See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/359328516

Pearson Correlation Coefficient-Based Performance Enhancement of Broad

Learning System for Stock Price Prediction

Article in IEEE Transactions on Circuits and Systems II: Express Briefs · May 2022
DOI: 10.1109/TCSII.2022.3160266

CITATIONS READS

43 517

5 authors, including:

Choujun Zhan
South China Normal University
102 PUBLICATIONS 1,144 CITATIONS

SEE PROFILE

All content following this page was uploaded by Choujun Zhan on 26 June 2022.

The user has requested enhancement of the downloaded file.

 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 69, NO. 5, MAY 2022 2413

Pearson Correlation Coefficient-Based Performance

Enhancement of Broad Learning System for
Stock Price Prediction
Guanzhi Li, Aining Zhang, Qizhi Zhang, Di Wu , and Choujun Zhan , Member, IEEE

Abstract—Accurate prediction of a stock price is a challeng- system [2]–[4]. However, the non-linearity, high noise and
ing task due to the complexity, chaos, and non-linearity nature of volatility of the stock market make the prediction of financial
financial systems. In this brief, we proposed a multi-indicator fea- systems take higher risks than to other complex systems [5]. In
ture selection method for stock price prediction based on Pearson
correlation coefficient (PCC) and Broad Learning System (BLS), addition, external factors such as national macro adjustments,
named the PCC-BLS framework. Firstly, PCC was used to select changes from the political situation and investor psychology
the input features from 35 features, including original stock make the prediction of the financial system extremely diffi-
price, technical indicators, and financial indicators. Secondly, cult [6], [7]. Data-driven methods show great potential for the
these screened input features were used for rapid information analysis and control of such complex systems [8], [9].
feature extraction and training a BLS. Four stocks recorded on
the Shanghai Stock Exchange or Shenzhen Stock Exchange were With the rapid development of data acquisition, storage
adopted to evaluate the performance of the proposed method. In and machine learning and big data techniques, researchers
addition, we compared the forecasting results with ten machine are beginning to use machine learning methods to try to
learning methods, including Support Vector Regression (SVR), capture nonlinear patterns of financial systems. Fischer and
Adaptive Boosting (Adaboost), Bootstrap aggregating (Bagging), Krauss predicted the S&P 500 trend from the up-down sig-
Random Forest (RF), Gradient Boosting Decision Tree (GBDT),
Multi-layer Perceptron (MLP), Convolutional Neural Network nals generated by LSTM [10]. Deep Neural Networks (DNNs)
(CNN), and Long Short-Term Memory (LSTM), Gated Recurrent also has shown remarkable results in machine learning prob-
Unit (GRU) and Broad Learning System (BLS). Among all algo- lems [11]–[13]. A hybrid model based on Naive Bayes
rithms used in this brief, the proposed model showed the best classifier and LSTM was used to predict the opening price
performance with the highest model fitting ability. of listed companies in China [14]. Thakkar et al. proposed
Index Terms—Broad learning system, machine learning, a method based on Pearson Correlation Coefficient (PCC) to
Pearson correlation coefficient, time series forecasting, complex initialize the edge weights of neurons between the input and
system. hidden layers of the vanilla neural network (VNN) [15]. In
conclusion, machine learning methods are good candidates for
volatility financial time series forecasting [16].
I. I NTRODUCTION Recent researches indicate that feature selection is essen-
ORECASTING stock prices is of significant importance
F for analyzing financial systems, as the stock market
reflects the economic situation of a country, even the world.
tial for improving the accuracy of stock market prediction.
Several feature selection approaches are taken to effectively
filter redundant or irrelevant features [17], [18]. Ni et al.
Accurate prediction of the stock market allows investors and hybridized the fractal feature selection method and Support
other stakeholders to understand the movement of the financial Vector Machine (SVM) to predict the daily trend of Shanghai
market. Effective trading strategies are then adopted to achieve Stock Exchange Index [19]. Yang et al. filtered the training
greater profits as well as returns with less risk [1]. In general, features by the Maximum Information Coefficient (MIC) and
the stock market is often analyzed by researchers as a complex built an ensemble forecasting model for stock price movement
prediction based on Support Vector Machine (SVM), RF, and
Manuscript received February 7, 2022; accepted March 11, 2022. Date Adaboost [20]. Huang and Tsai established a hybrid SOFM-
of publication March 17, 2022; date of current version May 3, 2022. This
work was supported in part by the Natural Science Foundation of Guangdong SVR based on filter feature selection to improve prediction
Province, China, under Grant 2020A1515010761, and in part by the Key Areas accuracy and predict the fluctuation of the Taiwan Index
Research and Development Program of Science and Technology Program (FITX) [21].
of Guangzhou under Grant 202103010005. This brief was recommended by
Associate Editor H. Li. (Corresponding author: Choujun Zhan.) This brief proposed a novel framework for forecasting the
Guanzhi Li, Aining Zhang, Qizhi Zhang, and Choujun Zhan are with the movement of complex financial systems called the Pearson
School of Computing, South China Normal University, Guangzhou 510641, correlation coefficient and Broad Learning System (PCC-BLS)
China (e-mail: [email protected]).
Di Wu is with the Department of ICT and Natural Science, Norwegian framework. The proposed method was used to predict the close
University of Science and Technology, 7491 Trondheim, Norway (e-mail: price of a stock listed in China for the day ahead. For verifying
[email protected]). the performance of the proposed model, ten machine learn-
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TCSII.2022.3160266. ing models were adopted for comparison. Experimental results
Digital Object Identifier 10.1109/TCSII.2022.3160266 indicate that the PCC-BLS shows more accurate results with
1549-7747 
c 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: SOUTH CHINA UNIVERSITY OF TECHNOLOGY. Downloaded on June 19,2022 at 12:02:08 UTC from IEEE Xplore. Restrictions apply.
 2414 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 69, NO. 5, MAY 2022

TABLE I
I NPUT F EATURES FOR M ACHINE L EARNING M ODELS

Fig. 1. The illustrative structure of a typical Broad Learning System.

Learning System (BLS) framework are introduced in detail.

Then, the experimental design and results are given in
Section III. Conclusion and future work are presented in
Section IV.

II. M ETHODOLOGY
A. Broad Learning System
Broad Learning System (BLS), based on the random vector
functional link neural network (RVFLNN), is a flat network
proposed by Chen and Liu [22]. The framework of a typical
BLS is shown in Fig. 1. The input features are first transformed
into mapped features by mapping functions and stored in the
feature nodes, which would be extended to enhancement nodes
through nonlinear activation functions and random weights. In
BLS, all the outputs of the enhancement and feature nodes are
connected to the output layer, and the connection weights are
derived by ridge regression of the pseudoinverse.
Considering the time series forecasting problem, the training
data set is represented by {(X, Y) | X ∈ RN×D , Y ∈ RN×1 }.
The mapping function first transforms the training samples
into mapping feature spaces in BLS. For n feature mappings,
each generates k nodes and this procedure can be expressed
as the following equation:
 
Zi = φ XWei + βei , i = 1, . . . , n, (1)
where weights Wei and bias term βei are randomly gener-
ated matrices with the proper dimensions. φ(·) is usually the
transformation of the activation function. The feature space of
training samples are defined as Z n = [Z1 , Z2 , . . . , Zn ] and the
the highest model fitting ability in terms of stock prediction. ith group of enhancement nodes is denoted as follow:
The main contributions of this brief are as follows.
• A large-scale dataset, containing multiple factors that Hj = ξ(Z n Whj + βhj ), j = 1, 2, . . . , m, (2)
affect the closing price, has been collected. The collection where weights Whj and bias term βhj are also generated
of features involves technical and financial indicators in randomly, and ξ(·) represents the activation function. And
addition to the original features (shown in Table I). we denote the outputs of the enhancement layer by Hm 
• PCC is adopted for feature selection analysis to improve
[H1 , H2 , . . . , Hm ]. Therefore, the output Y of BLS can be
BLS. The PCC-BLS model is proposed for forecasting represented as:
the closing price of listed companies in China.
• Ten machine learning methods, including ensemble learn- Y = [Z1 , Z2 , . . . , Zn | H1 , H2 , . . . , Hm ]W m
 
ing, RF, standard BLS and etc., have been adopted to = Zn | Hm W m
predict the closing prices of the adopted stock price and
= AW m , (3)
compare the performance with the proposed model.
The rest of this brief is organized as follows: In Section II, where A = |
[Z n H m ]. Wm
is the output weight connecting the
the proposed Pearson correlation coefficient (PCC) and Broad feature and enhancement nodes to the output layer. W m can

Authorized licensed use limited to: SOUTH CHINA UNIVERSITY OF TECHNOLOGY. Downloaded on June 19,2022 at 12:02:08 UTC from IEEE Xplore. Restrictions apply.
 LI et al.: PCC-BASED PERFORMANCE ENHANCEMENT OF BLS 2415

TABLE II
be rapidly approximated by the ridge regression. SZ.000776: T HE I NPUT F EATURES OF THE PCC-BLS M ODEL
 −1
W m = λI + AT A AT Y, (4)
where λ denotes a regularisation parameter to balance the
effect of the error term.
BLS allows output weights to be obtained quickly by pseu-
doinversion, which not only ensures efficient modeling but
also avoids structural redundancy. Therefore, the BLS provides TABLE III
SH.600019: T HE I NPUT F EATURES OF THE PCC-BLS M ODEL
an alternative approach to a deep learning framework, which
is typically subject to time-consuming training of numerous
parameters in a layer.

B. Pearson Correlation Coefficient

Pearson Correlation Coefficient, introduced by Pearson in
1895, is used to measure the strength and direction of the rela-
tionship between two variables [23]. The formula for deriving where xmin and xmax are the minimum and maximum value in
the PCC between variables xi (stock features) and y (stock the training dataset, respectively.
price)is shown below:
n
i=1 (xi − x̄)(yi − ȳ) B. Experimental Setup
rx,y =   , (5)
n 2 n In order to evaluate the forecasting performance of
(x
i=1 i − x̄) (y
i=1 i − ȳ) 2
the proposed PCC-BLS, ten machine learning models
The greater the absolute value of the rx,y is, the stronger the have been adopted for comparison. We employed SVR,
correlation. The value of PCC ranges from + 1 to −1. Here, Adaboost, Bagging, RF, GBDT, MLP, CNN, LSTM,
rx,y = 1 represents a completely positive linear correlation, GRU, and BLS approaches in the comparisons. Datasets
while rx,y = −1 stands for a completely negative linear corre- of listed companies with experimental stocks, containing
lation. On the other hand, rx,y = 0 means no linear correlation. GF Securities (sz.000776), BAOSTEEL (sh.600019), China
In this brief, we calculated the PCC between the features and Unicom (sh.600050), and Haitong Securities (sh.600837), span
the closing price of a stock. Features with an absolute value of from September 2010 to September 2021.
PCC greater than a threshold of 0.5 were used as input data. To evaluate the forecasting performance of each mode, we
used five evaluation criteria in the experiments: mean absolute
III. E XPERIMENT D ESIGN AND R ESULTS error (MAE), mean square error (MSE), coefficient of deter-
A. Data Description mination (R2 ), mean absolute percentage error (MAPE), and
adjusted coefficient of determination (R2adj ). Here, we assume
In this brief, we selected four stocks listed on the Shenzhen
the dataset has n samples and these criteria can be defined as
Stock Exchange or Shanghai Stock Exchange as illustrative
follows:
examples. Research data was collected from September 2010
1 
n
to September 2021, including approximately 2500 trading 2
MSE = ŷi − yi ,
days. The entire data set was divided into two parts. 80% of the n
i=1
dataset was used as the training dataset, while the remaining
1
n
20% was acted as the test dataset. MAE = ŷi − yi ,
Based on the knowledge of the stock market, we obtained n
i=1
data sets from multiple sources to improve the forecasting 100%  ŷi − yi
n
ability [24]. Historical data and financial indicators of each MAPE = ,
n yi
stock were fetched from BaoStock’s API interface. As the i=1

stock market index reflects the closing price of the company to (yi − ŷi )2
some extent, we collected the Shenzhen Securities Component R = 1− 
2
,
(yi − ȳ)2
Index (SZI) and Shanghai Securities Composite Index (SSEC)
(1 − R2 )(n − 1)
from the NetEase website. Finally, we derived technical indi- R2adj = 1 − , (7)
cators and integrated historical data into the machine learning n−k−1
models. The input variables used in this brief, including their where yi and ŷi represent the actual and predicted value,
formula or description, are summarized in Table I. As the respectively. ȳ is the average of the close price of the n
range of different input features varies inconsistently, these samples, and k denotes the number of features.
datasets were scaled into the range of [0, 1] through the Min-
Max normalization technique. xnorm is given by the following C. Correlation Analysis
equation: Tables II, III, IV, and V show the input features of the
x − xmin PCC-BLS for four experimental stocks. Here, we chose fea-
xnorm = , (6)
xmax − xmin tures with an absolute value of PCC greater than 0.5 as input

Authorized licensed use limited to: SOUTH CHINA UNIVERSITY OF TECHNOLOGY. Downloaded on June 19,2022 at 12:02:08 UTC from IEEE Xplore. Restrictions apply.
 2416 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 69, NO. 5, MAY 2022

TABLE IV TABLE VIII

SH.600050: T HE I NPUT F EATURES OF THE PCC-BLS M ODEL SH.600019: T HE E VALUATION VALUE OF D IFFERENT M ODELS

TABLE V
SH.600837: T HE I NPUT F EATURES OF THE PCC-BLS M ODEL

TABLE IX
SH.600050: T HE E VALUATION VALUE OF D IFFERENT M ODELS

TABLE VI
SZ.000776: T HE E VALUATION VALUE OF D IFFERENT M ODELS

of sh.600019, the PCC-BLS has 9% less MSE compared to

the second-ranked model BLS. The R2 of the proposed model
TABLE VII is 0.987, which is an outstanding fit. In the case of sh.600050,
SH.600837: T HE E VALUATION VALUE OF D IFFERENT M ODELS compared to the second-ranked model Bagging, PCC-BLS
achieves 14% decreases of MSE, and 0.3% increase of R2 ,
respectively. In the case of sh.600837, compared to the second-
ranked model Bagging, PCC-BLS achieves 4.8% decreases of
MSE, and 0.8% increase of R2 , respectively. The results show
that the PCC-BLS network achieves higher performance in the
stock price forecasting problem. The forecasting results of the
four stocks by the PCC-BLS model are shown in Fig. 2. In
addition, PCC-BLS also can save lots of training time. The
running time of training the PCC-BLS model for the four
stocks was about 1.14s, 1.34s, 1.95s, and 2.07s, respectively,
which is faster than most of the other models. For instance, the
features for the PCC-BLS. It can be seen that the stock price GRU model consumes 82.35s, 104.56s, 110.78s, and 102.41s,
data and the derived technical indicators of the experimental respectively, for training.
stocks are correlated with their closing prices to some extent.
In addition, there was also a strong correlation between the
close price of the experimental stock and the SSEC or SZI. IV. C ONCLUSION
Actually, we also adopt the threshold as 0.4∼0.6. Of the above The main purpose of this brief is to help investors make rea-
range of thresholds, setting the threshold at 0.5 gives the best sonable trading decisions by accurately predicting stock price
predictions. movements using machine learning methods. In this brief, we
proposed a new framework composed of PCC and BLS, and
D. Experimental Results applied this framework for a short-term prediction of stock
Tables VI, VII, VIII, and IX show the experimental results price recorded on the Shenzhen Stock Exchange or Shanghai
of the performance of the ten other models with the PCC-BLS Stock Exchange. As feature selection aims at selecting more
for four experimental stocks. It can be seen that the PCC- representative features for enhancing forecasting performance,
BLS outperformed the other models on the four experimental we adopted suitable input variables by PCC from 35 variables.
stock prediction results. In the case of sz.000776, compared The combinations of input variables were then input to the
to the second-ranked model BLS, PCC-BLS has 10% less BLS for training. Finally, we compared the proposed model
MSE. The R2 is improved from 0.945 to 0.950 through the with ten machine learning methods without feature selection
Pearson Correlation Coefficient screening feature. In the case on five evaluation metrics. The experimental results showed

Authorized licensed use limited to: SOUTH CHINA UNIVERSITY OF TECHNOLOGY. Downloaded on June 19,2022 at 12:02:08 UTC from IEEE Xplore. Restrictions apply.
 LI et al.: PCC-BASED PERFORMANCE ENHANCEMENT OF BLS 2417

R EFERENCES
[1] W. Lu, J. Li, J. Wang, and L. Qin, “A CNN-BiLSTM-AM method
for stock price prediction,” Neural Comput. Appl., vol. 33, no. 10,
pp. 4741–4753, 2021.
[2] C. K. Tse, J. Liu, and F. C. M. Lau, “A network perspective of the stock
market,” J. Empirical Financ., vol. 17, no. 4, pp. 659–667, 2010.
[3] X. F. Liu and C. K. Tse, “A complex network perspective of world
stock markets: Synchronization and volatility,” Int. J. Bifurcation Chaos,
vol. 22, no. 6, 2012, Art. no. 1250142.
[4] G. Chen, Y. Lou, and L. Wang, “A comparative study on controllability
robustness of complex networks,” IEEE Trans. Circuits Syst. II, Exp.
Briefs, vol. 66, no. 5, pp. 828–832, May 2019.
[5] M. Small and C. K. Tse, “Determinism in financial time series,” Stud.
Nonlinear Dyn. Econom., vol. 7, no. 3, pp. 1–29, 2003.
[6] J. Dong, W. Dai, Y. Liu, L. Yu, and J. Wang, “Forecasting chinese
stock market prices using baidu search index with a learning-based data
collection method,” Int. J. Inf. Technol. Decis. Making, vol. 18, no. 5,
pp. 1605–1629, 2019.
[7] D. Zhang and S. Lou, “The application research of neural network and
BP algorithm in stock price pattern classification and prediction,” Future
Gener. Comput. Syst., vol. 115, pp. 872–879, Feb. 2021.
[8] Q. Li et al., “Integrating reinforcement learning and optimal power dis-
patch to enhance power grid resilience,” IEEE Trans. Circuits Syst. II,
Exp. Briefs, vol. 69, no. 3, pp. 1402–1406, Mar. 2021.
[9] X. Zhang et al., “Braess paradox and double-loop optimization method
to enhance power grid resilience,” Rel. Eng. Syst. Safety, vol. 215, Nov.
2021, Art. no. 107913.
[10] T. Fischer and C. Krauss, “Deep learning with long short-term memory
networks for financial market predictions,” Eur. J. Oper. Res., vol. 270,
no. 2, pp. 654–669, 2018.
[11] B. N. G. Koneru and V. Vasudevan, “Sparse artificial neural networks
using a novel smoothed LASSO penalization,” IEEE Trans. Circuits Syst.
II, Exp. Briefs, vol. 66, no. 5, pp. 848–852, May 2019.
[12] L. Bai, Y. Zhao, and X. Huang, “A CNN accelerator on FPGA using
depthwise separable convolution,” IEEE Trans. Circuits Syst. II, Exp.
Briefs, vol. 65, no. 10, pp. 1415–1419, Oct. 2018.
[13] A. R. Mohamed, L. Qi, Y. Li, and G. Wang, “A generic nano-Watt power
fully tunable 1-D Gaussian kernel circuit for artificial neural network,”
IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 67, no. 9, pp. 1529–1533,
Sep. 2020.
[14] Q. Zhuge, L. Xu, and G. Zhang, “LSTM neural network with emo-
tional analysis for prediction of stock price,” Eng. Lett., vol. 25, no. 2,
pp. 167–175, 2017.
[15] A. Thakkar, D. Patel, and P. Shah, “Pearson correlation coefficient-
based performance enhancement of vanilla neural network for stock
trend prediction,” Neural Comput. Appl., vol. 33, pp. 1–16, Jul. 2021.
[16] B. Lim and S. Zohren, “Time-series forecasting with deep learn-
ing: A survey,” Philos. Trans. Roy. Soc. A, vol. 379, no. 2194, 2021,
Art. no. 20200209.
[17] C.-F. Tsai and Y.-C. Hsiao, “Combining multiple feature selection meth-
ods for stock prediction: Union, intersection, and multi-intersection
approaches,” Decis. Support Syst., vol. 50, no. 1, pp. 258–269, 2010.
[18] D. Liu, C. K. Tse, and X. Zhang, “Robustness assessment and enhance-
ment of power grids from a complex network’s perspective using
decision trees,” IEEE Trans. Circuits Syst. II, Exp. Briefs, vol. 66, no. 5,
pp. 833–837, May 2019.
[19] L.-P. Ni, Z.-W. Ni, and Y.-Z. Gao, “Stock trend prediction based on
fractal feature selection and support vector machine,” Expert Syst. Appl.,
Fig. 2. Actual and predicted price of the four experimental stocks using vol. 38, no. 5, pp. 5569–5576, 2011.
PCC-BLS. [20] J. Yang, R. Rao, P. Hong, and P. Ding, “Ensemble model for stock price
movement trend prediction on different investing periods,” in Proc. 12th
Int. Conf. Comput. Intell. Security (CIS), 2016, pp. 358–361.
[21] C.-L. Huang and C.-Y. Tsai, “A hybrid SOFM-SVR with a filter-
based feature selection for stock market forecasting,” Expert Syst. Appl.,
that PCC-BLS provided more accurate forecasting of future vol. 36, no. 2, pp. 1529–1539, 2009.
values than other prediction models mentioned above. [22] C. L. P. Chen and Z. Liu, “Broad Learning System: An effective and
However, in this brief, we are short on the practical appli- efficient incremental learning system without the need for deep architec-
ture,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 1, pp. 10–24,
cation of the proposed model. In the future, we will adopt Jan. 2018.
an actual stock investment strategy and evaluate the practi- [23] K. Pearson, “Contributions to the mathematical theory of evolution,”
cal application of the model through returns. Future research Philos. Trans. Roy. Soc. London A, vol. 185, pp. 71–110, Dec. 1894.
[24] A. Thakkar and K. Chaudhari, “Fusion in stock market prediction: A
will also focus on the comparison of different feature selection decade survey on the necessity, recent developments, and potential future
methods. directions,” Inf. Fusion, vol. 65, pp. 95–107, Jan. 2021.

Authorized licensed use limited to: SOUTH CHINA UNIVERSITY OF TECHNOLOGY. Downloaded on June 19,2022 at 12:02:08 UTC from IEEE Xplore. Restrictions apply.
View publication stats