,QWHUQDWLRQDO&RQIHUHQFHRQ$UWLILFLDO,QWHOOLJHQFHDQG%LJ'DWD

A Financial Statement Fraud Detection Model Based on Hybrid Data Mining

Methods

Jianrong Yao Jie Zhang Lu Wang*

School of Information Management
School of Information Management School of Information Management
and Engineering
and Engineering and Engineering
Zhejiang University of Finance and
Zhejiang University of Finance and Zhejiang University of Finance and
Economics
Economics Economics
Hangzhou, China Hangzhou, China Hangzhou, China
e-mail: y6310@163.com e-mail: zj_4018@163.com e-mail: wanglu_hit@126.com

Abstract—Financial statement fraud has been a difficult variables. In section 4, we describe our experimental setting
problem for both the public and government regulators, so and procedure and section 5 introduces the methods we used
various data mining methods have been used for financial in this research, including classification methods and feature
statement fraud detection to provide decision support for selection. Section 6 gives the results of the prediction.
stakeholders. The purpose of this study is to propose an Finally, we discuss the contribution of this study and propose
optimized financial fraud detection model combining feature future work in section 7.
selection and machine learning classification. The study
indicated that random forest outperformed the other four II. LITERATURE REVIEW
methods. As to two feature selection methods, Xgboost
performed better. And according to our research, 2 or 5 In the past, people used to use expert analysis to find
variables are more acceptable for models in this paper. fraududent financial statements. In this way people may not
fully analyzed the report data for its huge amount and wild
Keywords-machine learning; financial statements fraud range, which caused many shortcomings in judgment. In
detection; feature selection recent years, data mining method has been widely used in
fraud detection to reduce the errors caused by experts’
I. INTRODUCTION judgment, including Internet fraud detection[13-15],
Association of Certified Fraud Examiners (ACFE) telecommunciations fraud detection[16, 17], financial fraud
defines fraud as "A kind of misrepresentation or deception detection[18], and fraud in other areas. These frauds are
that an entity or individual makes knowing that the it could hidden in huge information, and traditional experts’ analysis
result in some unauthorized benefit." According to a study sometimes fails to take into account the whole, and the
conducted by the ACFE, financial statement fraud accounts application of data mining method solves this problem.
for about 10% of white-collar crime. Once fraudulent These data mining methods include SVM[5, 6, 19], RF[20],
accounting practices happened, various actions will be taken DT[4], ANN, LR[21], etc. Most previous stuies use only 1-2
to maintain a sustainable appearance. data mining techniques, without comparisons between each
Considering financial statement fraud bring huge other; And most people overlook the importance of feature
property damage to investors, a large number of researches selection, which is more important than we think to some
have been conducted on the this area using machine learning extent.
methods such as ANN[1-3] ,DT[4] , SVM[5, 6], and text As for those variables, previous studies usually picked
mining[7]. Meanwhile, other fraud problems like credit card them from annual reports. These variables need to be
fraud[8, 9], internal transaction fraud[10] and insurance comprehensive and representative enough to cover all
fraud[11] have also been investigated. Given the different aspects of the company's operations. For example, quick
characteristics of each type of financial fraud, specific ratio and liquidity ratio reflect the solvency; sales growth rate
methods have been developed[12]. This paper puts forward a and EPS represent the profitability; and we can see operating
hybrid detection model for financial statement fraud, and this capacity from inventory turnover and turnover of total capital.
model have the advantages of (1) combining the financial In addition, linguistic variables from MD&A are used for
and non-financial data, (2) using two feature selection emotional analysis[22]. Petr Hajek found that compared with
methods, and (3) easy to explain. non-fraudulent companies, fraudulent companies reported a
During the research, we chose 120 fraudulent financial slightly higher negative sentiment in their annual report. That
statements disclosed by CSRC during the period 2007–2016. is to say fraudulent companies are more likely to use
Then, according to the industry and size of these companies, negative words.
we found the non-fraudulent company for contract. The rest The number of fraudulent companies used in previous
of this paper is organized as follows. In the next section, we studies ranged from tens to thousands, and the patterns of
review the previous literature on financial statement fraud research in each country were slightly different. Most studies
detection. Section 3 introduces financial and non-financial have adopted a matching method to match the non-fraudulent

978-1-5386-6987-7/18/$31.00 ©2018 IEEE 57

companies with the fraudulent companies[4]. The year, TABLE II. NON-FINANCIAL VARIABLES
industry, and scale are mainly used as matching criteria.
Variables Variable Description
Fraudulent financial statement detection is a binary
classification problem so it has four possible classification x18 The proportion of the largest shareholder
results[22]. Of the four results, the cost of two types of x19 Board of directors
misclassification is different, so cost-sensitive learning was x20 Board of supervisors
used to solve this problem[23]. x21 the proportion of independent directors
Our research proposes a hybrid fraudulent financial x22 LHSR
statement detection model combining the PCA and Xgboost
to do feature selection, and then, SVM, RF, DT, ANN and C. Descriptive Statistics
LR were applied to construct the fraud detection model, and
the classification accuracy of each model was compared to TABLE III. DESCRIPTIVE STATISTICS ON ALL VARIABLES
determine the optimal model.
Min Max Mean f S.D.
III. DATA x1 0.02704 8.78993 1.06034 f 1.05680
Our cases of fraudulent financial reports were disclosed x2 -81.28303 1160.93152 24.42644 f 105.28405
by the China Securities Regulatory Commission (CSRC). x3 0.16113 8.79113 1.54670 f 1.16503
120 listed companies were involved during the period of x4 0.36027 3324.42767 54.67923 f 244.50699
2007–2016, so a set of 120 annual reports were used as our x5 7.63125 220.50679 54.05857 f 24.27745
sample. We extracted the financial and non-financial x6 0.00365 974.98648 11.94133 f 70.19266
information from the annual report for these may cover all x7 -20.83488 4.02214 0.47371 f 2.30085
aspects of a company. Then we identified the matched x8 -1405.53898 95.37726 -7.93551 f 118.54589
sample of non-fraudulent companies based on their industry x9 -3.33460 4.32000 0.27264 f 0.81329
and size. Thus, we finally have 240 firms (120 fraudulent x10 0.00431 9.10804 0.77142 f 0.86311
and 120 non-fraudulent). x11 -46.80306 46.18315 5.42923 f 7.95734
A. Financial Variables x12 -475.60620 58.76700 -0.74138 f 49.67481
x13 -1.20000 2.32000 0.20917 f 0.43805
In order to identify the various types of financial
x14 -2.50919 0.86654 0.22006 f 0.29058
reporting fraud, the selected collection of financial variables
x15 0.00008 0.65354 0.16174 f 0.13355
should cover as many aspects as possible. [24]provides a
theoretical support to use financial variables. Table I shows x16 -5534.31950 3809.30900 32.96119 f 554.23800
our choice of financial variables. x17 -4077.52524 2177.45421 -65.03040 f 465.20060
x18 1.27555 78.97457 33.39915 f 15.25724
TABLE I. FINANCIAL VARIABLES x19 1.00000 11.00000 5.15000 f 1.56932
x20 1.00000 5.00000 2.04583 f 0.73892
Variables Variable Description x21 0.12500 0.80000 0.38819 f 0.08750
x1 Quick ratio x22 0.00000 78.84930 22.01327 f 20.89372
x2 Sales growth rate
x3 Liquidity ratio Table III shows basic descriptive statistics of these
x4 Operation revenue / average account receivable chosen companies, from which we can see that the range of
x5 Rate of return on total assets several indicators is large, such as x8, x16 and x17.Due to
x6 Inventory turnover the reasons of high dimension, high repetition, and having
x7 Operating profit / income before tax noise, original data need to be preprocessed, which includes
x8 Net cash content of operating profit data cleaning, data transformation and data reduction. In this
x9 NCFPS paper we removed all the records which contain missing
x10 Turnover of total capital value and used sigmoid function to solve the problem of the
x11 Return on assets inconsistent orders of magnitude.
x12 Operating profit ratio
IV. EXPERIMENTAL SETTING
x13 EPS
x14 Asset quality index First, we used feature selection to reduce its
x15 Accounts receivable / current assets dimensionality. As we known, feature selection play an
x16 Growth rate of net profit important role in pre-processing[24]. In this paper, we used
x17 Growth rate of net cash flow of operating activities
PCA and Xgboost to do it. Principal component analysis
(PCA) is a statistical method. By orthogonal transformation,
B. Non-financial Variables a set of observations of possibly correlated variables
converted into a set of linearly independent variables, which
Non-financial variables, different from financial data, are
called principal component. As you can see from Fig. 1, six
related to the corporate governance structure. Table II shows
variables among all are of great importance.
our selection of the non-financial variables.

58
 model. ANN is inspired by the central nervous system of
animals, which is a model of information processing of
applied neurons. The distributed structure makes it have the
same robustness as human brain, while ANN also good at
self-learning, self-organizing. Random forest improves the
decision tree by using the voting mechanism of multiple
decision tree, it is an ensemble algorithm in essence. For
example, if there are N samples and M variables
(dimensions), the specific process of random forest is as
follows: (1) Determine a value m, which is used to indicate
how many variables each tree classifier chooses. (2) Collect
k samples from the data set and use them to create k tree
Figure 1. The result of feature selection using PCA. classifiers. In addition, k bags of external data are generated
to be used for testing later.(3) After entering the classified
Xgboost usually known as a trump card in a variety of sample, each tree classifier will classify it, and then all
competitions in the area of data mining. It also has the classifiers will determine the classification result according
function of feature selection. The order of importance of all to the majority rule.
the variables is showing followed in Fig. 2.
VI. EXPERIMENTAL RESULTS
At the beginning, we use machine learning methods to
explore with all the variables, the accuracy of SVM, RF, DT,
ANN, and LR are show in Fig.3. It is obviously that SVM
was better than others in this condition. Random forest has
the lowest accuracy among these methods.

Figure 2. The result of feature selection using Xgboost.

Second, we applied the five classification methods to two

feature subsets selected by PCA and Xgboost, therefore ten
models were created. At last, the results were compared and Figure 3. Accuracy with all the variables.
analyzed.
Second, in order to test five methods, we added the
V. CLASSIFICATION METHODS variables to the model one by one follow the order of
There are many classifiers available in data mining. importance from high to low. Thus, we do experiments with
Different methods can be used for more comprehensive and the most two important variables first, and then added the
multi-angle analysis of data. third, and go on. The results based on the importance of
Logistic regression is the regression analysis for binary variables provided by PCA (Fig.4) bellow indicate that with
variables. From the angle of implementation, logic variables’ growing, RF performs better and more stable.
regression is simple, easy to understand and realize. Its run
fast and the computational cost is relatively low. SVM puts
the original data into a higher dimension one using nonlinear
mapping, and then an optimal decision hyperplane is
established to maximize the distance between the two closest
samples of the plane on both sides of the plane. Compared
with other non-linear methods, SVM requires relatively few
samples, and its goal is to minimize structural risk. SVM
performs well in noisy financial data[5]. Decision tree model
is a tree structure that describes the classification of an
instance, consisting of nodes and directed edges. The biggest
advantage of decision trees is their interpretability to the
Figure 4. Results based on PCA.

59
 In the same way, we test data based on the variables
provided by RF, and results are in Fig. 5.
Random forest still performs good when the number of
selected variables get bigger. LR reach the highest accuracy
when the number of selected variables is 6, but it’s not stable.
Third, in order to test two feature selection functions and
to find which variables are of great importance in machine
learning, we plot the average value of five accuracy (Fig. 6).
We can find that when the number is 2 or 5, we may get a
satisfied result.
Figure 5. Results based on Xgboost.

Figure 6. Average value of Fig.4 and Fig.5.

At last, to compare these five methods more intuitively, random forest has the following advantages: (1) it is good at
we put all variables’ combination into Fig. 7, from which we processing high-dimensional data; (2) it may avoid
find that RF do the best performance. overfitting to some extent; (3) it has good robustness and
stable results.
There are still shortcomings in this article: the data is not
large enough that only Chinese companies contained;
variables can be more various and more innovative. The
practical problems are far more complicated than the ones in
the article, and more factors should be taken into
consideration when designing variables.
REFERENCES
Figure 7. Results of all variables’ combination. [1] Kirkos, E., C. Spathis, and Y. Manolopoulos, Data Mining
techniques for the detection of fraudulent financial statements.
Expert Systems with Applications, 2007. 32(4): p. 995-1003.
VII. DISCUSSION [2] Saha, S.C., C. Lei, and J.C. Patterson, Detecting the financial
statement fraud: The analysis of the differences between data mining
This article has three contributions: First, to calculate and techniques and experts’ judgments. Knowledge-Based Systems,
analyze the factors that affect the fraud behavior. According 2015. 89(9): p. 459-470.
to our research, the effect of x12 and x13 on the model is [3] Wang, L. and C. Wu, A Combination of Models for Financial Crisis
confirmed by both two feature selection methods. Operating Prediction: Integrating Probabilistic Neural Network with Back-
profit margin is the ratio of an enterprise’s operating profit Propagation based on Adaptive Boosting. International Journal of
and operating income. There's a big risk of financial fraud Computational Intelligence Systems, 2017. 10(1): p. 507.
when there's a big loss in the firm and a lower profit margin, [4] Kotsiantis, S., et al., Forecasting fraudulent financial statements
using data mining. Enformatika, 2006. 3(2): p. 104-110.
it would be reflected from indicators related to profit. EPS is
generally used to measure common profit level and [5] Pai, P.F., M.F. Hsu, and M.C. Wang, A support vector machine-
based model for detecting top management fraud. Knowledge-Based
investment risk, and to reflect the operating results of an Systems, 2011. 24(2): p. 314-321.
enterprise. In order to cover losses or exaggerate earnings, [6] Huang, S.Y., Fraud Detection Model by Using Support Vector
the companies have a tendency to increase EPS. Second, Machine Techniques. International Journal of Digital Content
consider the influence of the number of variables on the Technology & Its Applic, 2013.
model. The results indicate that 2 or 5 variables are better [7] Cecchini, M., et al., Making words work: Using financial text as a
than the others. Third, we compared the performance of five predictor of financial events. Decision Support Systems, 2011. 50(1):
machine learning methods and found that among them, p. 164-175.

60
[8] Bhattacharyya, S., et al., Data mining for credit card fraud: A
comparative study. Decision Support Systems, 2011. 50(3): p. 602-
613.
[9] Olszewski, D., Fraud detection using self-organizing map visualizing
the user profiles. Knowledge-Based Systems, 2014. 70(C): p. 324-
334.
[10] Jans, M., et al., A business process mining application for internal
transaction fraud mitigation. Expert Systems with Applications, 2011.
38(10): p. 13351-13359.
[11] Bermúdez, L., et al., A Bayesian dichotomous model with
asymmetric link for fraud in insurance. Insurance Mathematics &
Economics, 2008. 42(2): p. 779-786.
[12] Ngai, E.W.T., et al. The application of data mining techniques in
financial fraud detection:A classification framework and an
academic review of literature. 2013.
[13] Hoque, N., et al., Review: Network attacks: Taxonomy, tools and
systems. Journal of Network & Computer Applications, 2014. 40(1):
p. 307-324.
[14] Kumar, P.A.R. and S. Selvakumar, Detection of distributed denial of
service attacks using an ensemble of adaptive and hybrid neuro-
fuzzy systems. Computer Communications, 2013. 36(3): p. 303-319.
[15] Zhang, B., Y. Zhou, and C. Faloutsos. Toward a Comprehensive
Model in Internet Auction Fraud Detection. in Hawaii International
Conference on System Sciences, Proceedings of the. 2008.
[16] Olszewski, D., A probabilistic approach to fraud detection in
telecommunications. Knowledge-Based Systems, 2012. 26: p. 246-
258.
[17] Olszewski, D., Fraud Detection in Telecommunications Using
Kullback-Leibler Divergence and Latent Dirichlet Allocation.
Intelligent Data Analysis, 2011. 16(3): p. 467-485.
[18] Wang, L. and C. Wu, Business failure prediction based on two-stage
selective ensemble with manifold learning algorithm and kernel-
based fuzzy self-organizing map. Knowledge-Based Systems, 2017.
121: p. 99-110.
[19] Yeh, C.C., et al., A Hybrid Detecting Fraudulent Financial
Statements Model Using Rough Set Theory and Support Vector
Machines. Journal of Cybernetics, 2016. 47(4): p. 261-276.
[20] Liu, C., et al., Financial Fraud Detection Model: Based on Random
Forest. International Journal of Economics & Finance, 2015. 7(7): p.
178-188.
[21] Dechow, P.M., et al., Predicting Material Accounting Misstatements
*. Social Science Electronic Publishing, 2011. 28(1): p. 17–82.
[22] Petr Hajek, R.H., Mining corporate annual reports for intelligent
detection of financial statement fraud –A comparative study of
machine learning methods. Knowledge-Based Systems, 2017. 128: p.
139-152.
[23] Kim, Y.J., B. Baik, and S. Cho, Detecting financial misstatements
with fraud intention using multi-class cost-sensitive learning. Expert
Systems with Applications, 2016. 62: p. 32-43.
[24] Guyon, I. and A. Elisseeff, An Introduction to Feature Extraction.
2006: Springer Berlin Heidelberg. 1-25.

61