Research on the Risk Crisis Prediction of Enterprise Finance by Genetic Algorithm

Transcription

1 Research on the Rsk Crss Predcton of Enterprse Fnance by Genetc Algorthm Tngtng Ye Abstract Wth the development of the global economy, the competton between enterprses s gettng fercer and enterprses are facng ncreasng crses and challenges As a result, the predcton of rsk crss of enterprses becomes very mportant. The growng s of data wthn enterprses have caused great nconvenence to the rsk crss forecasts. In ths paper, the genetc algorthm s analyzed. Wth ths algorthm, a large of hgh-lattude data are reduced n dmenson. The genetc algorthm s also used to optmze the neural network. A genetc algorthm model s establshed by MATLAB. The experment proves that ths algorthm can effectvely optmze the BP neural network, and has an obvous early warnng effect on the fnancal rsk crss wth hgh accuracy, whch provdes a reference for the applcaton of the algorthm n the predcton of the fnancal rsk crss. Keywords BP neural network, crss predcton, fnancal rsk, genetc algorthm I. INTRODUCTION Enterprse s a knd of artfcal system and organzaton []. Wth the rapd development of the market economy, market competton s becomng ncreasngly ferce. Wth the development of nformaton technology, enterprses have begun to have more and more huge data, whch brought nconvenence to data processng. In the context of nternatonal economc ntegraton, the predcton of the rsk of corporate fnance crss s partcularly mportant, so we need to study an effectve and scentfc method to predct the rsk of corporate fnancal crss. L et al. [] proposed a fnancal crss predcton model, whch combned genetc algorthm wth support vector machne to use the fnancal nformaton spread on the network wth hgh accuracy. Expermental results showed that the proposed model could be used n fnancal and corporate governance and realze more accurate fnancal crss predcton. Hu et al. [3] establshed a new rsk ratng method based on default dstance and order statstcs, and then studed the fnancal crss forecast based on feature-weghted SVM model, and proved that the model had good performance n predctng the fnancal crss of lsted companes. Wang et al. [5] developed a FOA based Logstc fnancal rsk warnng model and proved ts effectveness through experments. Fan [7] proposed a weghted logstc regresson based fnancal rsk warnng model and found that t had favorable nonlnearty. Huang et al. [8] T. T. Ye s wth School of Busness Admnstraton, Unversty of Scence and Technology, Laonng, Anshan, 405, Chna (e-mal: tngtye@sna.com). establshed a SVM based fnancal warnng system for lsted companes and beleved that t could be extensvely appled. In ths study, genetc algorthm was analyzed and combned wth BP neural network as an evaluaton functon for predctng accuracy. Moreover a genetc algorthm model was establshed usng MATLAB for predctng the fnancal rsk crss. The valdty of ths model was verfed through experments. II. ENTERPRISE FINANCIAL RISK RELATED FACTORS Factors affectng the fnancal rsks of enterprses can be dvded nto nternal factors and external factors. The external factors are uncontrollable. Under the envronment of market competton, enterprses as a part of the market economy wll nevtably be affected by the whole market. Enterprses cannot determne the polces released by the government. In addton, the ncreasngly ferce market competton makes enterprse development unstable, and an enterprse may go up or down n a blnk. The fnancal condton of an enterprse may be affected once t become nferor n the competton, whch can nduce fnancal rsks, cause fnancal crss and even lead to bankruptcy. The nternal factors are also an mportant part of the fnancal rsk. In many cases, the wrong decson of the operator wll cause rreparable loss. Busness operaton s far-reachng; a small mstake wll cause huge reacton. Therefore enterprse operators can hardly absolve themselves from the blame n the problem of fnancal rsks. Moreover many enterprses have unreasonable captal structure, bad management and weak proftablty, whch can greatly ncrease the probablty and severty of fnancal rsks. III. ENTERPRISE FINANCIAL CRISIS The enterprse fnancal crss s manfested n the followng aspects: ()Management: defct caused by low sales ncome; proft declne caused by low effcency; captal flow nfluenced by decreased survval turnover; nsuffcent cash flow. ()Investment: Dversfed nvestment should be mplemented n lmted fnancal crcumstances, consderng coordnaton between corporate asset structure and captal structure as well as between proftablty and lqudty. Mere consderaton of proftablty would put the company n techncal dffcultes, and the mere consderaton of lqudty would leave the company dle most of ts propertes and lose proftablty. ISSN:

2 (3)Fnancng: It s an mportant factor for the development of enterprses to make profts through the use of other people's funds. In ths way, enterprses can rase the funds needed for development and mprove the rate of return on captal. Whle the captal return rate s mproved, fnancal rsks are also ncreasng. When the enterprse s n debt, t s easy to lose the ablty to repay due to poor management, whch causes the enterprse to go bankrupt. (4)Other aspect: In the process of operaton, enterprses are prone to large lawsuts and other emergences, from whch they are lkely to suffer huge losses and even go bankrupt. Genetc algorthm can reduce the dmenson of the fnancal data n enterprses and optmze the BP neural network predcton model, whch plays an mportant role n the predcton of corporate fnancal crss. IV. GENETIC ALGORITHM FEATURES Genetc algorthm s a new global optmzaton algorthm based on bologcal genetcs [9]. It can mprove the ndvdual ftness of the populaton through contnuous evoluton, thus achevng the purpose of fndng the optmal goal. Compared wth other algorthms, t has the followng features: ()Self-organzaton and adaptablty features. ()It can encode the ndvduals and process wth the encoded characters. (3)Extensve feature. (4)Parallelsm feature. (5)Expandablty [0]. Wth these features, t can be appled to functon optmzaton desgn, combnaton optmzaton desgn, producton schedulng optmzaton desgn and mage sgnal processng []. Genetc algorthm s a probablstc search algorthm whch can use encodng technology to encode ndvduals to form chromosomes. In ths process, the evolutonary process of the ndvduals composed of character strngs s smulated []. The man steps of the algorthm are: ()Code accordng to actual problems. ()Generate prmary populaton: randomly generate N ndvduals to generate prmary populaton. (3)Evaluaton of ftness value: Evaluaton of ftness of ndvduals of prmary populaton based on the target of optmzaton. (4)The genetc operator operatons of selecton, crossover and mutaton are carred out. (5)Form an evolvng populaton. (6)If the condtons are met, t goes to the next step; f not, t goes back to step 3. (7)Output the results. start Produce a prmary populaton Calculate ftness meet the optmzaton gudelnes? Y N Selecto n Crossov er Output the optmal soluton Fg. Genetc algorthm flow Mutaton A. Feature model based on genetc algorthm In ths paper, genetc algorthm s used n feature selecton, and the advantages of Flter model [3] and Wrapper model [4] are combned to form a hybrd selecton model whch can realze feature selecton of nformaton through flterng as well as feature selecton of data by sealng method so that data can be processed quckly whle ts qualty s ensured. B. Feature selecton based on hybrd selecton mode The man characterstc of the flterng method s that the evaluaton crteron s ndependent of the classfcaton algorthm and the relatonshp between the two characterstcs s not consdered [5]. Therefore, the selecton speed of the flterng algorthm s fast, but ts accuracy s not hgh. In ths paper, t s used for the screenng of the orgnal data. The man feature of the encapsulated method s that the evaluaton crteron s ndependent of the learnng algorthm of the classfer, whch can encapsulate the evaluaton algorthm n the learnng algorthm [6]. Therefore, the algorthm can select a subset of features wth hgh performance, whch s appled n ths paper for the fne screenng of the features after ntal screenng to ensure the performance of the feature subset. C. Double crtera evaluaton model Ths paper apples the double crtera evaluaton model and obtans a unon set of two results, whch not only rapdly screens targets but also prevents the randomness of a sngle crteron. The double crteron used n ths paper s nformaton gan and correlaton coeffcent. Informaton gan s nformaton measurement, and correlaton coeffcent s dependency measurement. () Informaton gan Entropy s the average amount of nformaton after the elmnaton of redundant nformaton, whch s generally referred to as nformaton entropy [7]. Informaton entropy manly uses the numercal form to measure the uncertanty of the value of a random varable. ISSN:

3 Suppose N to be a random varable, ( qn)to be the probablty when N= n, and the nformaton entropy of N can be expressed as HN ( ), then the equaton of HN ( )s as follows: = () Hn ( ) qn ( )log( qn ( )) The nformaton entropy HN ( )s not relevant to the value of N but ts probablty dstrbuton. Wth the ncrease of the uncertanty degree of the value of N, ts probablty dstrbuton ncreases, whch leads to the ncrease of ts nformaton entropy HN ( ), requrng more nformaton. When the value of N reaches the hghest degree of uncertanty, ts value of nformaton entropy wll also reach the maxmum. In ths case, the probablty of the random varable N s the same, so that each value has the same chance to appear. At ths moment, we cannot determne the specfc value of the random varable N but only choose one random value. But when X has only one value of x, the probablty dstrbuton of ths value s,.e., ( qn= ), where the value of X s defnte, and ts nformaton entropy s mnmzed. Condtonal entropy refers to the dependence of one varable on another varable, and when a varable s known, condtonal entropy s the uncertanty of another varable. If the random varable M s the known varable, then the condtonal entropy of the random varable N on M s: ( ) ( ) HN ( 丨 M)= - q m q nm log( qnm ( )) () j j j j Informaton gan refers to the nformaton obtaned after the elmnaton of uncertanty, whch s the dfference of nformaton entropy [8]. When nformaton gan s taken as the evaluaton crtera, t manly consders the nformaton amount ts feature provdes for the classfcaton. The more nformaton a feature provdes to the classfcaton, the larger the nformaton gan. If a feature s deleted, then the change n the overall amount of nformaton before and after the feature s the nformaton amount provded by the classfcaton and the changed part s the nformaton entropy of the feature. Informaton gan s defned as follows: ( ) IG N, M = H ( N) H ( N M ) (3) Informaton gan can most ntutvely show the amount of nformaton provded by the feature for the overall classfcaton and can be wdely appled to feature selecton. () In ths paper, Spearman correlaton coeffcent [9] s used to collect data wth uncertan dstrbuton. The correlaton between two features s tested by the unformty between them. When both varables ncrease or decrease smultaneously, these two varables are Spearman-related. The data n the data set s ranked accordng to ther sze. Average grade calculaton s performed on repeated data and whether the rank of two varables s relevant s checked. Then, the Spearman correlaton coeffcent can be expressed as: 6 a W P x l= ( ) xx ( ) (4) = a Where the rank correlaton coeffcent l refers to the magntude of the rank correlaton, when t s postve, t refers to postve correlaton, when t s negatve, t refers to negatve correlaton, when t s 0, t refers to zero correlaton (ts value s W between - to ). and P refer to the rank of n and m. V. BP NEURAL NETWORK MODEL BASED ON GENETIC ALGORITHM A network structure wth 4 nput layer nodes, 6 hdden layer nodes and output layer node s adopted n ths desgn. A. BP network weght based on genetc algorthm () Chromosome expresson and prmary populaton Codng s an mportant problem n genetc algorthms. Wth the ncrease of the complexty and the dmenson of data, the shortcomng of too long codng n bnary codng s exposed. To avod ths problem, we drectly apply floatng-pont encodng to the performance of soluton to mprove the effcency and accuracy of solvng problems. Randomly generate a populaton of X ndvduals, N { N, N, NN} ndvdual N { n n n } =, each =,, represents ntal weghts and x threshold dstrbuton of a neural network; each gene value represents a lnk weght or threshold for a neural network, then the length of the ndvdual s the sum of the of weghts and thresholds of the neural network, as follows: L= S* L+ L* L + L+ L (5) where S refers to the of nput layer, L refers to the of hdden layer, L refers to the of output layer. () Objectve functon and ftness functon The error sum of squares of the network output value and expected output value s an mportant performance of BP network. The smaller the square sum, the better the network performance. Mark the expected output of the output layer neuron of the th * group to be R, the actual output to be R ; A refers to the total E α s: sample and ts error sum of squares ( ) A * E( α ) = ( R R) (6) = ISSN:

4 Then, genetc algorthm can be appled to complete the search of network weghts. Substtutng each chromosome's weght and threshold nto the BP network to mnmze the error functon E ( α ) n the neural network and the target functon s defned as follows: mn E A ( R R) * ( α ) = = The ftness functon D( α ) of genetc algorthm s: D ( α) / E( α) (7) = (8) (3) Genetc operaton Each chromosome s calculated and proportonally assgned. Then, the possblty of descendants left s determned based on the probabltes of varous ftness degree. If there s an, then the ndvdual α, and ts ftness degree s F ( α ) probablty of beng selected s: ( α ) ( α ) Q= F f (9) / b = VI. SIMULATION RESULTS The model establshed above s appled for the smulaton experment under Matlab envronment and compared wth normal neural network model. The collected data were nput nto the normal neural network model and the results nclude msjudgment, accuracy and modelng tme. The data of 75 lsted companes durng were respectvely nput nto the tradtonal neural network and the optmzed BP neural network, among whch, the data of 60 lsted companes were taken as tranng and learnng and the rest was taken as predctng. The predcted results were compared wth the actual enterprse crss stuaton. The results are shown n Table Table Results of the optmzed neural network predcton model Non-ST companes ST companes Average accuracy Modelng tme Sample 9 55% 3.s Msjudgment 4 Accuracy rate 90% 0% Sample % 0.97s Msjudgment 0 Accuracy rate 00% 77.78% Sample %.85s Msjudgment 0 Accuracy rate 9.67% 00% Table Results of the tradtonal neural network predcton model Non-ST ST companes Average accuracy Modelng tme companes Sample % 63.s Msjudgment 6 Accuracy rate 45.45% 50% Sample % 5.s Msjudgment 4 Accuracy rate 66.67% 55.56% Sample % 7.6s Msjudgment Accuracy rate 8.8% 75% ISSN:

5 As shown n Table and, the accuracy of the predcton model ncreased each year and the average accuracy of the optmzed model was ncreased greatly. Each year's sample data ndvdual dfferences made the accuracy of each year dfferent. However, after the optmzaton of the genetc algorthm, BP neural network mproved the accuracy of fnancal rsk predcton and reached the expected target. The results showed that the predcton accuracy was the lowest n 04 and the accuracy was only 55% after optmzaton. In the pre-st data of lsted companes, the data for 06 dd not have much research value and usng the data of 03 to predct fnancal rsks n three years dd not have obvous advantages. Therefore, an effectve predcton perod started from 04. It was also found that the accuracy of 05 was the hghest, reachng 95.84%. Through the data of the three years, t was found that the modelng speed of the BP neural network after optmzaton mproved a lot. Wth the characterstcs of survval of the fttest n bology, genetc algorthm does not depend on the tradtonal gradent statstcs and drectly establshes the wnnng populaton, whch greatly mproves the modelng speed. In practce, t makes the predcton more convenent and fast. Corporate fnance can profoundly affect the lfe and death of a company, and the sooner a company can dscover the rsks, the sooner the busness decson makers can make respondng measures. Through experments, t can be concluded that the model proposed n ths paper can effectvely analyze the enterprse's fnancal rsk and avod the unknown rsk of the enterprse due to the fnancal stuaton. VII. DISCUSSION Wth the development of economc globalzaton, enterprses are facng ncreasngly larger challenges and rsks, whch ncrease the dffculty n decson makng. Fnancal crss s a progressve process. A small mstake may brng rreparable loss. If there s no good warnng system to help operators quckly and make approprate response, the enterprse s lkely to fall nto the fnancal crss. Many enterprses have been destroyed by large rsks whch are prevously small. Once fnancal crss appears, fnancal rsks wll flow from nternal to external f there s no effectve measure to control the rsks, whch can not only damage the nterests of enterprses but also greatly affect the economc condton of enterprse stakeholders and the whole regon [0]. In ths case, an effectve fnancal rsk early warnng model s necessary. An effectve fnancal early warnng system s hghly senstve to the omen of fnancal crss. It can warn operators tmely, and operators can take measures early and np fnancal crss n the bud. An effectve fnancal early warnng system can also analyze the causes of fnancal problems and help managers fnd a specfc way to avod t. More mportantly, fnancal early-warnng system can try to avod the same mstakes through a seres of analyss. Many effectve rsk warnng models have been verfed n the actual operaton. Tan et al. [] appled a warnng model whch combned quanttatve methods wth qualtatve methods to warn real estate enterprses and got favorable feedback. L et al. [] developed an early warnng model of fnancal rsks n core enterprses under supply chan envronment based on a large of theores and tested t through examples. In ths study, a genetc algorthm based BP neural network model was establshed for warnng fnancal rsks of enterprses. In the smulaton experment, t was found that the accuracy of the model was hgher than that of the tradtonal BP neural network. To make the fnancal rsk early warnng system work better, enterprses should combne qualtatve and quanttatve methods to make t more accurate n the actual operaton process. Tmely, effectve and accurate fnancal nformaton s also very mportant for the fnancal early warnng system; therefore enterprses should strengthen fnancal management and nternal control and establsh fnancal nformaton system to mprove the predcton accuracy and reduce loss. VIII. CONCLUSION Wth the development of the global economy, enterprses are facng more and more crses, so t s becomng more and more mportant to establsh the crss predcton model. Ths paper ntroduced genetc algorthm and establshed feature model. Combnng genetc algorthm model wth BP neural network, the enterprse fnancal rsk crss predcton model was establshed and the model was tested. Experments showed that the accuracy and speed of the predcton model were greatly mproved after optmzed by the genetc algorthm, whch provdes a reference for the applcaton of genetc algorthm n the enterprse fnancal rsk predcton. REFERENCES [] M. Mazen, The mplcatons of the acquston of a new natonalty for the rght of return of palestnan refugees, Asan J. Int. Law, vol. 5, no., pp , 05. [] C. K. L, D. Lang, F. Y. Ln, and K. L. Chen, The applcaton of corporate governance ndcators wth XBRL technology to fnancal crss predcton, Emerg. Markets Fnan. Trade, vol. 5, pp. S58-S7, 05. [3] W. H. Hu, F. Gao and C Huang, Fnancal crss predcton based on dstance to default and feature weghted support vector machne, Int. Confer. Nat. Comput., pp , 06. [4] Q. Lu, Y. J. Jn and Z. X. Yu, The Fnancal Early Warnng Model of Lsted Companes Based on Lasso Method and Logstc Regresson, Adv. Appl. Mathem., vol. 6, no. 4, pp , 07. [5] Y. Wang, Q. Y. J and J. Cha, An Research of Fnancal Early Warnng About Logstc Model Based on FOA, J. Syst. Sc. Mathem. Sc., vol. 37, no., pp , 07. [6] Y. Lang, Emprcal study on Chna's fnancal rsk early warnng under the new normal, Internatonal Conference on Industral Economcs System and Industral Securty Engneerng, pp. -7, 07. [7] X. Y. Fan, Fnancal Rsk Early Warnng Constrant Method Based on Weghted Logstc Regresson, Boletín Técnco, vol. 55, no., pp. 7-79, 07. [8] H. Huang, W. Jang and S. Wang, Research of Fnancal Early-warnng for Lsted Companes Based on SVM, Internatonal Conference on Computatonal Scence and Engneerng, pp. 78-8, 05. [9] H. S. Madraswala, Modfed genetc algorthm soluton to unt commtment problem, Int. Confer. Nascent Technol. Engn., pp. -6, 07. ISSN:

6 [0] A. L. Speddng, G. D. Fatta and M. Cannataro, A Genetc Algorthm for the selecton of structural MRI features for classfcaton of Mld Cogntve Imparment and Alzhemer's Dsease, IEEE Int. Confer. Bonform. Bomed., pp , 05. [] E. Jehamy, M. Salaün, M. Ney, and G Landrac, Applcaton of genetc algorthm for constraned metal plate lens profle optmzaton, Int. Symp. Antenn. Technol. Appl. Electromagn., pp. -4, 007. [] M. M. Smth and Y. S. Chen, A novel evolutonary algorthm for the homogeneous Probablstc Travelng Salesman Problem, IEEE/ACIS, Int. Confer. Comp. Informat. Sc., pp. -6, 06. [3] D. Messner, S. Reuter, E. Strgel, and K. Detmayer, Intersecton-Based Road User Trackng Usng a Classfyng Multple-Model PHD Flter, Intell. Transp. Syst. Magaz., vol. 6, no., pp. -33, 06. [4] X. Hou, B. R. Hodges, S. Negusse and C. Barker, A mult-model Python wrapper for operatonal ol spll transport forecasts, Comput. Sc. Dscov., vol. 8, no., pp , 05. [5] V. Aguarpuldo, M. Gestal, M. Cruzmonteagudo, J. R. Rabuñal, J. Dorado and C. R. Munteanu, Evolutonary computaton and QSAR research, Curr. Comp. - Aded Drug Desgn, vol. 9, no., pp. 3-5, 03. [6] J. Bergstra, N. Pnto, and D. D. Cox, SkData: data sets and algorthm evaluaton protocols n Python, Comput. Sc. Dscov., vol. 8, no., 04007, 05. [7] V. Zelevnsky, M. Horo and B. A. Brown, Informaton entropy, chaos and complexty of the shell model egenvector, Phys. Lett. B, vol. 350, no., pp. 4-46, 06. [8] C. Yn and J. X, Maxmum entropy model for moble text classfcaton n cloud computng usng mproved nformaton gan algorthm, Multmed. Tools Appl., pp. -7, 06. [9] M. T. Puth, M. Neuhäuser, G. D. Ruxton, Effectve use of Spearman's and Kendall's correlaton coeffcents forassocaton between two measured trats, Anm. Behav., vol. 0, pp , 05. [0] M. X. Hong, A study on fnancal rsk control of group enterprse, J. Chem. Pharm. Res., vol. 7, no. 4, pp , 05. [] X. R. Tan, H. D. Jang and J. L.Lu, The Research of Fnancal Rsk Warnng Model about Real Estate Enterprses n Chna, Adv. Mater. Res., vol , no. 3, pp , 04. [] L. L, L. Wang and L. Yu, Research of fnancal rsk warnng model of lsted companes n the Supply Chan envronment, Control Decson Conference, pp , 0. ISSN: