Application of Deep Learning in Financial Management

Research ArticleApplication of Deep Learning in FinancialManagement Evaluation

Wenlei Shi1 Lei Xu 2 and Dongli Peng3

1School of Accounting Shandong University of Finance and Economics Jinan 250014 Shandong China2School of Economics and Management Qilu Normal University Jinan 250000 Shandong China3School of Public Administration Hengshui University Hengshui 053000 Hebei China

Correspondence should be addressed to Lei Xu xuleiqlnueducn

Received 24 September 2021 Revised 12 October 2021 Accepted 13 October 2021 Published 3 November 2021

Academic Editor Tongguang Ni

Copyright copy 2021 Wenlei Shi et al +is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

+e competition among enterprises is becoming increasingly fierce +e research on the financial management evaluation modelis helpful for enterprises to find possible risks as soon as possible +is paper constructs the financial management evaluationmodel based on the deep belief network and applies the analytic hierarchy process to determine the weight of financialmanagement evaluation indicators which is compared with other classical deep learning evaluationmethods such as KNN SVM-RBF and SVM linear It has achieved an accuracy of more than 81 showing a satisfactory prediction effect which is of greatsignificance to formulate corresponding countermeasures strengthen financial management improve the capital market systemand promote high-quality economic development In addition aiming at the problem of abnormal financial data this paper usesthe new sample dataset obtained by principal component analysis for convolution neural networkmodel learning which enhancesthe prediction accuracy of the model and fully shows that deep learning is feasible in the research of financial managementprediction and there is still a lot of space to explore

1 Introduction

With the development of society financial intelligence in-creasingly affects our life and has a great impact on thetraditional financial work which is a topic of concern toenterprises all over the world By importing data into thedatabase or taking the existing data in the database as theanalysis object [1] financial intelligence processes the dataaccording to the financial management model and uses thehigh-speed and accurate computing power of the computerto quickly obtain the enterprise operation diagnosis reportso as to form a fast and reliable basis for business decision-making [2] As a popular direction in the computer fielddeep learning technology has been closely combined andapplied with the financial field Using reasonable deeplearning technology can solve the problem of efficient au-tomatic data analysis in the financial industry providevaluable prediction information for managers and providereliable early warning for healthy institutional operation [3]

+roughout the research trends of financial managementevaluation at home and abroad it mainly focuses on fi-nancial evaluation indicators and evaluation models In theselection of financial management evaluation indicatorsexisting studies mainly focus on which indicators can ac-curately predict enterprise crisis [4] It has experienced thecommon application stage of multidimensional indicatorsfrom single financial ratio indicators and multivariable fi-nancial ratio indicators to the combination of financialindicators and nonfinancial indicators [5] +e above studieshave achieved certain prediction results but the selection ofmultidimensional early warning indicators usually usesstatistical methods to test the normality and significance ofsample index data and then combined with manual dis-crimination based on professional ability to select indicators+e selection method is more complex and there is nounified conclusion on which indicators to select In theselection of early warning models early scholars often usedunivariate model logistic model discriminant analysis

HindawiScientific ProgrammingVolume 2021 Article ID 2475885 9 pageshttpsdoiorg10115520212475885

model and so on [6] With the development of informationtechnology a large number of open real data informationcan be obtained in the process of enterprise operation +esedata can often reflect the past risk status of the enterprise andeven some characteristics of the risk of the whole capitalmarket [7] +erefore scholars began to apply neural net-work support vector machine and other models based onartificial intelligence methods to financial crisis earlywarning Generally speaking the early warning model basedon artificial intelligence method has made great progress inthe mechanism and systematicness of early warning It notonly overcomes the limitations of the early warning modelbased on statistical method such as requiring data to obeynormal distribution and complex calculation and analysisbut also has strong fault tolerance and learning ability [8] Atpresent artificial neural network has developed to the stageof deep learning network which is characterized by self-learning and high dynamic adaptability +erefore deeplearning is also applied to the field of financial managementevaluation

Deep learning can complete a lot of regular simple andrepetitive work +e application of financial intelligencetechnology in the financial field has played an important rolein improving business efficiency reducing work errors andpreventing and controlling enterprise risks +is paperstudies the financial risk by establishing the financialmanagement evaluation model which is a quantitativemethod It is different from the qualitative analysis +ework is difficult and the accuracy is low+emodel researchhas higher reliability and uses the analytic hierarchy processto determine the enterprise evaluation index system +ismodel interprets the historical data links the characteristicsof the data with the financial situation of the enterprise andthen uses the existing data to analyze and predict the futureso as to ensure the accuracy of the prediction

2 Related Work

Ng et al proposed a new fuzzy CMAC (cerebellar modelarticulation controller) model based on reasoning compo-nent rules as a new method for bank fault classification andearly warning system [9] Artificial neural network is in-troduced into financial management evaluation for the firsttime Artificial neural networks not only deal with the lack ofdata errors but also allow for timely adjustment of internalcontrol parameters Erhan et al have proved throughpractice that the unsupervised training method can betterdescribe the complex functional relationship which pro-vides a good reference for financial risk early warning+rough the establishment of Yahoo information bulletinboard Jones studied the impact of network information onfinancial management by using the method of deep learning[10 11] +rough the establishment of online informationbulletin board Jones reduced the fluctuation of stock priceand found that investorsrsquo differences of opinion may reducerisk and increase turnover +rough the analysis of stockrelated blog information in support vector machineChoudhury found that the fluctuation of stock price will beaffected by blog content [12] Tetlock conducted a 24-year

analysis of 500 listed companies [13] and counted thenegative words in various reports in this period through themethod of deep learning On the basis of the relationshipbetween corporate income and stock income this paperpoints out that the negative words in the news reports oflisted companies can predict the decline of listed companiesrsquoprofits and reevaluate the listed companies Under thepressure of trading news content may capture some in-formation that is difficult to quantify which has a certainearly warning function in media reports Najafabadi et alabstracted the deep learning algorithm into data represen-tation through the hierarchical learning process [14] +ismethod has attracted great attention from data science It iswidely used to solve problems such as network securitymedical information national intelligence and marketingJason Kuen and chin poo Lee applied deep learning to therepresentation of visual tracking invariance [15] andachieved good results through strong spatiotemporal con-straints and stacking slow convolution tracking Yudistiraand Kurita [16] proposed that deep learning processor hasbecome the most potential solution to accelerate deeplearning algorithm and pointed out the disadvantage of lowefficiency of assembly instructions written by deep learningwhich puts forward higher requirements for the research ofdeep learning

To sum up although there are many researches on deeplearning at home and abroad covering a wide range of fieldsthey mainly focus on theory speech recognition and imagerecognition which are the feature judgment of known in-formation and lack of relevant empirical research Manyliterature studies show that deep learning can well describecomplex functions and financial management evaluationcan also be used as a complex function judgment but so fardeep learning has not been used in the research of financialmanagement evaluation Based on the previous researchcombined with the characteristics of financial managementthis paper applies deep learning to the evaluation model offinancial management

3 Evaluation Modelling Based on DeepBelief Network

31 Deep Belief Network Model Deep learning is a categoryof machine learning that focuses on neural networks Ma-chine learning can be applied to image speech patternrecognition weather prediction stock price prediction geneexpression and content recommendation +is is verysimilar to the recognition of images by the cerebral cortex+e deep learning model first extracts the low-level featuresfrom the original signal then obtains the higher-level fea-tures from the low-level features and then obtains thehigher-level expression In the face recognition system theoriginal signal is the pixel the low-level feature is the edge ofthe object the high-level feature is the contour composed ofedges and the high-level expression is the face +roughtraining with characteristic data the error is transmittedfrom top to bottom to fine tune the network Based on theparameters of each layer obtained in the first step the pa-rameters of the whole multilayer model are further

2 Scientific Programming

optimized Finally it classifies according to the high-levelcharacteristics and outputs the prediction results of themodel Deep belief network is one of the mainstream deeplearning algorithms [17] According to the Boltzmannmachine model of stochastic neural network the principlearchitecture of restricted Boltzmann machine [18] is shownin Figure 1

a (a1 a2 anv)T represents the offset vector of the

visible layer b (b1 b2 bnh)T represents the offset

vector of the hidden layer and W (wij) isin Rnhlowastnv is theweight matrix +e energy function in a deep confidencenetwork generated by multiple constrained Boltzmannmachines for any set of neurons with state vector (v h) isexpressed as follows

E(v h ∣ θ) minus 1113944

nv

i1aivi + 1113944

nh

j1bjhj

⎡⎢⎢⎣ ⎤⎥⎥⎦ minus 1113944

nv

i11113944

nh

j1aibjhjvi (1)

where nv is the number of all neurons v is the state vector his the state vector in hidden layer and nh is the number of allneurons in the hidden layer and θ ai bj wij1113966 1113967 denotes theadjustment factors that limit the Boltzmann machine ar-chitecture +rough the energy function defined in formula(1) we get the joint probability distribution of state (v h) asshown in

P(v h ∣ θ) Z(θ)minus 1 exp[minusE(v h|θ)] (2)

where the expression of Z(θ) is shown in

Z(θ) 1113944 exp[minusE(v h|θ)] (3)

where Z(θ) is the normalization parameter It can be seenthat in order to obtain p(v|θ) and p(h|θ) the key step is tocalculate the normalized parameter Z(θ) +e connection ofa DBN is guided and determined by generating weights fromtop to bottom RBMs is like a building block Compared withthe traditional and deeply layered sigmoid belief network itcan easily learn the connection weights +e probability ofactivation of a neural unit in the hidden layer can be cal-culated by the following formula [19]

P hj 1|v θ1113872 1113873 σ b2j + 1113944

i

2

radicviwij

⎛⎝ ⎞⎠ (4)

where σ(bull) indicates sigmoid activation function +e sto-chastic gradient algorithm is usually used to find themaximum value of 1113936

ti1 log P(vi|θ) +e momentum coef-

ficient is generally taken as (0 1) Intuitively it is understoodthat if the current gradient direction is the same as thegradient direction of the previous step the weight update ofthis step will be increased and if it is different the updatewill be reduced +e depth belief network model adopted isshown in Figure 2

32 Determination of Financial Management EvaluationIndex by Analytic Hierarchy Process +e evaluation of fi-nancial management objectives of logistics enterprises is acomplex systematic project which requires the establishment

middotmiddotmiddotmiddotmiddotmiddot

h v

hn

h4

h3

h1

h2

bn

b4

b3

b2

b1


vn

v3

v2

v1

an

a3

a2

a1

b isin Rh a isin Rv

W isin Rhtimesv

Figure 1 Restricted Boltzmann machine model




xmiddotmiddotmiddotmiddotmiddotmiddot

Input

Output

RestrictedBoltzmann Machine




P (h2 h3)~RBM

P (h1 h2)

P (x h1)

h2

h1

h3

Figure 2 Deep generative model

Scientific Programming 3

of a financial management evaluation system [20] ST meansldquospecial treatmentrdquo +e policy is aimed at those with ab-normal financial or other conditions +e addition of lowast STbefore the stock means that the listed company has sufferedlosses for three consecutive years and the exchange makes adelisting warning +e weight of financial managementevaluation indicators is determined by analytic hierarchyprocess Analytic hierarchy process decomposes the probleminto different constituent factors and gathers and combines thefactors according to different levels according to the corre-lation influence and subordinate relationship between thefactors to form a multilevel analysis structure model It is amodel and method for making decisions on complex systemsthat are difficult to be fully quantitative +e steps are asfollows First establish a financial management evaluationindex system In order tomake a correct evaluation of financialmanagement we should start from the corporate governancestructure establish the financial management evaluationsystem from the perspectives of management decision-makingand external environment +e corporate governance struc-ture management decision-making and external environmentevaluation are composed of some related elements See Fig-ure 3 for details

+en the judgment matrix A is constructed to obtain theweights ofU1U2 andU3 of the evaluation index system Asan example the calculation process is illustrated For theabove evaluation index system the expert group believesthat in the evaluation of financial management corporategovernance structure (U1) is more important than man-agement decision (U2) and external environment (U3) andthe management decision is more important than the ex-

ternal environment then A

1 2 312 1 213 12 1

⎡⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎦

Calculate the product Mi of each row element of thejudgment matrix and then calculate the nth root of MiFinally normalize the vector [W1 W2 W3]

T and calculatethe index weight Wi

W1 W1(1113936nl1 W1)

minus 1 0540 W2 W2 (1113936nl1 W2)

minus 1

0297 W3 W3(1113936nl1 W3)

minus 1 0163Calculate the maximum eigenvalue of judgment matrix

A | λmax|

A middot W

a11 a12 a1n

a21 a22 a2n

an1 an2 ann

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

W1

W2

Wn

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(5)

(AbullW)1 1times 05396 + 2times 02970 + 3times 016341624Similarly (AbullW)2 0894 (AbullW)3 04922 +e last step isconsistency testing When n 3 R1 058 the judgmentmatrix has satisfactory consistency +erefore the weight ofU1 U2 and U3 is [05396 02970 01634] Other indexweights can be calculated according to the above methodand the calculation results are shown in Table 1

When Cn 00048 Rn 090 and CR1 00054 thejudgment matrix has satisfactory consistency When

CI2 00192 RI2 058 and CR2 00562 the judgmentmatrix has satisfactory consistency When CI3 00268RI3 058 and CR3 00562 the judgment matrix has sat-isfactory consistency +e weight distribution data of indi-cators at each level are summarized in Table 2

33 Intelligent Detection Model of Financial Data Timeseries data is a data column recorded by the same unifiedindicator in chronological order All data in the same datacolumnmust be of the same caliber andmust be comparableAs IOT brings a large amount of time series data we need totime slice the time series data +e traditional time-seriesdata processing methods include median extreme valuedeviation variance year-on-year month-on-month andperiodic methods +ese methods can only roughly sum-marize the data and form a preliminary understanding +eoverlapping slicing method of sliding window is used in thispaper +is method sorts and counts the target dataaccording to the time sequence delimits the length and sizeof each window summarizes and calculates the character-istics of each time period window analyses different datawith the same dimension in continuous time periods andobtains the change trend of the target data In this paper itwill be solved by convolution neural network as shown inFigure 4

Firstly after the training data is processed by zero meanprincipal component analysis is carried out to reduce theinput parameters of convolutional neural network modeland reduce the correlation between input factors +en thenew sample dataset obtained by principal componentanalysis is used for convolutional neural network modellearning and the parameters in convolutional neural net-work are continuously adjusted by gradient descent Finallythe test sample data are applied to the model to verify theprediction accuracy of the model

Since this study only considers the impact of historicaldata on future enterprise finance a one-dimensional con-volutional neural network is adopted +e model includestwo convolutional subnet works as shown in Figure 5 +einput layer is m k-dimensional index data output fromformula (4) and the output layer is two classifiers +econvolution layer is used to extract different features of theinput layer the linear rectification layer is used to activateneurons in the network according to the linear function thepooling layer is used to reduce the data dimension and thefull connection layer is used to combine all local features andcalculate the final classification result

In this study the input data for one-dimensional con-volutional neural network is the k-dimensional orthogonalfeature Yij

prime of M companies +en three convolutionalsubnetworks are used to deeply learn the orthogonal featuredata of enterprises +e first convolutional network selects128 convolutional cores with the size of 1lowast 3 and the secondconvolutional network selects 128 convolutional cores withthe size of 1lowast 4 +e third subconvolution network selects128 convolution cores with size of 1lowast 5 +e calculationformula is as follows


X(l)

f W(l)

middot X(lminus 1)

+ b(l)

1113872 1113873 (6)

where X(l) and X(lminus1) are the neuron output values of layer land layer l minus 1 Wl is convolution kernel and b is offset+eactivation function adopts a modified linear unit For theinput x the weight vector is ω and the output with offset bis max (0 ωx + b) +e linear activation function simplysets the threshold to zero which greatly reduces thecomputational overhead Moreover compared with theexpensive operations (exponents etc) of sigmoid and tanhneurons relu can be activated through a simple zero

threshold matrix and is not affected by saturationMeanwhile L2 norm is used to normalize the fitting costas shown in

Financialmanagement

evaluation

CorporateGovernance

Structure

Managementdecisions

Externalenvironment

U1

U2

U3

Personalization of financialmanagement

Optimization of resourceallocation

Internal control mechanism

Early warning system

U11

U12

U13

U14

U21

U22

U23

U31

U32

U33

Project investment decision

Capital structure

Risk control decision

Economic environment

Financial environment

Legal environment

Figure 3 Comprehensive financial management evaluation system

Table 1 Index weight

U11 U12 U13 U14 WeightU11 1 2 3 4 06532U12 12 1 2 3 03841U13 13 12 1 3 02890U14 14 13 12 1 01206

Table 2 Weight distribution data

Index U1 U2 U3 U11 U12 U13 U14Weight 05396 02970 01634 04832 02717 01569 00882Index U21 U22 U23 U31 U32 U33Weight 06369 02583 01048 05278 03325 01397

Training data

Zero mean

principalcomponent analysis

Convolutional neural network model

Output predictionresults

Figure 4 Financial management model framework based onconvolutional neural network


C minusn 1113944

xj

1yi ln a

lj + 1 minus yi( 1113857ln 1 minus yi( 11138571113960 1113961 + n 1113944 λω2

(7)

where the first term represents the cross entropy cost thesecond term is the sum of squares of all weights added andthen the factor used λ2n to make quantitative adjustmentand λgt 0 is called the normalization parameter +e thirdconvolution subnetwork outputs to the full connection layerand then outputs the final binary result that is whether theenterprise is ST in this study the output result of ST en-terprise is 0 and that of non-ST enterprise is 1 [21] At thesame time this study uses the maximum pooling method topool local feelings and selects the flexible maximum methodto solve the problem of slow learning

4 Results and Safety Analysis

41 Data and Empirical Design In the future A-share mayform a two-way benign expansion of supply and demandand the regulatory authoritiesrsquo policies on its stock marketare also more effective and in place which is conducive tothe dynamic balance of supply and demand +ere are STsystem and lowast ST system in stock market From the per-spective of data availability and effectiveness it is a rea-sonable method to use enterprise stock ST or lowast ST as thesymbol of enterprise financial crisis

+is paper first selects the companies that are ST and lowastST (hereinafter referred to as ST companies) and then findsout the corresponding companies of each ST or lowast STcompany (hereinafter referred to as non-ST companies) inthe companies with normal financial conditions accordingto the industry and average total assets Use the financialindex data of ST companies and non-ST companies in theprevious years of 2016 to predict whether there will be afinancial crisis in 2016 (by ST or lowast ST) compare with theactual situation count the accuracy of the prediction andconduct empirical analysis +is paper selects a total of 3513companies +e reason for data normalization is that themeasurement units of each data are different and theprocessed data will be between 0 and 1 If the data is notnormalized the gradient descent is carried out in one unitso its descent step in each direction is the same Non-standardized data will cause the gradient to follow a zigzagroute in the direction perpendicular to the contour line

when the gradient decreases which will make the iterationvery slow In general normalization can make the order ofmagnitude of each stock index correspond to the length ofgradient decline [21]

+is paper has conducted four empirical analyses andthe selection of data quantity is shown in Table 3 +is paperhas conducted four empirical analyses and the selection ofdata volume is shown in Table 3 Taking the data of the firstfew years of 2016 as the training set and the data of the nextfew years of 2016 as the prediction set the output result of nofinancial crisis is 0 and the output result of financial crisis is1+e judgment result is recorded as x the actual situation ofthe company is recorded as y x and y are 0 or 1 the numberof companies in the prediction set is n and the calculationformula of accuracy Pa is

Pa 1 minus Nminus 1

|X minus Y|1113872 1113873 times 100 (8)

42 Outcome Evaluation Criteria Because each simulationwill randomly take an initial value the results of eachsimulation may be different +e experiments were con-ducted in four groups based on the size of the years of dataselected In general the nodes in the hidden layers have animpact on the prediction results If the number of hiddenlayer nodes is too small the network cannot have thenecessary learning ability and information processingability If too much it will not only increase the complexityof the network structure and make the network more likelyto fall into local minima in the learning process but alsomake the learning speed of the network very slow+e neuralnetwork structure has two hidden layers and the number ofthem can be determined by the following formula

L α2 +(m + n)12

log2 m(9)

where m and n represent the nodes of the output layer andthe input layer respectively α can be any value between 1and 10 +ese methods can only obtain feasible initial valuesfor the nodes of the hidden layer and this number usuallyneeds to be corrected during training and learning Gen-erally two methods of gradually increasing and graduallydecreasing are used to correct the number of nodes in the

Convolutionlayer

Linearrectifier layer Pool layer Convolution

layerLinear

rectifier layer Pool layer Full connectionlayer

Output layerHidden layerInput layer

Figure 5 Framework of one-dimensional convolutional neural network model


hidden layer to achieve the expected value and reduce theerror to a reasonable range +is study uses formula (9) tocalculate the hidden layer nodes and finally obtains thenodes of each network as shown in Table 4

From the results shown in Figure 6 the accuracy of themodel in predicting whether an enterprise will have crisiscan be maintained at more than 81 which shows that themodel based on deep learning has a general prediction effectfor A-share listed companies the more the years of data usedfor prediction the higher the accuracy of prediction

In order to test the effect of model training this studyinputs the test data into the model after the above trainingand observes the accuracy of the test data From Figure 7 itcan be seen that after training the accuracy of the model onthe training set is 791 and the accuracy on the test set is9134 +e results show that the accuracy of the test set isslightly higher than that of the training set which proves thatthe model has better generalization ability

43 Intelligent Analysis of Financial Data Based on DeepLearning In the pretraining stage each layer of RBM net-work is trained separately and unsupervised to ensure that thefeature vectors are mapped to different feature spaces andretain the feature information as much as possible It ispretrained by an unsupervised greedy layer-by-layer methodto obtain the weight In this process the data is input to thevisible layer to generate a vector V which is transmitted to thehidden layer through the weight W to obtain H In the lastlayer of DBN a BP network is set up to receive the outputeigenvector of RBM as its input eigenvector and train theentity relationship classifier supervised Moreover each layerof RBM network can only ensure that the weight in its ownlayer is optimal for the eigenvector mapping of that layer notfor the eigenvectormapping of the whole DBN+erefore thebackpropagation network also propagates the error infor-mation from top to bottom to each RBM layer and fine-tunethe whole DBN network +e process of RBM networktraining model can be regarded as the initialization of theweight parameters of a deep BP network +e naive Bayesianclassifier in this paper is realized by MATLAB programmingIn this section the financial data of traditional indicators are

Table 3 Empirical quantity statistics

ST Non-ST TotalNumber ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata

2 years Total of training setand test set

55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years

Total of training setand test set

17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368

Table 4 Network node setup

2 years 4 years 8 years 12 yearsFirst hidden layer 8 14 29 32Second hidden layer 4 9 10 10

8123

8259

8641

8708

4 years 8 years 12 years2 yearsEmpirical group

8123

8259

8641

8708

Cor

rect

rate

()

Correct rate

Figure 6 Empirical results

9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795

2 3 4 5 6 7 81TRAINING TIMES

727476788082848688909294

ACCU

RACY

Training dataTest data

Figure 7 +e accuracy of training and testing datasets changeswith the learning cycle


processed by the time-series construction method proposedabove and then input into the classifier After running theclassification effects of traditional models and various formsof models can be obtained +e operation results of eachmodel are shown in Figure 8

In Figure 8 the recall rate of the time series indexmodel inthe ratio form and the first relative value form is the highestIn terms of precision the time series index model in the formof difference is the best It can also be seen that the recall rateand precision rate show the law of one change and the other+e comparison of classification accuracy of DCNN underdifferent hidden layer structures is shown in Figure 9

According to Figure 10 hidden levels 1 2 and 3 showgood classification accuracy all reaching more than 91After 600 iterations the classification accuracy of the secondlevel reaches 9857 which is the maximum of the classi-fication accuracy +erefore the convolutional neural net-work model with 3-layer hidden layer structure has goodclassification accuracy

In order to further prove the effectiveness and superi-ority of convolutional neural network model this workcompares it with traditional classical machine learning earlywarningmethods+emethods used for comparison includek-nearest neighbor (KNN) support vector machineGaussian kernel (SVM-RBF) support vector machine linearkernel (SVM linear)+e basic parameter setting of the test isthe same as the above +e number of neighbors in KNN isset to 6 and the kernel function parameter of support vectormachine is set to 10 Making the weight smaller and smallerthen its corresponding loss function will be smaller andsmaller finally achieving our goal +e smaller the value ofthe loss function the more accurate the prediction is +eprediction accuracy of the model established in this studycan reach 8165 showing a good financial prediction effect+is is because convolutional neural network can betterlearn the correlation between various indexes and extract themost effective abstract features so as to ensure the accuracyof prediction results

5 Conclusion

In order to improve the accuracy of corporate financialmanagement evaluation this paper constructs a financialmanagement evaluation model based on deep learning anduses the data of A-share listed companies from 2007 to 2020to explore whether deep learning can build a generallyapplicable financial management model for listed compa-nies By using the data of different years it is also concludedthat the more the data years the model constructed in thispaper uses the higher the prediction accuracy +e testresults show that the intelligent analysis of financial ab-normal data based on deep learning is also effective andaccurate Finally the effectiveness and practicability of theintelligent analysis method are proved by an example

Data Availability

+e data used to support the findings of this study are in-cluded within the article

2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08

Recall ()Precision ()F-score

Figure 8 Comparison diagram of various forms of model iden-tification (1) Traditional model (2) Timing standard in the form ofdifference (3) Time series standards in ratio form (4) +e firstrelative value form of the timing standard (5) +e second relativevalue form of the timing standard

200 400 600 800 1000 1200 1400 1600 18000Number of iterations

090

092

094

096

098

100

Layer1Layer2Layer3

Figure 9 DCNN classification accuracy of different hidden layers

8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS

Figure 10 Comparison of prediction accuracy of differentmethods


Conflicts of Interest

All the authors do not have any possible conflicts of interest

Acknowledgments

+is work was supported by Key R amp D Plan of ShandongProvince (Grant no 2020RKB01367) and the National SocialScience Foundation (Grant no 20BJY033)

References

[1] O Tatari D C Lacouture and M J Skibniewski ldquoPerfor-mance evaluation of construction enterprise resource plan-ning systemsrdquo Journal of Management in Engineering vol 24no 4 pp 198ndash206 2008

[2] B V Samorodov O O Sosnovska and M O ZhytarldquoMethodical approach to the quantification of enterprise fi-nancial security levelrdquo Financial and credit activity Problemsof eory and Practice vol 1 no 32 pp 269ndash277 2020

[3] R Zajarskas and J Ruzevicius ldquoEvaluation of the effectivenessof the quality management system of the service enterpriserdquoEkonomika ir vadyba no 15 pp 857ndash864 2010

[4] S C Bpsmpcgtl ldquoAlgorithm OF complex evaluation OFenterprise investment enterpriserdquo TIME DESCRIPTION OFECONOMIC REFORMS no 1 pp 35ndash43 2018

[5] V Levytskyi ldquo+e optimization of system financial man-agement of enterprise based on the analysis of investments inits marketing activitiesrdquo Economic journal of Lesya UkrainkaVolyn National University vol 2 no 18 pp 101ndash108 2019

[6] S Mekadmi and R Louati ldquoAn evaluation model of usersatisfaction with enterprise resource planning systemsrdquoElectronic Journal of Information Systems Evaluation vol 21no 2 pp 143ndash157 2018

[7] A Meylis ldquoAnalysis and prevention of enterprise financialrisk under the new tax policyrdquo Open Journal of Business andManagement vol 7 no 4 pp 1943ndash1952 2019

[8] P S Rosa and I R Gartner ldquoFinancial distress in Brazilianbanks an early warning modelrdquo Revista Contabilidade ampFinanccedilas vol 29 no 77 pp 312ndash331 2018

[9] G S Ng C Quek andH Jiang ldquoFCMAC-EWS a bank failureearly warning system based on a novel localized patternlearning and semantically associative fuzzy neural networkrdquoExpert Systems with Applications vol 34 no 2 pp 989ndash10032008

[10] Z Wu and W Chu ldquoSampling strategy analysis of machinelearning models for energy consumption predictionrdquo inProceedings of the 2021 IEEE 9th International Conference onSmart Energy Grid Engineering (SEGE) pp 77ndash81 IEEEOshawa ON Canada Augugust 2021

[11] M Zhao A Jha Q Liu et al ldquoFaster Mean-shift GPU-accelerated clustering for cosine embedding-based cell seg-mentation and trackingrdquo Medical Image Analysis vol 71Article ID 102048 2021

[12] M D C H Sundaram A John and D D Seligmann ldquoCanblog communication dynamics be correlated with stockmarket activityrdquo Journal of Machine Learning Researchvol 11 no 9 pp 89ndash93 2008

[13] P C Tetlock M S Tsechansky and S Macskassy ldquoMorethan words quantifying language to measure firmsrsquo funda-mentalsrdquoe Journal of Finance vol 63 no 3 pp 1437ndash14672008

[14] M M Najafabadi F Villanustre T M KhoshgoftaarN Seliya R Wald and E A Muharemagic ldquoDeep learning

applications and challenges in big data analyticsrdquo Journal ofbig data vol 2 no 1 pp 1ndash21 2015

[15] A L Jones ldquoHave internet message boards changed marketbehaviorrdquo Info vol 8 no 5 pp 67ndash76 2006

[16] N Yudistira and T Kurita ldquoGated spatio and temporalconvolutional neural network for activity recognition to-wards gated multimodal deep learningrdquo EURASIP Journal onImage and Video Processing vol 2017 no 1 pp 1ndash12 2017

[17] M M Hassan M G R Alam M Z Uddin and S HudaldquoHuman emotion recognition using deep belief network ar-chitecturerdquo Information Fusion vol 51 pp 10ndash18 2019

[18] S Pirmoradi M Teshnehlab N Zarghami and S Arash ldquo+eself-organizing restricted Boltzmann machine for deep rep-resentation with the application on classification problemsrdquoExpert Systems with Applications vol 149 Article ID 1132862020

[19] P Save P Tiwarekar K N Jain and M Neha ldquoA novel ideafor credit card fraud detection using decision treerdquo Inter-national Journal of Computer Applications vol 161 no 13pp 6ndash9 2017

[20] Y Bai C Gu Q Chen J Xiao D Liu and S Tang ldquo+echallenges that head nurses confront on financial manage-ment today a qualitative studyrdquo International journal ofnursing sciences vol 4 no 2 pp 122ndash127 2017

[21] S Gupta T Gupta and G Shainesh ldquoNavigating fromprogramme loyalty to company loyaltyrdquo IIMB managementreview vol 30 no 3 pp 196ndash206 2018


model and so on [6] With the development of informationtechnology a large number of open real data informationcan be obtained in the process of enterprise operation +esedata can often reflect the past risk status of the enterprise andeven some characteristics of the risk of the whole capitalmarket [7] +erefore scholars began to apply neural net-work support vector machine and other models based onartificial intelligence methods to financial crisis earlywarning Generally speaking the early warning model basedon artificial intelligence method has made great progress inthe mechanism and systematicness of early warning It notonly overcomes the limitations of the early warning modelbased on statistical method such as requiring data to obeynormal distribution and complex calculation and analysisbut also has strong fault tolerance and learning ability [8] Atpresent artificial neural network has developed to the stageof deep learning network which is characterized by self-learning and high dynamic adaptability +erefore deeplearning is also applied to the field of financial managementevaluation

Deep learning can complete a lot of regular simple andrepetitive work +e application of financial intelligencetechnology in the financial field has played an important rolein improving business efficiency reducing work errors andpreventing and controlling enterprise risks +is paperstudies the financial risk by establishing the financialmanagement evaluation model which is a quantitativemethod It is different from the qualitative analysis +ework is difficult and the accuracy is low+emodel researchhas higher reliability and uses the analytic hierarchy processto determine the enterprise evaluation index system +ismodel interprets the historical data links the characteristicsof the data with the financial situation of the enterprise andthen uses the existing data to analyze and predict the futureso as to ensure the accuracy of the prediction

2 Related Work

Ng et al proposed a new fuzzy CMAC (cerebellar modelarticulation controller) model based on reasoning compo-nent rules as a new method for bank fault classification andearly warning system [9] Artificial neural network is in-troduced into financial management evaluation for the firsttime Artificial neural networks not only deal with the lack ofdata errors but also allow for timely adjustment of internalcontrol parameters Erhan et al have proved throughpractice that the unsupervised training method can betterdescribe the complex functional relationship which pro-vides a good reference for financial risk early warning+rough the establishment of Yahoo information bulletinboard Jones studied the impact of network information onfinancial management by using the method of deep learning[10 11] +rough the establishment of online informationbulletin board Jones reduced the fluctuation of stock priceand found that investorsrsquo differences of opinion may reducerisk and increase turnover +rough the analysis of stockrelated blog information in support vector machineChoudhury found that the fluctuation of stock price will beaffected by blog content [12] Tetlock conducted a 24-year

analysis of 500 listed companies [13] and counted thenegative words in various reports in this period through themethod of deep learning On the basis of the relationshipbetween corporate income and stock income this paperpoints out that the negative words in the news reports oflisted companies can predict the decline of listed companiesrsquoprofits and reevaluate the listed companies Under thepressure of trading news content may capture some in-formation that is difficult to quantify which has a certainearly warning function in media reports Najafabadi et alabstracted the deep learning algorithm into data represen-tation through the hierarchical learning process [14] +ismethod has attracted great attention from data science It iswidely used to solve problems such as network securitymedical information national intelligence and marketingJason Kuen and chin poo Lee applied deep learning to therepresentation of visual tracking invariance [15] andachieved good results through strong spatiotemporal con-straints and stacking slow convolution tracking Yudistiraand Kurita [16] proposed that deep learning processor hasbecome the most potential solution to accelerate deeplearning algorithm and pointed out the disadvantage of lowefficiency of assembly instructions written by deep learningwhich puts forward higher requirements for the research ofdeep learning

To sum up although there are many researches on deeplearning at home and abroad covering a wide range of fieldsthey mainly focus on theory speech recognition and imagerecognition which are the feature judgment of known in-formation and lack of relevant empirical research Manyliterature studies show that deep learning can well describecomplex functions and financial management evaluationcan also be used as a complex function judgment but so fardeep learning has not been used in the research of financialmanagement evaluation Based on the previous researchcombined with the characteristics of financial managementthis paper applies deep learning to the evaluation model offinancial management

3 Evaluation Modelling Based on DeepBelief Network

31 Deep Belief Network Model Deep learning is a categoryof machine learning that focuses on neural networks Ma-chine learning can be applied to image speech patternrecognition weather prediction stock price prediction geneexpression and content recommendation +is is verysimilar to the recognition of images by the cerebral cortex+e deep learning model first extracts the low-level featuresfrom the original signal then obtains the higher-level fea-tures from the low-level features and then obtains thehigher-level expression In the face recognition system theoriginal signal is the pixel the low-level feature is the edge ofthe object the high-level feature is the contour composed ofedges and the high-level expression is the face +roughtraining with characteristic data the error is transmittedfrom top to bottom to fine tune the network Based on theparameters of each layer obtained in the first step the pa-rameters of the whole multilayer model are further






E(v h ∣ θ) minus 1113944

nv

i1aivi + 1113944

nh

j1bjhj

⎡⎢⎢⎣ ⎤⎥⎥⎦ minus 1113944

nv

i11113944

nh

j1aibjhjvi (1)






P hj 1|v θ1113872 1113873 σ b2j + 1113944

i

2

radicviwij

⎛⎝ ⎞⎠ (4)






h v

hn

h4

h3

h1

h2

bn

b4

b3

b2

b1


vn

v3

v2

v1

an

a3

a2

a1

b isin Rh a isin Rv

W isin Rhtimesv






Input

Output





P (h2 h3)~RBM

P (h1 h2)

P (x h1)

h2

h1

h3






1 2 312 1 213 12 1

⎡⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎦



W1 W1(1113936nl1 W1)

minus 1 0540 W2 W2 (1113936nl1 W2)

minus 1

0297 W3 W3(1113936nl1 W3)


A | λmax|

A middot W

a11 a12 a1n

a21 a22 a2n

an1 an2 ann

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

W1

W2

Wn

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(5)










X(l)

f W(l)

middot X(lminus 1)

+ b(l)

1113872 1113873 (6)



Financialmanagement

evaluation

CorporateGovernance

Structure

Managementdecisions

Externalenvironment

U1

U2

U3





U11

U12

U13

U14

U21

U22

U23

U31

U32

U33


Capital structure




Legal environment



U11 U12 U13 U14 WeightU11 1 2 3 4 06532U12 12 1 2 3 03841U13 13 12 1 3 02890U14 14 13 12 1 01206



Training data

Zero mean






C minusn 1113944

xj

1yi ln a


(7)







Pa 1 minus Nminus 1

|X minus Y|1113872 1113873 times 100 (8)


L α2 +(m + n)12

log2 m(9)


Convolutionlayer


layerLinear











Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata


55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years


17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368



8123

8259

8641

8708


8123

8259

8641

8708

Cor

rect

rate

()

Correct rate


9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795


727476788082848688909294

ACCU

RACY








5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References




























E(v h ∣ θ) minus 1113944

nv

i1aivi + 1113944

nh

j1bjhj

⎡⎢⎢⎣ ⎤⎥⎥⎦ minus 1113944

nv

i11113944

nh

j1aibjhjvi (1)






P hj 1|v θ1113872 1113873 σ b2j + 1113944

i

2

radicviwij

⎛⎝ ⎞⎠ (4)






h v

hn

h4

h3

h1

h2

bn

b4

b3

b2

b1


vn

v3

v2

v1

an

a3

a2

a1

b isin Rh a isin Rv

W isin Rhtimesv






Input

Output





P (h2 h3)~RBM

P (h1 h2)

P (x h1)

h2

h1

h3






1 2 312 1 213 12 1

⎡⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎦



W1 W1(1113936nl1 W1)

minus 1 0540 W2 W2 (1113936nl1 W2)

minus 1

0297 W3 W3(1113936nl1 W3)


A | λmax|

A middot W

a11 a12 a1n

a21 a22 a2n

an1 an2 ann

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

W1

W2

Wn

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(5)










X(l)

f W(l)

middot X(lminus 1)

+ b(l)

1113872 1113873 (6)



Financialmanagement

evaluation

CorporateGovernance

Structure

Managementdecisions

Externalenvironment

U1

U2

U3





U11

U12

U13

U14

U21

U22

U23

U31

U32

U33


Capital structure




Legal environment



U11 U12 U13 U14 WeightU11 1 2 3 4 06532U12 12 1 2 3 03841U13 13 12 1 3 02890U14 14 13 12 1 01206



Training data

Zero mean






C minusn 1113944

xj

1yi ln a


(7)







Pa 1 minus Nminus 1

|X minus Y|1113872 1113873 times 100 (8)


L α2 +(m + n)12

log2 m(9)


Convolutionlayer


layerLinear











Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata


55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years


17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368



8123

8259

8641

8708


8123

8259

8641

8708

Cor

rect

rate

()

Correct rate


9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795


727476788082848688909294

ACCU

RACY








5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References



























1 2 312 1 213 12 1

⎡⎢⎢⎢⎢⎢⎣⎤⎥⎥⎥⎥⎥⎦



W1 W1(1113936nl1 W1)

minus 1 0540 W2 W2 (1113936nl1 W2)

minus 1

0297 W3 W3(1113936nl1 W3)


A | λmax|

A middot W

a11 a12 a1n

a21 a22 a2n

an1 an2 ann

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

W1

W2

Wn

⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣

⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦

(5)










X(l)

f W(l)

middot X(lminus 1)

+ b(l)

1113872 1113873 (6)



Financialmanagement

evaluation

CorporateGovernance

Structure

Managementdecisions

Externalenvironment

U1

U2

U3





U11

U12

U13

U14

U21

U22

U23

U31

U32

U33


Capital structure




Legal environment



U11 U12 U13 U14 WeightU11 1 2 3 4 06532U12 12 1 2 3 03841U13 13 12 1 3 02890U14 14 13 12 1 01206



Training data

Zero mean






C minusn 1113944

xj

1yi ln a


(7)







Pa 1 minus Nminus 1

|X minus Y|1113872 1113873 times 100 (8)


L α2 +(m + n)12

log2 m(9)


Convolutionlayer


layerLinear











Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata


55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years


17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368



8123

8259

8641

8708


8123

8259

8641

8708

Cor

rect

rate

()

Correct rate


9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795


727476788082848688909294

ACCU

RACY








5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References
























X(l)

f W(l)

middot X(lminus 1)

+ b(l)

1113872 1113873 (6)



Financialmanagement

evaluation

CorporateGovernance

Structure

Managementdecisions

Externalenvironment

U1

U2

U3





U11

U12

U13

U14

U21

U22

U23

U31

U32

U33


Capital structure




Legal environment



U11 U12 U13 U14 WeightU11 1 2 3 4 06532U12 12 1 2 3 03841U13 13 12 1 3 02890U14 14 13 12 1 01206



Training data

Zero mean






C minusn 1113944

xj

1yi ln a


(7)







Pa 1 minus Nminus 1

|X minus Y|1113872 1113873 times 100 (8)


L α2 +(m + n)12

log2 m(9)


Convolutionlayer


layerLinear











Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata


55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years


17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368



8123

8259

8641

8708


8123

8259

8641

8708

Cor

rect

rate

()

Correct rate


9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795


727476788082848688909294

ACCU

RACY








5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References
























C minusn 1113944

xj

1yi ln a


(7)







Pa 1 minus Nminus 1

|X minus Y|1113872 1113873 times 100 (8)


L α2 +(m + n)12

log2 m(9)


Convolutionlayer


layerLinear











Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata


55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years


17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368



8123

8259

8641

8708


8123

8259

8641

8708

Cor

rect

rate

()

Correct rate


9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795


727476788082848688909294

ACCU

RACY








5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References






























Amount ofdata

Number ofcompanies

Amount ofdata

Number ofcompanies

Amount ofdata


55 2805 55 2805 110 561055 2805 55 2805 110 5610110 5610 110 5610 220 11220


51 5151 51 5151 102 1030251 5151 51 5151 102 10302102 10302 102 10302 10506 20604


26 4576 26 4576 52 915226 4576 26 4576 52 915252 9152 52 9152 104 18304

12years


17 3842 17 3842 34 768417 3842 17 3842 34 768434 7684 34 7684 68 15368



8123

8259

8641

8708


8123

8259

8641

8708

Cor

rect

rate

()

Correct rate


9134 9134 9134 9134 9134 9134 9134 9134

81 802

867

791

832852

819

795


727476788082848688909294

ACCU

RACY








5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References




























5 Conclusion


Data Availability


2 3 4 510

20

40

60

80

100

120

0

01

02

03

04

05

06

07

08




090

092

094

096

098

100

Layer1Layer2Layer3


8732

8105

7691

7057

2000 4000 6000 8000 10000000ACCURACY ()

CNN

KNN

SVM-RBF

SVM-Linear

MET

HO

DS





Acknowledgments


References


























Acknowledgments


References
























Documents

Application of Deep Learning in Financial Management