Neurocomputing 55 (2003) 307 – 319 www. elsevier. com/locate/neucom Financial time series forecasting using support vector machines Kyoung-jae Kim? Department of Information Systems, College of Business Administration, Dongguk University, 3-26, Pil-dong, Chung-gu, Seoul 100715, South Korea Received 28 February 2002; accepted 13 March 2003 Abstract Support vector machines (SVMs) are promising methods for the prediction of ynancial timeseries because they use a risk function consisting of the empirical error and a regularized term which is derived from the structural risk minimization principle.
This study applies SVM to predicting the stock price index. In addition, this study examines the feasibility of applying SVM in ynancial forecasting by comparing it with back-propagation neural networks and case-based reasoning. The experimental results show that SVM provides a promising alternative to stock market prediction. c 2003 Elsevier B.
V. All rights reserved. Keywords: Support vector machines; Back-propagation neural networks; Case-based reasoning; Financial time series 1. Introduction Stock market prediction is regarded as a challenging task of ynancial time-series prediction.There have been many studies using artiycial neural networks (ANNs) in this area. A large number of successful applications have shown that ANN can be a very useful tool for time-series modeling and forecasting . The early days of these studies focused on application of ANNs to stock market prediction (for instance [2,6,11,13,19,23]). Recent research tends to hybridize several artiycial intelligence (AI) techniques (for instance [10,22]).
Some researchers tend to include novel factors in the learning process. Kohara et al.  incorporated prior knowledge to improve the ?Tel: +82-2-2260-3324; fax: +82-2-2260-8824.
E-mail address: [email protected] kaist. ac. kr (K.
-j. Kim). 0925-2312/03/$ – see front matter c 2003 Elsevier B. V. All rights reserved. doi:10.
1016/S0925-2312(03)00372-2 308 K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 performance of stock market prediction. Tsaih et al.
 integrated the rule-based technique and ANN to predict the direction of the S& P 500 stock index futures on a daily basis. Quah and Srinivasan  proposed an ANN stock selection system to select stocks that are top performers from the market and to avoid selecting under performers.They concluded that the portfolio of the proposed model outperformed the portfolios of the benchmark model in terms of compounded actual returns overtime. Kim and Han  proposed a genetic algorithms approach to feature discretization and the determination of connection weights for ANN to predict the stock price index. They suggested that their approach reduced the dimensionality of the feature space and enhanced the prediction performance.
Some of these studies, however, showed that ANN had some limitations in learning the patterns because stock market data has tremendous noise and complex dimensionality.ANN often exhibits inconsistent and unpredictable performance on noisy data. However, back-propagation (BP) neural network, the most popular neural network model, su ers from di culty in selecting a large number of controlling parameters which include relevant input variables, hidden layer size, learning rate, momentum term. Recently, a support vector machine (SVM), a novel neural network algorithm, was developed by Vapnik and his colleagues . Many traditional neural network models had implemented the empirical risk minimization principle, SVM implements the structural risk minimization principle.The former seeks to minimize the mis-classiycation error or deviation from correct solution of the training data but the latter searches to minimize an upper bound of generalization error. In addition, the solution of SVM may be global optimum while other neural network models may tend to fall into a local optimal solution. Thus, overytting is unlikely to occur with SVM.
This paper applies SVM to predicting stock price index. In addition, this paper examines the feasibility of applying SVM in ynancial forecasting by comparing it with ANN and case-based reasoning (CBR).This paper consists of yve sections. Section 2 introduces the basic concept of SVM and their applications in ynance. Section 3 proposes a SVM approach to the prediction of stock price index. Section 4 describes research design and experiments.
In Section 4, empirical results are summarized and discussed. Section 5 presents the conclusions and limitations of this study. 2. SVMs and their applications in ynance The following presents some basic concepts of SVM theory as described by prior research. A detailed explanation may be found in the references in this paper. 2. 1.Basic concepts SVM uses linear model to implement nonlinear class boundaries through some nonlinear mapping the input vectors x into the high-dimensional feature space.
A linear model constructed in the new space can represent a nonlinear decision boundary in K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 309 the original space. In the new space, an optimal separating hyperplane is constructed. Thus, SVM is known as the algorithm that ynds a special kind of linear model, the maximum margin hyperplane. The maximum margin hyperplane gives the maximum separation between the decision classes.
The training examples that are closest to the maximum margin hyperplane are called support vectors. All other training examples are irrelevant for deyning the binary class boundaries. For the linearly separable case, a hyperplane separating the binary decision classes in the three-attribute case can be represented as the following equation: y = w 0 + w 1 x 1 + w 2 x2 + w 3 x 3 ; (1) where y is the outcome, xi are the attribute values, and there are four weights wi to be learned by the learning algorithm. In Eq.
(1), the weights wi are parameters that determine the hyperplane.The maximum margin hyperplane can be represented as the following equation in terms of the support vectors: y=b+ i yi x(i) · x; (2) where yi is the class value of training example x(i), · represents the dot product. The vector x represents a test example and the vectors x(i) are the support vectors. In this equation, b and i are parameters that determine the hyperplane. From the implementation point of view, ynding the support vectors and determining the parameters b and i are equivalent to solving a linearly constrained quadratic programming (QP).
As mentioned above, SVM constructs linear model to implement nonlinear class boundaries through the transforming the inputs into the high-dimensional feature space. For the nonlinearly separable case, a high-dimensional version of Eq. (2) is simply represented as follows: y=b+ i yi K(x(i); x): (3) The function K(x(i); x) is deyned as the kernel function. There are some di erent kernels for generating the inner products to construct machines with di erent types of nonlinear decision surfaces in the input space.Choosing among di erent kernels the model that minimizes the estimate, one chooses the best model. Common examples of the kernel function are the polynomial kernel K(x; y)=(xy+1)d and the Gaussian radial basis function K(x; y) = exp(? 1= 2 (x ? y)2 ) where d is the degree of the polynomial kernel and 2 is the bandwidth of the Gaussian radial basis function kernel.
For the separable case, there is a lower bound 0 on the coe cient i in Eq. (3). For the non-separable case, SVM can be generalized by placing an upper bound C on the coe cients i in addition to the lower bound .
. 2. Prior applications of SVM in ynancial time-series forecasting As mentioned above, the BP network has been widely used in the area of ynancial time series forecasting because of its broad applicability to many business problems and preeminent learning ability. However, the BP network has many disadvantages including the need for the determination of the value of controlling parameters and the number of processing elements in the layer, and the danger of overytting problem.
310 K. -j. Kim / Neurocomputing 55 (2003) 307 – 319On the other hand, there are no parameters to tune except the upper bound C for the non-separable cases in linear SVM . In addition, overytting is unlikely to occur with SVM. Overytting may be caused by too much exibility in the decision boundary. But, the maximum hyperplane is relatively stable and gives little exibility . Although SVM has the above advantages, there is few studies for the application of SVM in ynancial time-series forecasting.
Mukherjee et al.  showed the applicability of SVM to time-series forecasting.Recently, Tay and Cao  examined the predictability of ynancial time-series including yve time series data with SVMs. They showed that SVMs outperformed the BP networks on the criteria of normalized mean square error, mean absolute error, directional symmetry and weighted directional symmetry. They estimated the future value using the theory of SVM in regression approximation. 3.
Research data and experiments 3. 1. Research data The research data used in this study is technical indicators and the direction of change in the daily Korea composite stock price index (KOSPI).Since we attempt to forecast the direction of daily price change in the stock price index, technical indicators are used as input variables. This study selects 12 technical indicators to make up the initial attributes, as determined by the review of domain experts and prior research . The descriptions of initially selected attributes are presented in Table 1. Table 2 presents the summary statistics for each attribute. This study is to predict the directions of daily change of the stock price index.
They are categorized as “0” or “1” in the research data. 0” means that the next day’s index is lower than today’s index, and “1” means that the next day’s index is higher than today’s index. The total number of sample is 2928 trading days, from January 1989 to December 1998. About 20% of the data is used for holdout and 80% for training. The number of the training data is 2347 and that of the holdout data is 581. The holdout data is used to test results with the data that is not utilized to develop the model.
The original data are scaled into the range of [ ? 1:0; 1:0]. The goal of linear scaling is to independently normalize each feature component to the speciyed range.It ensures the larger value input attributes do not overwhelm smaller value inputs, then helps to reduce prediction errors. The prediction performance P is evaluated using the following equation: P= 1 m m Ri i=1 (i = 1; 2; : : : ; m) (4) where Ri the prediction result for the ith trading day is deyned by Ri = 1 0 if POi = AOi ; otherwise; K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 Table 1 Initially selected features and their formulas Feature name %K Description Stochastic %K. It compares where a security’s price closed relative to its price range over a given time period.Stochastic %D.
Moving average of %K. Stochastic slow %D. Moving average of %D. It measures the amount that a security’s price has changed over a given time span.
Price rate-of-change. It displays the di erence between the current price and the price n days ago. Larry William’s %R. It is a momentum indicator that measures overbought/oversold levels. Accumulation/distribution oscillator. It is a momentum indicator that associates changes in price. 5-day disparity.
It means the distance of current price and the moving average of 5 days. 10-day disparity. Price oscillator.It displays the di erence between two moving averages of a security’s price.
Commodity channel index. It measures the variation of a security’s price from its statistical mean. Formula Ct ? LLt? n ? 100, where LLt and HHt HHt? n ? LLt? n mean lowest low and highest high in the last t days, respectively. n? 1 i=0 311 Refs.  %D %Kt? i n %Dt? i n  Slow %D Momentum n? 1 i=0   Ct ? Ct? 4 Ct ? 100 Ct? n ROC  Williams’ %R Hn ? Ct ? 100 Hn ? Ln Ht ? Ct? 1 Ht ? Lt Ct ? 100 MA5 Ct ? 100 MA10 MA5 ? MA10 MA5  A/D Oscillator  Disparity5  Disparity10 OSCP   CCI RSI Relative strength index.It is a price following an oscillator that ranges from 0 to 100. (Mt ? SMt ) [1,3] where Mt = (Ht + Lt + Ct )=3; (0:015 Dt ) n Mt? i+1 , and SMt = i=1 n n |Mt? i+1 ? SMt | .
Dt = i=1 n 100 100?  1 + ( n? 1 Upt? i =n)=( n? 1 Dwt? i =n) i=0 i=0 where Upt means upward-price-change and Dwt means downward-price-change at time t. Ct is the closing price at time t, Lt the low price at time t, Ht the high price at time t and, MAt the moving average of t days. 312 Table 2 Summary statistics Feature name %K %D Slow %D Momentum ROC Williams’ %R A/D Oscillator Disparity5 Disparity10 OSCP CCI RSIK. -j.
Kim / Neurocomputing 55 (2003) 307 – 319 Max 100. 007 100. 000 99. 370 102. 900 119. 337 100. 000 3. 730 110.
003 115. 682 5. 975 226. 273 100. 000 Min 0. 000 0. 000 0.
423 ? 108. 780 81. 992 ? 0. 107 ? 0.
157 90. 077 87. 959 ? 7. 461 ? 221. 448 0. 000 Mean 45.
407 45. 409 45. 397 ? 0. 458 99. 994 54. 593 0. 447 99.
974 99. 949 ? 0. 052 ? 5. 945 47. 598 Standard deviation 33. 637 28.
518 26. 505 21. 317 3. 449 33. 637 0.
334 1. 866 2. 682 1.
330 80. 731 29. 531 POi is the predicted output from the model for the ith trading day, and AOi s the actual output for the ith trading day, m is the number of the test examples. 3. 2. SVM In this study, the polynomial kernel and the Gaussian radial basis function are used as the kernel function of SVM.
Tay and Cao  showed that the upper bound C and the kernel parameter 2 play an important role in the performance of SVMs. Improper selection of these two parameters can cause the overytting or the underytting problems. Since there is few general guidance to determine the parameters of SVM, this study varies the parameters to select optimal values for the best prediction performance.
This study uses LIBSVM software system  to perform experiments. 3. 3. BP In this study, standard three-layer BP networks and CBR are used as benchmarks.
This study varies the number of nodes in the hidden layer and stopping criteria for training. In this study, 6, 12, 24 hidden nodes for each stopping criteria because the BP network does not have a general rule for determining the optimal number of hidden nodes. For the stopping criteria of BP, this study allows 50, 100, 200 learning epochs per one training example since there is little general knowledge for selecting the number of epochs.Thus, this study uses 146 400, 292 800, 565 600 learning epochs for the stopping criteria of BP because this study uses 2928 examples. The learning rate is 0. 1 and the momentum term is 0. 1.
The hidden nodes use the sigmoid transfer function and the output node uses the linear transfer function. This study allows 12 input nodes because 12 input variables are employed. K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 313 3.
4. CBR For CBR, the nearest-neighbor method is used to retrieve relevant cases. This method is a popular retrieval method because it can be easily applied to numeric data such as ynancial data.This study varies the number of nearest neighbor from 1 to 5. An evaluation function of the nearest-neighbor method is Euclidean distance and the function is represented as follows: n DIR = i=1 wi (fiI ? fiR )2 ; (5) where DIR is a distance between fiI and fiR , fiI and fiR are the values for attribute fi in the input and retrieved cases, n is the number of attributes, and wi is the importance weighting of the attribute fi . 4. Experimental results One of the advantages of linear SVM is that there is no parameter to tune except the constant C.
But the upper bound C on the coe cient i a ects prediction performance for the cases where the training data is not separable by a linear SVM . For the nonlinear SVM, there is an additional parameter, the kernel parameter, to tune. First, this study uses two kernel functions including the Gaussian radial basis function and the polynomial function. The polynomial function, however, takes a longer time in the training of SVM and provides worse results than the Gaussian radial basis function in preliminary test.
Thus, this study uses the Gaussian radial basis function as the kernel function of SVMs.This study compares the prediction performance with respect to various kernel parameters and constants. According to Tay and Cao , an appropriate range for 2 was between 1 and 100. In addition, they proposed that an appropriate range for C was between 10 and 100. Table 3 presents the prediction performance of SVMs with various parameters. In Table 3, the best prediction performance of the holdout data is recorded when 2 is 25 and C is 78.
The range of the prediction performance is between 50. 0861% and 57. 8313%. Fig. 1 gives the results of SVMs with various C where 2 is yxed at 25.Tay and Cao  suggested that too small a value for C caused under-yt the training data while too large a value of C caused over-yt the training data. It can be observed that the prediction performance on the training data increases with C in this study.
The prediction performance on the holdout data increases when C increases from 1 to 78 but decreases when C is 100. The results partly support the conclusions of Tay and Cao . Fig. 2 presents the results of SVMs with various 2 where C is chosen as 78. According to Tay and Cao , a small value of 2 would over-yt the training data while a large alue of 2 would under-yt the training data. The prediction performance 314 K. -j.
Kim / Neurocomputing 55 (2003) 307 – 319 Table 3 The prediction performance of various parameters in SVMs C Training data Number of hit/total number (a) 1 10 33 55 78 100 (b) 1 10 33 55 78 100 (c) 1 10 33 55 78 100 (d) 1 10 33 55 78 100 (e) 1 10 33 55 78 100 2 Holdout data Hit ratio 82. 9566 98. 4117 99. 8167 100 100 100 59. 0104 61. 515 63. 3476 64. 0195 64.
7526 65. 73 58. 2773 59. 2547 60. 6597 61.
7593 62. 1258 62. 3091 58. 0941 59.
438 59. 5602 59. 9878 60. 171 60. 4765 57.
2999 58. 66 59. 3158 59.
4991 60. 11 59. 8656 Number of hit/total number 305/581 296/581 291/581 295/581 293/581 293/581 319/581 331/581 330/581 334/581 336/581 332/581 331/581 325/581 335/581 324/581 322/581 326/581 323/581 323/581 325/581 331/581 333/581 333/581 320/581 317/581 322/581 324/581 325/581 329/581 Hit ratio 52. 4957 50. 9466 50.
0861 50. 7745 50. 4303 50. 4303 54. 9053 56. 9707 56. 7986 57. 4871 57.
8313 57. 1429 56. 9707 55. 938 57. 6592 55. 7659 55.
4217 56. 1102 55. 5938 55. 5938 55. 938 56. 9707 57. 315 57.
315 55. 0775 54. 5611 55.
4217 55. 7659 55. 38 56.
6265 =1 1358/1637 1611/1637 1634/1637 1637/1637 1637/1637 1637/1637 2 = 25 966/1637 1007/1637 1037/1637 1048/1637 1060/1637 1076/1637 = 50 954/1637 970/1637 993/1637 1011/1637 1017/1637 1020/1637 = 75 951/1637 973/1637 975/1637 982/1637 985/1637 990/1637 = 100 938/1637 962/1637 971/1637 974/1637 984/1637 980/1637 2 2 2 on the training data decreases with 2 in this study. But Fig. 2 shows the prediction performance on the holdout data is stable and insensitive in the range of 2 from 25 to 100. These results also support the conclusions of Tay and Cao .
Figs. 3 and 4 present the results of the best SVM model for the training and the holdout data, respectively. K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 315 Fig.
1. The results of SVMs with various C where 2 is yxed at 25. Fig.
2. The results of SVMs with various 2 where C is yxed at 78. Figs. 3(a) and 4(a) represent data patterns before SVM is employed. Two di erent colors of circles are two classes of the training and the holdout examples. Figs. 3(b) and 4(b) show the results after SVM is implemented.
The two classes are represented by green and red bullets.In addition, this study compares the best SVM model with BP and CBR. Table 4 gives the prediction performance of various BP models. In Table 4, the best prediction performance for the holdout data is produced when the number of hidden processing elements are 24 and the stopping criteria is 146 400 or 316 K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 (a) Before SVM is implemented (b) After SVM is implemented Fig.
3. Graphical interpretation of the results of SVM for the training data: (a) before SVM is implemented and (b) after SVM is implemented. a) Before SVM is implemented (b) After SVM is implemented Fig. 4. Graphical interpretation of the results of SVM for the holdout data: (a) before SVM is implemented and (b) after SVM is implemented. 292 800 learning epochs. The prediction performance of the holdout data is 54. 7332% and that of the training data is 58.
5217%. For CBR, this study varies the number of retrieved cases for the new problem. The range of the number of retrieved cases is between 1 and 5.
However, the prediction performances of these yve experiments produce same results.Thus, this study uses the prediction performance when the number of retrieved cases is 1. The prediction accuracy of the holdout data is 51. 9793%. For CBR, the performance of the training K.
-j. Kim / Neurocomputing 55 (2003) 307 – 319 Table 4 The results of various BP models Stopping criteria (epoch) 146 400 Number of hidden nodes 6 12 24 6 12 24 6 12 24 Prediction performance for the training data (%) 58. 1552 58. 6439 58.
5217 58. 1552 58. 6439 58. 5217 58. 1552 58. 1552 58.
1552 317 Prediction performance for the holdout data (%) 52. 8399 53. 3563 54. 7332 52. 8399 53.
3563 54. 7332 52. 399 52. 8399 52. 8399 292 800 565 600 Table 5 The best prediction performances of SVM, BP, and CBR (hit ratio: %) SVM Training data Holdout data 64. 7526 57. 8313 BP 58.
5217 54. 7332 CBR 51. 9793 Table 6 McNemar values (p values) for the pairwise comparison of performance BP SVM BP 1642 (0. 200) CBR 4.
654 (0. 031) 0. 886 (0. 347) data is ignored because the retrieved case and the new case are the same in the training data. Table 5 compares the best prediction performances of SVM, BP, and CBR. In Table 5, SVM outperforms BPN and CBR by 3. 0981% and 5.
852% for the holdout data, respectively.For the training data, SVM has higher prediction accuracy than BPN by 6. 2309%. The results indicate the feasibility of SVM in ynancial time series forecasting and are compatible with the conclusions of Tay and Cao . The McNemar tests are performed to examine whether SVM signiycantly outperforms the other two models. This test is a nonparametric test for two related samples and may be used with nominal data. The test is particularly useful with before-after measurement of the same subjects .
Table 6 shows the results of the McNemar test to compare the prediction performance of the holdout data.As shown in Table 6, SVM performs better than CBR at 5% statistical signiycance level. However, SVM does not signiycantly outperform BP. In addition, Table 6 also shows that BP and CBR do not signiycantly outperform each other. 318 K. -j. Kim / Neurocomputing 55 (2003) 307 – 319 5. Conclusions This study used SVM to predict future direction of stock price index.
In this study, the e ect of the value of the upper bound C and the kernel parameter 2 in SVM was investigated. The experimental result showed that the prediction performances of SVMs are sensitive to the value of these parameters.Thus, it is important to ynd the optimal value of the parameters. In addition, this study compared SVM with BPN and CBR.
The experimental results showed that SVM outperformed BPN and CBR. The results may be attributable to the fact that SVM implements the structural risk minimization principle and this leads to better generalization than conventional techniques. Finally, this study concluded that SVM provides a promising alternative for ynancial time-series forecasting. There will be other research issues which enhance the prediction performance of SVM if they are investigated.The prediction performance may be increased if the optimum parameters of SVM are selected and this remains a very interesting topic for further study. The generalizability of SVMs also should be tested further by applying them to other time-series. Acknowledgements This work was supported by the Dongguk University Research Fund.
References  S. B. Achelis, Technical Analysis from A to Z, Probus Publishing, Chicago, 1995.  H. Ahmadi, Testability of the arbitrage pricing theory by neural networks, in: Proceedings of the International Conference on Neural Networks, San Diego, CA, 1990, pp. 85 –393.  J.
Chang, Y. Jung, K. Yeon, J. Jun, D. Shin, H. Kim, Technical Indicators and Analysis Methods, Jinritamgu Publishing, Seoul, 1996.
 C. -C. Chang, C. -J. Lin, LIBSVM: a library for support vector machines, Technical Report, Department of Computer Science and Information Engineering, National Taiwan University, 2001, Available at http://www. csie.
edu. tw/? cjlin/papers/libsvm. pdf.  J.
Choi, Technical Indicators, Jinritamgu Publishing, Seoul, 1995.  J. H. Choi, M. K. Lee, M. W.
Rhee, Trading S& P 500 stock index futures using a neural network, in: Proceedings of the Annual International Conference on Artiycial Intelligence Applications on Wall Street, New York, 1995, pp. 63–72.  D. R. Cooper, C. W. Emory, Business Research Methods, Irwin, Chicago, 1995.
 H. Drucker, D. Wu, V. N.
Vapnik, Support vector machines for spam categorization, IEEE Trans. Neural Networks 10 (5) (1999) 1048–1054.  E. Gi ord, Investor’s Guide to Technical Analysis: Predicting Price Action in the Markets, Pitman Publishing, London, 1995.  Y. Hiemstra, Modeling structured nonlinear knowledge to predict stock market returns, in: R.R.
Trippi (Ed. ), Chaos & Nonlinear Dynamics in the Financial Markets: Theory, Evidence and Applications, Irwin, Chicago, IL, 1995, pp. 163–175.  K.
Kamijo, T. Tanigawa, Stock price pattern recognition: a recurrent neural network approach, in: Proceedings of the International Joint Conference on Neural Networks, San Diego, CA, 1990, pp. 215 –221. K.
-j. Kim / Neurocomputing 55 (2003) 307 – 319 319  K. Kim, I.
Han, Genetic algorithms approach to feature discretization in artiycial neural networks for the prediction of stock price index, Expert Syst. Appl. 19 (2) (2000) 125–132.  T.
Kimoto, K.Asakawa, M. Yoda, M.
Takeoka, Stock market prediction system with modular neural network, in: Proceedings of the International Joint Conference on Neural Networks, San Diego, CA, 1990, pp. 1– 6.  K. Kohara, T. Ishikawa, Y. Fukuhara, Y.
Nakamura, Stock price prediction using prior knowledge and neural networks, Int. J. Intell.
Syst. Accounting Finance Manage. 6 (1) (1997) 11–22.
 S. Mukherjee, E. Osuna, F.
Girosi, Nonlinear prediction of chaotic time series using support vector machines, in: Proceedings of the IEEE Workshop on Neural Networks for Signal Processing, Amelia Island, FL, 1997, pp. 11–520.  J. J. Murphy, Technical Analysis of the Futures Markets: A Comprehensive Guide to Trading Methods and Applications, Prentice-Hall, New York, 1986.  T.
-S. Quah, B. Srinivasan, Improving returns on stock investment through neural network selection, Expert Syst. Appl. 17 (1999) 295–301.
 F. E. H. Tay, L.
Cao, Application of support vector machines in ynancial time series forecasting, Omega 29 (2001) 309–317.  R. R. Trippi, D. DeSieno, Trading equity index futures with a neural network, J. Portfolio Manage. 19 (1992) 27–33.
 R. Tsaih, Y.Hsu, C. C. Lai, Forecasting S& P 500 stock index futures with a hybrid AI system, Decision Support Syst. 23 (2) (1998) 161–174.
 V. N. Vapnik, Statistical Learning Theory, Wiley, New York, 1998.  I. H. Witten, E. Frank, Data Mining: Practical Machine Learning Tools and Techniques with Java Implementations, Morgan Kaufmann Publishers, San Francisco, CA, 2000.  Y. Yoon, G. Swales, Predicting stock price performance: a neural network approach, in: Proceedings of the 24th Annual Hawaii International Conference on System Sciences, Hawaii, 1991, pp. 156 –162. 24] G. Zhang, B. E. Patuwo, M. Y. Hu, Forecasting with artiycial neural networks: the state of the art, Int. J. Forecasting 14 (1998) 35–62. Kyoung-jae Kim received his M. S. and Ph. D. degrees in Management Information Systems from the Graduate School of Management at the Korea Advanced Institute of Science and Technology and his B. A. degree from the Chung-Ang University. He is currently a faculty member of the Department of Information Systems at the Dongguk University. His research interests include data mining, knowledge management, and intelligent agents.