Proceedings of International Joint Conference on Neural Networks, Atlanta, Georgia, USA, June 14-19, 2009

Intelligent Stock Trading System based on SVM Algorithm and

Oscillation Box Prediction
Qinghua Wen, Zehong Yang, Yixu Song, Peifa Jia

Abstract-The stock market is considered as a high complex accumulated in stock investment which we believed are
and dynamic system. Many machine learning and data mining critical to design a workable stock expert system.
technologies are used for stock analysis, but it still leaves an In our previous work [7], an expert system which learn
open question about how to integrate these methods with the trading strategy by markov network probabilistic model from
plentiful knowledge and techniques accumulated in stock high-level representation of time series, i.e. turning points
investment which are critical to the successful stock analysis. In and technical indicators was developed, which is based on the
this paper, we propose an intelligent stock trading system by
idea of price trend turning point in technical analysis. In this
combining support vector machine (SVM) algorithm and box
theory of stock. The box theory believes a successful stock paper, the objective is to exploit the common used box theory
buying/selling generally occurs when the price effectivley of stock and develop an intelligent stock trading system based
breaks out the original oscillation box into another new box. In on oscillation box prediction.
the system, support vector machine algorithm is utilized to There are two modules in our trading system: oscillation
make forecasts of the top and bottom of the oscillation box. box prediction module and trading strategy module. The
Then a trading strategy based on the box theory is constructed support vector machines algorithm is utilized to make
to make trading decisions. The different stock movement forecasts of the top and bottom of the oscillation box. Then
patterns, i.e, bull, bear and fluctuant market, are used to test the trading strategy based on the box theory is constructed to
feasibility of the system. The experiments on S&P500
make trading decisions. In the experiments, we investigated
components show a promising performance is achieved.
the performance of the supposed system on individual stocks
with different movement patterns, test the average rate of
I. INTRODUCTION
profit of nearly all the stocks in S&P 500 components and

T he study of stock market is a hot topic, because if

successful, the result will transfer to fruitful rewards. A
lot of significant work has been done in this field. Some
compared the performance with the buy-and-hold strategy
and other trading system. The experiments showed the
proposed system achieved a promising performance in term
of them are to predict the future trend or price by taking the of profitability, in contrast with 6% average rate of return by
stock price as a time series. Artificial neural networks (ANN) buy-and-hold strategy over the testing 400 trading day, our
are claimed particularly well suited for finding accurate system got a nearly 27 percent average rate of return.
solutions in an environment characterized by complex, noisy, The advantage of the system lies in two aspects. First, the
irrelevant or partial information [1]. So many works focus on trading tragedy is more reasonable which is based on the box
applying different neural networks into stock prediction [2-4, theory but overcomes the problem of poor performance of
14]. Other methods include support vector machine (SVM) [5, box theory in bearish market. Second, instead of trying to
6], hidden markov model (HMM) [13], reinforcement predict the closing price as other literatures did, the system
learning algorithm [15]. forecasts the synthetic time series, i.e. the upper bound and
Another high-level research is to develop decision support the lower bound of the price oscillation box which suffer less
system which helps to make trading decision. A typical work from noises than the rough closing price series. This makes
is [8] in which a decision support system was proposed by the forecasts easier and more accurate.
utilizing the neural network to make a forecast of the TOPIX The rest of the paper is organized as follows: Section 2
in the future and then using the genetic algorithm to find an describes basic design of the stock trading system and the
effective way to deal. In more recent work [9], based on the support vector machine estimators. Section 3 shows the
mechanisms of group decision making and cooperative experiments and the corresponding analysis. Finally some
learning, the proposed system combined of the Wilson's concluding remarks are drawn in section 4.
extended classifier system and neural network modules, the
investment strategy was learned from the cooperative II. INTELLIGENT STOCK TRADING SYSTEM
learning of 50 agents with different input features and
strategy. The problem of all of these methods was they didn't A. Design ofthe Stock Trading System
use or used little the practical knowledge and techniques Due to continuous changes of strength between buying
side and selling side, the stock price time series moves
generally in a certain range in a period of time. The running
Qinghua Wen, is with the Tsinghua University, Beijing, 100084 China range is called price oscillation box in box theory. If the
(corresponding author to provide phone: 086-10-62777703; e-mail: strength of buying or selling beat that of the another, the price
wenqh06@ mai1s.tsinghua.edu.cn).
will effectively break out the upper bound or the lower bound

978-1-4244-3553-1/09/$25.00 ©2009 IEEE 3341

of the original oscillation box, then enter another price If next trade = buy
oscillation box and build the new stable running range. The
If
IC.-i·1 ~ a -
and L. is in uptrend
time of the breakout is supposed to be the best time to buy or I I
C, I
sell the stock [16]. The critical problem is how to identify the
price box and how to confirm the breakout is effective. If sellprice - C, ~ tp
1) Oscillation Box Prediction then Buy, buyprice == C,
In our system, we take the highest value Hi and lowest
next trade = sell
value L, of the closing price C n-day in the future from the
ith day as the estimation of the two bounds in the ith day.
else If
IC.-H·I ~ a
I I
-
and H. is in downtrend
C, I

Hi == max(Ci,Ci+l,···,Ci+n_l) If C, - buyprice ~ rjJ

(1)
L, == mineCi' Ci+ 1 , · •• , Ci+n-1 ) then Sell, sellprice == C,
next trade =buy
Then we obtain two time series H, L corresponding to the
rough price series C . The reason of the definition is twofold. If buyprice - Ci ~ ()

In the first, it can transform the estimation of the price box buyprice
into time series regression problem when maintains the basic then Sell, sellprice == C,
meaning of price range, then we can employ the data mining next trade =buy
method to solve it. Second, by the definition, it facilitates the where (J" is the trading rate which is proportional with the
construction of trading strategy. Forecasting H, L is equal to accuracy of SVMmax and SVMmin. When the accuracy is
make a regression for the function which is mined from the high, (J" could be smaller, and vice verse. The role of (J" is to
historical information. regulate the trade frequency. The smaller the value, the fewer
the number of trading. tp, ¢ are used to filter out the false
operation to make sure the trading is more profitable. () is
stop-loss rate to minimize the loss when make a wrong
where Hi' L, represent the forecasts of the Hi' L i, X, yare buying operation. It is also especially useful in bearish market.
the factors related to the change of Hi' L, . The constraints of the trend of Hi' i, are in order to make sure
The forecast module employs support vector machine the true trend reverse happened. The idea of the trading
algorithm. Two estimators based on SVM algorithm called strategy can be illustrated in Fig.I.
SVMmax and SVMmin are used to estimate Hi' L, One of the major criticisms of the original box theory is
respectively. We train the two forecast modules by the sliding
window method. 32,------.-----.--------,--------.--------.-------.--------.--------.--------.-------,

2) Trading Strategy ~ oscillation box

30
By the definition of H, L , we take the relationship /'-------1 ------,
between C, and Hi' Li as the basic design of the trading 28
rJ I
strategy. If the current price effective breaks out the upper r-/-.-.--/=I _.../_O~II~O~O.:..---------1 •
bound of original oscillation box, it means the price will start 26
/' I . I \.1 _
I I I
an uptrend and start to form another price box. The original I I
resistance line, i.e. top of the original box will transfer to an 24
I I
important support line, i.e. bottom of the new box. That is to \ 1 . ~ J
22 - c l o s e price
say the current price will be the valuable estimation of bottom - - maximum of future 3D-day close price
-------minimum of future 3D-day close price
of the new oscillation box. According to the definition of
10 20 30 40 50 60 70 80 90 100
Hi .L, ' the current closing price C, will be close to L, at this
time. The same is true for lower bound breakout. But it have Fig. 1. ' t 'means buy, ' + ' means sell, or else do nothing.
been noted not all C, close to the two bounds corresponds to
a box breakout. In fact, at every time the price oscillates to the many studies state it can only work in strong bull markets and
box top and low, regardless breakout following or not, C, is that if the market is bearish or trading sideways this method
near to Hi or L, . Fortunately, even with no breakout doesn't work. In this study, we introduce stop-loss rate and
short-term trading when the price moves within box, the loss
happening, it is also a good time for short-time trading to win
can be minimized. The experiments in Section 4 shows the
the price difference in that the current price is in local extreme
system performs well even in bearish market.
value and we can add constraints to regulate the frequency of
short-time trading. Based on the relationship between C, and B. SVMmax and SVMmin
Hi,Li, we define the trading strategy as follows. In previous work, J. Cao and Francis E. H. Tay [5] show
that support vector machines method is significantly better
than other neural network in financial time series forecasting.

3342
Y.K. Bao et al [6] conclude the support vector machines is a Where C >0 is a prescribed parameter to determine the
robust technique for stock index regression. In our system, trade-off between the flatness of f (x) and the amount up to
SVM algorithm is utilizing to build the two estimators which deviations larger than e are tolerated. To solve this
SVMmax and SVMmin to forecast the top and low of the box optimization problem, we construct a Lagrange function and
respectively. introducing a dual set of Lagrange multipliers
1) Overview ofSVM algorithm Ai,Ai* and 17i ,17i*
The SVM algorithm is based on statistical learning theory 1 11 11
T
and structural risk minimization principle [10]. It was first L = "2WT w+ c~ (J; + J;') - ~ A;(&+ J; - y; + w cI>(xJ + b)
developed by Vapnik and his co-workers in the early 1990s to
solve the classification problem. With the introduction of loss
function such as Vapnik's e -insensitive loss function,
-'i=lI. Ai'(& + s; - i + w cI>(xJ + b) - 'i=lI. (17;8; + 17i' J i' )
Y
T
(7)

Huber's loss function, SVM have been extended to solve where Ai' Ai* ,17i ,17i* ~ O. This function has a saddle point
regression estimation problems and applied successfully in which corresponds to the solution of the optimization
time series forecasting, non-linear modeling and optimal problem. Solving (4) we get the optimal w,b andf(x):
control problem[ 11]. The basic idea of support vector
regression is to map the input space into a high-dimension w' = 'i=lI. (A; -A;')cI>(XJ
space by a non-linear mapping which is accomplished (8)
implicitly by the trick of kernel function and then to do a * 1~ *T (*)
b ==-L...J(Yi-W cD (xi)=t&), O<Ai <C
linear regression in the new feature space. n i=l
Given a time series samples {Xi,Y i} Xi E RI'Yi E R,
i == 1,2, .. " n where Xi is the input feature variable, Yi is the f(x) = 'i=lI. (A; -A;')cI>(XJcI>(x )+b' (9)
target value. SVM regression algorithm first maps the data to
The Lagrange multipliers Ai' Ai* in (8) are sparse. Only when
a high-dimension feature space r using a Xi is outside or on the & -insensitive tube, they are non-zero.
mapping cD: R l ~ r . Then in the high-dimension, we make a
These points are called support vector. The training points in
regression to form a linear function
the tube are useless for f (x) . The dot products in f (x) is
computed in the new high dimensional space which is usually
f(x, w) == w T cD (x) + b (3) intractable. This problem is tackled by substituting the dot
products with a kernel function which satisfies Mercer's
in the condition of minimizing the sum of empirical risk and conditions. Any symmetric kernel function satisfying
the complexity term w 2 which enforce flatness in feature
II 11 Mercer's condition corresponds to a dot product in some
space. Where WE T is the weight vector, b is a bias. The feature space [11], then the computation of dot products in the
linear function in the high dimensional feature space high dimensional transforms to the simpler computation in
corresponds to the non-linear function in the original lower the original low feature space.
dimensional feature space [10]. In e - SV regression [11], 2) Parameter selection
our goal is to find a function f (x) that has at most e In our forecast module, the radial basis function (RBF) is
deviation from the targets Yi for all the training samples, at used as the kernel function of the two SVM estimators. The
the same time is as flat as possible. That is to find (3) under dynamics of stock price series are nonlinear, so it is
the optimization problem: intuitively believed that using nonlinear kernel functions
. .. 1 T could achieve better performance. Another advantage ofRBF
mInImIze - w w
2 kernel is it tends to give good performance under general
(4) smoothness assumptions and especially useful when no
. {Yi -wTcD(xi)-b ~ s
subject to additional knowledge of the data is available [5]. It defines as
T
w cD (Xi ) + b - Yi ~ s (10).
2
Considering the existence of data outside the & -insensitive II Xi - X j 11 )
tube K (Xl' , Xl' ) == exp( 2 (10)
2g
In SVM algorithm, the selection of the user-prescribed
IY - T
( w cD (x) + b) I~ s (5)
free parameters C,g plays an important role on its
performance. C is referred to as the regularized constant to
we introduce slack parameter 6i , 6i* to cope with the balance the empirical error tolerated and the regularization
otherwise infeasible constraints of the optimization problem term to control the capacity of the estimated function, they
(4), s; 6i* are defined as together make up the risk function of SVM. g is the
. . . -1 wT w + C~
mmumze L...J (vi
s: s" )
+ Vi bandwidth of the RBF kernel. Improper selections can cause
2 i=l the problem of overfitting or underfitting.
(6) The parameter C, g are optimized by cross-validation and
. {Yi -wTcD(xi)-b < &+6i
subject to grid search. The training set is divided into five folds. One
wTcD(xi)+b- Yi < &+6i* fold was taken as the validation set, others taken for training.

3343
The grid point with the best accuracy of predicting is used as investigated . All of the experiments run in the MATLAB
the value of the two parameters. environment.
3) Inputfeature
A. Typical stock movement and their trades
Due to the estimated values related to the next n-day prices,
so intuitively, the input data should at least contain the Generally, the same trading strategy may have a different
information of n-day before. n is set to 30 according to performance in different stock movements. In this section, we
experimental result as section 3.2. For SVMmax (SVMmin), test the trading system on different stock movements to
the input data selected in the system totally include 8*n examine the profitability in different environment. They are
features: fluctuant and bull market, fluctuant and bear market, overall
bull market, and overall bearish market.
c; MAk , sst.,ROCk' FastKk, SlowKk , SlowDk, Fig.2 shows a stock movement in a fluctuant and bull
H k_n(Lk_J k = i,i-l, . .. i-n market and long period trade (almost 4 years). Testing data
where MA k .ssr, ,ROCk , FastK k , Slo wKk , S lo wD k are set are sampled from March 27, 1990 to March 7, 1994 of
technical indicators computed from closing prices C . The ALCOA INC. The training set is constructed as section
computation of these arguments can be found in [12]. ILB.4). With trading rate (J set to 0.05 and stop-loss rate ()
H k -n' L k _ n are the historical values of the estimated set to 14% and rp,¢ set to 0, 5 respectively, our trading
parameters. All of the input data is scaling to [-1, 1]. system is able to profit up to 121.63% while the market gains
4) Training about 15.3%. The MSE and SCC for SVMmax are 0.00661
The sliding window validation is a train and test technique and 0.7109 while for SVMmin they are 0.00842 and 0.6677.
which is much more suitable for time-series data that is slow The selection of parameters a,B,cp,r/J is set manually through
varying or non-stationary . So the sliding window method is experiments .
employed to train the two estimators. We divide the whole Fig.3 shows a stock movement in a fluctuant and bear
data set into overlapping training-test set. The length of market and short period trade (less than 2 years). Testing data
window is 1050 in which the first 1000 data points are taken are sampled from Jun 13, 1970 to Jan 11, 1972 of ALCOA
as the training set, the 50 data points following are taken as INC. With trading rate (J set to 0.12 and stop-loss rate () set
the test set. to 15% and rp,¢ set to 0, 2 respectively, the profit is up to
C. Performance Evaluation 21.04% while the stock loss is about 11.3%. The MSE and
The performance of the stock trading system is evaluated SCC for SVMmax are 0.00890 and 0.8370 while for SVMm in
by the rate of profit which calculated by (11). In the trading they are 0.00842 and 0.6677.
process, suppose $1000 initial fund and trade all stocks at In other word, our system can outperform buy-and-hold
each operation. A 0.5 percent transaction cost for each strategy in the fluctuant market.
transaction is assumed. Movement in Fig. 4 is an overall bull stock movement and
400 days short trade. Test data are sampled from Mar 17,
,r ,r. final return - initial fund 10001
rate oj profit = x / 0 (11) 2004 to Oct 17, 2005 of AES CP INC. With trading rate
initial fund (J set to 0.09 and stop-loss rate 0 set to 15% and rp,¢ set to
The mean squared error (MSE) and squared correlation 0, 2 respectively, the experiment result shows our trading
coefficient (SCC) is used to measure the accuracy of SVM system achieves a perfect excellent trade with profit 139.5%
regression. MSE reflects the local fitness of the regression while the stock gains about 80.6%. The result benefits from
and SCC reflects the global fitness. The higher the values of the good regression of the two SVM estimators. The MSE
SCC, the better global fitness achieved. MSE and SCC and SCC for SVMmax are 0.00662 and 0.8806 while for
defined as follows SVMmin they are 0.00776 and 0.90364.
1~ • 2 Movement in Fig. 5 is an overall bear stock movement and
MSE = - L. (Yi - Yi )
N i =1 400 days short trade. Testing data are sampled from Mar 17,
N

[L(Yi - Yi )(Yi - Yi' )f (12)

SCC= N i =1 n
L(Yi - Yi )2L(Yi' - Yi')2
;=1 i =1

where Y i is the actual output and Yi' is its estimate, )Ii' s; are
their averages respectively.

III. EXPERIM ENTS AND R ESULTS

In order to test the feasibility of the trading system, we

experiment the proposed system in two aspects: performance
on typical stock movements, i.e. bull market, bear market,
fluctuant market and the average performance on stocks of
S&P 500 components . The width of the oscillation box is also Fig. 2. Trading log (' t ': buy, , I ': sell) in a fluctu ant and bull market

3344
 10,--- -,-- - ,...-- --,-- - ..,..-.--.,--,.- - -,-- ---,,--- -,
In order to determine n, we investigate the average rate of

.. t I
r
,
profit by varying n from 3 to 50. The experiment is carried
on 50 stocks (the alphabetically first 50 stocks in S&P 500
components) over 500 trading days from June 15, 2005 to
}
. ! ,f
, , l 4 June 11,2007. The result shows in Table 1.
" l- ~,
.t
J .~ I/ nI'\.t
J ~ v
!J
~ TABLEt

,,
I AVERAGE PERFORMANCE OF TRADI NG 50 STOCKS BY VARYING N
.\ . .. n
Average Average Average see Average see
.~. ' ~J 3
pro fit (%)
35.5
number o f trade
9.8
ofSVMmax
0.96 91
ofSVMmin
0.9593
~ ... 5 39 .49 8.7 0.9452 0.9252
" 8 40.79 8.1 0.9141 0.8816
10 39 .83 7.7 0.8888 0.852 3
Fig. 3. Trading log in a fluctuant and bear market 15 34 .35 7.3 0.8431 0.7862
20 35.8 6 8.4 0.7924 0.7324
30 37.88 8.8 0.70 63 0.6096
40 37. 0 1 7.9 0.623 5 0.5232
50 34.89 7.7 0.5629 0.4 62 8

From the result , There are two peaks of average rate of

profit, n = 8 and 30, in which the best performance achieved.
Further studying the movements of the 50 stocks, the
performance is better with n = 8 when the movement of the
stock price is more fluctuant over the test period, the better
performance is achieved with n = 30 when the stocks change
gently .
C. Trading on S&P 500
1) Experiment setup
Fig . 4. Trading log in overall bull mark et
Most existing trading systems justify their performance on

]
several selected stocks or stock index. However, this method
is unreliable to validate the effectiveness of the system
because of lack of generalization. In the experiments, we
select 442 stocks of the S&P 500 compon ents which have
over 3000 daily data and evaluate the average performance of
., those stocks. The experiment is based on the daily price data
from Mar 18,2004 to Oct 17,2005, namely 400 trading days.
-" ~....I ..... t"
f In this period the S&P 500 index raised 6.57%. For every
'" stock, we implement the following step.
II
•"j In the oscillation box prediction stage, the testing set with
total 400 trading days is divided into 8 subsets . The time
"'0 10 100 110 200 m lOt )SO '00
period of the subsets used in training and testing are listed in
TIftNIIl1Uy)
Table n.
Fig. 5. Tra ding log in overall bear market
TABLE lI
2004 to Oct 17, 2005 of TRIBUNE INC. With trading rate AVERAGE PERFORMANCE OF TRADING 50 STOCKS BY VARYING N

CY set to 0.01 and stop-loss rate 0 set to 15% and rp,¢ set to Training set Testing set
1 3/2212000 - 3/1712004 3/1812004 - 5/27/2004
0, 2 respectively, our trading system losses only 7.6% when 2 6/ 2 /2000 - 5/2 7/2004 5/28/2004 - 8/1 0/2004
the stock lose totally 35.2% in the period, The MSE and see 3 8/ 14/2000 - 8/10/2004 8/1 1/2004 - 10/20 /2004
for SVMmax are 0.00157 and 0.9921 while for SVMmin they 4 10/24 /2000 - 10/20 /2004 10/2 1/2004 - 12/3 1/2004
5 1/ 5 12001 -12/3112004 1/ 3 12005 - 3/15 /2005
are 0.00218 and 0.8336.
6 3/2012001 - 3/1512005 3/1612006 - 5125/2005
B. Width ofthe oscillation box n 7 5/3 112001 - 512512005 5/2612005 - 8/19 /2005
8 8/ 10/200 1 - 8/19 /2005 8/22/2005 - 10/17 /2005
The width of the oscillation box n is an important
parameter of the proposed trading system need to be
The cost C is set to 550 and 350 for SVMmax and SVMmin
determin ed. Intuitively, the n smaller, the higher the
respectively. The parameter g of RBF kernel function is set
prediction accuracy of the two estimators SVMmax and
to 0.00002125 for both estimators . The stop-loss rate () is set
SVMmin may be got, but the trade decision may be less
to 15% and rp,¢ set to 0, 2 respectively.
optimal due to the forecast s concerning less days ahead vice
verse.

3345
 In the trading stage, suppose $1000 initial fund and trade 3) System Comparison
all stocks at each operation. A 0.5 percent transaction cost for Due to lack of benchmark tests for stock trading system,
each transaction is assumed. most of the existing stock trading system experiment on
2) Result Analysis different kind of financial data with different experimental
The evaluation criteria of the experimental result include scheme and the performances vary considerably. This makes
the total number of stocks underperformed the Buy-and-Hold the system comparison difficult.
method over the test period, the number of non-profitable In this part, we compare the performance of our system
stock and the average profit of return over the test period. with trading system which used turning point confirming and
The statistical result shows in Table III. The average profit probabilistic reasoning method [7] and has a similar
of the proposed system is 27.23% while the profit of experimental scheme with us. The trading system first builds
buy-and-hold strategy is 5.77% for the testing 400 trading the turning points and technical indicators according to
days. That means the profit of the trading system is much technical analysis method, then learn trading strategy by
better than the S&P 500 index and outperforms the probabilistic reasoning from the two high-level
buy-and-hold strategy. For convenience of the experiments, it representations of financial series. We utilize the following
is needed to note that the value of user-prescribed free formula to calculate the performance. Natural Gain is the rate
parameters in both SVM estimator are same for all stocks in of profit by Buy-and-Hold method.
training process, which is obviously not optimal for all of . Average rate of profit (13)
ratio == - - - - - - - - -
them. Natural Gain
TABLE III Table V compares the results achieved using various
AVERAGE PERFORMANCE OF TRADING 422 STOCKS FOR400 DAYS investment strategies. Our system based on SVM and box
Market Pattern Bull Bear Total theory achieved a comparatively better performance than
Number of stock 224 198 422 turning point confirming & probabilistic reasoning method.
Less than Buy-and-Ho1d 39 o 39
Hence, in terms of profitability, the system proposed in this
Less % 17.41% 0% 9.24%
Number of loss o 14 14 study can bring investors relatively better rewards.
Loss % 0% 7.07% 3.32%
Average Profit 39.11 % 13.79% 27.23% TABLE V
Average Profit of Buy-and-Ho1d 26.47% -21.04% 5.77% THEPERFORMANCE OF DIFFERENT TRADING SYSTEM
Method Support vector machine Turning point confirming +
+ Box theory Probabilistic reasoning *
In the total 422 stocks, only 14 stocks don't profit due to Data set S&P 500 components S&P 500 components
the overall downtrend during the test period in which there is Testing set 3/18/2004~ 10/171 2005 5/3/2002~ 11/25/2005

little changes to profit. Comparing with buy-and-hold method, Number of

422 454
trading stocks
39 stocks are underperformed due to mainly the overall
Number of
uptrend over the test period in which obviously the trading days
400 900
buy-and-hold is the best trade strategy. Average rate
27.23% 43.6%
Due to the good performance of the box theory in bull of profit
Natural Gain 5.77% 18.15%
market, all the bullish stocks investigated are profited during
Ratio 4.72 2.40
the period of time and the average profit in bull market is as *The experimental result quoted from [7], no transaction cost is assumed
high as 39.11 %. We also can find even in the bear market, the
system receives 13.79% average profit when the IV. CONCLUSION AND FUTURE WORK
buy-and-hold method has an average loss of 21.04%. That
means the system in bearish market still has a good In this paper, we present an implementation of intelligent
performance. stock trading system based on oscillation box prediction by
The rates of profit of some individual stocks list in Table combining support vector machines and box theory of stock
IV. investment. The support vector machine algorithm is used to
forecast the upper and lower bound of the oscillation box
TABLE IV respectively. The trading strategy is based on box theory rule.
THE PERFORMANCE OF 20 ARBITARY SELECTED STOCKS IN S&P 500 In order to examine the feasibility, we test the system on
Name Profit Profit of Name Profit Profit of
typical stock movements and S&P 500 components. The
(%) buy-and-ho1d (%) buy-and-ho1d
(%) (%) experiments show a promising performance of the system and
LU 76.22 -43.33 NVLS 38.76 25.12 it outperforms buy-and-hold strategy.
MOT 22.86 8.64 ABC 17.12 4.94 The prediction algorithms with a single algorithm are often
BMC 17.90 -3.38 BBBY 2.86 -19.75 fragile because of the complexity of the stock movement, and
ESRX 5.07 -39.81 C -16.58 -67.05
ROK 10.92 6.49 RDC 0 -56.95 there is a wide acceptance of benefit of the synergy effect. In
R 15.41 32.71 K 24.05 20.11 the system, the accuracy of the forecasts for the upper and
ETN 42.07 24.35 CMX 31.66 20.49 lower bound of the oscillation box is also the bottleneck of the
PPG 28.31 7.32 AMD 48.54 32.98 performance. From the future research point of view, the
AIG 23.78 23.57 ABS 1.39 -5.43
DOV -16.72 -35.60 MTB 23.76 -2.29
clear extensions to be investigated is to explore more robust
estimators to improve the accuracy of the forecasts by

3346
combining other soft computing techniques together. The
self-learning and self-adjustment of system parameters such
as trading rate and stop-loss rate is also a necessary future
work.

REFERENCES
[1] Grudnitski. G., Osburn, L., "Forecasting S&P and gold futures prices:
an application of neural networks," Journal ofFutures Market, vo1.l3,
pp. 631-643, 1993.
[2] Tenti P., "Forecasting foreign exchange rates using recurrent neural
networks," Applied Artificial Intelligence, vo1.l0, pp. 567-581,1996.
[3] Kuan C.M., Liu T., "Forecasting exchange rates using feedforward and
recurrent neural networks," Journal ofApplied Econometrics, vo1.l 0(4),
pp.347-364, 1995.
[4] Chen A.S., Leung M.T., "Regression neural network for error
correction in foreign exchange forecasting and trading," Computers
and Operations Research, vo1.31, pp.l 049- 1068, 2004.
[5] L. l. Cao, Francis E., H. Tay, "Support Vector Machine With Adaptive
Parameters in Financial Time Series Forecasting," IEEE Trans.on
neural networks, vo1.l4(6), pp.1506-l5l8, Nov. 2003.
[6] Y. Bao, Z. Liu, L. Guo, W. Wang, "Forecasting Stock Composite
Index by Fuzzy Support Vector Machines Regression," in Proc. ofthe
Fourth Int. Con! on Machine Learning and Cybernetics, Guangzhou,
2005, pp.3535- 3540.
[7] D. Bao, Z. Yang, "Intelligent stock trading system by turning point
confirming and probabilistic reasoning," Expert Systems with
Applications, vo1.34 (1), pp.620-627 , 2008
[8] Baba N., Inoue N., Yan Yanjun, "Utilization of Soft Computing
Techniques for Constructing Reliable Decision Support System for
Dealing Stocks," in Proc. of the 2002 Int. Joint Conf. on Neural
Networks, Hawaii, 2002, pp:2l50 - 2155
[9] lung-Bin Li, An-Pin Chen, "Refined Group Learning Based on XCS
and Neural Network in Intelligent Financial Decision Support System,"
in Proc. of the Sixth Int. Conf. on Intelligent Systems Design and
Applications, 2006, vo1. 2, pp:925 - 930
[10] Vapnik V. "The nature of statistical learning theory," New York:
Springer, 1995.
[11] Muller K. R., Smola A., Ratsch G., Scholkopf B., Kohlmorgen,
l.,Vapnik V., "Predicting time series with support vector machines," In:
Proc. ofthe 7th Int. Conf. on Artificial Neural Networks, Berlin, 1997,
pp. 999-1004.
[12] Pring Martin, "Technical analysis explained (4th ed.)," McGraw-Hill
Company, 2002.
[13] Hassan M. R., Nath B., "Stock market forecasting using hidden Markov
model: a new approach," In: Proc. of 5th Int. Con! on Intelligent
Systems Design and Applications, 2005, pp.192-l96.
[14] X. Lin, Z. Yang, Y. Song, "The Application of Echo State Network in
Stock Data Mining," Lecture Notes in Computer Science, vo1. 5012,
pp.932-937, 2008.
[15] J. W. Lee, "Stock price prediction using reinforcement learning," In:
Proc. ofIEEE Int. Symposium on Industrial Electronics, Pusan, South
Korea, 2001, vo1. 1, pp.690-695.
[16] Nicolas Darvas, "How I Made Two Million Dollars in the Stock
Market, " BN Publishing, 2007.

3347