Research on corporate financial fraud detection has established a comprehensive framework of indicators. Early studies primarily focused on financial metrics, including profitability, solvency, growth potential, operational efficiency, and cash flow [4][5]. However, as market environments have grown more complex and fraudulent practices more sophisticated, reliance on financial metrics alone has proven insufficient for effective risk identification. Consequently, researchers have expanded the scope of fraud detection to include non-financial indicators, like board structure or executive team characteristics [6][7]. These indicators offer insights into the quality of corporate governance. For instance, Liu etal. [8] demonstrated that the shareholding ratio of major shareholders serves as a significant predictor of fraudulent behavior in Chinese firms.

Beyond the tabular-style metrics, textual corporate annual reports have emerged as a critical resource for fraud detection. Annual reports, as the primary medium for corporate disclosure, are instrumental for investors in assessing financial risks. Studies reveal that firms engaging in financial misconduct often employ a dual-fraud strategy, manipulating both financial data and report narratives [9]. For example, Xu & Zhang [10] found that high-risk firms tend to adopt an overly optimistic tone in their reports to obscure irregularities. Similarly, Wang etal. [11] found a negative correlation between a firm’s financial status and report readability, suggesting that obfuscation is a common tactic used conceal financial distress. Zhang etal. [12] incorporated readability metrics into fraud prediction models, achieving a 26.33% improvement in F1-score, underscoring the predictive value of textual.

Furthermore, an increasing number of advancements have deepened the analysis of the report content. Brown etal. [13] and Craja etal. [14] extracted direct fraud signals from report paragraphs, while Wu and Du [1] input text feature vectors to LSTM models to achieve a 94.98% prediction accuracy. Liu etal. [15] further integrated latent semantic features of reports with accounting metrics, demonstrating the synergistic benefits of multimodal data fusion.

Methodologies for fraud prediction can generally be classified into three categories: statistical models, machine learning approaches, and deep learning techniques. Statistical models, such as logistic regression (LR), are valued for their predictive power and strong interpretability; however, they often encounter difficulties when dealing with high-dimensional or nonlinear data relationships [16]. By contrast, machine learning models are well-suited for capturing complex patterns in data. Notably, ensemble methods—such as random forests (RF), XGBoost, and LightGBM—have delivered outstanding performance in fraud detection tasks. For instance, Ali etal. [17] optimized an XGBoost model to achieve an accuracy rate of 96.05%, while Zhao & Bai [18] combined LR with XGBoost to further enhance predictive performance. Additionally, Yadav [19] demonstrated that deep neural networks (DNNs), when applied to textual data with thorough feature engineering, can achieve accuracy of up to 95%. Even so, the black-box nature of machine learning models limits their adoption in practical applications, as most studies prioritize accuracy over interpretability [17]. Moreover, ensemble methods predominantly rely on structured data, leaving textual information underexplored [15].

Breakthroughs in NLP have significantly expanded the possibilities for fraud detection research. For example, Wu & Du [1] utilized word embeddings to vectorize annual reports, providing a robust foundation for subsequent analyses. Craja etal. [14] applied hierarchical attention networks to extract nuanced semantic features from financial texts. Building on these advances, Bhattacharya & Mickovic [20] fine-tuned BERT for analyzing 10-K reports, achieving a 15% performance improvement over traditional models. Additionally, Wang etal. [21] proposed a multimodal model that incorporates attention mechanisms to jointly analyze textual and financial data, thereby offering deeper insights into the factors influencing fraud risk.

Despite these advancements, three critical gaps persist in the literature. First, regarding multimodal data fusion, although Wang etal. [21] demonstrated that multimodal models outperform unimodal approaches, there is a paucity of research examining the effects of data clustering within such frameworks. Second, in terms of model interpretability, previous studies have not systematically explored the specific mechanisms through which various predictors impact model outputs [22]. Third, regarding textual depth, current analyses of annual reports primarily focus on lexical, sentiment, and readability features, with limited attention to more nuanced semantic representations.

To address these gaps, this study integrates advanced deep learning models, multimodal data fusion, and explainable machine learning methodologies. This approach not only aims to improve the accuracy of fraud risk predictions but also seeks to elucidate the key textual and financial determinants underlying these risks. By enhancing both semantic modeling and interpretability, this research aspires to provide methodological innovations for the development of more effective fraud early-warning systems.

This study aims to develop a multimodal, data-driven model for predicting corporate fraud by integrating predictive factors across three dimensions: financial indicators, organizational structure, and annual reports. While financial indicators and organizational structure are represented as tabular data, annual reports are processed as textual data. The primary objective is to uncover latent fraud-related signals embedded within annual reports and to improve prediction accuracy through the integration of these diverse data sources. To this end, we first perform a comprehensive analysis of annual report content to establish a semantic description framework. Next, we combine features derived from financial indicators and organizational structure to construct the predictive model. Furthermore, explainable machine learning techniques are applied to quantitatively evaluate the contribution of each feature to fraud risk, thereby clarifying the model’s reasoning process and providing actionable guidance for regulatory authorities and stakeholders.

The research workflow is depicted in Figure 1. Feature extraction is conducted in the initial stage: financial and organizational structure indicators are obtained from the CSMAR database (www.gtarsc.com). Meanwhile, annual reports are parsed to generate four categories of textual features: tone, readability, content similarity, and semantics. These multimodal features are subsequently integrated into the predictive modeling process. Focusing on ensemble tree models, we utilize three widely adopted algorithms—LightGBM, XGBoost, and CatBoost—alongside LR as a baseline. After model training, hyperparameter optimization, and evaluation, the model with the best performance is selected. Finally, the SHAP (Shapley Additive Explanations) method is employed to interpret the model, assessing feature importance and their relationships with fraud risk, with particular attention given to fraud-related elements identified within annual reports.

Figure 1. Research workflow for financial fraud detection.

3.2.1. Semantic Features of Annual Reports

Annual reports, required by regulators, reveal a company’s operational status. Because worsening financial conditions often trigger corporate fraud, annual reports may contain signals of fraudulent behavior. To identify these signals systematically, we analyze the content of the report—focusing exclusively on the Management Discussion and Analysis (MD&A) section, which is authored by management and discusses causes, risks, and outlooks. The MD&A’s language reflects management’s mindset and strategy; under performance pressure or fraud incentives, it often grows more complex, ambiguous, or overly positive, indicating possible information manipulation. Therefore, the MD&A is a key textual source for detecting corporate fraud and management intent.

The MD&A section is lengthy but adheres to standardized disclosure rules, so we can split the section into nine definite thematic modules: 1) overall operational overview, 2) core business analysis, 3) non-core business analysis, 4) asset and liability status, 5) investment activities, 6) significant asset & equity transactions, 7) analysis of major subsidiaries and affiliates, 8) structured entities, and 9) future outlook. Each module is extracted using keyword-based regular expressions and truncated to fit the 512-character input limitation of the BERT model.

Subsequently, we conduct semantic encoding and dimensionality reduction for each module. For semantic encoding, we utilize the pre-trained BERT-Base-Chinese model, which features a 12-layer Transformer encoder with 12 self-attention heads per layer (Table 1 presents details on fine-tuning parameters). BERT generates a 768-dimensional semantic vector for each module, which is then reduced to between 3 and 20 dimensions through Principal Component Analysis (PCA) to enable integration with structured data features. The optimal dimensionality is determined by model performance across various settings. This reduction in feature dimensionality not only alleviates issues associated with high-dimensional data but also improves computational efficiency and enhances the model’s interpretability.

Table 1. BERT fine-tuning parameters.

3.2.2. Tone and Readability

The tone of an annual report corresponds to its textual sentiment features. In this study, we consider two indicators: the overall tone of the report and the tone of the MD&A section, both obtained from the CNRDS Annual Report Sentiment Database (www.cnrds.com). Consistent with the method proposed by Zeng etal. [23], the tone value is calculated using the Loughran-McDonald Financial Sentiment Dictionary. Specifically, as shown in Equation (1), the tone is determined based on the counts of positive words (POSword) and negative words (NEGword) within the text.

Readability, a linguistic feature that reflects text complexity, is typically inversely related to text length and the frequency of technical terms. While financial research uses various methods to assess readability, we simplify measurement by using the MD&A file size as a proxy, since text length correlates closely with cognitive load—longer texts require greater processing effort and thus indicate higher reading difficulty. In financial reports, which follow standardized formats, excessive length often signals structural complexity or redundancy, implying lower readability. Given this study’s focus on textual content across a large sample of financial documents, file size is employed as a simple and accessible readability proxy.

3.2.3. Content Similarity

Content similarity measures the degree of change in annual reports across consecutive fiscal years and is typically calculated using cosine similarity (see Equation (2)). Higher similarity indicates fewer new disclosures [24], a pattern that previous research has associated with increased fraud risk (e.g., Qian & Zhu [25], 2020). In this study, cosine similarity is computed using document vectors generated by Doc2Vec, a widely used technique that effectively captures contextual information for representing texts in similarity analyses. In Equation (2), V t represents the text vector for the annual report in year t.

3.3.1. Ensemble Tree-Based ML Algorithms

In this study, financial fraud detection models are constructed using logistic regression (LR), XGBoost, LightGBM, and CatBoost. Logistic regression serves as the baseline model due to its simplicity and interpretability; however, it is limited in capturing complex, nonlinear relationships. In addition, the performance of gradient-boosted decision tree (GBDT) algorithms is evaluated. As noted by Chen & Guestrin [26], GBDT integrates a set of weak learners through iterative optimization to minimize loss functions. Previous studies have demonstrated that tree-based ensemble methods consistently outperform deep neural networks on structured tabular data [22], primarily due to their ability to model nonlinear relationships, robustness during training, and adaptability in terms of parameter sensitivity, computational efficiency, and handling of missing values.

The GBDT variants employed in this study each offer distinct advantages: XGBoost enhances generalization and computational speed through regularization and parallelization; LightGBM reduces prediction error using histogram-based algorithms and leaf-wise tree growth; and CatBoost offers robust handling of categorical features while delivering high precision and stability.

3.3.2. Interpretable ML with SHAP

For the financial fraud detection model built using ensemble tree algorithms, we use the highly interpretable SHAP method to analyze the importance and mechanisms of various features. Rooted in Shapley value theory from game theory [27], SHAP decomposes model predictions into weighted contributions from input features, thereby enhancing the interpretability of complex machine learning models. Under the SHAP framework, the prediction output y i for a given sample x i can be expressed as a linear combination of all feature contributions, as shown in Equation (3):

Here, y b a s e represents the global mean of the target variable (baseline value), x i j denotes the j-th feature of sample x i , and f ( x i j ) corresponds to the SHAP value of the feature, reflecting its marginal contribution to the prediction. A positive SHAP value indicates a promotive effect on fraud risk, while a negative value suggests an inhibitory effect. By quantifying these directional impacts, SHAP elucidates the model’s decision logic. As illustrated in Figure 2, the baseline value y b a s e serves as the initial reference (e.g., average predicted probability). Arrows depict deviations from this baseline, e.g., feature x i 1 exerts a positive contribution ( f ( x i 1 ) > 0 ), elevating the prediction above the baseline, while x i 3 exhibits a negative contribution, reducing the prediction. The final prediction results from the algebraic sum of all feature contributions. SHAP provides post-hoc explanations applicable to various machine learning algorithms, particularly excelling in interpreting nonlinear decision processes of tree-based ensemble models.

Figure 2. SHAP Feature attribution diagram.

The target variable in this study is corporate fraud risk, while the predictors include financial indicators, organizational structure features, and dimensions derived from annual reports. SHAP analysis is employed to quantify the marginal contributions of each predictor to fraud risk, identify early warning indicators significantly associated with fraudulent activities, and provide deeper insights into the underlying mechanisms of corporate fraud.

4.1.1. Data Samples

This study utilizes violation records from the CSMAR database, focusing on firms listed on the Shanghai and Shenzhen Stock Exchanges between 2015 and 2020. Special attention is given to five prevalent forms of financial fraud: fictitious profits, inflated assets, false entries, major omissions, and disclosure misstatements. To ensure sample consistency and avoid confounding factors, financial firms are excluded due to their unique operational characteristics and higher financial risk. The resulting dataset comprises 1226 firms involved in financial misconduct, accounting for a total of 2652 fraud cases.

For the control group, non-fraudulent firms are selected from the CNRDS ESG-R database, which has provided ESG (Environmental, Social, and Governance) ratings for all Chinese A-share listed companies since 2007. Employing the ESG ratings as the criterion for non-fraud control firms has clear theoretical and empirical support. The governance dimension comprehensively reflects a company’s standards in risk management, internal control, financial reporting quality, and information transparency. A higher governance score typically indicates a sound governance structure and effective oversight mechanisms, thereby reducing the likelihood of fraud. Cohen etal. [28] found that governance quality directly affects the reliability of financial reporting; Dechow etal. [29] showed that weak governance significantly increases fraud risk; and Tamimi and Sebastianelli [30] further demonstrated that higher scores under the ESG framework are associated with greater information transparency and lower fraud risk. Governance-related ESG indicators, including risk management capability, financial reporting quality, and disclosure transparency, have been proven to be relevant for fraud detection [31].

In the study, non-fraudulent firms are matched to fraudulent firms in a 1:1 ratio by year and industry. Additional criteria require non-fraudulent samples to have higher governance scores, no violation records from 2015 to 2020, and not be under ST (Special Treatment) status. These systematic matches yield 2652 non-fraudulent firms, thereby providing a robust foundation for comparative analysis.

4.1.2. Feature Variables

1)FinancialIndicators

Nineteen financial indicators across five dimensions—solvency, profitability, growth potential, operational efficiency, and cash flow capacity—are used (see Table 2). Solvency indicators assess default risk, profitability measures earning ability, growth reflects expansion, operational efficiency captures asset utilization, and cash flow evaluates management of liquidity. Unusual variations in these indicators, such as sharp shifts in profitability or cash flow, can signal potential financial fraud.

Table 2. Financial indicators (19 Items).

2)OrganizationStructureFeatures

Organizational structure related features are classified into two main categories: ownership concentration and top management characteristics (see Table 3). Ownership metrics evaluate the proportion of shares held by major shareholders, with particular emphasis on the concentration among the top five shareholders, a key indicator for fraud detection in China (Qian & Luo [7]). Additionally, a higher shareholding by the largest shareholder serves as an effective deterrent to fraud, as it strengthens shareholder oversight and control [32]. Top management characteristics encompass variables such as dual roles, management shareholdings, and other related attributes.

Table 3. Features related to corporate organizational structure (11 Items).

3)AnnualReportCharacteristics

Using the method from Section 3.2, we extract semantic features, tone, readability, and content similarity from the MD&A (see Table 4).

Table 4. Features related to corporate annual report.

Following the overall research framework (see Figure 1), three main experiments were conducted. First, semantic features were extracted from MD&A texts using BERT-based classification models across nine thematic modules to predict corporate fraud risk. The resulting theme-specific semantic vectors were then subjected to dimensionality reduction, yielding compact representations for the entire MD&A section (denoted as TextFeature_i, where i = 1 ⋯ n ).

In the second experiment, feature fusion was performed by combining these semantic vectors with financial indicators, organizational structure metrics, and additional features related to tone, readability, and content similarity from annual reports. An ensemble tree-based model was optimized using grid search to achieve the best predictive performance.

The third experiment employed the interpretable SHAP method on the best-performing model to clarify the contribution of each feature to fraud risk, thereby identifying key indicators of fraudulent behavior. Model performance was evaluated using accuracy, F1-score, and ROC-AUC.

4.3.1. Semantic Features of MD&A

Table 5. MD&A content-based corporate fraud detection model performance.

Table 5 presents the fraud prediction performance of BERT models across nine MD&A thematic modules. The Significantasset&equitytransactions module achieves the highest predictive performance (ROC-AUC = 0.697), followed by Overview (ROC-AUC = 0.695) and Futureoutlook (ROC-AUC = 0.670). This performance hierarchy can be explained by the informational value of each section: 1) Significantassetandequitytransactions documents substantial transactions with direct financial implications; 2) The Overview synthesizes the company’s comprehensive status, effectively encapsulating management’s core messaging; and 3) FutureOutlook conveys development expectations and profitability projections, offering critical insights into management intentions and potential fraud motives. Conversely, the Investmentactivities module demonstrates relatively weaker predictive power (ROC-AUC = 0.632), suggesting its limited relevance to fraud risk assessment.

Semantic vectors (CLS outputs) from the three top modules—Overview, SignificantAssetandEquityTransactions, and FutureOutlook—were dimensionally reduced to remove redundancy and emphasize key information. These reduced vectors were evaluated with XGBoost, LightGBM, CatBoost, and logistic regression (LR); CatBoost performed best and was selected for the final model. Finally, the reduced semantic vector from the best-performing module (SignificantAssetandEquityTransactions) was concatenated with other annual report features and corporate financial and organizational structure indicators to form a multi-source feature vector, which was fed into CatBoost (see Table 6).

4.3.2. Integrated-Data Financial Fraud Detection Model

Table 6 presents the results of corporate fraud prediction using three ensemble tree-based algorithms and LR, which incorporate features from multiple levels, including financial and organizational structure indicators, semantic features derived from MD&A, as well as tone, similarity, and readability. The ensemble tree algorithms consistently outperform the baseline LR model across all evaluation metrics. CatBoost achieves the highest performance, with ROC-AUC scores of 0.859 for models built on the SignificantAsset&EquityTransactions module. The corresponding dimension of semantic features for the module is 7, demonstrating that even low-dimensional semantic representations derived from the MD&A section provide valuable information for detecting financial fraud.

Table 6. Performance of integrated-data fraud detection models.

4.3.3. Feature Importance Analysis

We utilized the optimal CatBoost model for SHAP analysis, incorporating semantic features from the SignificantAsset&EquityTransactions module alongside 41 additional features. Figure 3 presents the top 20 features ranked by SHAP values, where the X-axis represents feature importance and different colors distinguish between feature categories.

Figure 3. SHAP-based feature importance analysis (Top 20 features).

Among the top 20 features, financial indicators constitute 50%, with organization structure and annual report features each accounting for 25%. Financial indicators have a median rank of 7, significantly higher than the other two categories, underscoring their dominant role in fraud detection—particularly metrics related to operational efficiency (e.g., accounts receivable turnover, Actrcbto; total asset turnover, Totastto) and profitability (e.g., return on equity, Roe_1).

Within the realm of organization structure, the shareholding ratio of the largest shareholder (ShrHolder1) and the board’s shareholding ratio (BShrRat) are particularly noteworthy. The former reflects the concentration of corporate power, while the latter indicates both the alignment of interests between the board and the company and the board’s capacity for independent oversight and decision-making. These findings highlight the importance of power distribution and interest alignment in predicting corporate fraud risk.

Among annual report features, similarity (Similariy), tone (LM_tone2), and two semantic indicators (TextFeature₆ and TextFeature₁) rank among the most important variables. Similarity reveals year-over-year changes in report content, while tone indicates management’s confidence in performance and outlook. Although semantic features are less influential than tone or similarity, they still meaningfully aid in fraud detection, suggesting that annual reports encode information closely tied to corporate fraud.

4.3.4. Analysis of Impact Factors

The SHAP beeswarm plot (Figure 4) visually demonstrates the relationship between the top 20 features and financial fraud risk. The x-axis represents SHAP values, indicating both the magnitude and direction of each feature’s impact. Positive SHAP values suggest an increased fraud risk, while negative values indicate a decreased risk.

Figure 4. SHAP beeswarm plot of top 20 features.

This study demonstrates that integrating semantic extraction, data fusion, and multimodal feature incorporation can substantially improve corporate fraud prediction. The fusion-based model, which combines financial indicators, organizational structure variables, and annual report features, achieves a notable ROC-AUC of 0.859. While financial indicators remain the core predictors, the addition of organizational and textual information significantly enhances predictive accuracy and provides deeper insights into fraud detection.

Notably, ensemble tree models exhibit a clear advantage over traditional logistic regression, thanks to their ability to capture the complex, nonlinear patterns inherent in real-world fraud data. Consistently attaining ROC-AUC scores above 0.85, these ensemble approaches provide a more robust basis for fraud detection in complex environments.

The integration between textual and tabular data further strengthens model performance. Leveraging BERT for semantic extraction from MD&A sections, followed by dimensionality reduction via PCA, enhances both computational efficiency and the model’s sensitivity to fraud-related signals. Analysis shows that disclosures in the Overview, Significantasset&equitytransactions, and Futureoutlook sections of annual reports are particularly predictive of fraudulent activity, which suggests that companies engaged in fraud often focus manipulation on key narrative and transactional elements within their reporting.

Beyond overall predictive success, model interpretability—enabled by SHAP analysis—provides important forensic insights into the relationship between individual features and fraud risk. Financial metrics stand out as primary indicators: reduced operating efficiency, profitability, and growth are all closely tied to a heightened risk of fraud. Regarding organizational structure, a higher concentration of shares held by major shareholders is negatively correlated with fraud risk, supporting the idea that concentrated ownership can deter fraudulent actions [33]. In contrast, board shareholding has a nuanced effect. Moderate ownership aligns directors’ interests with shareholders’, reducing agency problems and fraud risk. But excessive ownership can have the opposite effect: large stakes increase board control, weaken external oversight and the balancing role of independent directors, and raise information asymmetry and opportunities for manipulation. Directors with substantial holdings may conceal poor performance or manipulate financials to protect share prices or personal wealth. Empirical studies support this nonlinearity: Jensen and Meckling [34] argue managerial ownership reduces agency conflict only up to a point; Larcker etal. [35] find fraud risk rises when insider ownership exceeds an optimal range; and Yang etal. [36] report that excessive ownership alignment can introduce decision biases. Overall, board ownership and fraud risk appear to follow an inverted U-shaped relationship: moderate ownership improves governance, while excessive ownership concentrates power and increases fraud risk.

Linguistic and behavioral signals embedded in annual reports also add another layer of valuable insight. Key features such as year-over-year content similarity, tone, semantics, and readability help surface indications of hidden or manipulative practices. Large changes in report content or a shift away from an optimistic tone often accompany increased fraud risk, perhaps as management attempts to obscure irregularities. Module-level analysis highlights the particular importance of the Overview, Significantasset&equitytransactions ,and Futureoutlook, while variations in readability provide additional red flags. Overall, how management adjusts report language and complexity can reflect underlying intentions and psychological states. These patterns in annual report content—whether in narrative style, explanation, or structure—offer meaningful clues, allowing for earlier and more accurate identification of potential corporate fraud.

This study presents new evidence supporting the effectiveness of machine learning ensemble tree algorithms in identifying corporate financial fraud. By employing an innovative method for semantic feature extraction, joined with feature dimension reduction, we modularized extensive annual report texts. This approach not only facilitated the seamless integration of textual and tabular information but also offered a practical solution for managing complex multimodal financial data.

Such an approach has meaningful implications for practice: extracting module-level risk signals lets regulators and practitioners efficiently screen many disclosures and target fraud-related content. For example, if the FutureOutlook text closely repeats the prior year while financial analysis shows a sharp drop in profitability and cash flow, the system can auto-trigger a risk alert to prompt a focused investigation, enabling earlier detection and more precise allocation of regulatory resources. The study also extends risk identification to linguistic behavior, examining how tone, and readability reflect management’s psychological and behavioral tendencies during report preparation. Quantifying these linguistic features aids investors and regulators in assessing corporate transparency and integrity culture and provides a basis for regulators to refine disclosure-review standards, making oversight more scientific, dynamic, and evidence-based.