Financial Fraud Identification Model of Listed Companies based on

Time-Series Information

LILI WANG

Lyceum of the Philippines University Manila Campus,

Manila 1002,

PHILIPPINES

Abstract: - The aim of this research is to establish a high-precision financial fraud identification model for

model has a high effect on the processing of financial data, and the error between the data after dimensionality

reduction and the original data is small. In addition, the clustering effect of the model also shows a clear pattern

of fraud. In practical application, the accuracy rate of this model is as high as 94.89%, showing high accuracy

and recall rate, and its F1 value is 87.08%, showing its feasibility and effectiveness in practice. The results

highly prove that the performance of the financial fraud identification model proposed in this study is excellent,

and it has a wide application prospect in the future.

Key-Words: - Temporality, Finance, Fraud, Local linear embedding, Recognition clustering, Time-Series

Information, Financial index.

finance. Financial fraud has a serious impact on

investors, companies and the market as a whole.

Therefore, establishing an accurate and reliable

financial fraud identification (FFI) model is crucial

to protect investors' interests and maintain market

stability. The study aims to investigate the

identification model of financial fraud in listed

companies grounded on time-series information

indicators. Time-series information indicators

include financial indicators, financial statements and

main methods include logistic regression, decision

tree, random forest, etc. The advantage is that

powerful machine learning technology can

[2], used the empirical research method to study the

impact of digitalization on financial inclusion

through questionnaire survey and interview of BOP

population, combined with economic model. Its

advantage is that it combines both qualitative and

quantitative research methods to obtain

comprehensive and in-depth insights. Through the

combination of temporal information indicators, the

study hopes to understand and identify financial

fraud more comprehensively. The innovation of the

effectively distinguish between financial fraud and

non-fraud samples. Some studies have established

multiple linear regression models and support vector

machine classification models, combined with

which helps to reveal the characteristics and rules of

financial fraud.

The advantage of the research method is to use

the financial indicators of time series to investigate

the dynamic characteristics of the company's

financial data over time, which can more accurately

reflect the high performance of the model. Using

K-means algorithm, the pattern of financial fraud

can be identified, and some inherent characteristics

and laws of fraud can be further revealed. The

temporal information indexes, by introducing

temporal information indexes, the research provides

an in-depth understanding of the evolution of

financial fraud over time, which makes the model

more accurate and generalizable; two of which is the

adoption of SVM classifier as the core algorithm of

the FFI model, which has strong learning and

generalization capabilities, and helps to better

classify and predict; The third is that by combining

the K-means algorithm for clustering analysis, the

is to select the relevant time-series information

indicators, combined with SVM classifier and

K-means algorithm to construct a FFI model for

listed companies; the third part is to validate the

data downgrading of the established model through

the dataset, respectively, as well as the identification

performance; the fourth part is to summarize and

analyze the results of the study and put forward the

current study also has shortcomings and gives the

future research. Deficiencies and gives future

research directions.

The fraud of financial statements includes four types

of violations: fictitious profits, fictitious assets, false

records and general accounting improper treatment.

Non-financial statement fraud includes 12 kinds of

behaviors such as delayed disclosure, investment

violation, insider trading, illegal guarantee and so on.

The data set is divided into training set and test set,

in which the training set accounts for 70%, and the

test set accounts for 30%. The performance of the

constructed model was verified through

Financial statements of listed companies play an

important role in business as a communication tool

to convey the financial position of a company for a

specific accounting period to convey information to

the users of the statements. Many scholars have

proposed different models and methods to identify

financial fraud by analyzing financial data and other

related information. [4], proposed a new fraud risk

assessment model for listed companies that

fraud risk assessment model. The model was applied

to the fraud risk assessment of listed companies in

China, and the results showed that the model can

effectively assess the fraud risk of listed companies

and has higher accuracy, recall, and F-1 measure

compared to the traditional probabilistic model.

evaluate the influence of intellectual capital (IC) on

financial statement fraud of listed companies in

Tehran Stock Exchange (TSE). Therefore, the

research hypotheses were tested using logistic

regression model with integrated data technique. In

investing in ICs and their components, fraud in the

financial statements of commercial companies is

reduced.

The study by, [6], team aimed to analyze the

obligations of companies listed on the Athens Stock

Exchange and to identify and manage the business

risks, as well as disclose these risks to investors.

Effective risk identification and management

protects the business and adds value to shareholders

and other interested parties. Over the past few years,

study assesses the risk of disclosure through a

content analysis of the 2005-2011 annual reports of

non-financial companies listed on the Athens Stock

Exchange.

[7], study aims to integrate two streams of

literature related to fraud, namely, relationship and

overconfidence, to construct a fraud triangulation

framework that explains the mechanisms of fraud

commissioning and detection. A binary probit model

was used for the analysis. The study found that

leads to accounting and disclosure fraud and

facilitates the detection of disclosure fraud. It

facilitates the detection of disclosure fraud.

Relationships inhibit fraud in management and

disclosure but exacerbate fraud detection in the

presence of fraud; overconfidence leads to fraud in

state-owned and non-state-owned firms and

facilitates fraud detection in state-owned firms.

[8], study aimed to explore managers'

willingness to manage corporate surplus through

business activity intervention. Using a sample of

companies listed on the Stock Exchange of Thailand,

the suspected firms engage in real activity

manipulation to avoid losses or to smooth the firm's

earnings. Given that business activity interventions

are difficult to detect, market regulators in emerging

countries need to monitor such practices and

introduce legislation relating to this form of

corporate fraud.

for victims. The research team conducted a survey

experiment and found that a brief online educational

education intervention is effective in reducing

susceptibility to financial fraud.

listed companies based on time-series information

has made some progress. However, there are still

some challenges and problems, such as the

reliability of the data and the accuracy of the model.

Therefore, future research needs to be further

explored and improved to enhance the effectiveness

and accuracy of FFI.

Series Information

financial fraud, and the basis for FFI is the selection

of abnormal financial indicators. Indicator selection

of good and bad degree can determine the

effectiveness of the FFI model, so the conditions for

the selection of indicators should be satisfied with

the fraud problem itself related indicators, indicators

that can sensitively reflect fraudulent behavior,

indicators that can comprehensively reflect the

various types of fraud, indicators that are operable,

as well as the frequency of the relevant indicators.

statements relies mainly on the ethical standards of

managers, robust audits of financial statements, and

audit reports and opinions issued by auditors, [10],

[11], [12]. However, most financial statement frauds

are committed with the awareness or consent of

management. As the speed growth of computer

technology, various fields have entered the era of

big data and artificial intelligence, and machine

learning is broadly used because it can process large

amounts of data quickly and efficiently. Building a

FFI model based on temporal information, and the

specific flow of the model is shown in Figure 1.

Fig. 1: FFI model construction based on time series

information

The study has selected 26 indicators among the

financial and non-financial aspects of listed

companies, which include many aspects such as

indicators are expressed through numbers. Among

them, the indicators selected for the study are shown

in Table 1.

Table 1. Selection of financial and non-financial

indicators of listed companies

movements in financial data to identify fraud.

Taking this idea further, it may be possible to

.

Locally Linear Embedding (LLE) is an algorithm

capable of obtaining a low-dimensional

representation from a locally linear manifold of any

dimension, [13]. This algorithm has two uncertain

parameters: one is the diameter of the identified

neighbors and the other is the reduced dimension.

The LLE method, on the other hand, mainly utilizes

between the sampling points and then using the LLE

method with restriction, the minimum neighboring

right between the sampling points is found and then

the minimum neighboring right between the

sampling points is found. The LLE algorithm is

shown in Figure 2, [15].

solution process of this method, the first one is: that

neighboring points fall in or close to the locally

linear blob region, and then it describes the local

geometrical structure of those blobs by building the

linear coefficients for that data point and its

neighboring points to characterize the local

geometric structure of these pockets. The number of

rows with reconstruction errors is measured by the

cost function, which is formulated as shown in

formula (1), [16], [17], [18].

Subsequently the

is

obtained by minimizing the embedding cost

function and its expression is shown in formula (3).

low-dimensional space.

applications. The VC dimension is an important

measure of the complexity of a functional problem,

and the larger the dimension, the more complex the

problem, [19]. By improving the support vector

machine, the solution of the support vector machine

is made independent of the data dimensionality,

[20].

and a line in the middle is a classification function

that completely separates the two types of samples,

which is a linear classification function called

"hyperplane". If a linear function can accurately

separate the samples, then the samples are linearly

separated, and vice versa, the samples are

nonlinearly separated. However, not every sample is

linearly separable, so the data needs to be

transformed to a higher dimension so that they are

also better linearly separable. The core of the

SVM classification process starts with the need to

give the training set

, and the optimization

(4)

denotes the mapping

function, which maps the training data

dimensions;

and

denotes the model bias term. The study

transforms the optimization problem of formula (4)

into a dyadic problem to be solved as shown in

formula (5).

(5)

multiplier;

positive matrix; and

is

(6)

estimation of classifier accuracy by using k-fold

cross-validation (K-CV). First, the study divides

these samples into K samples, tests each sample,

and then tests the remaining K-1 samples to obtain

K samples. Under K-CV, we use the average

algorithm can effectively prevent the phenomenon

of "over-learning" or "under-learning", and the

conclusions obtained are more reliable. It is

assumed that there is a sample set

,

and the algorithm is trained in the sample set to get

a prediction function, in which the prediction can be

represented by

. K-fold cross-validation is to

taken as the training set. The prediction results of

the data in the sample set can be expressed by

proposes a classification algorithm based on

K-means. Using the k-mean method, multiple

targets are divided into multiple clusters, and the

resulting clusters have high cluster similarity and

small cluster features. Cluster similarity is a

measure of the average distance between objects in

a cluster, which can be used as both the center of

mass and the center of the cluster. The method is

used in the clustering process by randomly selecting

k targets that represent the mean or center of the

shown in formula (8).

subset is used as the training set. Repeat k times,

and the final model evaluation result is the average

probability of overfitting and improve the predictive

performance of the model in new data. Due to the

possibility of data skewness in financial fraud data,

warnings. The Accuracy indicator shows the

model's ability to accurately predict. The Recall

indicator focuses on the proportion of true financial

fraud detected by the model to all financial fraud,

and a high recall rate can identify the majority of

financial fraud. The F1 value can balance the

importance of recall and precision while also

improving the performance of the model. In

financial fraud detection, a higher F1 value indicates

that the model can better identify financial fraud and

7 Performance Analysis of FFI Model

for Listed Companies based on

Temporal Sequence Information

The study selected the financial data of 250 listed

companies to validate the LLE algorithm. In this

step, we use the original data without normalized

processing for dimensionality reduction, and the

results are shown in Figure 4.

Fig. 4: Unnormalized data downscaling

In Figure 4, the most critical parameter is the

number of neighboring points k. The experiment

obtains different results by changing the parameter k.

When the number of neighbor points takes the value

increased, the curve dimension will have 1 or 2

inflection points, and at this time, it is impossible to

Fig. 5: Dimensionality reduction of normalized data

increasing dimensionality, indicating that the

financial data used for the experiment were

successfully downgraded, and that the residual

values of the data become smaller as the

true intrinsic dimension; after the turning point, the

deviation between the raw dimension and the true

intrinsic dimension is very small. From Figure 5(b),

it can be seen that the residual curves become flat

and the residuals are almost the same at dimensions

greater than 3.

categorized into two main groups according to their

financial status: 155 with good creditworthiness and

95 with credit default risk. The study therefore set a

label "1" for companies with good creditworthiness;

and a label "-1" for companies with poor

creditworthiness. Once the data variables were

Fig. 6: Graph of the results of SVM parameter

selection

method of parameter optimization, which achieves

an accuracy of up to 82.254% by 5-fold

cross-validation. The experiments visualize the

output of SVM classification as a 2D problem and

the visualization results are shown in Figure 7.

Fig. 7: SVM classification visualization

only be drawn if the financial data label is +1 or -1.

Figure 7 successfully classifies the listed companies

in the test dataset into two categories, in which the

first category of data is clearly delineated from the

divides the enterprise financial indexes into ten

grades AAA, AA, A, A, BBB, BB, B, CCC, CC, C

CCC-D. In the face of the "ST" category the

company needs to be a loss for several years, it can

not be considered, and non-ST enterprises can not

be considered. In the face of the "ST" category of

enterprises that need to lose money in consecutive

years can not be considered, the non-ST enterprises

are clustered and divided into BB and B poles. The

specific classification results are shown in Figure 8.

Fig. 8: Distribution of data points for the clustering

K-means algorithm

of A-share listed companies that are not losing

distribution points will be overlapped and crossed,

but the distribution clusters are far away from each

other, which has a good classification effect.

model performance through a test set, while logistic

regression model, random forest model, and

XGBoost model were used for comparative analysis.