Unveiling Financial Fraud: A Comprehensive Review of Machine Learning and Data Mining Techniques

For reviewing Fraud Detection Techniques in Financial Statements, we have taken 4 review papers and 31 articles.

Ashtiani and Raahemi [4] have raised 2 questions in their paper which were attempted to answer. The questions reviewed in the paper are:

(RQ1) Which financial statement-related datasets and fraud detection strategies were used in work?

(RQ2) Gaps, Suggestions and tends for further study in this field?

The articles may be found using search terms such as financial statements, fraud, machine learning, artificial intelligence, data mining, according to search algorithms that have been established. The aforementioned keywords have been used to search five well-known digital libraries. Total 187 articles were found, and filtrations were made to remove duplicates and adopted a snowballing method in addition to the earlier automated searches of digital libraries. The completed publications were sorted into groups based on comparable investigations, both in terms of methodology and datasets.

For the question RQ1, the papers considered were between 1995 and 2019 and were limited to conference proceedings and peer-reviewed journal papers. Fewer papers were from conference proceedings, and the majority were journal articles (32%).

As for the question RQ2, the studies confirmed that could provide new paths for next scholars to work in this field. Classification and regression techniques were also extensively used.

Ali et al. [5] used Systematic Literature Review (SLR) method that covers mainly 3 stages, planning review, review conduct, and reporting the review. In the 2nd stage Research Questions are Analysed and reviewed.

1. What are the most common financial frauds addressed by ML approaches?

2. What popular ML-based approaches are used to detect financial fraud?

3. What evaluation metrics are used to detect financial fraud?

4. What are the next research directions, trends, and gaps in the field?

When searched for the related articles 287 articles were found similar and finally 93 articles were selected for review that relates to research questions. The searched articles ranged between 2010 to 2021.

Ali et al. [1] analyzed data from 950 MENA businesses. Manufacturing, technology, telecommunications, energy, real estate, and insurance are some of the sectors that are represented. By identifying fraud on a collection of sample firms, the research seeks to construct an improved model for financial fraud prediction.

During the preprocessing phase, missing values were filled in with within-country mean values instead of deleting them one by one. The data was then normalized using the MinMaxScaler to make sure that all the features were on the same scale. To better understand the dataset, detailed descriptive statistics were gathered, and outliers were found and thrown out using the Isolation Forest method. The Gini importance value for each feature was used to figure out how important it was, and the most important features were chosen based on how well they related to the target variable.

Shen et al. [6] use knowledge graph models in conjunction with conventional features to study correlation information on fraud in financial statements detection. They also learn novel representations that are enhanced with feature embedding of different financial categories (FSFD). Features may be built as knowledge graphs with correlation relations serving as edges and features serving as nodes. Correlation types create feature relations. Research indicates that adding correlation information to financial feature representations enhances SVM and K-NN classification performance, but not as much as decision trees and logistic regression (Kernel).

In the preprocessing methods, fraudulent financial records are gathered, and a set of financial features are chosen to be analyzed. The Pearson correlation coefficient is used to find out how closely two or more financial variables are related. This information is then used to make a feature graph that can be used to spot both fake and real financial statements. Then, the correlations found in both types of statements are used to make a shared latent feature graph. This method finds positive, negative, and irrelevant correlations, which are very important for knowing how different financial factors are related.

Zhou et al. [2] recommended learning and representing topological characteristics in the financial system network into low-dimensional dense vectors using the graph embedding technique Node2Vec. This made it possible for the deep neural network to predict and categorize data samples from a large dataset in an efficient and intelligent manner. This approach provided an intelligent and distributed Big Data method for detecting Internet financial fraud. The strategy processes the massive dataset in parallel using Hadoop clusters and Apache Spark GraphX. The experimental findings demonstrate that by raising accuracy, recall, F1-Score, and F2-Score, the suggested strategy may increase the effectiveness of identifying financial fraud on the Internet.

The goal of this study by Albizri et al. [7] is to investigate, categorise, and synthesise current scholarly work in a variety of scientific fields on data analytics as well as machine learning towards financial fraud detection. A few research questions are generated and answered in this paper.

RQ1: What are the prime research topics and phases covered by data analytics and machine learning research for detecting financial statement fraud?

RQ2: How do researchers propose to investigate financial statement fraud using data analytics and machine learning?

RQ3: How can the findings of RQ2 be presented to guide future academic research?

The search terms were financial statement fraud, financial fraud detection, management fraud, and fraud examination. The above questions were researched in 60 papers. The following research topics were chosen.

•Creating Fraud Detection Models.

•Investigating and Predicting Fraud-Related Factors.

•Analytical Tools for Fraud Prediction.

•Literature review.

For RQ2, these papers are also divided into the following categories based on their research method:

•Analytical (Design Science, Secondary Data Based).

•Behavioural (Survey, Experimental, Field Study).

•Conceptual (Theoretical, Literature Review, Discussion).

Deng [8] use the Nave Bayes classifier to develop an FFS detection model. He chose 44 FFS based on auditing reports and 44 non-FFS based on a particular set of criteria from Chinese listed firms between 1999 and 2002 served as the experiment's training data set. Similarly, from 2003 to 2006, the data set for testing includes 73 FFS and 99 non-FFS. He trains the model with the training data set and then applies it to the testing data set, yielding good experimental results.

One of the preprocessing steps used in the study was data normalization, which lessened the effect of the different dimensions of factors. This is achieved by adjusting the sample data according to a specific formula to ensure uniformity across the dataset. Additionally, the selection of variables for the input vector is based on prior research and includes 26 candidate financial ratios, which are chosen to enhance the model's predictive capabilities. The study also emphasizes the importance of identifying significant variables that correlate with fraudulent and non-fraudulent financial statements to improve classification accuracy.

Chen et al. [9] suggest a two-step study method that uses both financial and non-financial factors to find an organization's scam early warning system. In this study, the approach includes selecting the data for a stepwise regression analysis, finding the TTF's significant variable after screening, and using that variable as the input for both the logistic regression and SVM. The study compares and looks at the data at the end in order to get better results with FFS classification.

The FFS enterprises from 1998 to 2012 were used as research samples. 66 enterprises were chosen from the Taiwan Economic Journal Data Bank's listed and OTC companies (TEJ). A total of 132 study samples are matched using the one-by-one pair approach, including 66 typical firms.

Chen [10] suggested a number of data mining methods, such as ANN, SVM, BBN, and DT. This investigation investigates OTC and lists Taiwanese businesses that, between 2002 and 2013, released fraudulent financial statements.

The study employs normalization of selected variables as a preprocessing method to prepare data for modeling. This step ensures that the variables are on a similar scale, which is crucial for the performance of many machine learning algorithms. Random sampling without repetition is conducted to create a training dataset, which helps in reducing bias and ensuring that the model is trained on a representative sample of the data. Additionally, rigorous tenfold cross-validation is utilized to test classification accuracy, providing a robust evaluation of the model's performance.

Goel and Uzuner [11] related publications on financial statement fraud detection that use natural language processing (NLP) to analyze textual information. In order to expand the scope of qualitative textual analysis and highlight the importance of emotion in fraud detection, this work looked at sentiments that manifest as linguistic expressions that are observable in written text. There hasn't been much prior study on the function of sentiment in fraud detection. They want to fill this research void by providing an alternative, rational approach to evaluating the qualitative data from annual reports and determining if sentiment is significant for identifying fraud.

Additionally, they selected only the fraudulent businesses that satisfied the following five requirements for the final sample: The claimed fraudulent activity should have an effect on the selected firms' annual reports, or 10-Ks; (2) the selected companies' 10-Ks should be retrieved from the Analysis, Retrieval (EDGAR) database,  and Electronic Data Gathering; and (3) the original 10-Ks, not the updated 10-Ks, should be downloaded from EDGAR; (4) the MD&A section of the chosen companies must be included in the 10-Ks; and (5) the chosen companies should have Between 1994 and 2012, this led to the detection of 180 fake firms.

As part of the preprocessing steps, all tables with financial information were taken out of the text of yearly reports' Management Discussion and Analysis (MD&A) sections. Also, all the numerical information that was in the MDAs' text was taken out so that only the qualitative material could be seen. This way of doing things made sure that the research focused on the meaning and language without using numbers, which could change the results. There were 360 MDAs in the final MDA text corpus, which included both fake and real entries.

Huang et al. [12] developed a method to detect and extract FFR topology patterns. In particular, the proposed method uses pseudo- and non-spurious samples to train GHSOM pairs with the same training parameters, and takes advantage of the unsupervised nature of learning and the evolving hierarchical structure of GHSOM to test hypotheses about counter-relationships between groups. This study also presents (1) a topological pattern-based classification rule to detect FFR and (2) a competitive-expert to capture the key features of cheating behavior. Empirical results from 762 144 publicly traded annual financial reports in Taiwan.

The following sources were used between 1992 and 2008 to detect false statements: Major Securities Violations and Indictments via the Securities and Futures Investor Protection Centre, and the lawsuit filed by Tadisian Juwan. Law and regulation search engine.

Before it can be used, the suggested method needs a number of data preparation steps, such as variable measurement, sampling, and choosing which variables are important to feed into the Growing Hierarchical Self-Organizing Map (GHSOM) model. Usually, the matched-sample approach is used in empirical studies that look at fraudulent financial reporting (FFR). There are many tools, like discriminant analysis and logistic models, that can help you choose the important factors. These tools help you find the most relevant features for the GHSOM input.

Kirkos et al. [13] investigated the efficacy of facts Using mining (DM) classification techniques, one may identify items connected to fake financial statements (FFS) and identify businesses that generate them. using statistics Mining strategies to discover control fraud may want to help auditors in their paintings. This observe investigates the detection of fraudulent monetary statements using choice trees, Neural Networks. The input vector consists of financial statement ratios. The 3 fashions' performance is as compared.

The information pattern protected data from seventy-six Greek production firms (no economic organizations had been included). The 38 FFS businesses have been paired with 38 non-FFS businesses. these companies had been categorized as non-FFS because there has been no indication.

The study employed a supervised discretization method for preprocessing, which utilizes class information to improve classification accuracy. This method was chosen after testing various discretization techniques, including equal depth and equal width. Supervised entropy-based discretization is particularly effective as it defines intervals that are more likely to enhance classification performance. Due to software limitations, value discretization was necessary to eliminate the effects of outliers, albeit at the cost of some information loss. This preprocessing step is crucial for the effective application of Data Mining techniques in detecting fraudulent financial statements.

Li et al. [14] discovered financial statement fraud by using data mining algorithms with financial and linguistic elements taken from yearly 10-k filings. Financial statement fraud is detected using a distance weighted discrimination (DWD) model when the sample size is less than the sample dimensions. In terms of generality, this model does well in HDLSS settings. To improve feature selection and parameter optimization for classifiers such as DWD, Back Propagation. We also used genetic algorithms. The suggested GA-based DWD model showed promise as a tool for identifying fake financial statements as it outperformed existing GA-based classification models in terms of classification accuracy while requiring less input data.

To verify that the data was current and reliable, the authors gathered annual report 10-k filings from publicly traded corporations that were accessible via the Securities and Exchange Commission's (SEC) website's Electronic Data-Gathering, Analysis, and Retrieval (EDGAR) database.

Over 111 companies' 2007 yearly financial reports were obtained from the EDGAR database. They thought that companies that weren't in the AAERs weren't necessarily dishonest. About 21.05 percent of the 57 firms we kept for our project were fake.

The F-score technique is used for filter-based feature selection, measuring the discrimination of features. Data standardization, such as z-score scaling, is applied to handle financial data with different units and scales. The combination of financial and textual features is also utilized to enhance the detection of financial statement fraud.

Wu and Du [15] aim to combine textual information from management comments in the annual reports of 5130 Chinese listed firms with numerical elements extracted from financial accounts to produce a better method for identifying financial fraud. First, we construct a financial index system that incorporates indices that have traditionally been ignored from prior research. The annual filings of Chinese-listed companies are then extracted using word vectors. Strong deep learning systems are then used, and the outcomes are compared with combination, text, and numerical data.

The preprocessing methods for financial fraud detection involve several key steps to ensure data quality and model performance. Data cleaning is performed to remove noise, distortion, or extreme values from the dataset. Missing values are addressed by removing samples with incomplete attributes. Scaling and standardization techniques are applied to bring features to a similar scale, enhancing model suitability. Chinese word segmentation is utilized to transform unstructured text into a numeric format, facilitating algorithmic processing. The Jieba package is selected for effective Chinese word segmentation during text preprocessing.

Andayani and Wuryantoro [16] aim to explain how corporate social responsibility, financial statement fraud detection, and sound company governance affect the decline in false financial statements. The 53 papers that were gathered from scientific journals and subjected to qualitative content analysis yielded results that could not be broadly applied. Quantitative studies usually describe the detection of fraud structure of financial statements through responsive elements in the data to the report frauds detection category system, whereas qualitative studies employ studies to explore more theories regarding fraud alongside additional interdisciplinary concepts over a longer period.

To eliminate the chance of deceptive financial statements, preventive actions are essential. Because there are still opportunities for fraudsters to exploit, the system for good corporate governance, which strengthens governance standards, only sometimes results in a reduction of false financial statements. The governance components subjected to quantitative analysis continue to provide a range of outcomes. In this instance, some areas have managed to put in place an efficient system of good governance to combat fraud, while others have not.

Sabatian and Hutabarat [17] set out to look into the part of the Fraud Triangle that makes it hard to spot fake bank records. The financial records of the Cigarettes and Cosmetics subsectors from the Indonesia stock exchange from 2016 to 2018 were used in this study. This study makes use of requirements-primarily based totally purposive sampling to acquire thirty samples of facts. analysis of records the usage of logistic linear regression. in step with the findings, rationalisation had a significant impact on economic statement fraud. economic stability, outside strain, non-public financial want, financial desires, vain monitoring, and enterprise nature, alternatively, haven't any touching on economic assertion fraud.

This study looked at seven cigarette subsector groups and five cosmetics subsector groups that were listed on the Indonesia stock market between 2016 and 2018. This study uses a method called "purposeful sampling" and has several conditions, such as:

1. Organizations indexed at the IDX inside the cigarette and cosmetics subsectors.

2. The employer became indexed at the IDX 12 months before the research length (2015) and persisted to be listed at the IDX in the course of the studies period (2016-2018).

3. Businesses which have an audited Annual document from 2016 to 2018.

4. The enterprise changed into now not delisted sooner or later of the statement duration.

As samples, the researchers received 18 annual reviews from cigarette groups and 12 annual critiques from beauty organizations. Documentation strategies in conjunction with downloading annual reviews from the website www.idx.co.identity are used to accumulate information. The IBM SPSS Statistic 25 software program software and logistic regression analysis techniques had been used to build up the information.

Sorkun and Toraman [18] examine how data mining techniques may be used to identify fraud in online ledgers using financial information. To do this, data sets were constructed using 72 sample e-ledgers along with a rule-based control system. From these, error percentages were calculated and marked. The tagged e-ledgers created statements of finances that were taught on nine differentiating characteristics using various data mining techniques. The training approach made use of Artificial Neural Networks, Decision Tables, M5P Trees, J48 Trees, Support Vector Machine, K-Nearest Neighbour algorithm, Linear Regression, and Decision Stump. The acquired outcomes are contrasted and explained.

Finding fraud in financial statements is made easier by choosing the right features during the preprocessing step, according to the study. Although there isn't a clear agreement on the best features to use, this shows that different ratios have been mentioned in the research. A rule-based e-ledger control program is used to label financial statements. For rule violations, a fraud score running from 0 to 100 is presented. For successful fraud detection, the preprocessing includes choosing 9 unique features from the bank statements.

Othman [19] conduct an evaluation on contemporary era-primarily based absolutely methods for detecting economic statement fraud. The motive of this paper is to explain the annoying conditions of predicting an extraordinary fraud event and to offer a know-how of the diverse records-mining-primarily based completely techniques for detecting economic declaration fraud. because con artists are becoming nimbler and are always coming up with a look at gives instructions for future studies in detecting the evolution of fraudulent economic reporting.

Anisykurlillah et al. [20] checked out the real-world evidence that shows how rationalization, industry type, outside pressure, financial goals, and financial security affect financial statement fraud. Don't forget to consider institutional ownership as a moderating factor. The LQ45 was made up of 58 widely traded companies on the Indonesia Stock Exchange in the 2016–2018 population study. With SPSS and deliberate sampling, descriptive and regression analysis were done on 29 companies.

Yadav and Sora [21] develop an improved FSF identification technique for qualitative data in financial reports using deep neural networks. Text is first pre-processed by tokenization, lemmatization, and filtering. The Harris Hawks Optimization (HHO) algorithm is then used to choose the features. Finally, a Deep Neural Network-Based Deer Hunting Optimization (DNN-DHO) algorithm is used to check if there is a fraud report in the financial accounts. Financial statement datasets in a Python environment were used to test the proposed FSF detection algorithm. When compared to conventional classifiers the new technique yields output with a high classification accuracy of 96%. Additionally, it yields superior outcomes across all performance indicators.

The preprocessing methods in text mining include filtering, lemmatization, and tokenization. Filtering involves removing unnecessary characters and stop words from the text, which are common words that do not contribute significant meaning.

Lemmatization breaks down words into their base or root form, which helps make the text data more consistent. Tokenization breaks the text into smaller components, such as phrases, words, or symbols, which are known as tokens. These preprocessing steps are crucial for enhancing the quality of the data before applying further analysis or classification techniques.

Omeir et al. [22] use two well-known fraud detection models created by Beneish [23] and Dechow et al. [24]. This article compares the predictive accuracy of financial fraud for Iranian firm statements between these two models. They start by attempting to determine the statistical description linked to the Beneish [23] and Dechow et al. [24] models' first and fourth quartiles. The t-test and variance analysis are then used using SPSS software to assess the forecasting capabilities of the models. They conducted an 11-year study, from 2009 to 2019, on 197 firms.

Zhao and Bai [25] propose a new method for detecting and forecasting financial fraud among publicly traded groups based totally on system studying. They accumulated 18,060 transactions as well as 363 monetary signs, 362 economic variables, and a category variable. They first ignored nine indicators that had nothing to do with economic fraud before processing the missing data. Subsequently, they identified 13 indicators out of 353 indicators that significantly impact financial fraud. These indicators are only dependent on a few distinctive selection patterns and the frequency of functions occurring across all algorithms. After that, the researchers developed three ensemble models, five single classification models, and ensemble styles with a voting classifier to anticipate economic fraud facts of publicly listed agencies. These models included the LR, RF, XGBOOST, SVM, and DT. After that, they chose the best single version out of five gadget learning algorithms as well as the enjoyable ensemble version from all hybrid models. The most helpful model parameters have been identified by comparing many version assessment metrics and using the grid seek technique.

The preprocessing methods involved handling missing values and outliers in the dataset. Indicators with more than 70% missing values were deleted, while those with 40% to 70% missing values were filled with 0. For indicators with less than 40% missing values, the mean filling method was applied. After preprocessing, 240 effective indicators were extracted, ensuring no missing values remained. Standardization was used to scale the data between 0 and 1 to mitigate accuracy issues due to varying scales of indicators.

To discover fraudulent monetary statements, Indrati and Claraswati [26] employ the idea of fraud diamond. To stumble on monetary announcement fraud, the modified Jones model is used. The changed Jones version is used to calculate the enterprise's accumulated income and receivables from credit income. The authors use the receivables ratio as a proxy variable from the character of the enterprise on this observe, ensuing inside the modified Jones version being the excellent research version for detecting monetary statement fraud. This study's population includes all assets and actual estate sector businesses that had been listed at the Indonesia inventory trade among 2015 and 2019. From 2015 to 2019, the sample consisted of 20 companies in the belongings and actual property sectors indexed on the Indonesia stock exchange (100 organization information with a 5-yr observation length). For statistical techniques consisting of more than one linear regression and speculation checking out, SPSS version 26 is used. economic stability, goal, and auditor exchange don't have any effect on economic announcement fraud, in line with this observe. meanwhile, fraudulent economic statements are inspired via external pressure, enterprise nature, and general accruals.

Humphrey et al. [27] analyze the conceptual framework for financial statement fraud detection used by the Fraud Pentagon. The study investigates the connection between fraud pentagon and financial statement fraud. In order to gather relevant and current information on the elements of the fraud pentagon pertinent to financial statements fraud, the current study used a library research technique. Based on the literature search, financial statement fraud is significantly impacted by the following components of the fraud pentagon: Competence (differences in the number of independent directors on the board and the company's directorship); Arrogance (frequent CEO images and CEO duality); Pressure  Opportunity (industry nature and ineffective monitoring); and Rationalization (auditor switching a This paper suggests that the fraud pentagon should be supported by legislation to combat financial statement fraud.

Malik [28] combines three different machine learning algorithms with two feature selection techniques (correlation and wrapper) using data from the Auditor General of India between 2015 and 2016 to find the best algorithm for distinguishing between fraudulent and non-fraudulent firms. Machine learning was implemented via a data science life cycle. The results and analysis demonstrated that machine learning algorithms significantly outperformed traditional fraud detection methods.

Alvadain et al. [29] present a new machine learning method to predict financial fraud. Various machine learning classifiers are fed transaction level features from a synthetic database of 6,362,620 transactions. Correlations between different characteristics were also investigated. An adversarial system of conditional generation is also used for Table data to generate an additional 5000 data samples (CTGAN).

The database, titled "Synthetic Financial Database for Fraud Detection," was obtained from the Kaggle database.

The study used a synthetic set of financial transactions with 6,362,620 samples, with a big difference between transactions that were false and those that were not. To fix this problem, a Conditional Generative Adversarial Network for Tabular Data (CTGAN) was used to make about 5,000 examples of fraudulent deals. Next, the dataset was divided into two groups: training and testing. The ratio of training to testing was 70:30, which meant that 70% of the data was used for training and 30% for testing. Heatmaps and scatter density plots were used to look at the patterns of association between features and get a better understanding of the data.

Rabade [30] examine and compile the corpus of information about the identification of intelligent fraud in business financial accounting. This study looks at several datasets being considered data mining methodologies. This research illuminates how to select the best approach for various types of datasets while keeping speed, accuracy, and cost in mind.

The primary objective of the study conducted by Saleh et al. [31] was to present scientific evidence regarding the correlation between causes of fraud and false financial statements. The research delved into extensive details on the utilization of Altman's z-score and Dechow f-score in detecting false financial statements among Jordanian industrial owners. Spanning from 2015 to 2019, the study considered false financial statements, while the nature of the firm, external pressure, financial security, and financial objectives were considered separate fraud elements. To assess the theories proposed in the study, a methodological model employing a multiple regression procedure was employed for research analysis. Although certain triangle fraud variables were found to have no association with fraudulent financial reports, other factors exhibited a strong correlation with fraud.

Chukwuma et al. [32] look for signs of accounting fraud that may be used to identify businesses that are more likely to falsify financial statement reporting and support the investigation of risky organizations. It is suggested to use a thorough forensic data analysis technique that incorporates all necessary phases of a data-driven methodology. To identify accounting fraud, the method uses machine learning models, logistic regression models, and financial ratio analysis. Because machine learning models are accurate, comprehensible, and inexpensive, they are very helpful in identifying fraudulent activity.

The preprocessing methods for the study include data cleaning and pre-processing to ensure accuracy and readiness for analysis. This involves removing any missing or duplicate data and correcting any errors in the financial statement data collected from publicly available sources. These steps are crucial as they prepare the dataset for subsequent analyses, such as financial ratio analysis, logistic regression modeling, and machine learning modeling, which rely on high-quality data to produce reliable results. Proper preprocessing enhances the effectiveness of the detection of accounting fraud by ensuring that the data used is accurate and relevant.

Jan [3] conducted a comprehensive analysis of both financial and non-financial data obtained from TWSE/TEPx listed companies spanning from 2001 to 2019. Their study encompassed a sample of 153 companies, out of which 51 were found to have engaged in financial statement fraud, while the remaining 102 companies were free from such fraudulent activities. To construct effective models for detecting financial statement fraud, the researchers employed two robust Remarkably, the empirical findings consistently demonstrated the superior performance of the LSTM model across all performance indicators. Notably, the LSTM model achieved an impressive accuracy rate of up to 94.88 percent, making it the most widely utilized performance indicator in this context. The study used a fake set of financial transactions with 6,362,620 samples, with a big difference between transactions that were false and those that were not. To fix this problem, a Conditional Generative Adversarial Network for Tabular Data (CTGAN) was used to make about 5,000 examples of fraudulent deals. Next, the dataset was divided into two groups: training and testing. The ratio of training to testing was 70:30, which meant that 70% of the data was used for training and 30% for testing. Heatmaps and scatter density plots were used to look at the patterns of association between features and get a better understanding of the data. The study used a fake set of financial transactions with 6,362,620 samples, with a big difference between transactions that were false and those that were not. To fix this problem, a Conditional Generative Adversarial Network for Tabular Data (CTGAN) was used to make about 5,000 examples of fraudulent deals. Next, the dataset was divided into two groups: training and testing. The ratio of training to testing was 70:30, which meant that 70% of the data was used for training and 30% for testing. Heatmaps and scatter density plots were used to look at the patterns of association between features and get a better understanding of the data.

The study employs data normalization and standardization as preprocessing methods, utilizing the MinMaxScaler to scale features within a range of 0 to 1. Random data splitting is conducted to create training and test datasets, with 75% of the data used for training and 25% for testing. The independent and dependent variables are defined, with the dependent variable being a dummy variable indicating financial statement fraud. The data matrix is reshaped to meet the input requirements of the models, ensuring compatibility with the RNN and LSTM architectures.

Cheng et al. [33] create a financial statement fraud detection model that considers missing values and imbalanced classes. First, missing values are removed using pairwise and listwise deletion. Second, it uses nonlinear distance correlation to find relevant qualities and proposes three combined attribute selection techniques. Third, it makes use of both oversampling and under sampling to rectify the unequal class distribution. Lastly, a collection of practical rules is produced by using rule-based classifiers. In actuality, a list of dishonest businesses is used in this research to get data on financial statement fraud.

The preprocessing methods utilized in the study include handling missing values and addressing imbalanced classes. Two techniques for handling missing values were employed: listwise deletion and pairwise deletion. Listwise deletion removes all instances with missing values, while pairwise deletion retains more instances by only deleting those with missing data for selected attributes. To address the imbalanced class problem, the study applied under sampling and oversampling techniques. Oversampling increases the number of records in the minority class, while under sampling reduces the majority class to achieve balance.

Antawirya et al. [34] studied the fraud pentagon's parts to find examples of false financial statements. The companies that were studied were all listed on the Indonesian Stock Exchange in the financial field between 2015 and 2018. After deliberate selection of the study sample, the data were looked at using multiple regression analysis. The study's findings support the fraud pentagon theory by showing that genuine financial statements can be found using parts of the model.

The fraud pentagon idea is examined by Ariyanto et al. [35] in relation to financial statements that are false. There is no way to test every piece directly. But there are other options. The pressure component is substituted with a personal financial necessity. The industry is starting to have an opportunity-driven character. The external auditors' attributes, together with the director transition, all point to competence and rationalisation. One measure of hubris is the frequency with which CEOs appear in pictures. Testing was done on pharmaceutical businesses that were listed on the Indonesian stock market between 2015 and 2019. They chose the samples using a purpose-driven sampling approach. Panel data regression is used to analyse the data. The analysis's conclusions indicate that some features of the sector have a favourable effect on false financial reporting. A sign of dishonest financial reporting may be a shift in senior management roles, such as directors. The falsified financial accounts of Indonesian pharmaceutical businesses are unaffected by factors such as the number of CEO appearances in images, personal financial need variables, or the calibre of external auditors. For the purpose they have used Dechow et al. [36] and Skousen et al. [37] models.

Ye et al. [38] evaluate many classification models for identifying financial statements that are fake. (FFS). Samples from earlier research were not handled realistically particular area. Thus, to learn unbalanced data, Random Forest is utilised in combination with SMOTE sampling. Some other useful performance measures are also provided. The experimental dataset spans the years 2007–2017 and includes 11726 publicly accessible Chinese financial declarations, 1314 of which the CSRC suspected of being fraudulent. The Random Forest approach is outperformed by Artificial Neural Networks (ANN), Logistics Regression (LR), Support Vector Machines (SVM), CART, Decision Trees, Bayesian Networks, Bagging, Stacking, and Adaboost.

The study uses SMOTE (Synthetic Minority Over-sampling Technique) to prepare the data. This technique mixes over-sampling and under-sampling to make the classifier work better. By setting k=5, random seed=1, and N=500 based on the sample imbalance ratio of 1:7.9, a new training sample of 7884 fraudulent and 10412 non-fraudulent instances is made, giving the system a fraud ratio of 1:1.32. In addition, the mean is used to replace missing numerical attributes, and the mode is used to replace missing word attributes. This works better than Listwise Deletion.

Sawangarreerak and Thanathamathee [39] examined financial ratios from Thai Stock Exchange financial statements for possible fraud trends using financial ratio separation that were connected to false financial statements were found by them. This research is unique in that it examined the probability of fraud for every financial item using mathematics. Six financial elements connected to fraud were also found by this study: Gross profit is the first, primary business revenue is the second, primary company revenue in relation to total assets is the third, capitals and reserves are the fourth, and capitals and reserves are the fifth. Five is the ratio of long-term debt to total capital and reserves; six is the ratio of accounts receivable to main company income. Three additional financial elements set this research apart from others.

The dataset was processed using RapidMiner Studio version 9.8, applying binning discretization methods to classify the data. Two types of binning were utilized: equal width and equal frequency, with varying numbers of bins (3, 5, and 10 for equal width; 3, 5, and 8 for equal frequency). The comparison of these methods aimed to determine the most suitable approach for identifying associated patterns in financial items. Ultimately, binning with equal width using five bins was deemed appropriate for detecting fraudulent financial statements.

Each of the models this paper talks about have its own pros and cons. It's easy to understand and explain linear regression, and it can be adjusted to keep it from overfitting. Using stochastic gradient descent, it is also easy to add new data to linear models. When there are non-linear relationships, linear regression doesn't work very well. They aren't naturally adaptable enough to handle more complicated patterns, and it can be hard and take a long time to add the right interaction terms or polynomials.

Decision trees can learn connections that aren't linear and can handle outliers pretty well. In real life, ensembles work really well and have won a lot of traditional (not deep learning) machine learning events. Individual trees that are not limited can keep growing until they remember the training data, which can lead to overfitting. Ensembles, on the other hand, can help with this.

At the moment, deep learning is the best way to do things in some areas, like computer vision and speech recognition. Deep neural networks are great at processing picture, audio, and text data, and batch propagation makes it easy to add new data to them. The number and structure of their layers can be changed to fit different kinds of situations, and the fact that they have hidden layers makes feature engineering less necessary. Because they need a lot of data, deep learning algorithms aren't generally good for general-purpose tasks. When it comes to standard machine learning problems, tree ensembles usually do better than them. In addition, they take a lot of computing power to learn and a lot more knowledge to tune (set the architecture and hyperparameters).

This means that each training report is saved by nearest neighbours algorithms, which are "instance-based." After that, they guess what will happen with new observations by looking for the most similar training observations and adding up their scores. These methods use a lot of memory, don't work well with data that has a lot of dimensions, and need a meaningful distance function to figure out how similar two sets of data are. Most of the time, it's better to use your time to train regularized regression or tree ensembles.

Regression based on logic The outputs make sense in terms of probability, and the method can be tweaked to keep it from fitting too well. With stochastic gradient descent, it's easy to add new data to logistic models. When there are multiple or non-linear choice boundaries, logistic regression doesn't work as well as it could. They aren't adaptable enough to easily show relationships that are more complicated.

There are a lot of different kernels for SVMs that can be used to model decision boundaries that are not linear. Also, they don't tend to overfit, especially in spaces with a lot of dimensions. But SVMs use a lot of memory, are harder to tune because choosing the right kernel is so important, and don't work well with bigger datasets. Random forests are generally better than SVMs in the business world right now.

Naive Bayes (NB) models work fairly well in real life, especially for how simple they are, even though the assumption of conditional independence rarely holds true. They are simple to use and can grow as your information does. Because they are so simple, NB models are often beaten by models that have been properly trained and tuned using the methods above.

K-Means is by far the most famous clustering algorithm. It's quick, easy to use, and surprisingly flexible if you prepare your data ahead of time and add features that are useful. It won't always be easy for the user to say how many groups they want. Also, K-Means will not make good clusters if the real underlying clusters in your data are not spherical.

The best thing about hierarchical clustering is that the groups aren't always supposed to be round. It also works well with bigger numbers. You have to pick the number of groups (i.e., the level of the hierarchy to "keep" after the algorithm is done), just like with K-Means.

DBSCAN doesn't depend on globular groups, and it can handle more users. Also, not every point has to be put into a cluster. This makes the clusters less noisy, which could be a weakness based on your use case. The user has to change the hyperparameters 'epsilon' and 'min_samples,' which determine how dense the groups are. You can tell a lot about these hyperparameters by how DBSCAN works.

Ali et al. [5] used the SLR method, this is a thorough process for gathering and examining all research that addressed certain research issues. It is used to find and compile data that is focused on certain concerns in order examine the reasoning behind the results and judgements of reviewers. Review reporting, review execution, and review planning are the three main phases that this SLR research addresses.

For RQ1 only Financial Statement Fraud is taken for consideration and the result is displayed in Table 2.

A lot of researchers have looked into how to use ML methods for RQ2 to find financial fraud. Artificial Neural Networks (ANN), Decision Trees, Genetic Algorithms (GA), Fuzzy Logic (FL), Support Vector Machines (SVM), Hidden Markov Models (HMM), K-nearest neighbours (KNN), Bayesian models (BN), Decision Trees, and Logistic Regression are some of these. Table 3 lists the methods that were used and how often they were used.

Previous studies have shown the effectiveness of SVM, NN, DT, and LR in the job of fraud detection, according to Ali et al. [1]. The models SVM, DT, and LR were selected to compare with XGBoost. Since Small along with medium-sized data sets get the greatest results from decision tree-based algorithms, they selected the RF and AdaBoost algorithms from the ensemble techniques. The comparison phase techniques were tested, and the models were trained using the Scikit-learn Python package.

Extreme Gradient Boosting (XGBoost) is the main technique used in the study to find Financial Statement Fraud (FSF). The Synthetic Minority Oversampling Technique (SMOTE) is used to improve the representation of the minority class and fix the mismatch between classes. For comparison, different machine learning methods are used, such as AdaBoost, Random Forest, Decision Tree, Support Vector Machine, and Logistic Regression (LR). Putting these methods into practice is done in Python, with the Scikit-learn tool being used for training and testing models.

Shen et al. [6] used the Pearson's correlation coefficient to analyse correlation information among financial parameters. Knowledge graphs use entities and their interactions to represent information. They postulate that the link between financial elements in knowledge graphs has statistical qualities, and that the hierarchical information in these graphs may have an impact on the identification of financial statement fraud. They start by randomly choosing financial characteristics and generating vector space representations for every financial statement. Once they have made a shared latent feature network with correlation links, they use the ANALOGY method to learn how to embed financial features. They use matrix-vector multiplication to project the financial records onto the feature latent space as the third step.

To find financial statement fraud (FSFD), this study uses both traditional machine learning classifiers and knowledge representation learning methods. We use five standard models to compare them: naive Bayes with kernel (NB), Support Vector Machines (SVM), k-nearest neighbour (K-NN), Decision Trees (DT), and Logistic Regression (LR). This model, called ANALOGY, learns how to embed financial features from a shared latent feature graph that is made up of correlations between financial features. Using the Pearson correlation coefficient to measure correlation and make feature graphs for both fraudulent and non-fraudulent financial statements is part of the approach.

Zhou et al. [2] proposes an intelligent and distributed bigdata approach to increase the effectiveness of Internet financial fraud detections. One of the biggest Internet financial service providers in China provided the first trial dataset. The dataset includes 192586 data samples, including 4375 fraud samples, after pre-processing. More than sixty data columns make up the dataset, including beginning amount, currency, income level, payment history, balance sheet, sale status, and more. The machine learning algorithms Node2Vec, DeepWalk, and SVM are compared in experiment groups, and the outcomes are assessed.

The paper proposes an intelligent and distributed Big Data approach for Internet financial fraud detection, which includes four main modules: data preprocessing, normal data feature extraction, graph embedding, and prediction module. The graph embedding module utilizes the Node2Vec algorithm to learn and represent the topological features of vertices in the network graph into low-dimensional dense vectors. The approach is implemented on Apache Spark GraphX and Hadoop to process large datasets in parallel, enhancing the classification effectiveness of deep neural networks. The methodology emphasizes the use of structural equivalence and homophily properties of Node2Vec for better feature representation.

Albizri et al. [7] used a variety of machine learning techniques, such as neural networks, support vector machines (SVM), Bayesian belief networks, decision trees, and regression-based algorithms, to identify fraud.

The study methods are put into four main groups in the paper: analytical, behavioural, conceptual, and design science. A large number of the papers (72) used analytical methods, mostly statistical analysis with secondary data or the design science study paradigm. Surveys, experimental schemes, and field studies were all examples of behavioural methods. Theoretical discussions and book reviews were all part of conceptual methods. The study aims to summarize what has already been written, find gaps, and suggest a framework for further work in the area of detecting fraud in financial statements.

To test how well the model could classify things, Deng [8] used ten variables. Six of these variables had high values in both the training and testing data sets. The FFS classification accuracy rates are 100% with ten variables and with six variables. The average accuracy rates with six variables are 95.45% and with ten variables they are 86.36 percent. The FFS classification accuracy rates in the testing data set are 89.04 percent with six variables and 83.56 percent with ten variables. The average accuracy rates with ten variables are 92.44%, and with six variables they are 90.70%. The results of the experiments show that the variables in Table 1 are very important and that the Nave Bayes classifier is good at recognizing FFS.

The financial and nonfinancial variables are tested for using as input variables in logistic regression and SVM by Chen et al. [9] using the stepwise regression screening approach. The research then looks at all model training and assessment techniques. Lastly, the research evaluates the benefits and limitations of the categorization correct ratio and provides suggestions in light of the analytical findings. Stepwise regression is used in the study to filter the variables and keep the research variables that have the most impact. The three models—Decision Tree (DT), Logistic Regression, and Support Vector Machines—have their prediction accuracy computed (SVM).

The study uses a mixed method that combines several ways to predict fake financial statements (FFS). To find the most important factors, it uses stepwise regression. These variables are then fed into logistic regression, support vector machine (SVM), and decision tree (DT) models. Out of these, the decision tree C5.0 model does the best job of classifying, with a 93.94% success rate. It then moves on to the logistic regression and SVM models, which are 83.33% and 78.79% accurate, respectively. Through variable screening and model comparison, the method tries to improve the percentage of correct classifications.

In this work, Chen [10] used tenfold cross validation, which is often regarded as cautious in the academic sector, along with a two-stage statistical treatment (data mining methods are applied). DT is utilised to determine which variables are the most significant and representative since this research has a greater number of variables. In this work, the DT variable selection programme is SPSS Clementine, and the variable selection tools are CART and CHAID.

Decision Trees (DT), Bayesian Belief Networks (BBN), Support Vector Machines (SVM), and Artificial Neural Networks (ANN) are some of the data mining methods used in the study to find fake financial records. A two-stage statistical treatment method is used, which is thought to be safer than earlier one-stage methods. Tenfold cross-validation is used in the study to make the classification models more accurate. The Chi-squared Automatic Interaction Detector (CHAID) and Classification and Regression Trees (CART) are also used to choose variables in the study.

Goel and Uzuner [11] study uses natural language processing (NLP) and machine learning to look at the role of emotion in finding fraud. Techniques for gathering information and rating features were used to figure out which traits are most crucial for finding scam. To automatically find mood traits, they used a lexicon-based method, which is also sometimes called a dictionary-based method.

A methodology for supervised machine learning called SVM was used to create a fraud sentiment classification model, using the Waikato Environment for Knowledge Analysis (WEKA) platform the fraud emotion classification model was made. They also used a linear kernel and the sequential minimal optimization method to make the SVM available in WEKA.

By dividing the total number of categorized occurrences by the number of properly identified instances, the accuracy of the fraud sentiment classification model was determined.

Sentiment lexicons are lexical resources containing words with their associated semantic orientation. Sentiment lexicons provide pre-made sets of words with labels indicating sentiments. which may be used to match input words in the text to establish their semantic orientation. They evaluated a number of sentiment dictionaries, lexicons, and tools for sentiment analysis before choosing just those which were either robust across numerous domains or specialized to a particular domain for automated sentiment feature detection.

They use the financial emotion lexicon by Loughran and McDonald (henceforth, LM), which is specialized to our industry and contains terms that are often seen in financial material. Next, I retrieved sentiment characteristics from genuine and fraudulent MD&A corpora for each of these LM categories using DICTION.

They explore the function of more complex linguistic features in identifying sentiment features that can be used to differentiate between genuine and fraudulent MD&As.

Natural language processing (NLP) and machine learning are used in the study to look at how people feel about fraud identification. It uses part-of-speech (POS) and sentiment lexicon-based features to tell the difference between fake and real Management Discussion and Analysis (MD&A) parts. To find the most useful features for finding scam, feature ranking methods like chi-squared (χ²) and information gain are used. We use the WEKA platform to build a support vector machine (SVM) classification model and a 10-fold cross-validation method to check how well the model works. The study looks at how people feel about things based on their polarity, subjectivity, and passion.

Huang et al. [12] suggested a unique use of unsupervised neural networks for dichotomous data (fraudulent and non-fraudulent, for example) efficiently. In order to investigate the dual GHSOM technique is suggested for general pattern detection of dichotomous data. The non-fraud-central geographical hypothesis is addressed first, followed by the fraud-central spatial hypothesis, using two dual GHSOMs.

•The sum of the variances between a node's weight vector and all of the transmitted input data is known as the node's mean quantization error, or MQE.

•Train each individual map: During each map's training phase, input data is loaded one at a time. The winner is the node that has the least distance between its weight vector and the freshly provided input data after all nodes' distances are computed.

•Horizontally expand every single map: Every map expands independently until, by Eq. (1), a new, separate node across it and its neighbor in the column or row if the stop condition is not fulfilled.

Avg(MQE) <  τ₁  × MQE_p                         (1)

•Extend or abolish the hierarchical framework: The following layer will be developed using the node whose MQEi is bigger than τ2 × MQE0. The hierarchy continues in this manner until every leaf node satisfies the stop requirement given in Eq. (2).

MQE_i < τ_{2 ×} MQE₀                                          (2)

Data pre-processing activities like variable measurement, sampling, and the identification of important variables to serve as GHSOM input variables must be completed before to putting the suggested strategy into practice. Traditionally, matched-sample designs are used in FFR empirical investigations. There are several helpful techniques for choosing important variables, such the logistic model and discriminant analysis.

The dual GHSOM technique has two mechanisms: (1) a feature-extraction mechanism that extracts FFR-related patterns/features for decision support, and (2) a classification mechanism that identifies fraudulent samples.

The study employs a novel dual GHSOM (Growing Hierarchical Self-Organizing Map) approach for detecting fraudulent financial reporting (FFR) and feature extraction. This method utilizes both fraudulent and non-fraudulent samples to train dual GHSOMs under identical parameters, allowing for the examination of topological patterns among their subgroups. Additionally, the approach incorporates a classification mechanism to identify fraudulent samples and a feature-extraction mechanism to capture FFR-related patterns. Principal Component Analysis (PCA) is also utilized for dimensionality reduction and feature selection.

Based on artificial intelligence and statistics, Kirkos et al. [13] give a lot of different ways to group things into groups. This study uses three methods that are well-known for their ability to sort data into groups. Some of these ways are neural networks, decision trees, and Bayesian belief networks.

Three different models were made, each using a different method. Sipina Research Edition software was first used to build the Decision Tree model.

The authors built the Neural Network model in the second experiment. They used Nuclass 7 Non-Linear Networks for Classification to create a multi-layer feed-forward perceptron network.

In the third experiment, they built a Bayesian Belief Network (BBN). Our Programme was TheBN Power Predictor.

The Bayesian Belief Network model produced the best results in a 10-fold cross-validation procedure, correctly classifying 90.3% of the validation sample. The accuracy rate of the neural network model was 80%, compared to 73.6 percent for the decision tree model. All the models had a reduced Type I error rate.

The paper employs three primary Data Mining classification techniques to detect fraudulent financial statements (FFS): Decision Trees, Neural Networks, and Bayesian Belief Networks. Decision Trees are utilized for their ability to represent acquired knowledge and extract IF-THEN classification rules. Neural Networks, particularly backpropagation networks, are chosen for their efficiency in handling noisy data and their popularity in prediction tasks. Bayesian Belief Networks are applied to capture dependencies among attributes and improve classification accuracy through conditional independence tests. The performance of these methods is compared based on their predictive accuracy in identifying FFS.

Li et al. [14] looked into Traditional multivariate analysis is useless when there is sample size, and classification techniques, such as support vector machines, encounter "data piling." Marron et al. proposed distance weighted discrimination (DWD) as a feasible approach to address HDLSS problems.

DWD could be improved by addressing the data piling problem from the distinct standpoint of classifier performance via the projection of samples in the normal vector's direction. With samples weighted according to distance, the goal of the DWD optimization problem is to minimize the sum of the reciprocals of the residuals. In HDLSS scenarios, a linear classifier achieves good performance. GA-based classification models are utilized in this work for feature selection and parameter optimization.

Three cross-validations were performed. To evaluate the classifier's performance, 5-fold cross validation was used to calculate the average accuracy of the training data for each fold.  Next, the testing sets prediction accuracy is calculated using the optimal parameter and chosen characteristics. Three equal-sized subsets of the raw dataset D are randomly selected to serve as independent testing sets T, while the other two subsets are used as training sets k = D - T.

Several classification techniques are employed to model the link between the input samples and the fraud outcomes. Optimizing parameters and feature subsets using genetic algorithms leads to better classifier performance.

The classifiers' accuracy determines the fitness score. Using the training and testing sets, are applied using individually produced parameters and chosen features.

The paper employs a combination of financial and textual features extracted from 10-k filings for detecting financial statement fraud. It utilizes the Distance Weighted Discrimination (DWD) model, which is effective in high dimension low sample size (HDLSS) contexts. A genetic algorithm (GA) is applied for feature selection and parameter optimization to enhance the performance of classifiers. The methodology includes grid search for optimal parameter settings and F-score technique for filter-based feature selection. Various classification models, including DWD, Support Vector Machine, Neural Networks, and Decision Trees, are compared for their effectiveness.

Different types of machines learning models, including random forest, SVM, XGB (eXtreme Gradient Boosting), ANN, and deep learning models, including CNN, LSTM, GRU, and transformer, were used with real-world data by Wu and Du [15] to come up with classification results. These algorithms classify fraud based on financial, non-financial, and textual factors found in the Business Situation Discussion & Analysis (BSDA) sections of publicly traded companies' yearly reports. The study showed that GRU got a score of 94.49 percent and LSTM got a score of 93.98 percent.

The paper employs deep learning methods for financial fraud detection, specifically utilizing Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models. It combines both numerical and textual data from Chinese listed companies' annual reports as input for the models. The study also incorporates techniques for Chinese word segmentation and embedding processing to transform unstructured text into a numeric format suitable for algorithmic processing. Additionally, the paper evaluates model performance using metrics such as AUC, sensitivity, specificity, and F1-score.

Andayani and Wuryantoro [16] Content analysis is the process of methodically reading a text, taking note of its recurring themes, and then extrapolating inferences on the meaning and purpose of the document (Hall & Wright, 2008). To get a more complete view of the governance (or set of procedures) and policies in place inside the organization, a qualitative content analysis was conducted (GCG). In addition to GCG, businesses should now use fraud detection systems and social responsibility.

The three system categories that comprise the coding scheme are social responsibility, governance systems, fraud detection, and faked financial statements. Every article emphasizes reading holistically and is coded. Additionally, there were debates about a number of papers where it was difficult to come to an agreement on whether the data and analysis were adequate.

The research employs qualitative content analysis as its primary methodology, which involves systematically reading and analyzing documents to identify consistent characteristics and derive conclusions about their purpose and meaning. This method was applied to a collection of 53 research articles from both Indonesia and abroad, focusing on themes related to financial statement fraud, governance (GCG), and corporate social responsibility (CSR). The coding scheme utilized in the analysis links falsified financial statements with governance mechanisms, fraud identification, and social responsibility, ensuring a holistic reading of the articles.

Sabatian and Hutabarat [17] As a stand-in for the Fraud Financial Statement, utilised the Benesh M-Score Formula.

The Fraud Triangle, which has three components—Pressure, Opportunity, and Rationalization—is used as the independent variable in this research. The concept of "opportunity" is separated into two categories: ineffective monitoring, represented by the proportion of independent commissaires (BDOUT), and industry nature, represented by the ratio of total inventory (INVENTORY). TATA, or the Total Accrual Ratio, is a stand-in for "rationalization."

A method called "purposive sampling" was used to look at 7 companies in the cigarette industry and 5 companies in the cosmetics industry that were traded on the Indonesia Stock Exchange between 2016 and 2018. Companies have to have been on the IDX since 2015 and have audited yearly reports for the time period of the study. Documentation methods, like downloading yearly reports from the IDX website, are used to gather data. To test theories about financial statement fraud, the study uses IBM SPSS Statistic 25 software and logistic regression analysis.

Sorkun and Toraman [18] use several types of machines learning to train the data. Specifics Data from the Balance Sheet and the Income Table made by 72 e-ledger were used during the training phase. Artificial Neural Networks (ANN), Decision Tables, Decision Stump, M5P Trees, K-Nearest Neighbor (KNN), Support Vector Machines (SVM), and Linear Regression were all used in the training process. These methods were put to the test in two different ways, and the results showed how well they worked. Right-sized. The Normalized Root Mean Squared Error (NRMSE) measure is used to figure out how well these methods work.

Apart from these investigations, three sets of labels were created and evaluated using the ranges of 0–40 is regarded as low, 41–80 as medium, and 81–100 as high. This grouping produced a fresh labelled data set, which was used to test several categorization techniques. Linear regression and M5P trees were substituted with measurements are used to assess the success rates of these techniques' results.

Several data mining techniques were used in the study to look for signs of fraud in financial statements drawn from e-ledgers. Artificial Neural Networks (ANN), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Decision Stump, M5P Tree, Random Forest, and J48 Tree are some of the methods that are used. There were two different methods used: one put fraud numbers on financial statements so that regression analysis could be done, and the other put statements into three groups: low, medium, and high fraud. Metrics like Normalized Root Mean Squared Error (NRMSE) and True Positive Rate (TP-Rate) were used to judge how well these methods worked.

The limited predictability of unusual occurrences, current technology-based techniques for identifying financial statement fraud, and upcoming research in identifying changing false financial reporting are all covered by Othman [19].

It is plausible to argue that catastrophic rare fraud occurrences are unlikely to occur given the limited predictability of rare events. On the other hand, offenders are still vital in order to alert interested parties to red flags of fraudulent activity early on. The shortcomings of earlier detection algorithms could be gradually improved with ongoing attempts to create a fraud detection model. Future fraud forecasts may become more accurate as a result of this.

The paper discusses various technology-based methods for detecting financial statement fraud, including data mining techniques such as regression, neural networks, decision trees, and Bayesian belief networks. Logistic regression is highlighted as a commonly used statistical method for predicting fraud patterns. The Bayesian classifier is noted for its effectiveness in assessing financial misstatement risks. Additionally, the random forest model is recognized for its accuracy and ability to handle large datasets. The study emphasizes the need for innovative techniques to keep pace with evolving fraud tactics.

Moderated regression analysis is used by Anisykurlillah et al. [20] to evaluate the model's hypothesis, assuming that the model's classical assumptions are followed beforehand. In this investigation, an absolute value difference test was used. The difference between the standardized absolute values of the two independent variables is found using the absolute value difference test. The next step determines the direction of the moderating effect of institutional ownership on dependent variable.

The study chose 29 companies from a group of 58 publicly traded companies on the Indonesia Stock Exchange using purposive sampling. These companies were exactly those that made up the LQ45 index from 2016 to 2018.The data were analyzed with SPSS, which did both descriptive and regression studies. To test the hypothesis about how independent factors affect financial statement fraud, moderated regression analysis was used, with institutional ownership acting as a moderating variable. In order to make sure the regression model was correct, classical assumption tests like normality, multicollinearity, autocorrelation, and heteroscedasticity tests were also run.

Yadav and Sora [21] use the four-step procedure for detecting financial statement fraud: of text classification. We present here DNN-DHO approaches to improve the DNN model for FSF detection. A performance comparison is made between the suggested technique and other classifiers.

Omeir et al.’s [22] approach of systematic deletion is utilized for sampling. The analysis was made with a number of fraud factors, and a sample of all the firms in the statistical population throughout the study period with one of those variables is selected for this purpose, with the other companies being removed.

To find the necessary variables pertaining to the financial statements of the firms under examination, they consulted the financial statements available in the Tehran Stock Exchange's computerized archive. Part of the information required came from the widely used database software developed. The most current updated accounting standards provided the data.

The M-score, or Beneish [23] model, was created by Prof. Messod D. Beneish to identify financial fraud. To assess the possibility of financial or earnings manipulation in a company's financial statements, the model uses eight financial measures.

The research employs two primary fraud detection models: the Beneish model [23] and the Dechow model [24], to compare their precision in predicting financial statement fraud in Iranian companies. A sample of 197 companies was analyzed over an 11-year period from 2009 to 2019, utilizing SPSS software for statistical analysis, including t-tests and variance analysis. The methodology includes identifying companies with specific instances of fraud as noted in their audit reports, which were then selected for further analysis. The study also examines the predictive accuracy of both models against actual cases of fraud confirmed by judicial verdicts.

Zhao and Bai [25] use the four-step procedure for detecting financial statement fraud: text pre-processing, feature selection, feature extraction, and text classification. We present here DNN-DHO approaches to improve the DNN model for FSF detection. A performance comparison is made between the suggested technique and other classifiers, such as DNN.

The Synthetic Minority Oversampling Technique (SMOTE) is used in this work to fix sample imbalance in financial data. It uses several feature selection methods to find the most important indicators that affect financial fraud. In the end, 13 key indicators were chosen. Five single classification models are created by the authors. These are XGBoost (XGBOOST), Decision Tree (DT), Random Forest (RF), and Logistic Regression (LR). They also use ensemble models with a vote classifier to improve the accuracy of their predictions. Performance measures like accuracy, recall, precision, and AUC are used to judge the models.

Indrati and Claraswati [26] Jones Model as one of their calculation models to calculate earnings management. Discretionary and non-discretionary accruals are the two categories into which the Modified Jones Model separates accruals. The acknowledgment of accruals of unbound and unregulated income or expenses that are the decision of management policy is known as discretionary accruals. The quality of financial statements may deteriorate if non-discretionary accruals are not followed, even when they are legitimate accruals that adhere to generally accepted accounting rules. The Modified Jones Model is utilised since credit income is reflected on an accrual basis. The study's element fraud diamond makes use of the receivable ratio as a proxy variable. The Modified Jones Model is therefore a highly effective technique for identifying financial statement fraud.

The study employs the Modified Jones Model to measure financial statement fraud, which distinguishes between discretionary and non-discretionary accruals. A purposive sampling technique is utilized to select a sample of 20 property and real estate companies listed on the Indonesia Stock Exchange from 2015 to 2019, resulting in 100 company data observations. Multiple linear regression analysis is conducted using SPSS version 26 for hypothesis testing, allowing the examination of relationships between independent variables and fraudulent financial statements.

Using the fraud pentagon theory as a tool, Humphrey et al. [27] looked into financial statement fraud. The fraud triangle model, which was used in the research, was investigated in accordance with Cressey's 1953 aspect of the fraud triangle. This feature focuses on situations when dishonest persons in positions of trust have betrayed trust. In 2004, Wolfe and Hermanson developed the fraud diamond model, which included the capability component in addition to the fraud triangle. The study also looked at the shortcomings of the previously mentioned model as discovered via information research of relevant and recent literature. According to the study's findings, pressure indicators such as institutional ownership, financial stability, financial aim, and external pressure are reliable and sufficient for spotting potential financial statement fraud. The research also examined opportunity and discovered that it depends on several elements, such as the industry and the size of the audit committee. The research findings validated the significance of these variables in mitigating and identifying financial statement fraud inside a business. The study also found that rationalization helps identify potential financial statement fraud. This is demonstrated by factors such as accrual levels and the number of auditors who switch. Additionally, the study demonstrates that factors like the shift in corporate directorship, which has a big impact on determining the likelihood of financial statements, can be used to estimate competency.

The study uses a library research method and focuses on looking at relevant and already published literature about the five parts of fraud that are involved in financial statement fraud. It looks at different parts of the fraud pentagon, like chance, pressure, rationalization, competence, and arrogance, and how they have big impacts on financial statement fraud. The study is divided into parts that explain what the study is about, talk about financial statement fraud, look at the fraud pentagon, and finally, say what they found about how these things relate to financial statement fraud.

Malik [28] used the data science life cycle to implement a conceptual framework for machine learning (DSLC). They asked two questions in the first phase:

How much can be predicted about fraud detection using machine learning techniques?

Which algorithms work best for this kind of task?

The second phase included gathering data from the UCI machine repository's publicly available Indian audit dataset of businesses. Between 2015 and 2016, the dataset was obtained from the Auditor General Office (AGO) in India. There are 776 occurrences and 26 numerical attributes in all. Enterprises from 14 different industries are included in the collection. Table 3 enumerates the industries and the number of companies in each.

Using Weka software, an exploratory analysis of the dataset was carried out in the third step. Because risks were unrelated to the prediction model, they were eliminated. The numeric attribute Money value had one missing value, which the exploratory analysis revealed. They used the ReplaceMissingValues filter to replace it with the mean.

The fourth phase involved NB, SVM, and RF are the three different modelling strategies that are used before proceeding with the experimental setup and choosing features for dimensionality reduction.

Alwadain et al. [29] used machine learning to guess how much financial crime there was. They used a conditional generative adversarial network for tabular data (CTGAN) to make over 5,000 cases after looking at a fake dataset of financial transactions. After teaching 27 machine learning algorithms, they carefully looked at how well the best model did.

To predict financial fraud using machine learning methods, the study used a made-up set of financial transactions. The Conditional Generative Adversarial Network for Tabular Data (CTGAN) was used to make about 5,000 more examples. With the usual settings and implementations, a total of 27 machine learning classifiers were trained and tested. Over and over, 10-fold cross-validation was used to make sure the evaluation was strong. The sample was split into training and testing groups with a 70:30 ratio. It was thoroughly tested and found that the model that worked best, XGBoost, was the most accurate.

Research trends, data sources, and ML/DM approaches were recognized by Rabade [30]. There are certain restrictions and questions about legitimacy, though. Utilizing an unbalanced data set and other supervised machine learning algorithms, the most effective and acceptable technique for identifying fraud. Following a comparison examination, they found that the K-Nearest Neighbour method outperforms Logistic Regression and Nave Bayes for detecting fraudulent transactions based on metrics like Precision/Recall and F1-Score.Owing to its benefits like increased dimensionality and precision, the Random Forest algorithm can be a wise option. Both regression and classification issues can be handled by it. Subsequent investigations have to concentrate on enhancing the precision of the models through the incorporation of intricate factors, utilization of unorganized data, and optimal training.

The paper explores various machine learning and data mining techniques for financial fraud detection, emphasizing their application in analyzing corporate financial accounts. It discusses the use of algorithms such as decision trees, logistic regression, K-Nearest Neighbour, and Random Forest for identifying fraudulent activities. The Random Forest algorithm is highlighted for its ability to handle both classification and regression tasks effectively, while K-Nearest Neighbour is noted for its superior performance in precision, recall, and F1-Score compared to other methods.

The study also emphasizes the importance of using historical data and confusion matrices to evaluate the performance of these algorithms.

A sample of manufacturing companies registered on the Amman Stock Exchange (ASE) between 2015 and 2019 was used by Saleh et al. [31]. 53 companies were included in the sample. The secondary data used in the analysis were taken from yearly reports that were released on (ASE). After subtracting the results that did not fit the study's criteria and were not part of the research period, there were 238 financial findings in all.

The research employs a thorough analytical method to determine firm bankruptcies and, as a result, financial statement fraud by using the Altman z-score model. According to the model's creator, according to research made 95% of firms' financial collapses may be accurately predicted one year prior to their demise. accuracy drops to 72% after two years and to 52% after three.

According to Altman's revised model, the likelihood of a business filing for bankruptcy is either too low or too high, depending on whether the z-score was accepted above 2.9 or below 1.2.

Dechow model states that a company's financial statements are thought to be clean if the F-Score is less than one (1); financial statements are thought to be clean if the F-Score is less than one If the F-Score is greater than one (>1), the company is probably lying. If the F-Score is zero, the company is probably telling the truth. (F-Score = 1), which means that there is a 1 in the chance that a certain event will end in a certain way, no matter what else may happen. As long as the F-score (FScore>1) is greater than 1, the expected probability may be greater than the unconditional probability. One more thing it could mean is that the company's books have been changed.

To test the study's theories, the research uses a methodological model with a multiple regression process. It mainly looks at the Altman Z-score model and the Dechow F-score model, which are two well-known mathematical models for finding fraud. The sample is made up of 53 manufacturing companies that are listed on the Amman Stock Exchange. Secondary data came from the companies' yearly reports. As part of the analysis, financial statements are put into groups based on how likely they are to fail, which makes it possible to look at fraud signs in a structured way. Financial ratios are also used in the study to look for fraud and judge the security of the finances.

Chukwuma et al. [32] 's research methodology includes the following steps:

1. Data collection: A sample of companies' financial statements will be collected from publicly available sources. Financial ratios, financial statement items, and information on accounting fraud cases will be included in the data.

2. Data cleaning and pre-processing: To ensure that the collected data is accurate and ready for analysis, it will be cleaned and pre-processed. This will include the removal of any missing or duplicate data as well as the correction of any errors.

3. Financial ratio analysis: For each company in the sample, financial ratios will be calculated. The purpose of these ratios is to find anomalous trends in financial data that might point to fraud.

4. Logistic regression modelling: Models of logistic regression will be trained to estimate the probability of fraud for every organisation in the sample using the financial ratio data. The models will be calibrated through the use of regularisation and cross-validation techniques in order to enhance their predictive performance.

5. Machine learning modelling: Machine learning models will be developed to forecast the probability of fraud for every organisation in the sample using the financial ratio data. The models will be calibrated through the use of regularisation and cross-validation techniques in order to enhance their predictive performance.

6. Model evaluation: To see how well the logistic regression and machine learning models work, we will look at their accuracy, precision, recall, and the area under the receiver operating characteristic curve (AUC-ROC). We will also look at how easy the models are to understand and how much they cost.

7. Data analysis: To find important patterns and trends in the data, the results of the financial ratio analysis, logistic regression modelling, and machine learning modelling will be evaluated. The results will be put to use in identifying businesses that are more prone to falsify financial statement reports and in supporting the investigation of higher-risk enterprises.

The study employs a thorough forensic data analytic approach that includes several key methods for detecting accounting fraud. Data collection involves gathering financial statement data from publicly available sources, including financial ratios and fraud case information. Financial ratio analysis is conducted to identify abnormal patterns in financial data that may indicate fraudulent activity. Logistic regression models are utilized to predict the likelihood of fraud based on financial ratios, with calibration techniques to enhance performance. Machine learning models are also trained on the financial ratio data to improve fraud detection accuracy and interpretability.

In order to evaluate financial statement fraud, this study, referenced as Jan [3], selected 18 criteria that are frequently employed. The set comprises four non-financial variables, often referred to as corporate governance variables, and fourteen financial variables. To build an ideal model, deep learning and optimization techniques like RNN and LSTM are applied. 60% dataset. The dataset's parameters are fitted during the learning process, and continuous optimization leads to the optimal prediction model. In order to validate the status and convergence with consistent accuracy, a random sample of 15% of the total data is used to create the validation dataset. Adjusting the hyper-parameters prevents overfitting. The test dataset, or the remaining 25% of the data, is used to assess the model's performance—that is, its ability to identify and generalize.

The study employs two deep learning algorithms: Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) for detecting financial statement fraud. RNN is utilized for processing sequential data, capturing relationships between data points, and preserving important past information during modeling. LSTM, an advanced version of RNN, is used for its superior performance in handling long-term dependencies in data. The research samples data from 153 TWSE/TEPx listed companies, combining both financial and non-financial data for model training and validation. Performance indicators such as accuracy, precision, sensitivity, and ROC curve AUC values are used to evaluate the models.

Cheng et al. [33] outline the six-step research approach that is being suggested. The following are the specific steps:

Getting the data ready: Thus, a survey of the literature served as the foundation for this research in order to identify the pertinent elements of financial statement fraud. Next, the collected dataset was assembled to include a total of 72 properties. Only 437 quarterly reports (181 fraudulent as well as 256 non-fraudulent) were kept after the 72 financial fraud characteristics were computed from the acquired financial fraud dataset. This was necessary since the query returns data had many missing values.

Dealing with null values Numerous missing variables in the gathered financial fraud dataset will have an impact on the categorization outcomes. The two approaches utilized in this study to address missing values are as follows.

(a) Pairwise deletion: After obtaining significant attributes through the use of five attribute selection techniques, the data from the chosen attributes was kept.

(b) Listwise deletion: Prior to applying the attribute selection technique to identify the crucial attributes, they eliminated every missing value from the financial fraud dataset that had been gathered.

Handling classes unevenly Unbalanced classes can be handled by many techniques; however, this study employed the SMOTE approach and random oversampling as two oversampling strategies to address the imbalanced class issue. In order to produce balanced classes, there are more minor class records when using the oversampling approach. By comparison, the minority class has almost the same number of entries as the majority class when the under sampling strategy is used.

Evaluation is a common technique for determining a model's efficacy, and in this step, 10-fold cross-validation is used to do so. After categorization, the efficacy evaluation metrics can be computed using a confusion matrix.

To deal with missing values in the dataset, the study uses both listwise and pairwise elimination. To find the important characteristics, it uses three combined attribute selection methods, one of which is a nonlinear distance correlation method. Both under sampling and oversampling techniques are used in this study to fix the imbalanced class problem. Random oversampling and the SMOTE technique are used especially. To make a useful set of rules for finding financial scams, rule-based classifiers like decision trees and random forests are used.

The study's population was defined by Antawirya et al. [34] as all financial sector businesses that were listed on IDX between 2015 and 2018. The method for gathering data is purposeful sampling. There are 47 firms that satisfy the requirements based on the sample selection criteria. To evaluate the data, multiple linear regression analysis was used.

In the present research, Ariyanto et al. [35] examine fake financial reports using Dechow et al. [36] 's fraud score model. "The F-score model consists of the sum of two variables, namely accrual integrity and financial performance," according to Skousen et al. [37]. A stand-in for the variable of personal financial status is insider ownership of shares (OSHIP). An insider's cumulative percentage ownership in a corporation is determined by multiplying their shares by the total number of shares that are outstanding. The present study employs total inventory divided by total sales of the research year reduced by total inventory divided by total sales of the preceding year as a proxy for the industry's characteristics. Additionally, an ordinal scale (dummy variable) is used to evaluate the calibre of independent auditors. Companies audited by the Big Four firms—Ernst & Young, Deloitte Touche Tohmatsu, KPMG, and Pricewaterhouse Coopers—are coded 1 while companies audited by other firms receive code 0. The proxy for competence in relation to changes in corporate directors is examined in the present research. An ordinal scale (dummy variable) is used in this assessment, with code 1 denoting a change in the company's directing roles over the observed time. If there hasn't been a shift in the company's directing roles over the observed time, it is coded 0. The total number of CEO images on the business's annual financial report for the study year is a measure of arrogance. The data was analysed using panel data regression analysis.

The study chose a group from all the financial companies listed on the Indonesia Stock Exchange from 2015 to 2018 using a method called "purposeful sampling. “There were 47 businesses that met the requirements for the study. Multiple linear regression analysis was used to analyze the data and figure out how the parts of the fraud pentagon relate to fake financial statements. The goal of the study was to find out how well the fraud pentagon model can find theft in financial statements.

Ye et al. [38] use the Random Forest technique to identify FFS. In practice, it is not uncomplicated to utilise a single basic classifier due to the large class imbalance issue. Moreover, the data shown here indicates that the Random Forest is not a good option for logistic regression in this scenario. SMOTE was selected in order to reduce biases resulting from a greater probability of past fraud. When dealing with an unbalanced dataset, it is also suggested to include performance measurements like precision, recall, F-Measure, Kappa Statistics, and AUC to accuracy.

The Random Forest algorithm is used in this study to find financial statement fraud (FFS). This algorithm is good at learning from data that isn't balanced. By making fake samples for the minority class, SMOTE (Synthetic Minority Over-sampling Technique) is used to fix the class mismatch. Normalizing numeric characteristics and imputation for missing values are part of the method and are used to improve the quality of the data. In order to judge how well the model works, performance measures like recall, F-Measure, Kappa Statistics, and AUC are used. The study also looks at how well Random Forest works compared to other predictors, such as ANN and logistic regression.

Association rule mining was used by Sawangarreerak and Thanathamathee [39] to identify fraudulent trends in financial statements from the Thai Stock Exchange.

The study uses a method with three key steps: processing and discretizing the data, using FP-Growth to find related patterns, and looking for and analyzing fraud patterns in financial statements. As part of data preprocessing, binning discretization sorts the information into different groups that make it easier to analyze. Association rule mining is used to find patterns of fraud in financial records from the Stock Exchange of Thailand for this study. Another idea from the study is to use different data analysis techniques, like clustering, to help find fraud trends that are linked.