Highlights in Business, Economics and Management MSIE 2022

Volume 3 (2023)

Support Vector Machine-based Stock Prediction Analysis

Jingdong Yang
School of Management and Economic, The Chinese University of Hong Kong, Shenzhen 518000,
China
Abstract. Changes in the stock market are closely related to the dynamics of the market
development and economic research of the whole country. Correctly predicting the stock price trend
is not only beneficial for investors to make correct investment management decisions but also of
great importance to promote the effective allocation of resources and improve market effectiveness.
To this end, the empirical prediction and analysis of a stock development market using a support
vector machine show that the accuracy of stock prediction using a support vector classification
machine is as high as about 90%, indicating that the accuracy of stock prediction based on support
vector machine is ideal and meaningful.
Keywords: Support Vector Machines; Stock Forecasting.

1. Introduction
In recent years, artificial intelligence technology (big data, data mining, machine learning
technology) has developed rapidly, and the resulting data analysis models have mapped low-
dimensional indistinguishable data to high-dimensional distinguishable data through various spatial
transformations and dimensional transformations, greatly compensating for the human brain's lack of
analysis of non-linear laws. We can use historical asset prices, trading data, etc. as information input
to the model, analyze the characteristics of the data, and use various classification and regression
prediction algorithms in the field of machine learning to construct trading strategies. In actual
secondary market securities trading, stocks have time-series characteristics, so recurrent neural
networks are often used for stock price prediction. However, due to the weak ability to portray time
series data with a long memory, the general recurrent neural model is prone to gradient explosion and
overlearning, which makes training very difficult. In contrast, support vector machine algorithms do
not increase the computational complexity when mapping to higher dimensional spaces because they
use kernel functions to overcome the problems of dimensional catastrophe and nonlinear
differentiability [1]. In addition to this, support vector machines overcome the problem that traditional
neural networks can only find local optimal solutions and discover the global minimum of the
objective function using known efficient algorithms, and also have better generalization capabilities.

2. Support Vector Machine Overview

Support vector machine is a supervised learning algorithmic model for analyzing data in
classification and regression analysis in machine learning, proposed by the Soviet scholar Vladimir
N. Vapnik in 1964 [2]. The core idea is to derive the maximum interval hyperplane that completely
separates the two classes of samples, which is a global convex optimization problem. In practice, it
is called linear support vector machine and nonlinear support vector machine for the reason of linear
separability of training samples or not [3], while the optimal solutions of support models can be
broadly classified into three categories: (1) hard interval maximization (2) soft interval maximization
(3) kernel function and nonlinear interval maximization [4].
Where the strict mathematical definitions of linearly differentiable and hyperplane are and are two-
point sets (sets of points) in an n-dimensional Euclidean space. If there exists an n-dimensional vector
w and a real number b such that all points belonging to the points x have, and for all points belonging
to the points y then we have. Then we call linearly divisible, that is, hyperplanes in n-dimensional
Euclidean spaces of dimension equal to n-1 [5].
(1) Hard interval maximization

12
Highlights in Business, Economics and Management MSIE 2022
Volume 3 (2023)

When the training samples are linearly divisible, assuming that the sample data set is , ,
1,2, … , , ∈ , ∈ 1, 1 , then any hyperplane can be represented by 0 to be
represented. Then in the n-dimensional space, the distance from the point x to the line 0
is the distance of ‖ ‖
, ⋯ . Assume that the hyperplane , can
classify the training samples correctly, i.e., for , ∈ , if 1 , then we have
0 ; if 1, then we have 0 . So far we have the upper and lower hyperplanes of the
maximally spaced hyperplane, and the distance of each support vector to the hyperplane can be
written as ‖ ‖
, then the maximization of the distance can be expressed as max 2 ‖ ‖
,
max ‖ , thus constructing the Lagrangian function. min max , , ‖ ‖ ∑ 1
, ‖ ,
. . 0 , and then use the strong duality transformation to solve. max ∑
∑ ∑ . .∑ 0, 0 . [6]
(2) Soft interval maximization
In practical applications, fully linearly divisible samples are rare, so soft spacing is introduced.
Compared with the harsh conditions of hard spacing, soft spacing allows individual sample points to
appear inside the spacing band, and this part of the sample points cannot satisfy the constraints1
0 .In order to measure the degree of soft spacing, we introduce a relaxation variable
for each sample , such that 0, the 1 0 , the optimization objective
becomes min ‖ ‖ ∑ . . , 1 0 , 0 ,
1,2, … , . Thus, the Lagrangian function is constructed as. min max , , , , ‖ ‖
, , ,
∑ ∑ 1 ∑ . . 0, 0 ,which can then be
solved by using the strong duality transformation.
max ∑ ∑ ∑ . .∑ 0, 0, 0 . [7]
(3) Kernel function and nonlinear interval maximization
Hard and soft spacing refers to the complete linear separability of samples or linear separability of
most samples, but it will encounter sample points that are not linearly separable. For continuous
samples, the solution to this situation is to map the samples that are linearly indistinguishable in the
original space to a higher dimensional space, make the sample points linearly distinguishable in the
higher dimensional space, and then construct the optimal classification hyperplane, which is a
nonlinear SVM [8]. If the original sample points are denoted by x and denotes the new vector
obtained after mapping x to a higher dimensional feature space, then the segmentation hyperplane
can be expressed as . Thus, the pairwise problem of nonlinear SVM is
transformed into. max ∑ ∑ ∑ . .∑ 0,
0, 0 . The kernel function is a function that calculates the inner product of two vectors
in the space after the implicit mapping, and it does this by first doing the inner product on the feature
vector and then using the function k to change it, which helps to avoid the calculation in the high-
dimensional space and greatly simplifies the problem solution, which is equivalent to doing the kernel
mapping on the vector first and then doing the inner product, which shows that the introduction of
the kernel function reduces our computation on the one hand, and reduces our storage data on the
other hand memory usage. The common kernel functions are: linear kernel function ( ,
), polynomial kernel function ( , ), and Gaussian kernel function

( ). [9]
,

13
Highlights in Business, Economics and Management MSIE 2022
Volume 3 (2023)

3. Stock Data Analysis

The data in this paper are obtained from the CSMAR database, and representative sample stocks
from three different industries are selected as the research objects, with a period from October 15,
2018, to October 14, 2022, with a total of 3,135 entries, and the data set is split into a training set and
a test set with a ratio of 75% and 25%, respectively, and some of the raw data are shown in Table 1
below. The raw variables in the data set are transaction date, opening price, closing price, minimum
price, maximum price, volume, price change, increase or decrease, etc. The data meaning of this paper
TradingDate represents the date, Code for the stock code, OpenPrice for the opening price of the day,
HighPrice for the highest price of the day, ClosePrice for the closing price of the day, LowPrice for
the lowest price of the day, Volume for the volume of the day, diff for the difference between the
closing price and the opening price that ClosePrice - Openprice is obtained, up for 1 represents up,
up for -1 represents down.

Table 1. Table of some raw data

TradingDate OpenPrice ClosePrice HighPrice LowPrice Volume Amount
2018/10/15 38.555 37.348 39.092 37.147 19564292 111620218
2018/10/16 37.281 36.141 37.751 35.94 21409682 117750310
2018/10/17 36.745 36.879 37.214 35.806 24031624 131320280
2018/10/18 36.544 35.739 36.678 35.538 17711684 95611246
2018/10/19 35.001 36.678 36.946 34.666 22024600 119053631
2018/10/22 37.08 39.226 39.896 36.879 39740925 229006742
2018/10/23 39.226 38.22 39.695 37.885 23898185 138072131
2018/10/24 38.22 38.22 39.159 38.086 17971668 103214803
2018/10/25 36.812 38.086 38.22 36.544 20079863 111420962
2018/10/26 38.421 37.885 39.025 37.616 17219956 98225670
2018/10/29 37.482 38.488 39.159 37.415 23351723 134120252

3.1 Pre-processing
Data pre-processing is a key step in the modeling process. In most cases, the raw data obtained
from the data acquisition stage may have missing values or outliers, so preprocessing is required.
Preprocessing data not only improves the prediction of the model, but also reduces the arithmetic
power loss, shortens the computation time, and subsequently optimizes the model efficiency [10].
The basic process adopted in this paper includes: applying python's data preprocessing function
IsNull() to detect missing values, and no missing values are found, and performing max-min
normalization on data such as volume.
3.2 Modeling
The SVM model of the Sklearn library from the Python third-party library is called, and the linear
kernel function, polynomial function, sigmoid function, and RBF function are called to construct the
SVM model separately, to analyze the effect of different kernel functions on the stock prediction of
the SVM model.
There are many ways to evaluate models, but both classification and regression problems can be
broadly based on two methods. In stock price forecasting, the mean square error MSE and the
coefficient of determination R2 can be used; in stock price trend forecasting, the F1 value and the
ROC curve are used [11].
In this paper, firstly, different kernel functions are evaluated based on four dimensions: training
set score, test set score, test set mean square deviation, and test set R2 score, and the specific scores
are shown in Figure 1. From the comparison of these four data, the linear kernel function and RBF
function outperform the polynomial function and sigmoid function.

14
Highlights in Business, Economics and Management MSIE 2022
Volume 3 (2023)

Table 2. Some raw data scoring of the kernel functions constructed

Linear Kernel
Training set scoring 0.997256515
Test set scoring 0.979423868
Standard Deviation of the test set 0.082304526
R2 of the test set 0.917694079
Gaussian kernel function
Training set scoring 1
Test set scoring 0.699588477
Standard Deviation of the test set 1.201646091
R2 of the test set -0.20166644
sigmoid Kernel
Training set scoring 0.504801097
Test set scoring 0.489711934
Standard Deviation of the test set 2.041152263
R2 of the test set -1.04118683
Polynomial kernel function
Training set scoring 0.68861454
Test set scoring 0.674897119
Standard Deviation of the test set 1.300411522
R2 of the test set -0.30043355

Secondly, this paper evaluates the model by the different areas under the ROC curve, the larger
the area under the ROC curve, the higher the accuracy. And from the comparison of ROC curves, the
inflection points of ROC curves of different kernel function SVM models are different, among which
the SVM model with linear kernel function has the highest inflection point, and the specific curve is
shown in Figure 1.

Figure 1. ROC curves of different kernel functions constructed

Finally, from the overlap curves between the predicted and true values of each model, the SVM
model with linear as the kernel function has the highest degree of agreement between the predicted
and true values, as shown in Figure 2.

15
Highlights in Business, Economics and Management MSIE 2022
Volume 3 (2023)

Figure 2. Overlap curve between predicted and true values constructed

The model evaluation code logic is as follows.

def evaluate(X_train, X_test, y_train, y_test):
# Model name
#Linear kernel functions
clf3 = svm.SVC(kernel='linear')
clf3.fit(X_train, y_train)
y_pred3 = clf3.predict(X_test)
print("Linear kernel function:")
print("Training set score:", clf3.score(X_train, y_train))
print("Test set score:", clf3.score(x_test, y_test))
print("Test set mean squared error:", metrics.mean_squared_error(y_test,
y_pred3.reshape(-1, 1)))
print("Test set R2 score:", metrics.r2_score(y_test, y_pred3.reshape(-1, 1)))
# Gaussian kernel function
clf1 = svm.SVC(kernel='rbf', C=10, gamma=0.1, coef0=0.1)#kernel function type,
gamma kernel coefficient, number of degree polynomial kernel functions, penalty factor for C error
term
clf1.fit(X_train, y_train)
y_pred1 = clf1.predict(X_test)
print("Gaussian kernel function:")
print("Training set score:", clf1.score(X_train, y_train))# The best possible score is 1.0,
which can be a negative number.
print("Test set score:", clf1.score(x_test, y_test))
print("Test set mean squared error:", metrics.mean_squared_error(y_test,
y_pred1.reshape(-1, 1)))
print("Test set R2 score:", metrics.r2_score(y_test, y_pred1.reshape(-1, 1)))

# sigmoid kernel function

clf2 = svm.SVC(kernel='sigmoid', C=2)
clf2.fit(X_train, y_train)
y_pred2 = clf2.predict(X_test)
print("sigmoid kernel function:")
print("Training set score:", clf2.score(X_train, y_train))
print("Test set score:", clf2.score(x_test, y_test))
print("Test set mean squared error:", metrics.mean_squared_error(y_test,
y_pred2.reshape(-1, 1)))

16
Highlights in Business, Economics and Management MSIE 2022
Volume 3 (2023)

print("Test set R2 score:", metrics.r2_score(y_test, y_pred2.reshape(-1, 1)))

# Polynomial kernel functions

clf = svm.SVC(kernel='poly', C=2)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
print("Polynomial kernel function:")
print("Training set score:", clf.score(X_train, y_train))
print("Test set score:", clf.score(x_test, y_test))
print("Test set mean squared error:", metrics.mean_squared_error(y_test,
y_pred.reshape(-1, 1)))
print("Test set R2 score:", metrics.r2_score(y_test, y_pred.reshape(-1, 1)))
#The larger the area under the ROC curve, the higher the accuracy
fig, ax = plt.subplots(figsize=(12, 10))
lsvm_roc = plot_roc_curve(estimator=clf3, X=X_test, y=y_test, ax=ax, linewidth=1,
label='linear')
ssvm_roc = plot_roc_curve(estimator=clf2, X=X_test,
y=y_test, ax=ax, linewidth=1, label='sigmoid')
psvm_roc = plot_roc_curve(estimator=clf, X=X_test,
y=y_test, ax=ax, linewidth=1, label='poly')
rsvm_roc = plot_roc_curve(estimator=clf1, X=X_test,
y=y_test, ax=ax, linewidth=1, label='rbf')
ax.legend(fontsize=12)
plt.show()
#print(y_pred1)
# Draw the closing price trend and the true and predicted value charts
plt.subplot(3, 1, 1)
# Closing price chart
plt.title("Closing Price Trend")
plt.plot(X_test1['TradingDate']. iloc[-20:], X_test1['ClosePrice']. iloc[-20:])
plt.xticks(rotation=30)
plt.subplot(3, 1, 3)
plt.title("Forecast-Real Graph")
plt.plot(X_test1['TradingDate']. iloc[-20:], y_pred3[-20:], color='red', label='predicted
value')
plt.plot(X_test1['TradingDate']. iloc[-20:], X_test1['up']. iloc[-20:], color='blue',
label='true value')
plt.xticks(rotation=30)
plt.legend()

plt.show()

return 0

4. Main Conclusion
Stock prices fluctuate frequently and, in many cases, quite dramatically. For a more stable stock
market, it is necessary to reduce uncertainty for investors and funders and to enable stock forecasts
to be applied more flexibly between investors and funders in order to achieve an optimal resource
allocation market.

17
Highlights in Business, Economics and Management MSIE 2022
Volume 3 (2023)

In this paper, SVM is applied to stock market forecasting. SVM is used to select the kernel function
and optimize the parameters, and then find the optimal model to measure the stock trend. Based on
the fast convergence and high accuracy of the SVM model, the SVM model can predict the stock data
well and make the prediction results very close to the actual values. Kernel function selection and
kernel parameter selection have a very important impact on the learning and prediction performance
of SVM. Different kernel functions and kernel parameters are directly related to the accuracy of the
operation results. svm has good accuracy in predicting stock prices and provides a meaningful
analysis tool for general investors.

5. Summary
Stock data are chaotic and disorderly. The original variables of the dataset used in this paper are
trading date, opening price, closing price, minimum price, maximum price, volume, and increase/
decrease rate. The SVM algorithms using different kernel functions for stock prediction are different.
From the experimental analysis results, the LINEAR kernel function is more suitable for stock
prediction analysis.

References
[1] Li, Hang. Statistical learning methods. Beijing: Tsinghua University Press,2012.
[2] Xixi Xu. Research on stock price prediction and investment strategy based on support vector machine.
Xi'an: Northwest University Academic, 2018.
[3] Shahi Tej Bahadur, Shrestha Ashish, Neupane Arjun, Guo William. stock Price Forecasting with Deep
Learning: A Comparative Study. Mathematics,2020,8(9).
[4] Zhang, X.-G. On statistical learning theory and support vector machines. Journal of Automation,
2000,01:36-46.
[5] Gu Yaxiang. Ding S. F. Advances in support vector machine research. Computer Science, 2011,02:14-17.
[6] Tang, Inv. Research on support vector machine algorithm based on statistical learning theory. Huazhong
University of Science and Technology, 2005.08.
[7] Chen, Yi-Bing. Zhang L. L., Nie G. L. et al. financial time series forecasting based on improved support
vector regression machine. Mathematical Practice and Perceptions, 2012, 42(04): 38-44.
[8] Jun Wei. An efficient support vector machine parameter optimization algorithm. Computer Technology
and Development. 2015,25(12):97-104.
[9] Liu M, Zhou S, Wu H. A new hybrid kernel function support vector machine. Computer Applications,
2009, 29(S2):167-168.
[10] Yang XW, Hao ZF. Algorithm design and analysis of support vector machines. Beijing: Science Press,
2013.
[11] Ren, Y. F. Some research on SVM model improvement. Nanjing University of Posts and
Telecommunications, 2013.

18