Using SVM with Financial Statement Analysis for Prediction of Stocks

Transcription

1 Communcatons of the IIMA Volume 7 Issue 4 Artcle Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks Shuo Han Department of Management Scence and Engneerng Unversty of Scence and Technology Bejng Rung-Chng Chen Department of Informaton Management Chaoyang Unversty of Technology Follow ths and addtonal works at: Recommended Ctaton Han, Shuo and Chen, Rung-Chng (2007) "Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks," Communcatons of the IIMA: Vol. 7: Iss. 4, Artcle 8. Avalable at: Ths Artcle s brought to you for free and open access by CSUSB ScholarWorks. It has been accepted for ncluson n Communcatons of the IIMA by an authorzed admnstrator of CSUSB ScholarWorks. For more nformaton, please contact scholarworks@csusb.edu.

2 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks Shuo Han Department of Management Scence and Engneerng Unversty of Scence and Technology Bejng, Rung-Chng Chen Department of Informaton Management Chaoyang Unversty of Technology ABSTRACT At present, there are many techncal analyses for predcton n stock market. However, the techncal ndces are fluctuated wth the quantty of stock exchanges. The fnancal ndces are more relable, nonvolatle and vald compared wth the techncal ndces. In ths paper, we propose an orgnal and unversal method by usng SVM wth fnancal statement analyss for predcton of stocks. We appled the SVM to construct the predcton model and select Gaussan radal bass functon (RBF) as the kernel functon. The expermental results show our method not only mprove the accuracy rate, but also meet the dfferent stockholders expectatons. INTRODUCTION Support vector machne (SVM) s a useful technque for data classfcaton (Burges, 998), regresson (Smola et al, 998) and predcton (Müller et al, 997). Prevously there has been a lot of study usng artfcal neural network (ANN) n these areas, especally n the feld of predcton. However, n the stock market, because the data often has enormous noses and complex dmensonalty, the ANN method has some lmtatons (Km, 2003). Recently, SVM has been successfully used n the feld of predcton. SVM can treat hgher dmensonal data better even wth a relatve low amount of tranng set. Further more, t can present a good ablty of generalzaton for complex model (Thssen, 2003). Some applcatons by SVM to predct the stock market have been ssued, but the degree of accuracy rate and the acceptablty of certan predcton are measured by the predctors devaton from ther own experences or the neffectve data (Huang, 2005). In the feld of predcton for stock market, the most mportant thng s to mprove the predcton accuracy rate (Huang et al, Chen et al, 2006). However, lttle study has justfed the sutablty of stock market predcton by SVM. In ths paper, we propose an orgnal and unversal method by usng SVM wth fnancal statement analyss for predcton of stocks. Commonly there are many techncal analyses for predcton n stock market. But these techncal ndces such as RSI, BIAS, etc. appear to fluctuate wth the quantty of stock exchanges. Compared wth the techncal ndces, the fnancal ndces from the fnancal statement are much more relable, nonvolatle and vald. The goal of ths paper s to mprove the accuracy rate of predcton and to meet dfferent knds of stockholders expectatons. Ths paper s organzed as follows. In secton 2, we wll brefly explan the theory of SVM and some concepts from fnance and accountng. In secton 3, the methodology s gven. In secton 4, the experment and the expermental result s shown. In secton 5, we wll present the concluson and some suggestons. Communcatons of the IIMA Volume 7 Issue 4

3 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks THEORY SVM The support vector machne (SVM) s a type of learnng machne that s based on statstcal theory and t s a popular technque for classfcaton. In order to perform bnary devaton, the SVM uses a hgh dmenson space to fnd a hyper plane where the error rate s mnmal. The methodology of SVM can be stated brefly as follows: Gven a tranng set of nstance-label pars (x, y ), =,..., l where x R n and y {, } l, the support vector machnes requre the soluton of the followng optmzaton problem: l T mn w w+ c ξ subject to w, b, ξ 2 y = T ( w z + b) ξ 0, ξ, =,..., l. The tranng vector x are mapped nto a hgher dmensonal space by the functon φ as z = φ(x ). C > 0 s the penalty parameter of the error term. Thus the problem s solved as follows (Chang et al, 2007): mn α F T T ( α ) α Qα e α = 2 T subject to y α = 0, (2) 0 α C, =,..., l Where e s the vector of all, C > 0 s the upper bound, Q s an l by l postve semdefnte matrx, Q j =y y j K(x, x j ), and K(x, x j )= φ(x ) T φ(x j ) s the kernel functon. Thus the decson functon s: sn w = l T ( w ( x) + b) = sn α yk( x, x) ϕ + b (3) = α y ϕ l = ( x ) () In ths paper, the kernel functon of the research s called Gaussan kernel (Keerth, 2003): ~ 2 ( ~ x x K x, x) = exp( ) 2 2σ (4) LIBSVM s a lbrary for SVM. The goal of LIBSVM s to produce a model whch predcts target value of nstant datas n the testng set whch are gven only the attrbutes and let the users can easly use SVM as a model. Often n LIBSVM the kernel functon s radal bass functon (RBF), and the γ s called kernel parameter (Hsu, 2003). It s represented as follows: Concepts of Fnance and Accountng 2 ( x, x ) = exp( γ x x ), γ > 0 K (5) j j As mentoned above, there are many techncal analyses for stock market predcton. The stockholders often make decsons by applyng the techncal ndces such as RSI, BIAS, MACD, PSY, KDJ, etc. But these techncal ndces are easly fluctuatng wth the quantty of stock exchanges. Besdes, for some man-made reasons, these techncal ndces are not very trustful. Consequently there are some lmtatons to these applcatons when the stockholders want to evaluate ther target stocks or make decsons. The stock companes wll release ther fnancal statements for each fnancal year to publc. In that case, we can get Communcatons of the IIMA Volume 7 Issue4

4 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks some useful fnancal ndces from the target fnancal statement. For nstance, current assets, non-current assets, total assets, shareholders equty, share captal, current labltes, non-current labltes, man busness ncome, proft from man busness, gross proft, earnngs before nterest and tax, net proft, net cash ncome per share, earnngs per share, book value per share, adjusted book value per share, return on equty, the nterests of shareholders rato, gross margn, net proft growth rate, shareholders rate, net rate of return on captal-growth, man busness ncome growth and man proft growth. The stockholders can process dfferent evaluaton by selectng the related account elements they needed. In ths way, the stockholders are able to do the relatve fnancal analyses. For example, by selectng the net proft margn, gross proft margn, net profts on assets, return on assets, net proft rato, return on equty and growth rate of earnngs before tax, we can do the proftablty analyss. By selectng the man busness ncome growth, net proft growth rate, total asset growth rate, shareholders rate and man proft growth, we can do the growth rate analyss. By selectng the debt to total assets rato, debt to total assets rato, long term debt rato and fxed rato, we can do the fnancal structure analyss. By selectng the current rato, quck rato, current cash debt coverage rato, the nterests of shareholders rato, current labltes rato and gross gearng, we can do the solvency analyss. By selectng the nventory turnover, asset turnover, recevables turnover, man busness cost rato, operatng expense rato, admnstratve expense rato and fnance expense rato, we can do the operatng effcency analyss. In the stock market, the operatng status s the pvotal factor for the stock value. It could be reflected by the fnancal statement. Therefore, analyzng and researchng the account elements from the fnancal statement are very mportant. If we can do the advanced analyses for these tems, we wll be more aware of the fnancal and operatng status of those stock companes. Certanly t wll be helpful for stockholders makng decsons to buy the target stocks. Besdes, compared wth the techncal ndces, these ndces are much more relable, nonvolatle and vald. For the common stockholders, they can not evaluate and predct the target stock by analyzng all the fnancal ndces as the professonals. They often concern more on the prmary fnancal ndces such as earnng per share growth rate, book value per share growth rate, return on equty, net proft growth rate, shareholders rate, net rate of return on captal-growth and the man busness ncome growth. If the common stockholders can ntegrate the professonal advces wth the prmary ndexes analyss, they wll be more confdent n makng decsons of the stock nvestment. The goal of ths paper s to process a new method by usng SVM wth fnancal statements analyss for predcton of stocks to mprove the accuracy rate and meet dfferent stockholders expectatons. Normally, some professonals or experts wll predct the trend of stocks n future by ther own feld of knowledge, nformaton and technques, and release the results to publc. We mprove the accuracy rate of predcton based on these professonal results combned wth our results of prmary fnancal ndces. In that case, the results wll be more relable. We hope the method would provde a better way of applcaton n the real world, whch wll be explaned n the followng sectons. Data Collecton METHODOLOGY The research data used n ths study s selected from the fnancal statements released by the stock companes of the stock exchange n Shangha and Shenzhen. By the endng of ths study, the numbers of the fnancal statements released by the stock companes s 25, so the research data are brand-new and dependable. The prmary ndces we selected are Earnngs Per Share (EPS), Book Value Per Share (BVPS) and Net Proft Growth Rate (NPGR). Based on the results released by the professonals and experts, Outstandng Achevement Growth Rate (OAGR), whch we have classfed two choces, + and -, and denoted as SVM. Consequently, n the experment, we combned the SVM wth three prmary ndces respectvely. Then we can generate eght knds of predcton modules. The data set and the predcton modules are shown n Table and Table 2. Communcatons of the IIMA Volume 7 Issue4

5 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks Table : The data set (usng SVM to present OAGR and dfferent fnancal ndces). SVM OAGR EPS BVPS NPGR Table 2: Predcton modules. Experment SVM OAGR EPS BVPS NPGR In ths paper, we wll compare the eght dfferent knds of expermental results from the eght modules. The frst module (Experment : SVM and OAGR) s the basc module. Other seven expermental results wll compared wth t. In that case, for dfferent stockholders, they can select dfferent tems to compare. Because the frst expermental results s only reflected the accuracy rate of the professonal or experts, not concludng the stockholders analyses. We defne the frst result as the basc result. Dfferent analyzer can get dfferent results by selectng the target tems they need. Theoretcal Model Normally the professonals or experts wll predct the trend of target stocks based on the Outstandng Achevement Growth Rate (OAGR), and the results wll be classfed to fve levels. From good to bad, they are makng proft, favorable of outstandng achevement, advsng, makng up the defcts, defct. In ths paper, our method s based on the hypothess that all the stockholders are sensble and ratonal people. Commonly they wll select the stocks belongng to the level of makng proft or level of favorable of outstandng achevement for the target stocks. Other Communcatons of the IIMA Volume 7 Issue4

6 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks levels wll be dened. In ths way, we defne the former choce as class +. Smlarly, we defne the latter choce of denyng the other levels of stocks as class -. Thus, the method wll be sutable of the SVM. And then, we wll combne the SVM wth some prmary fnancal ndces we selected to make predcton. We hope that not only wll the accuracy rate be mproved, but also wll the dfferent fnancal ndces selected by the stockholders be presented dfferent results for ther needs. In ths paper, we compare the accuracy rate of the eght experments. The expermental results can help us to make decson whether we choose the target stocks or not. Besdes, by analyzng the accuracy rate, we can evaluate the effectveness of the prmary ndces we selected. If the ndces are not useful to mprove the accuracy rate of predcton, we can reevaluate or regulate other fnancal ndces from the fnancal statement untl we fnd the best combnaton for the predcton of stocks. In ths way, for dfferent stockholders, they can choose ther favorte fnancal ndces to satsfy ther dfferent predctons. The theoretcal model s shown n Fgure. Fgure : Theoretcal model. Researchng Methodology As mentoned above, we apply two methods to accomplsh the experment. The frst method s mplemented as follows. For the 25 samples of stocks, we dvde them to two parts. Every part has the 50% of the samples. Besdes, the way we dvde the samples at random. The way of ths method s cross-valdaton (Pandya, 995). Frstly, we choose the frst part as the tranng set, and the second part s the testng set. We produce a model by the tranng test, and then we predct the target value n the testng set. The expermental result s Accuracy_AB. On the contrary, n the second tme, we choose the second part of the expermental samples as tranng set, the frst part of the samples Communcatons of the IIMA Volume 7 Issue4

7 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks s testng set. Smlarly, we can get the expermental results as Accuracy_BA. After that, we use the k-mean method to get the fnal accuracy rate whch s called accuracy_. Ths method s shown n Fgure 2. The advantage of ths method s that we can avod the devaton of accuracy rate from only dong the experment once, the expermental results wll be more practcal and dependable. Fgure 2: Mehod_. The dfferent of the second method and the frst one s the proporton of the two parts of samples. In the second method, we dvde the sample as the eghty-twenty rules. It means, n the frst part whch s called part A s contented 80% of the samples. The second part whch s called part B s contended 20% of the samples. We choose the part A as the tranng set, and choose the part B as the testng set. Also we dvde the samples at random. In order to compare wth the former method, we do the experment by ths way twce. We get the expermental results accuracy and accuracy2. Then, the accuracy_2 by the k-means s the fnal expermental result. Ths method s shown n Fgure 3. Fgure 3: Method_2. EXPERIMENT In ths paper, we appled the SVM to construct the predcton model and select Gaussan radal bass functon (RBF) as the kernel functon. Usng SVM wth fnancal statement analyss as mentoned above, we can calculate all the expermental results shown n Table 3. From these results, we can obvously fnd that after choosng some fnancal ndces as the parameters, the accuracy rate s better than only predctng by the Outstandng Achevement Growth Rate (OAGR) gven by the professonals and experts. Whatever we apply method_ or method_2, the results of experment 2, 5, 8, 7 and 6 are better than experment whch s the basc experment. That s to say, f we choose some fnancal ndces correctly, we wll mprove the accuracy rate of predcton combned wth the predcton results gven by the experts such as the rankng of OAGR. Communcatons of the IIMA Volume 7 Issue4

8 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks Table 3: Expermental results of dfferent composton. Experment SVM OAGR EPS BVPS NPGR Accuracy_ Accuracy_ % 8% % 86% 76.0% 79% 74.9% 85% 80.47% 86% 77.70% 86% 78.5% 85% 78.90% 85% Table 4: Expermental results sorted by accuracy rate. Method The number of the experments (%) Method_ (50%-50%) 2 (84.45) 5 (80.47) 8 (78.90) 7 (78.5) 6 (77.70) (77.69) Method_ (80%-20%) (86) (86) (86) (85) (85) (85) 3 (76.0) (8) 4 (74.9) 3 (79) In addton, we wll fnd that n Table 4, there are some smlartes after we sorted the results based on the accuracy rate. For example, the experment 2 and the experment 5 both have the best results of accuracy rate. That means, we select the Earnngs Per Share (EPS) as the parameter, or we select both Earnngs Per Share (EPS) and Book Value Per Share (BVPS) as the parameters. We can get the better results for our goal of predcton. In the future tasks of predcton, we only choose the Earnngs Per Share (EPS) and Book Value Per Share (BVPS) as the parameters are enough. Meanwhle, we have to acknowledge that the results of experment 3 or experment 4 are somewhat worse than experment. That s to say, t s not so much of only addng the Book Value Per Share (BVPS) or only addng the Net Proft Growth Rate (NPGR) as usng nothng parameters. In that case, unless we have some specal analyss, we had better not use these ndces separately. However, after analyzng the results of experment 5, 6, 7 and 8, we can obvously fnd that the results by combnng those three fnancal ndces n any patterns are better than experment. Consequently, there are some assocaton rules among these fnancal ndces n the fnancal statement, because the fnancal ndces are restrcted relatvely. Communcatons of the IIMA Volume 7 Issue4

9 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks CONCLUSION In ths paper, we proposed a new method by usng SVM wth fnancal statement analyss for predcton of stocks. There has some advantages comparng our method to the common techncal analyses for predcton n stock market. The expermental results show that there s a hgher accuracy rate of predcton than usng SVM to predct only, because the fnancal ndces as the expermental parameters are drectly selected from the fnancal statements whch are released n dfferent perods by the stock companes to publc, and the expermental results are much more relable, nonvolatle and vald. In addton, we can dscover some assocaton rules among the fnancal ndces n the fnancal statement, and evaluate some specal stocks, because the fnancal ndces are restrcted relatvely. It may be useful n the practcal of the advanced analyss of stock predcton. Based on ths method, not only can we mprove the accuracy rate of predcton, but also can we create an orgnal and unversal method for stock predctons, because the fnancal ndces can be selected dfferently to meet the dfferent stockholders expectatons. In that case, the stockholders are able to avod the subjectve dscrepances of predcton and make decsons to buy the target stocks. However, each method has ts own advantages and dsadvantages. In ths paper, we choose the Outstandng Achevement Growth Rate (OAGR) as the basc predctve standard whch s evaluated by the professonals or experts, but there are some lmtatons for the method. Sometmes, we have to consder the data manpulated by people and the useless nformaton from the professonal analyzers own perspectves. Besdes, the ranges of the expermental samples are not large enough. In that case, we can not descrbe the overall status of the stock market so that the accuracy rate s gong to be mproved n further experments. For the followng research, we wll stll focus on the stock market. After comparng the expermental results wth the practcal stuaton, we wll do some advanced studes based on ths method. Communcatons of the IIMA Volume 7 Issue4

10 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks REFERENCES Burges, C. J. C., (998). A tutoral on support vector machnes for pattern recognton. Data Mnng and Knowledge Dscovery 2, Chang, C. C., & Ln, C. J. (2007). LIBSVM: a Lbrary for Support Vector Machnes. Avalable at ntu.edu.tw/ cjln/lbsvm. Chen, W. H., & Shh, J. Y., (2006). A study of Tawan s ssuer credt ratng systems usng support vector machnes. Expert Systems wth Applcatons, 30, Huang, C. L., & Tsa, C. Y., (2006). Usng SOM-SVR wth Flter Feature Selecton for Predcton of Tawan Stock Index Future, CSIM IMP 2006 Tawan. Huang, W., Nakamor, Y., & Wang, S. Y., (2005). Forecastng stock market movement drecton wth support vector machne. Computers & Operatons Research 32, Hsu, C. W., Chang, C. C., & Ln, C. J., (2003). A Practcal Gude to Support Vector Classfcaton. Avalable at Keerth, S. S., & Ln, C. J., (2003). Asymptotc Behavors of Support Vector machne wth Gaussan Kernerl. Neural Computaton, 5, Km, K. J. (2003). Fnancal tme seres forecastng usng support vector machnes. Neurocomputng, 55, Avalable at Müller, K. R., Smola, S. A., Rätsch, G., Schölkopf, B., Kohlmorgen, J., & Vapnk, V., (997). Predctng tme seres wth support vector machnes, n: W. Gerstner, A. Germond, M. Hasler, J.-D. Ncoud (Eds.), Proceedngs of ICANN 97, Sprng LNCS327, Berln, Pandya, A. S., & Macy, R. B., (995). Pattern Recognton wth Neural Networks n C++. IEEE Press. Smola, A. J., & Schölkopf, B., (998). A tutoral on support vector regresson. NeuroCLOT Techncal Report NC-TR-98-03, Royal Holloway College, Unversty of London, UK. Avalable at Thssen, U., Brakel, R. V., Wejer, A. P., Melssen, W. J., & Buydens, L.M.C., (2003). Usng support vector machnes for tme seres predcton. Chemometrcs and Intellgent Laboratory Systems, 69, Communcatons of the IIMA Volume 7 Issue4

11 Usng SVM wth Fnancal Statement Analyss for Predcton of Stocks Communcatons of the IIMA Volume 7 Issue4