Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 31 (2014) 406 – 412

2nd International Conference on Information Technology and Quantitative Management, ITQM

2014

A SVM Stock Selection Model within PCA

Huanhuan Yua, Rongda Chenb,* , Guoping Zhangc
a
School of Finance, Zhejiang University of Finance & Economics, Hangzhou, 310018, China
b
School of Finance, Zhejiang University of Finance & Economics, Hangzhou, 310018, China
c
School of Economics and International Trade, Zhejiang University of Finance & Economics, Hangzhou, 310018, China

Abstract

In the financial market, well-performing stocks usually have some specific features in financial figures. This paper introduces a
machine learning method of support vector machine to construct a stock selection model, which can do the nonlinear
classification of stocks. However, the accuracy of SVM classification is very sensitive to the quality of training set. To avoid the
direct use of complicated and highly dimensional financial ratios, we bring the principal component analysis (PCA) into SVM
model to extract the low-dimensional and efficient feature information, which improves the training accuracy and efficiency as
well as preserve the features of initial data. As empirical results show, based on support vector machine, within PCA after norm-
standardization, the stock selection model achieves the entire accuracy of 75.4464% in training set and of 61.7925% in test set.
Further, the PCA-SVM stock selection model contributes the annual earnings of stock portfolio to outperforming those of A-
share index of Shanghai Stock Exchange, significantly.

© 2014 Published by Elsevier B.V. Open access under CC BY-NC-ND license.

© 2014 The Authors. Published by Elsevier B.V.
Selection and peer-review under responsibility of the Organizing Committee of ITQM 2014.
Selection and peer-review under responsibility of the Organizing Committee of ITQM 2014.

Keywords: machine learning; stock selection; principal components analysis; support vector machine

1. Introduction

Stock has always been one of the most popular investment instruments in financial markets. Investors and
researchers are devoting themselves to study out a method that can select accurately the stocks with favorable future

* Corresponding author. Tel.: +860571-85750010; fax: +860571-85212001.

E-mail address: [email protected].

1877-0509 © 2014 Published by Elsevier B.V. Open access under CC BY-NC-ND license.
Selection and peer-review under responsibility of the Organizing Committee of ITQM 2014.
doi:10.1016/j.procs.2014.05.284
 Huanhuan Yu et al. / Procedia Computer Science 31 (2014) 406 – 412 407

return to be constituents of investment portfolio. Guo and Zhang1, Kuo et al.2 and Tsumato et al.3 develops several
method to forecast stock prices or pick qualified ones from large sample. However, some traditional stock selection
models usually face challenges when dealing with highly dimensional and nonlinear sample data for the reason that
stock selection is a kind of determination with multi objectives and multi restrictions, along with the highly
dimensional and huge financial data. The machine learning-based theory, Artificial Neural Network (ANN), can
capture the regular patterns hidden behind the complex and high-dimension data through its machine learning 4,5.
Although ANN performs better than traditional methods, it has lots of defects at the same time, such as the difficulty
to determine network structures, the problem with local minimum points and the over-fitting. Vapnik 6 proposed a
new machine learning-based method called Support Vector Machine (SVM), which can better handle the high-
dimension data avoiding the defects of ANN. SVM applies widely in many fields because of its particular
advantages. A lot of researches, domestic and abroad, use SVM to predict stock prices or reversal points, as in Yeh
et al.7 and Huang8. But it’s seldom to establish a stock selection model by SVM, and specifically rare in domestic.
This paper applies SVM into domestic stock market to establish an effective selection model. We treat financial
ratios of listed companies in A-share of Shanghai Exchange as original data, and then use the principal components
analysis (PCA) to preprocess them. First, we established a stock selection model (PCA-SVM) that recognizes high-
return stocks when utilized SVM theory to train the training set. Second, apply PCA-SVM on test set to forecast the
high-return stocks in the next year and do a comparison between the forecast and the actual to illustrate effectiveness
of the established stock selection model.

2. Principal components analysis (PCA)

Financial ratios of a listed company include earning ability, growth ability, solvency ability and so on. Each
ability contains many sub-ratios. If all the ratios were used as inputs in the training set, it would result in redundancy
and low efficiency; even decrease the quality of empirical results. New variables can be created through
transformation of original variables. Number of variables is less and most information is still retained. These new
variables are called principal components.

2.1. Definition of principal components

Principal components can be expressed as follows:

­Y T
D1 ˜ X D11 X 1  D12 X 2   D1n X n
° 1
°
°Y2
T
D ˜X D 21 X 1  D 22 X 2   D 2n X n
®
2
, (1)
°
° T
°̄Yn Dn ˜ X D n1 X 1  D n 2 X 2   D nn X n

where X i is the original variable, Yi is the principal component and D i is the coefficient vector respectively.
T
Di can be estimated by maximizing Var (Yi ) with the constraint conditions of D i ˜ D i 1 and

¦
T
Cov(Yi , Yj ) D ˜ ¦˜D i
i 0, j 1, 2, , i  1 , where (V ij )nun is the covariance matrix of X .

2.2. Selection of principal components

The covariance matrix of X ( X1 , X 2 , , X n )T , ¦ (V ij )nun , is a symmetric non-negative definite matrix.

Therefore it has n characteristic roots O1 , O2 , , On , and n characteristic vectors.
Suppose O1 t O2 t t On t 0 and the orthogonal unit eigenvectors are e1 , e2 , , en . The i th principal component
of X1 , X 2 , , X n can be expressed as follows:
408 Huanhuan Yu et al. / Procedia Computer Science 31 (2014) 406 – 412

Yi ei1 X1  ei 2 X 2   ein X n , i 1, 2, ,n (2)

ei ˜ ¦ ˜ ei Oi and Cov(Yi , Yj ) ei ˜ ¦ ˜ ei
T T
with Var (Yi ) 0, i z j . The first p principal components’
accumulated conribution rate is

p n
ACR( p) ¦O / ¦O
i 1
i
i 1
i (3)

which represents the explanation power for original data of the principal components extracted by PCA method.
Generally, an ACR of 85% is at least required, or the PCA method would be thought as unsuitable for losing too
much original information.
Since the covariance matrix is sensitive to the order of magnitudes of data, we need to standardize the data first.
There are two method of standardization in common use:
• Norm-standardization: X ij * ( X ij  X j ) / s j , X j is the mean and s j is the standard deviation.
• Mean-standardization: X ij * X ij / X j , X j is the mean.

3. Support vector machine

3.1. Linear classification of SVM

Linear classification of SVM is realized through solving for the optimal separating hyper-plane when the training
set is linear separable. If the mingled two classes ( C1 , C2 ) of a sample can be separated correctly with the linear
function ( H 0 ) in a two-dimension plane, this sample is treated as linear separable.
Suppose the training set is {( x1 , y1 ),( x2 , y2 ), ,( xn , yn )} , where xi is sample information vector ( xi is the
coordinate vector in a two-dimension plane), yi Y {1, 1} and +1 represents class C1 , -1 represents class C2 . If
T
the linear separating hyper-plane H 0 : w ˜ x  b 0 separates the training set correctly, it is equivalent with the
T T
situation: when yi 1 , w ˜ xi  b t 1 ; when yi 1 , w ˜ xi  b d 1 . If the distance of two data cluster of the
sample, D* , is maximized, this hyper-plane is called the optimal separating hyper-plane in this classification case.
Define D* d  d ,

T
dr min{ w ˜ xi  b w} (4)
i , y r1

T
By substituting w ˜ x  b r1 in (4), we can obtain D* d  d 2 || w || and the problem is transformed to get
the w minimizing || w || . ( b can be calculated by substituting sample points with w known)

Additionally, to avoid the situation that distance between the two parallel hyper-planes is maximized while
effective classification is not realized, we must pose constraints on this optimization problem as follows:

T
yi (w ˜ xi  b) t 1  [i , 0 d [i d 1 . (5)

[ i is the slack variable to tolerate the outliers. And a penalty factor C is also introduced into the objective
function to reflect losses for tolerating the outliers. Training a SVM model, i.e. solving the optimization problem,
will lead to a quadratic programming problem, as shown in (6).
 Huanhuan Yu et al. / Procedia Computer Science 31 (2014) 406 – 412 409

­ n
1 n n
°max ¦O i  ¦¦ OiO j yi y j  xi , x j !
2i 1 j 1
°
°
i 1

® s.t. 0 d Oi d C , i 1, 2, ,n (6)
° n
°
°̄
¦O y
i 1
i i 0

*T
¦O
*
Suppose O * is the solution of (6) and thus the optimal hyper-plane is w ˜ x  b* 0 , where w i
*
yi xi and
*
b can be calculated by the contraints of (5)..

3.2. Nonlinear classification of SVM

Linear classification of SVM we talked about in the prior section can be only applied when sample is linear
separable. In this section, an improved nonlinear SVM method is proposed to solve the complicated and high-
dimensional financial ratios.
A kernel function M is very important here because it can map the original date into high-dimensional space H ,
i.e. M : R o H ; x o M ( x) , which can let the data can be linear separable in H . Then an optimal separating hyper-
n

plane discussed in prior section can be obtained to do the classification.

Suppose the training set is {( x1, y1 ),( x2 , y2 ), ,( xn , yn )} , xi is the highly dimensional information vector of the
sample and yi Y {1, 1} . A quadratic programming similar with (8) is obtained through mapping M :

­ n
1 n n

°max ¦ O  2 ¦¦ O O y y
i i j i j  M ( xi ),M ( x j ) !
°
°
i 1 i 1 j 1

® s.t. 0 d Oi d C , i 1, 2, ,n (7)
° n
°
°̄
¦O y
i 1
i i 0

To solve (7), M : R o H ; x o M ( x) is needed to know, so we choose Gauss radial based kernel function (RBF)
n

to get the inner product value as k ( x, y)  M ( x), M ( y) ! directly without searching for the complex M .

4. Data selection

Table 1. Financial ratios and sample stocks information

Sample stock Earnings ability A Activity ratio B Shareholder return C
Turnover of accounts EPS c1
2009, 677 stocks EBIT a1
receivable b1 Price-to-book ratio c2
ROA a2
Turnover of inventory b2 Common stock profitability c3
2010, 679 stocks ROE a3
Turnover of current assets b3 P/CF c4
Cash ratios D Growth ratios E Risk level F Solvency ratios G
Quick ratio g1
EBIT-to-Cash ratio d1 Financial leverage f1
Debt-to-Asset ratio g 2
Cash-to-Assets ratio d 2 Growth of total assets e
EBIT/Interest ratio g 3
Operating ratio d 3 Operating leverage f 2
EBIT/Fixed charge ratio g 4

This paper selects 7 categories of financial ratios of companies in A-share Shanghai Stock Exchange from their
annual reports of 2009 and 2010. The detailed financial indexes chosen are shown in Table 1. Our objective is to
410 Huanhuan Yu et al. / Procedia Computer Science 31 (2014) 406 – 412

separate the high-return stocks from the low ones according to their features hidden inside the financial ratios, thus it
is necessary to label each stock with the return characteristic. After statistical analysis, all the companies have
announced their annual report before 1th/May in 2009 and 2010. Therefore we label the stock as +1 if its return
ranks the first 25% of all the sample stocks, i.e. yi 1 and yi 1 for the rest stocks. Labels of a part of sample are
presented in Table 2.

5. Stock selection of model and analysis

5.1. Extraction of training set based on PCA method

Financial ratios of 677 stocks in 2009 are the original data. We apply PCA to extract the principal components
satisfying the condition of ACR t 85% . Since our sample is large, if we apply PCA on all of the ratios of 677 stocks
directly, we would lose the local information and the effect of dimension reduction is also smaller. Thus we do PCA
extraction one time for every 40 sample stocks. The training set is in Table 2.

Table 2. Training set of SVM nonlinear classification (part of 677 stocks)

Earnings Activity Shareholder Growth Solvency y

Stock code Cash ratios Risk levels
ability ratios return ratios ratios

PCA with norm-standardization

600069 -1.6114 -0.9830 -0.4337 -1.0664 -0.4253 0.7874 0.1431 1
600070 0.5249 -0.3005 -0.8563 -0.5438 -0.0903 -0.1103 0.0136 -1
600071 2.1843 0.1875 -1.5191 1.1364 -0.6570 -1.7170 0.7624 1
PCA with mean-standardization
600069 0.8222 -1.3006 0.8049 1.0620 -0.9571 0.3681 1.8768 1
600070 4.6133 1.0647 -0.3712 -1.1497 0.8309 1.6046 1.5020 -1
600071 7.0948 1.1286 -0.7982 0.2286 0.2485 -0.2133 2.0515 1

5.2. SVM stock selection model and analysis

The total scores obtained in the prior section combined with return labels of sample stocks constitute the
complete training set of SVM. By applying the nonlinear classification of SVM introduced in section 3 on the
training set, we can obtain the optimal separating hyper-plane. If we use this hyper-plane on test set, stocks in test
set can be classified into the high-return part and the low-return part. It can be seen as a prediction of stocks’ future
return characteristic. The accuracy of classification and prediction is presented in Table 3.

Table 3. Accuracy of SVM nonlinear classification

Method used Mean-standardization PCA-SVM Norm-standardization PCA-SVM

Whole accuracy a 88.6905% 75.4464%

Training Accuracy of +1 a 100% 58.5366%
Accuracy of -1 a 85.0394% 80.9055%
Whole accuracy b 69.1943% 61.7925%
Test Accuracy of +1 b 10.1266% 24.5283%
Accuracy of -1 b 88.8421% 74.2138%

Training and testing of SVM proceed with Livsvm 3.1 in Matlab. To achieve the best generalization ability, the
optimal penalty factor C and the coefficient V in RBF is determined by Grid Searching method.
 Huanhuan Yu et al. / Procedia Computer Science 31 (2014) 406 – 412 411

By observing Table 3, we can find that the accuracy of mean-standardization PCA-SVM for label +1 in training
set is 100%. However, the accuracy of the same label in test set is only 10.1266%. It is the over-fitting phenomenon
that too many support vectors were used to explain the training set, which could has a good classification effect on
training set while a bad effect on predictions. The accuracy of norm-standardization PCA-SVM is obviously better.
For further analysis, we construct an equal weighted portfolios with stocks selected by PCA-SVM and do a
comparison between the accumulated return (ACR) gained by this model and the A-share index of Shanghai Stock
Exchange. The comparison is presented in Fig.1. It manifests that PCA-SVM has higher accumulated return over the
A-share index, which means SVM classification method is accurate and highly efficient when dealing with complex
and highly dimensional data.

Fig.1. Comparison between PCA-SVM and A-share index of

Shanghai Stock Exchange

6. Conclusions

Support Vector Machine is commonly used to train the time-series data of stocks for price forecasting. In this
paper, SVM is employed to generate an optimal separating hyper-plane in high-dimensional space based on the
training set. To increase the accuracy and efficiency of SVM classification model, we apply PCA to process the
original data. Finally, the empirical result has suggested that the return of stocks selected by PCA-SVM is
apparently superior to A-share index.
Information features of financial ratios of companies vary with their industries. We believe that the quality of
training set can be improved if we apply PCA on each industry separately. Additionally, it is quite meaningful for
achieving higher returns if stocks could have different weights according to their risk-return characteristics when
portfolios are constructed.

Acknowledgments

This research was supported by the National Natural Science Foundation of China (Grant No. 71171176).

References

1. Ming Guo, Yuan-Biao Zhang. A Stock Selection Model Based on Analytic Hierarchy Process. Factor Analysis and TOPSIS//The International
Conference on Computer and Communication Technologies in Agriculture Engineerin. 2010. p. 466-469.
2. Kuo R.J., Chen C.H.& Hwang Y.C. A Intelligent Stock Trading Decision Support System Through Integration of Genetic Algorithm based
Fuzzy Neural Network and Artificial Neural Network. Fuzzy Sets and Systems. 2001; 118: 21-45.
3. Tsumato S., Slowinski S., Komorowsk J. & Grzymala-Busse J.W. Lecturenotes in Artificial Intelligence. The fourth international conference
on rough sets and current trends in computing. 2004.
412 Huanhuan Yu et al. / Procedia Computer Science 31 (2014) 406 – 412

4. E.L. de Faria, Marcelo P. Albuquerque, J.L. Gonzalez, J.T.P. Cavalcante, Marcio P. Albuquerque. Predicting the Brazilian Stock Market
Through Neural Networks and Adaptive Exponential Smoothing Methods. Expert Systems with Application. 2009; 36:12506-12509.
5. Yudong Zhang, Lenan Wu. Stock Market Prediction of S&P 500 via Combination of Improved BCO Approach and BP Neural Network.
Expert Systems with Applications. 2009; 36: 8849-8854.
6. Vladimir N. Vapnik. Statistical Learning Theory. Publishing House of Electronics Industry. 2004.
7. Chi-Yuan Yeh, Chi-Wei Huang, Shie-Jue Lee. A multiple-kernel support vector regression approach for stock market price forecasting. Expert
Systems with Applications.2011; 38: 2177-2186.
8. Pengpeng Huang. Prediction of the Turnover Points in Stock Trend Based on Support Vector Machine. College of Software, Fudan University.
2010.