Construction and Application of the Online Finance Credit Risk Rating Model Based on the Artificial Neural Network - Hindawi.com

Page created by Linda Shelton
 
CONTINUE READING
Hindawi
Discrete Dynamics in Nature and Society
Volume 2021, Article ID 6926216, 11 pages
https://doi.org/10.1155/2021/6926216

Research Article
Construction and Application of the Online Finance Credit Risk
Rating Model Based on the Artificial Neural Network

 Yufeng Mao,1,2 Zongrun Wang,1 Xing Li ,3 Chenggang Li,4 and Hanning Wang2
 1
 Business School, Central South University, Changsha 410083, China
 2
 School of Big Data Application and Economics, Guizhou University of Finance and Economics, Guiyang 550025, China
 3
 School of Management, Nanchang University, Nanchang 330031, China
 4
 New Structure Finance Research Center, Guizhou University of Finance and Economics, Guiyang 550025, China

 Correspondence should be addressed to Xing Li; xingli@mail.gufe.edu.cn

 Received 17 August 2021; Revised 6 September 2021; Accepted 11 September 2021; Published 30 September 2021

 Academic Editor: Ahmed Farouk

 Copyright © 2021 Yufeng Mao et al. This is an open access article distributed under the Creative Commons Attribution License,
 which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

 The low-cost, highly efficient online finance credit provides underfunded individuals and small and medium enterprises (SMEs)
 with an indispensable credit channel. Most of the previous studies focus on the client crediting and screening of online finance.
 Few have studied the risk rating under a complete credit risk management system. This paper introduces the improved neural
 network technology to the credit risk rating of online finance. Firstly, the study period was divided into the early phase and late
 phase after the launch of an online finance credit product. In the early phase, there are few manually labeled samples and many
 unlabeled samples. Therefore, a cold start method was designed for the credit risk rating of online finance, and the similarity and
 abnormality of credit default were calculated. In the late phase, there are few unlabeled samples. Hence, the backpropagation
 neural network (BPNN) was improved for online finance credit risk rating. Our strategy was proved valid through experiments.

1. Introduction cost, and robustness, and obtained the weight and score of
 each index through the analytic hierarchy process (AHP)
Online finance is a low-cost, highly efficient means to and efficiency coefficient method (ECM). Kamboh et al. [16]
provide services and attract consumers based on internet put forward several countermeasures to the online finance
platforms, making the financial system in China much more credit risk: enhancing the organizational management of
inclusive. Online finance can operate in various modes, online finance enterprises, controlling the risk of credit
including online money management, online payment, and businesses within a reasonable range, and optimizing the
online consumer finance [1–5]. Among them, online finance focus of the risk control department of online finance
credit provides underfunded individuals and small and platforms.
medium enterprises (SMEs) with an indispensable credit Some scholars have analyzed the credit transformation
channel [6–9]. To keep the risk within control, risk rating of banks in the context of online finance development
and control are necessary in the early and late phases after [17–21]. Considering the regulating effect of online finance,
the launch of online finance credit products [10–14]. Xu et al. [22] evaluated the influence of green credit mea-
Therefore, the suitable rating of the online finance credit risk sures on bank performance, constructed an evolutionary
can effectively promote the performance of online finance game model for the online finance enterprise and loan
enterprises in risk management, safe operation, and income enterprise, and verified that online finance can effectively
generation. weaken the influence of green credit on bank performance.
 Purohit and Kulkarni [15] constructed an evaluation Fayyaz et al. [23] conducted big data analysis on the financial
system for the online finance credit risk control of SMEs, data before, during, and after online finance credit products
which covers four primary indices such as portfolio, quality, are issued to SMEs, summarized the big data warning
2 Discrete Dynamics in Nature and Society

conditions (including risk process and internal evaluation Suppose there exists a dataset of n samples
application system, banking operation system, data ware- TS � (a1 , b2 ), . . . , (ak , bk ), . . . , (an , bn ) , where ai ∈ A is a
house platform, risk data market, capital metering, and sample in the dataset. The first k samples of the dataset are
decision support system), and constructed a risk manage- labeled: TSk � (ak+1 , bk+1 ), . . . , (an , bn ), . . . , (an , bn ) . The
ment and warning system based on the analysis results. remaining n− k samples are labeled:
 Most of the previous studies focus on the client crediting TSv � (ak+1 , bk+1 ), . . . , (an , bn ), . . . , (an , bn ) . For TSv , the
and screening of online finance. Few have studied the risk potential default samples and normal samples were deter-
rating under a complete credit risk management system mined based on the calculated similarity and abnormality of
[24–28]. Capable of solving most nonlinear mapping the credit default.
problems, backpropagation neural network (BPNN) can
correctly mine the deep relationships between the indices m
 2.1. Default Similarity. Suppose dataset TS � ai i , which is
and values of online finance credit risks. By virtue of the equal to 1, possesses two parameters: a hidden variable C �
superiority of the artificial neural network in nonlinear (c1 , . . . cm ) characterizing the cluster labels of samples and a
mapping, this paper introduces the BPNN to the risk rating class parameter Ψ(Ψ1 , . . . , Ψl ) of the model. Note that
of online finance credit products. The main contents of this cm ∈ {1, . . . , L}, and Ci � l indicates that class i has l
research are as follows: after the launch of an online finance members. According to the Bayesian theory, the probability
credit product, there are few manually labeled samples and distribution satisfies GV(Ψ, c|TS) ∝ GV0 (Ψ)GV0 (c)GV(TS
many unlabeled samples in the early phase. To adapt to this |Ψ, c).
situation, Section 2 designs a cold start method for the credit The index data in this research are all multidimensional
risk rating of online finance and calculates the similarity and continuous variables obeying the multidimensional
abnormality of the credit default. In the late phase, there are Gaussian random distribution. Let Ψ be expectation λ. Then,
few unlabeled samples. Hence, Section 3 improves the wolf GV(ai |Ψ, c) ∼ M(vci , Θ). Conjugate prior
pack algorithm (WPA) based on the belief learning model GV0 (Ψl ) ∼ M(0, ε2 J) can be applied to GV0 (Ψ) Let
and relies on the improved WPA to optimize the BPNN. The ml , l ∈ {1, . . . , L} be the number of data in class l. Then, the
improved BPNN was adopted to rate the online finance cluster prior GV0 (cl ) of each sample can be generated by the
credit risk, which effectively avoids the subjectivity in weight Dirichlet process mixture model:
assignment. The proposed model was proved effective ml
through experiments. ⎪
 ⎧
 ⎪ GV ci � l|c− i � ,
 ⎪
 ⎪ m − 1+β
 ⎪
 ⎨
 ⎪ (1)
2. Cold Start of the Risk Rating Model in the ⎪
 ⎪ β
 ⎪
 ⎪
 Early Phase ⎩ GV ci � L + 1|c− i � .
 m− 1+β
Online finance credit risk rating is a key step in the risk The first step is to randomly initialize parameters c and
control of online finance credit. Traditional online finance Ψ. Each ci can be sampled by
credit risk rating models need to be constructed based on
 GV ci � l|TS, Ψ, c− i ∝ GV ci � l|c− i GV ai |c, Ψ 
enough historical labeled samples. If there are not enough
such samples, the modeling personnel of the early-phase risk ml 1 T −1 (2)
 − (1/2)(ai − λl ) ai − λl 
rating model must fully understand the risk control rules of � ξ/2 1/2
 e .
 m − 1 + β (2π) |Θ|
the credit product. New methods are necessary to realize the
cold start of the online finance credit risk model. The probability of establishing a new class L + 1 can be
 After the launch of an online finance credit product, calculated by
there are few manually labeled samples and many unlabeled
 GV ci � L + 1|TS, Ψ, c− i ∝ GV ci � L + 1|c− i GV ai |c, Ψ 
samples in the early phase. To adapt to this situation, this
paper designs a cold start method for online finance credit
risk rating. By machine learning, the proposed method can � GV ci � L + 1|c− i GV0 ΨL+1 GV ai |ci , Ψ, ΨL+1 ξΨL+1
obtain useful samples out of many unlabeled samples and
 1/2
provide sufficient data for subsequent training of the online β 1 Θ′ −1 ′ −1
 Θ− 1 )ai
 e− (1/2)ai (Θ Θ Θ −
 T

finance credit risk rating model based on the neural network. �
 m − 1 + β (2π) ε |Θ|1/2
 ξ/2 ξ
 Let a � Sξbe the sample space of online finance credit
risk rating; B � {− 1, 0, +1} be the label space; and b be the −1
 1
 where Θ′ � 2 I + Θ− 1 .
indicator of sample labeling. If b � +1, the corresponding ε
sample is labeled “client default,” i.e., payment is not (3)
completed on time; if b � 0, the sample is labeled “normal,”
i.e., payment is completed on time; if b � − 1, the sample is When new data are added to class l, the parameter Ψl of
not labeled. that class should be updated by
Discrete Dynamics in Nature and Society 3

 m potential default sample. If many unlabeled samples are not
 GV Ψl |TS, c ∝ GV0 Ψl GV ai |c, Ψ normalized, the LT (XB (a)) value will be relatively high. To
 i�1
 prevent the problem, this paper adopts the mean path length
 � GV0 Ψl GV ai |c, Ψ of unsuccessful searches in the binary search tree for nor-
 ili − l malization. If there are M samples, then

 − (1/2ε2 )λTl λl − (1/2) (ai − λl )
 T
 − 1 ai − λl (M − 1)
 d(M) � 2XB(M) − 2 . (10)
 ∝e ili − l M

 ∼ M λ′l, Θ′l Let δ be the Euler–Mascheroni constant. Then, the
 number of harmonics XB (M) in formula (10) can be esti-
 − 1 m mated by
 ⎝λ′ � Θ′⎛
 where⎛ ⎝ ⎠ ′ 1
 ⎞ − 1 ⎠
 ⎞
 l l ai , Θl 2 J + Θ , dl � cc,l .
 il − l
 ε i�1 XB(M) � ln(M) + δ. (11)
 i

 (4)
 The normalized default abnormality VEH (a) can be
 Let λ be the cluster mean. Then, the distance from sample calculated by
a to a cluster can be measured by Euclidean distance:
 ������������� VEH(a) � 2LT(XB(a))/d(M) . (12)
 dist(a, λ) � (a − λ)T (a − λ). (5)
 Formula (12) shows the following:
 v
 If a ∈ TS , the minimum distance from a to the center of (1) If LT (r (a)) approximates d (m), VEH (a) ap-
a cluster of normal samples, i.e., the similarity between a and proaches 0.5, and the corresponding sample is not an
a class of normal samples, can be calculated by obvious default sample
 DAL(a) � minli�1 dist a, λi . (6) (2) If LT (r (a)) approximates 0, VEH (a) approaches 1,
 and the corresponding sample is a default sample
 The similarity between a and a class of default samples (3) If LT (r (a)) approximates m − 1, VEH (a) ap-
can be calculated by proaches 0, and the corresponding sample is a
 DLT(a) � minτi�1 dist a, λi . (7) normal sample

 There are very few client default samples in the scenario
of online finance credit risk rating. However, the model is 2.3. Sample Recognition and Weighting. SEM (a) and VEH
expected to recognize potential default samples more ac- (a) should be considered comprehensively to recognize
curately than normal samples. Comparing the results of potential default samples and normal samples from the
formulas (6) and (7), the similarity SEM(a) between a and a unlabeled dataset. Let ω be the weight that balances the
class of default samples can be modified as relative importance of SEM (a) and VEH (a). Then, the total
 score of online finance credit risk samples can be calculated
 DLT(a) by
 SEM(a) � . (8)
 DLT(a) + DAL(a)
 Formula (8) shows the following: RD(a) � ωVEH(a) +(1 − ω)SEM(a). (13)

 (1) If DLT(a) ≫ DAL(a), SEM(a) approximates 1, and Let φ be the threshold for a sample to be a potential
 the corresponding sample is a default sample default sample. To judge a sample as a potential default
 (2) If DLT(a) ≈ DAL(a), SEM(a) approximates 0.5, sample, the total score must satisfy
 and the corresponding sample is not an obvious
 default sample RD(a) ≥ ϕ. (14)
 (3) If DLT (a)
4 Discrete Dynamics in Nature and Society

 The mean α of normal samples RD (a) can be calculated O1
by
 qij
 |TS−l 1 | A1 qjk
 1 O2 H1
 α � − 1 RD ai . (17)
 TS i�1
 k B1
 This paper sets the weight of each labeled sample to 1. A2
 Input O3 … Output
For potential default samples, the higher the total score, the Hn
higher the confidence of its risk rating. The weight of these …
samples can be configured by Bn
 Am …
 RD(a) Ol
 q(a) � . (18)
 maxa RD(a)
 Input Hidden Output
 For normal samples, the higher the total score, the lower layer layer layer
the confidence of its risk rating. The weight of these samples
can be configured by Figure 1: Topology of the BPNN.

 maxa RD(a) − RD(a)
 q(a) � . (19) The BPNN training can be trained in two stages: forward
 maxa TS(a) − mina RD(a)
 propagation and error backpropagation. The network
 Finally, the loss function can be minimized by training can be detailed as follows.
 Based on the sample set of evaluation indices, the
 qi k bi , g ai + μS(q). (20) structure of input and output layers is determined for the
 i
 neural network. To initialize the connection weights and
 biases, the activation function g of the hidden layer is
3. Late-Phase Risk Rating Model generally a sigmoid function:
 1
3.1. Model Construction. BPNN needs to be trained by g(a) � . (23)
sufficient samples. During the training, the error is back- 1 + e− a
propagated through the network and used to iteratively During the forward propagation of information, the
correct weights and thresholds. The regularity and infor- output Oj of each hidden layer node can be calculated by
mation of the training data are memorized by the nodes in
 m
the network. During evaluation, the BPNN can output risk ⎝ q a + f ⎞
 ⎠
 Oj � g⎛ ij i j , j � 1, 2, . . . , k. (24)
values once the indices are imported into the network. Since i�1
there are few unlabeled samples in the late phase after the
launch of an online finance credit product, the BPNN was The output bj of each output layer node can be calculated
adopted to rate online finance credit risks, which effectively by
avoids the subjectivity in weight assignment. Figure 1 shows k
the topology of the BPNN. bl � Oj qij + εl , l � 1, 2, . . . , n. (25)
 Let Q and f be the connection weights and biases of the j�1
artificial neural network (ANN), respectively. Then, the
input of hidden layer node j can be solved by The error σ l of each output layer node can be calculated
 m by
 O � qji · ai + fj � qj A + fj . (21) σ l � Bl − Pl , l � 1, 2, . . . , n. (26)
 i�1

 The output of hidden layer node j can be calculated by Let ζ be the learning rate. During error backpropagation,
the activation function g of the hidden layer: the weights can be updated by
 m
 ⎝ q · a ⎞
 bj � g Oj � g⎛ ⎠ � G q A . (22)
 ji i j
 i�1

 n
 q∗ij � qij + ζOj 1 − Oj a(i) qjk σ k , i � 1, 2, . . . , m; j � 1, 2, . . . , k, q∗jl � qjl + ζOi σ l , j � 1, 2, . . . , k; l � 1, 2, . . . , n.
 l�1
 (27)
Discrete Dynamics in Nature and Society 5

 The bias can be updated by

 n
 fj � fj + ζOj 1 − Oj qjl σ l , j � 1, 2, . . . , k, εl � εl + σ l , l � 1, 2, . . . , n. (28)
 l�1

 The training should be terminated when the error rea- To optimize the response to future behaviors, every
ches the preset threshold or the number of iterations reaches inside individual will choose the strategy with the maximum
the maximum. probability of the optimal response in phase τ + 2.
 The BPNN was optimized with the WPA improved
 based on the belief learning model in the following steps:
3.2. Algorithm Optimization. The WPA is a stochastic
 Step 1: initialize algorithm parameters, such as pack
probabilistic search algorithm, capable of finding the opti-
 size, number of scout wolves, scout wolf movement
mal solution at a high probability. Paralleled search is an-
 factor, scout wolf movement distance factor, maximum
other feature of the algorithm: the search can start from
 number of iterations, and pack update ratio, and
multiple points at the same time to improve the algorithm
 randomly initialize the pack. Complete the topology
efficiency; the points do not affect each other. The traditional
 design of the BPNN.
BPNN converges slowly and easily falls into the local
minimum. This paper improves the WPA based on the belief Step 2: import the wolf positions and training samples
learning model and relies on the improved WPA to optimize into the neural network. Let NT be the total number of
the initial connection weights and biases of the BPNN. weights and biases and bi − pi be the output error of the
Figure 2 shows the process of the WPA. ith node. Compute the absolute error between expec-
 Figure 3 explains the idea of neural network optimiza- tation and prediction by formula (33), and take it as the
tion with the WPA improved by the belief learning model. order concentration function for individual wolves in
Firstly, a belief learning model was constructed for the smelling prey:
virtual game between multiple individuals. Let N be the NT
 
number of outside individuals and D(d1 , d2 , . . . , di , . . . dl ) WR � bi − pi . (33)
be the set of strategies of these individuals. The belief weight i�1
for another outside individual j to choose strategy di can be
calculated by Select the head wolf, scout wolves, and ferocious wolves
 τ in turn based on the calculated concentrations.
 ⎨ fj di + 1, selected,
 ⎧
 fτ+1
 j d �
 i ⎩ τ (29) Step 3: based on the results of the belief learning model,
 fj di , not selected. determine the most probable future wandering direc-
 tion for scout wolves. Let Bi and BHW be the distance of
 Formula (31) shows that if the said outside individual j an individual wolf i and the head wolf to the prey,
chooses strategy di in phase τ, then the belief weight should respectively. If Bi is greater than BHW or the number of
be increased by 1 in phase τ + 1. The probability for indi- iterations surpasses the maximum number Φmax of
vidual j to choose strategy di obeys the following correlation wanderings, each individual wolf will compute its ex-
with the belief weight of phase τ + 1: pectation for the next movement according to the
 historical wanderings of other individuals and deter-
 fτ+1
 j di mine the next wandering direction.
 λτ+1
 j di � . (30)
 Ji�1 fτ+1
 j di Step 4: compute the distance between individuals x and
 y by
 The expectation for inside individual j to choose strategy TS TS 
 τ+1 1 
di can be derived from the cost υτ+1
 j (di /λj + 1(di )) of an ED(x, y) � sxy − syt ∘ tN � · maxy − mint .
outside individual to choose a strategy: t�1 TS · θ t�1
 (34)
 di
 υτ+1 di � υτ+1
 j
 ⎝
 ⎛ ⎠•λτ+1 d .
 ⎞
 j i (31)
 j∈PE λτ+1
 j di Let [Mint , Maxt ] be the interval of the tth weight or bias
 to be optimized and cf be the step length of a ferocious
 In phase τ + 2, the probability for an inside individual to wolf to approach the head wolf. Then, the ferocious
choose strategy di can be calculated by wolf can approach the prey by
 e υ ( di )
 τ+1 l l
 t ht − sit 
 GVτ+2 di � . (32) sl+1 � s l
 + c · 
 y l
 . (35)
 i∈PE eυ (di )
 it it
 ht − slit 
 τ+1
6 Discrete Dynamics in Nature and Society

 Head wolf

 Pack Distance
 information calculation

 Ferocious Perceiving Scout
 wolf peer wolf
 information

 Environmental

 Smelling prey

 Determining
 exploration

 Wandering
 wandering
 direction
 Prey
 Figure 2: Process of the WPA.

 Belief learning
 model
 Wandering Online
 BPNN finance
 WPA credit risk
 Weights and optimization
 biases updates rating
 BPNN

 Figure 3: Idea of neural network optimization with the WPA improved by the belief learning model.

 Formula (35) shows that if Bi > BHW , then BHW � Bi , 4. Experiments and Results’ Analysis
 and the ferocious wolf will replace the head wolf; if
 Bi < BBHW , then the ferocious wolf will continue to Receiver operating characteristic (ROC) is a curve drawn
 approach the prey until tI ≤ tN . by a series of different dichotomy methods, with true
 Step 5: randomly choose a number Θ from [− 1, 1]; positive (TP) as the ordinate and false positive (FP) as the
 define cd as the besieging step length of individual wolf abscissa. Unlike traditional evaluation metrics, ROC
 i. Then, the state of each individual wolf during the metric allows the existence of intermediate state, which
 besieging can be updated by reflects the actual situation, and applies to a wide range.
 By this metric, the test results can be divided into several
 t l l 
 sl+1 l
 
 it � sit + Θ · cd · Ht − sit . (36) sequential classes for statistical analysis. This paper
 adopts the AUC of the ROC curve to evaluate the per-
 Step 6: set the individual closest to the prey as the head formance of the online finance credit risk rating model.
 wolf, and allocate the prey to the pack according to the The number of labeled samples in the early phase after the
 contribution to foraging. Finally, remove the inferior launch of an online finance credit product was set to 20,
 individuals from the pack. and the proportion of potential default samples was
 adjusted from 0 to 0.5. Figure 5 shows the influence of
 Step 7: after the end of iterations, import the head wolf
 that proportion on the AUC. With the growing pro-
 position as the initial values of weights and biases into
 portion of potential default samples, the AUCs of STSSLA
 the neural network, and complete the network training.
 and our algorithm increased to a certain extent. Mean-
 Figure 4 illustrates the flow of the late-phase online fi- while, the AUC of isolation forest algorithm changed very
nance credit risk rating model. The relevant parameters have little because the algorithm is an unsupervised method
been introduced in preceding parts. immune to labels.
Discrete Dynamics in Nature and Society 7

 Model initialization

 Topology design of
 the BPNN
 Obtaining the objective
 function value of the BPNN
 Outputting head
 wolve position Optimization of initial
 weights and biases
 Computing the expectation
 Yes
 of wandering
 Termination condition
 satisfied? Computing training
 error
 No
 Bi>BHW Or Φ>Φmax No

 Yes Prey allocation and inferior Updating weights and
 individual elimination biases

 Calling
 No
 Termination condition
 Updating head
 satisfied?
 wolf position
 Yes
 Bi>BHW? Yes
 Outputting evaluation
 result
 Approaching the
 Yes prey
 tI>tN?

 No

 Figure 4: Flow of the late-phase online finance credit risk rating model.

 0.7

 0.6

 0.5

 0.4
 AUC

 0.3

 0.2

 0.1

 0
 0 0.1 0.2 0.3 0.4 0.5
 Proportion of default samples
 Isolation forest algorithm
 Our algorithm
 Self-trained semi-supervised
 learning algorithm (STSSLA)
Figure 5: Influence of the proportion of potential default samples in labeled samples on the AUC. Note: AUC is short for area under the
curve.
8 Discrete Dynamics in Nature and Society

 0.7

 0.6

 0.5

 0.4

 AUC
 0.3

 0.2

 0.1

 0
 0 0.01 0.025 0.05 0.075 0.1
 Proportion of default samples
 Isolation forest algorithm
 Our algorithm
 STSSLA
 Figure 6: Influence of the proportion of potential default samples in unlabeled samples on the AUC.

 Table 1: Prediction errors and time consumptions of different models.
Serial number Model MSE MTC Serial number Model MSE MTC
 BPNN 4.8763∗10− 2 19.7854 BPNN 5.1812∗10− 3 20.8921
1 GA-BPNN 3.4221∗ 10− 2 14.1387 2 GA-BPNN 4.1974∗10− 6 21.4613
 Our model 3.1986∗10− 7 11.9341 Our model 8.3761∗ 10− 7 30.4897
 BPNN 2.5405∗10− 5 22.7932 BPNN 3.2439∗10− 7 21.6495
3 GA-BPNN 3.8873∗10− 6 18.9246 4 GA-BPNN 5.7946∗10− 5 22.2056
 Our model 6.9249∗10− 8 12.1624 Our model 9.3632∗10− 7 22.1632
 BPNN 4.2126∗10− 4 21.7843 BPNN 2.0643∗10− 2 28.1989
5 GA-BPNN 4.1934∗10− 6 20.6415 6 GA-BPNN 6.3786∗10− 3 28.3894
 Our model 9.1507∗10− 7 16.7964 Our model 8.3894∗10− 7 27.4891

 Suppose there are 20 labeled samples in the early phase traditional BPNN randomly initializes weights and thresh-
after the launch of an online finance credit product, in- olds, which lowers efficiency and accuracy. Experimental
cluding 10 normal samples and 10 potential default samples. results show that our model overcomes this defect of the
Then, the proportion of potential default samples in unla- traditional BPNN and effectively improves the prediction
beled samples and test samples was adjusted from 0 to 0.1. performance.
Figure 6 shows the influence of that proportion on the AUC. Figures 7–9 provide the output error curve of our model
With the growing proportion of potential default samples, during the training on different samples, MSE curves of our
the AUC of STSSLA increased to a certain extent, while that model, and the scatterplot of predictions and expectations
of our algorithm and isolation forest algorithm declined. on different samples. It can be seen that our model can
 To verify the effectiveness of the late-phase credit risk effectively rate the postphase online finance credit risk.
rating model for online finance credit products, this paper Figure 10 compares the ROCs of our model before and
applies the traditional BPNN and fuzzy BPNN separately on after optimization. This paper adjusts the ratio of the loss of
our samples for 200 iterations and compares their search mistaking normal clients as default clients to the loss of
results with the solution obtained by our model. Table 1 mistaking default clients as normal clients. As shown in
compares the prediction errors and time consumptions of Figure 10, our model performed stably at different ratios.
these models. The convergence accuracy and efficiency of The AUCs of the original model and the improved model
each model were measured by mean squared error (MSE) stabilized at 0.7231 and 0.8277, respectively. The optimized
and mean time consumption (MTC), respectively. model obviously has better prediction ability.
 As shown in Table 1, our model achieved a lower MSE in Table 2 presents the confusion matrix of sample veri-
a shorter MTC than the BPNN and GA-BPNN in rating the fication with different misjudgment ratios. Table 3 presents
postphase online finance credit risk. This is because the the confusion matrix of time verification with different
Discrete Dynamics in Nature and Society 9

 0.2
 0.15
 0.1
 0.05

 Output error
 0
 -0.05
 -0.1
 -0.15
 -0.2
 -0.25
 -0.3
 0 20 40 60 80 100 120
 Sample number
 Figure 7: Output error curve on different samples.

 1

 0.01
 MSE

 0.0001

 0.000001

 1E-08
 0 5 10 15 20 25
 Number of iterations
 Training Optimal
 Verification Target
 Test
 Figure 8: MSE curve of our model.

 60

 50
 Network prediction

 40

 30

 20

 10

 0
 0 20 40 60 80 100
 Sample number
 Prediction
 Expectation
 Figure 9: Scatterplot of predictions and expectations on different samples.
10 Discrete Dynamics in Nature and Society

 1 1

 0.8 0.8
 True positive rate

 True positive rate
 0.6 0.6

 0.4 0.4

 0.2 0.2

 0 0
 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
 False positive rate False positive rate
 AUC=0.7231 AUC=0.8277
 (a) (b)

 Figure 10: ROCs of the (a) original model and (b) improved model.

 Table 2: Confusion matrix of sample verification with different misjudgment ratios.

v 0.3 0.5 1.0 2.0
Actual/predicted 0 1 0 1 0 1 0 1
0 592 34 527 108 462 167 235 389
1 634 2,879 427 3,085 236 3,271 85 3,428

 of credit default were calculated. For the late phase, the
Table 3: Confusion matrix of time verification with different BPNN was improved to rate the online finance credit risk.
misjudgment ratios.
 Through experiments, the AUC curves were plotted with
v 0.3 0.5 1.0 2.0 different proportions of potential default clients in labeled
Actual/predicted 0 1 0 1 0 1 0 1 and unlabeled samples, and the prediction errors and time
0 96 3 97 5 81 19 32 66 consumptions were compared between the BPNN, GA-
1 156 1,915 67 2,001 27 2,042 11 2,057 BPNN, and our model. The results show that our model can
 effectively rate the postphase credit risk of online finance.

misjudgment ratios. As expected, the number of normal Data Availability
clients being misjudged decreased with the growing loss,
while the number of default clients being misjudged in- The data used to support the findings of this study are
creased with the loss. Therefore, the confusion matrix available from the corresponding author upon request.
outputted by our model changes with the demands of online
finance credit businesses. Conflicts of Interest
 All in all, the above experimental results testify the
superiority of our model in sample recognition accuracy, The authors declare that they have no conflicts of interest
operating stability, and prediction power. regarding the publication of this paper.

5. Conclusions References
 [1] X. Wanqin Yang and W. Yang, “Prediction and analysis of
This paper explores the online finance credit risk rating
 literature loan circulation in university libraries based on RBF
based on the neural network. Specifically, the research pe- neural network optimized model,” Automatic Control and
riod was divided into an early phase and a late phase by the Computer Sciences, vol. 54, no. 2, pp. 139–146, 2020.
launch of an online finance credit product. For the early [2] X. Zhu, “Deep learning modelling of systemic financial risk,”
phase, a cold start method was developed for the credit risk Revue d’Intelligence Artificielle, vol. 34, no. 2, pp. 137–141,
rating of online finance, and the similarity and abnormality 2020.
Discrete Dynamics in Nature and Society 11

 [3] A. Nagurney, “Networks in economics and finance in net- [19] M. Leow and J. Crook, “Intensity models and transition
 works and beyond: a half century retrospective,” Networks, probabilities for credit card loan delinquencies,” European
 vol. 77, no. 1, pp. 50–65, 2021. Journal of Operational Research, vol. 236, no. 2, pp. 685–694,
 [4] R. Dekkers, R. de Boer, L. M. Gelsomino et al., “Evaluating 2014.
 theoretical conceptualisations for supply chain and finance [20] K. Choi, G. Kim, and Y. Suh, “Classification model for
 integration: a Scottish focus group,” International Journal of detecting and managing credit loan fraud based on individ-
 Production Economics, vol. 220, p. 107451, 2020. ual-level utility concept,” ACM SIGMIS-Data Base: The
 [5] F. Liu and Y. You, “A big data-based anti-fraud model for DATABASE for Advances in Information Systems, vol. 44,
 internet finance,” Revue d’Intelligence Artificielle, vol. 34, no. 3, pp. 49–67, 2013.
 no. 4, pp. 501–506, 2020. [21] A. Sokolov, R. Webster, A. Melatos, and T. Kieu, “Loan and
 [6] J. Lopez-Jimenez, J. L. Gutierrez-Rivas, E. Marin-Lopez, nonloan flows in the Australian interbank network,” Physica
 M. Rodriguez-Alvarez, and J. Diaz, “Time as a service based A: Statistical Mechanics and Its Applications, vol. 391, no. 9,
 on white rabbit for finance applications,” IEEE Communi- pp. 2867–2882, 2012.
 cations Magazine, vol. 58, no. 4, pp. 60–66, 2020. [22] Z. Xu, X. Cheng, K. Wang, and S. Yang, “Analysis of the
 [7] L. A. Sauls, “Becoming fundable? Converting climate justice environmental trend of network finance and its influence on
 claims into climate finance in Mesoamerica’s forests,” Cli- traditional commercial banks,” Journal of Computational and
 matic Change, vol. 161, no. 2, pp. 307–325, 2020. Applied Mathematics, vol. 379, Article ID 112907, 2020.
 [8] R. Chandra and S. Chand, “Evaluation of co-evolutionary [23] M. R. Fayyaz, M. R. Rasouli, and B. Amiri, “A data-driven and
 neural network architectures for time series prediction with network-aware approach for credit risk prediction in supply
 mobile application in finance,” Applied Soft Computing, chain finance,” Industrial Management & Data Systems,
 vol. 49, pp. 462–473, 2016. vol. 121, no. 4, pp. 785–808, 2020.
 [9] H. Ghoddusi, G. G. Creamer, and N. Rafizadeh, “Machine [24] A. Leng, G. Xing, and W. Fan, “Credit risk transfer in SME
 learning in energy economics and finance: a review,” Energy loan guarantee networks,” Journal of Systems Science and
 Economics, vol. 81, pp. 709–727, 2019. Complexity, vol. 30, no. 5, pp. 1084–1096, 2017.
[10] A. Smolyak, O. Levy, L. Shekhtman, and S. Havlin, “Inter- [25] J. Juma and D. Gichoya, “Artificial neural network based
 dependent networks in economics and finance-a physics expert system for loan application evaluation: case of Kenya
 approach,” Physica A: Statistical Mechanics and Its Applica- commercial bank,” in Proceedings of the 2013 IST-Africa
 tions, vol. 512, pp. 612–619, 2018. Conference & Exhibition, pp. 1–11, Nairobi, Kenya, May 2013.
[11] M. Karimi and H. Saberi Nik, “A piecewise spectral method [26] D. Maček, I. Magdalenić, and N. B. ReCep, “A systematic
 for solving the chaotic control problems of hyperchaotic fi- literature review on the application of multicriteria decision
 nance system,” International Journal of Numerical Modelling: making methods for information security risk assessment,”
 Electronic Networks, Devices and Fields, vol. 31, no. 3, Article International Journal of Safety and Security Engineering,
 ID e2284, 2018. vol. 10, no. 2, pp. 161–174, 2020.
[12] A. Sanford and I. Moosa, “Operational risk modelling and [27] P. Wetzel and E. Hofmann, “Supply chain finance, financial
 organizational learning in structured finance operations: a constraints and corporate performance: an explorative net-
 Bayesian network approach,” Journal of the Operational Re- work analysis and future research agenda,” International
 search Society, vol. 66, no. 1, pp. 86–115, 2015. Journal of Production Economics, vol. 216, pp. 364–383, 2019.
[13] A. Taufik and N. I. Soesilo, “The impact of loan to value to [28] P. R. Srivastava, Z. J. Zhang, and P. Eachempati, “Deep neural
 property credit growth sustainability in Indonesia,” IOP network and time series approach for finance systems: pre-
 Conference Series: Earth and Environmental Science, vol. 716, dicting the movement of the Indian stock market,” Journal of
 no. 1, Article ID 012091, 2021. Organizational and End User Computing, vol. 33, no. 5,
[14] L. Han, L. Han, and H. Zhao, “Study and application of credit pp. 204–226, 2021.
 scoring models to appraisal of the loan to Chinese companies
 with uncertain linguistic information,” International Journal
 of Applied Cryptography: International Journal of Advance-
 ments in Computing Technology, vol. 4, no. 6, pp. 43–49, 2012.
[15] S. Purohit and A. Kulkarni, “Credit evaluation model of loan
 proposals for Indian Banks,” in Proceedings of the 2011 World
 Congress on Information and Communication Technologies,
 pp. 868–873, Mumbai, India, December 2011.
[16] U. R. Kamboh, Q. Yang, M. Qin, and S. Rauf, “Uncertainty
 cost analysis of heterogeneous wireless network based on loan
 repayment approach,” in Proceedings of the 2017 9th Inter-
 national Conference on Advanced Infocomm Technology
 (ICAIT), pp. 170–175, Chengdu, China, November 2017.
[17] S. Chakraborty, J. Gaeta, K. Dutta, and D. Berndt, “An analysis
 of stability of inter-bank loan network: a simulated network
 approach,” in Proceedings of the 50th Hawaii International
 Conference on System Sciences, Hilton Waikoloa Village, HI,
 USA, January 2017.
[18] G. Bhat, S. G. Ryan, and D. Vyas, “The implications of credit
 risk modeling for banks’ loan loss provisions and loan-
 origination procyclicality,” Management Science, vol. 65,
 no. 5, pp. 2116–2141, 2019.
You can also read