, C., \u0026amp; Shewell, P. (2023). Predicting financial statement manipulation in South Africa: A comparison of the Beneish and Dechow models. \u003cem\u003eCogent Economics \u0026amp; Finance\u003c/em\u003e,\u003cem\u003e 11\u003c/em\u003e(1). https://doi.org/10.1080/23322039.2023.2190215 \u003c/li\u003e\n\u003cli\u003eMitchell, R., Frank, E., \u0026amp; Holmes, G. (2022). GPUTreeShap: massively parallel exact calculation of SHAP scores for tree ensembles. \u003cem\u003ePeerJ Comput Sci\u003c/em\u003e,\u003cem\u003e 8\u003c/em\u003e, e880. https://doi.org/10.7717/peerj-cs.880 \u003c/li\u003e\n\u003cli\u003eMunkhdalai, L., Munkhdalai, T., Namsrai, O.-E., Lee, J. Y., \u0026amp; Ryu, K. H. (2019). An empirical comparison of machine-learning methods on bank client credit assessments. \u003cem\u003eSustainability\u003c/em\u003e,\u003cem\u003e 11\u003c/em\u003e(3), 699. \u003c/li\u003e\n\u003cli\u003eMuthukrishnan, R., \u0026amp; Rohini, R. (2016). LASSO: A feature selection technique in predictive modeling for machine learning. 2016 IEEE international conference on advances in computer applications (ICACA), \u003c/li\u003e\n\u003cli\u003eOmar, N., Johari, Z. A., \u0026amp; Smith, M. (2017). Predicting fraudulent financial reporting using artificial neural network. \u003cem\u003eJournal of Financial Crime\u003c/em\u003e,\u003cem\u003e 24\u003c/em\u003e(2), 362-387. \u003c/li\u003e\n\u003cli\u003eParaschiv, F., Schmid, M., \u0026amp; Wahlstr\u0026oslash;m, R. R. (2023). Bankruptcy prediction of privately held SMEs using feature selection methods. \u003cem\u003eAvailable at SSRN 4588767\u003c/em\u003e. \u003c/li\u003e\n\u003cli\u003ePeng, C.-L., Chen, W.-K., \u0026amp; Wei, A.-P. (2021). Teaching CAPM for a Pre-Finance Graduate Program at the STEM Undergraduate Level: Linear Algebra Perspective. \u003cem\u003eMathematics\u003c/em\u003e,\u003cem\u003e 9\u003c/em\u003e(14). https://doi.org/10.3390/math9141668 \u003c/li\u003e\n\u003cli\u003ePham, K., Kim, D., Park, S., \u0026amp; Choi, H. (2021). Ensemble learning-based classification models for slope stability analysis. \u003cem\u003eCatena\u003c/em\u003e,\u003cem\u003e 196\u003c/em\u003e. https://doi.org/10.1016/j.catena.2020.104886 \u003c/li\u003e\n\u003cli\u003eRane, N., Choudhary, S., \u0026amp; Rane, J. (2023a). Explainable Artificial Intelligence (XAI) approaches for transparency and accountability in financial decision-making. \u003cem\u003eAvailable at SSRN 4640316\u003c/em\u003e. \u003c/li\u003e\n\u003cli\u003eRane, N., Choudhary, S., \u0026amp; Rane, J. (2023b). Explainable Artificial Intelligence (XAI) Approaches for Transparency and Accountability in Financial Decision-Making. \u003cem\u003eSSRN Electronic Journal\u003c/em\u003e. https://doi.org/10.2139/ssrn.4640316 \u003c/li\u003e\n\u003cli\u003eRavisankar, P., Ravi, V., Rao, G. R., \u0026amp; Bose, I. (2011). Detection of financial statement fraud and feature selection using data mining techniques. \u003cem\u003eDecision support systems\u003c/em\u003e,\u003cem\u003e 50\u003c/em\u003e(2), 491-500. \u003c/li\u003e\n\u003cli\u003eShoetan, P. O., Oyewole, A. T., Okoye, C. C., \u0026amp; Ofodile, O. C. (2024). Reviewing the role of big data analytics in financial fraud detection. \u003cem\u003eFinance \u0026amp; Accounting Research Journal\u003c/em\u003e,\u003cem\u003e 6\u003c/em\u003e(3), 384-394. \u003c/li\u003e\n\u003cli\u003eSomayyeh, H. N. (2015). Financial ratios between fraudulent and non-fraudulent firms: Evidence from Tehran Stock Exchange. \u003cem\u003eJournal of Accounting and Taxation\u003c/em\u003e,\u003cem\u003e 7\u003c/em\u003e(3), 38-44. https://doi.org/10.5897/jat2014.0166 \u003c/li\u003e\n\u003cli\u003eSonkavde, G., Dharrao, D. S., Bongale, A. M., Deokate, S. T., Doreswamy, D., \u0026amp; Bhat, S. K. (2023). Forecasting Stock Market Prices Using Machine Learning and Deep Learning Models: A Systematic Review, Performance Analysis and Discussion of Implications. \u003cem\u003eInternational Journal of Financial Studies\u003c/em\u003e,\u003cem\u003e 11\u003c/em\u003e(3). https://doi.org/10.3390/ijfs11030094 \u003c/li\u003e\n\u003cli\u003eTharwat, A. (2020). Classification assessment methods. \u003cem\u003eApplied Computing and Informatics\u003c/em\u003e,\u003cem\u003e 17\u003c/em\u003e(1), 168-192. https://doi.org/10.1016/j.aci.2018.08.003 \u003c/li\u003e\n\u003cli\u003eTibshirani, R. (1996). Regression shrinkage and selection via the lasso. \u003cem\u003eJournal of the Royal Statistical Society Series B: Statistical Methodology\u003c/em\u003e,\u003cem\u003e 58\u003c/em\u003e(1), 267-288. \u003c/li\u003e\n\u003cli\u003eWang, W., \u0026amp; Liu, F. (2020). \u003cem\u003eResearch on Identification of Fraud in Financial Reporting Based on Adaptive Group Lasso-Logistic Regression\u003c/em\u003e 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), \u003c/li\u003e\n\u003cli\u003eWarren, J. D., Moffitt, K. C., \u0026amp; Byrnes, P. (2015). How big data will change accounting. \u003cem\u003eAccounting Horizons\u003c/em\u003e,\u003cem\u003e 29\u003c/em\u003e(2), 397-407. \u003c/li\u003e\n\u003cli\u003eXu, Y., Ye, W., Song, Q., Shen, L., Liu, Y., Guo, Y.,\u0026hellip;Sun, X. (2022). Using machine learning models to predict the duration of the recovery of COVID-19 patients hospitalized in Fangcang shelter hospital during the Omicron BA. 2.2 pandemic. \u003cem\u003eFrontiers in Medicine\u003c/em\u003e,\u003cem\u003e 9\u003c/em\u003e, 1001801. \u003c/li\u003e\n\u003cli\u003eZainudin, E. F., \u0026amp; Hashim, H. A. (2016). Detecting fraudulent financial reporting using financial ratio. \u003cem\u003eJournal of Financial Reporting and Accounting\u003c/em\u003e,\u003cem\u003e 14\u003c/em\u003e(2), 266-278. https://doi.org/10.1108/jfra-05-2015-0053 \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Ensemble machine learning, Financial statement misstatements, Fraud detection, LASSO, Random Forest, XGBoost, SMOTE, Bayesian optimization, Stratified cross-validation","lastPublishedDoi":"10.21203/rs.3.rs-7360630/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7360630/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis study develops a machine learning\u0026ndash;based framework for detecting material misstatements in the financial statements of Vietnamese listed companies. Using 10,286 firm-year observations from 2016\u0026ndash;2023, the research applies two ensemble algorithms, Random Forest (RF) and Extreme Gradient Boosting (XGBoost), to a binary classification task based on audit-adjusted profit discrepancies. To address data imbalance and improve prediction reliability, the Synthetic Minority Over-sampling Technique (SMOTE) is applied within a stratified cross-validation procedure, while Bayesian optimization tunes hyperparameters to enhance generalization performance. Both RF and XGBoost achieved high predictive accuracy (~\u0026thinsp;0.839) and strong discriminative power (AUC-ROC\u0026thinsp;~\u0026thinsp;0.91), outperforming logistic regression. Model interpretability was improved through the Least Absolute Shrinkage and Selection Operator (LASSO), which selected key financial and non-financial predictors from over 50 variables. RF\u0026rsquo;s feature importance analysis further highlighted the influence of listing exchange characteristics, prior misstatement history, and forward-looking performance indicators. The proposed framework offers auditors and regulators a scalable, data-driven tool for risk-based audit planning and regulatory oversight\u0026mdash;particularly valuable in emerging markets with limited confirmed fraud data.\u003c/p\u003e","manuscriptTitle":"Detecting financial misstatements in emerging markets: a machine learning approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-01 16:09:46","doi":"10.21203/rs.3.rs-7360630/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"dc99acc1-a335-448d-8d2e-4b7dcc19458b","owner":[],"postedDate":"October 1st, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":55091151,"name":"Physical sciences/Engineering"},{"id":55091152,"name":"Physical sciences/Mathematics and computing"}],"tags":[],"updatedAt":"2025-12-04T08:38:17+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-01 16:09:46","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7360630","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7360630","identity":"rs-7360630","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}