Received: 20 February 2021

DOI: 10.1049/cit2.12059

ORIGINAL RESEARCH
--Revised: 12 June 2021 Accepted: 25 July 2021 CAAI Transactions on Intelligence Technology

Stock market prediction using deep learning algorithms

Somenath Mukherjee1 | Bikash Sadhukhan2 | Nairita Sarkar2 | Debajyoti Roy2 |

Soumil De2

1
Nazrul Center of Social and Cultural Studies, Kazi Abstract
Nazrul University, Asansol, India
The Stock Market is one of the most active research areas, and predicting its nature is an
2
Department of Computer Science and epic necessity nowadays. Predicting the Stock Market is quite challenging, and it requires
Engineering, Techno International New Town
(Techno India College of Technology), Kolkata, intensive study of the pattern of data. Specific statistical models and artificially intelligent
India algorithms are needed to meet this challenge and arrive at an appropriate solution.
Various machine learning and deep learning algorithms can make a firm prediction with
Correspondence minimised error possibilities. The Artificial Neural Network (ANN) or Deep Feed‐
Bikash Sadhukhan, Department of Computer forward Neural Network and the Convolutional Neural Network (CNN) are the two
Science and Engineering, Techno International New
Town (Techno India College of Technology), 1/1, network models that have been used extensively to predict the stock market prices.
Service Rd, DG Block(Newtown), Action Area I, The models have been used to predict upcoming days' data values from the last few days'
Newtown, Chakpachuria, West Bengal 700156, data values. This process keeps on repeating recursively as long as the dataset is valid.
Kolkata, India.
Email: [email protected] An endeavour has been taken to optimise this prediction using deep learning, and it has
given substantial results. The ANN model achieved an accuracy of 97.66%, whereas the
CNN
model achieved an accuracy of 98.92%. The CNN model used 2‐D histograms generated
out of the quantised dataset within a particular time frame, and prediction is made on that
data. This approach has not been implemented earlier for the analysis of such datasets. As
a case study, the model has been tested on the recent COVID‐19 pandemic, which caused
a sudden downfall of the stock market. The results obtained from this study was
decent enough as it produced an accuracy of 91%.

KEYW ORDS
artificial neural network, convolutional neural network, nifty, stock market

1 | INTRODUCTION crucial role in predicting the stock prices and, hence, deter-
mining an accurate result.
The stock market refers to the collection of markets and ex- A large portion of the Indian stock market trading occurs
change centres where economic activities like buying, selling, on two stock exchanges: the Bombay Stock Exchange
and deploying shares of publicly held companies take place. (BSE) and the National Stock Exchange (NSE). Both the
Such financial practices are conducted through institutionalised exchanges pursue the same trading method, trading hours, and
formal exchanges through over‐the‐counter marketplaces that settlement process. BSE has more than 5000 listed firms,
function under a defined set of regulations. The stock market is whereas its counterpart NSE has about 1600 enlisted firms.
a very dynamic and uncertain field, so the stock market's Out of all the listed firms on the BSE, only about 500 firms
prediction naturally becomes a burning topic. Due to the constitute more than 90% of its market capitalisation; the rest
advancement of computational power in recent times, pre- consists of highly illiquid shares. Almost all the major and
dicting the stock market has been much faster and accurate. giant trade firms of India are enlisted on both exchanges. Both
Artificial Intelligence and machine learning models play a exchanges compete for the order flow that results in reduced
costs, market

-
This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial‐NoDerivs License, which permits use and distribution in any medium, provided the
original work is properly cited, the use is non‐commercial and no modifications or adaptations are made.
© 2021 The Authors. CAAI Transactions on Intelligence Technology published by John Wiley & Sons Ltd on behalf of The Institution of Engineering and Technology and Chongqing
University of Technology.
82 CAAI Trans. Intell. Technol. 2023;8:82–94. wileyonlinelibrary.com/journal/cit2
 MUKHERJEE ET
■ 8
efficiency, and innovation. The presence of arbitrageurs keeps
even during extreme market fluctuations like in the COVID‐19
the prices on the two stock exchanges within a very close
pandemic during March and April 2020.
range.
The whole work is categorised into several sections. The
The stock market is frequently used as a sentiment
Related Work and the experimental dataset are discussed in
indicator and can have an impact on GDP (gross domestic
Sections 2 and 3, followed by Section 4, which presents the
product). GDP is a metric that measures an economy's
proposed model used to generate the predictions. Section 5
total output of goods and services. As the stock market
gives a detailed description of the system configuration used
rises and falls, so does economic sentiment [1]. People's
for this study. The empirical and graphical results of the pre-
spending fluctuates in response to changes in emotion,
dictions, along with a case study, are shown in Section 6.
which drives GDP growth [2]. The stock market, on the
Section 7 argues the advantages of the CNN method over
other hand, can have both positive and negative effects on
the backpropagation ANN method and finally Section 8
GDP. GDP is usually expressed as a percentage increase
provides the concluding remarks.
from one period to the next. For example, if the quarter‐ to‐
quarter growth rate is 2%, that means the economy
increased by 2% on an annualised basis in that quarter. It
is important to remember that India's economy is mostly 2 | RELATED WORK
based on consumption. Private consumption spending, or
money spent by the residents on goods and services, has Several researchers implemented their work to provide accu-
accounted for approximately three‐fifths of the Indian rate solutions to this dynamic problem and have proposed
economy through the years [3]. various methods for predicting the stock market. Jayanth Balaji
Figures 1 and 2 describe the growth of the Indian Gross [6] performed a deep learning method to predict a
Domestic Product (GDP), which is analysed through company's stock price using 14 different deep learning
machine learning algorithms. It shows the predictions on the methods. Similar work is implemented using Artificial Neural
GDP, leading to a better understanding of any country's Networks (ANN) by Tsong Wuu Lin [7, 8]; his work tried to
socio‐ economic situation. Machine learning and deep learning maximise the profitability using this model [9]. Autoregressive
algo- rithms [4] play a vital role in predicting a nation's models are powerful models for predicting the stock market,
economy. they give a strong insight on time series analysis and make
These predictions are essential for making correct financial very accurate predictions [10, 11]. Sentiment analysis is also
decisions [5]. The motive of this study is to provide accurate one of the strong ways to predict the stock market. Social
predictions of the stock market values for consecutive days, media analytics plays a vital role in sentiment analysis.
ARIMA model helps in

FIGURE 1 The pattern of growth of the Indian Gross Domestic Product (GDP) analysed through machine learning
 8 ■ MUKHERJEE ET

FIGURE 2 The time‐series data that has been used in this study

sentiment analysis and predicting time series data [12–14]. has been implemented, using a convolutional neural network
Sentiment analysis can also be implemented by using deep (CNN) model. This model generates 2‐D histograms and
learning models like CNN and LSTM [15]. For better accuracy feeds them to the neural network. Such an approach in-
and better use of features, deep learning methods are preferred creases the efficiency of training as well as prediction.
over supervised machine learning methods. Boosted Decision LSTMs and Transformer models are used to generate
Tree model [16, 17]and ELSTM model [18, 19] give valuable context from sequences which are long. The efficiency of
insight on the empirical results of the predictions but are un- the LSTM models and Transformer models are significantly
able to provide accurate results during fluctuating scenarios reduced when the sequence length is reduced. [26–29]. The
like in the COVID‐19 case. Along with deep learning sequence length used in the CNN model is considerably
algorithms, a powerful model known as convolution neural lesser than the ones used for LSTM and transformer
network (CNN) is advantageous and accurate to solve models. The problem that has been addressed here is a
problems of this genre [20, 21]. CNN plays a significant regression problem but a model that has mostly been used
role in the fields of image processing, pattern recognition,
for classification problems has been used for this purpose,
and in analysing time series data. Traditionally, a
that is without implementing known Recurrent Neural
Convolutional Neural Network input is often a 2‐D image,
Networks like LSTM and Transformers, an endeavour has
but that does not mean this model cannot be used for
predicting time series data [22]. There are two ways to pre‐ been taken to implement the same using CNN so that we
process the data, one way is to convert the 1‐D‐input data can prove that we can obtain results of similar stature too.
into a 2‐D matrix, and the other way is to take advan- tage It has given accurate and concrete results even during
of the 1‐D function to help do the convolutional extreme stock market fluctuations like the COVID‐19
computation [23]. pandemic situation in March and April 2020. As compared
Based on the above discussions, an endeavour has been to two recent works [30, 31], our proposed model gives an
taken to implement deep learning methods to predict the accuracy of 97.66% using Backpropagation Artificial Neural
nature of stock market prices. National Stock Exchange Network and 98.92% using Convolutional Neural Networks.
(NSE) [24]stock market dataset is used for predicting the
stock market values [25]. As the National Stock Exchange
(NSE) dataset provides short settlement cycles and very 2.1 | Backpropagation on feed‐forward
high transaction time, the National Stock Exchange neural network
dataset can be computationally efficient for analysis
purposes. Two conventional machine learning approaches are The most fundamental unit of an Artificial Neural Network
implemented. The first approach uses the backpropagation is an artificial neuron that takes in multiple inputs,
algorithm on a simple artificial neural network. While this multiplies them by the weights assigned, and adds a bias. The
approach is quite efficient in predicting future price values, result is fed to an activation function whose output
another approach determines to what
 MUKHERJEE ET
■ 8
level the neuron will be triggered, and the corresponding value
Activation value at hidden layer 2:
is passed on to the neurons in the next layer.
Figure 3 describes the model of an artificial neuron
a3 ¼ φðz3Þ
where
{ x1, x2, x3,…, xn } is the set of inputs, { w1, w2, w3, …,wn, } Output layer:
is the set of weights associated with the input nodes, φ is the
activation function, and y is the output that can be calculated y ¼ w3a3 ð4Þ
by applying the activation function over the net input.
The following equation governs the firing of an artificial Once the output is generated, a cost function is evaluated
neuron: to analyse the difference between the actual and predicted
�� ! outputs.
Xn i þ ð1Þ Cost/Loss function:
y ¼φ xi w bias
!

i¼ �
1
1 XN
ypredicted – y �2
xi is the neurons' inputs in the previous layer, and wi are the CostðC Þ ¼ actual ð5Þ
corresponding weights. φ is the activation function. N i¼1
Considering an ANN with two hidden layers along with an
input layer and an output layer, the equations would be as where, N is the number of inputs' combinations that have been
follows: fed forward through the network.
Notations: This cost function is used to backpropagate the errors
and in turn tweak the weights and biases present in the
k
� A represents the matrix containing the output of the network to minimise the cost function for better prediction.
neurons after activation at layer k. The concept of gradient descent is used to achieve it. The
k
� W represents the weight matrix in‐between layer k‐1 and k. sensitivity of the cost function concerning a particular
k parameter is analysed by calculating the cost function's gradient
● b represents the bias matrix at hidden layer k.
� Z krepresents the matrix containing the neuron value at for that parameter. The parameter is then tweaked
layer k. accordingly.
∂C
Input layer: xupdated ¼ x − η ð6Þ
∂x
x ¼ a1 where, η signifies the learning rate, C signifies the cost func-
tion, and x signifies any parameter (weight or bias) in the
Neuron value at hidden layer 1:
network.
z2 ¼ w1x þ b1 ð2Þ This process of forwarding propagation and back-
propagation continues until the cost function is minimised and
the model can give accurate results.
Activation value at hidden layer 1:
In this project, the backpropagation algorithm is imple-
mented on a deep feed‐forward neural network to predict
a2 ¼ φðz2Þ the upcoming days' time‐series data value from the previous
two days' data values.
Neuron value at hidden layer 2:

z3 ¼ w2a2 þ b2 ð3Þ 2.2 | Convolutional neural network (CNN)

A convolutional neural network (CNN, or Conv‐Net) is a type
of deep neural network [33, 34] that is most typically used
to analyse and conduct Artificial Intelligence operations on
visual images.
Because of their shared‐weights architecture and trans-
lation invariance qualities, they are also known as shift
invariant or space invariant artificial neural networks
(SIANN). They have image and video recognition
applications, recommender systems, image classification,
medical image analysis, and nat- ural language processing.
The name ‘‘convolutional neural network’’ signifies that
FIGURE 3 An artificial neuron
the neural network uses a mathematical operation known as
convolution. Convolution is a specialised kind of linear
oper- ation. Convolutional networks are simply neural
 8 ■
networks that
MUKHERJEE ET
 MUKHERJEE ET
■ 8
employ convolution instead of general matrix multiplication in
days, passed to a convolution neural network to predict the
at least one of their layers.
time‐series data value (stock prices) of the following day. The
A convolutional neural network consists of an input and an
convolution neural network (CNN) model gives a new angle to
output layer and multiple hidden layers. The hidden layers of a
look at stock market data and gives greater accuracy in the
CNN typically consist of a series of convolutional and pooling
prediction. The system configuration for implementing these
layers arranged in a particular pattern that is eventually flat-
models is discussed in the section below.
tened and followed by a fully connected network. These layers
are referred to as the hidden layers because the activation
function and final convolution mask their inputs and
3 | EXPERIMENTAL DATASET
outputs. The final convolution, in turn, often involves
backpropagation to weight the end product more accurately.
The experimental dataset that has been used is the National
The convolution layers consist of filters or kernels of
Stock Exchange (NSE) stock market dataset, specifically the
spe- cific dimensions smaller than the image dimensions,
NIFTY price index, ranging in the time frame of April 2008
passed over the entire image to extract the features. The filters
to April 2018, collected from the National Stock Exchange
may be pre‐determined like the ones to detect edges and
(NSE) India website [32]. The recent dataset of NSE India
various shapes or may be learnt by the model in the training
process. A stride length is also applied to the image along with from 1 November 2019 to 5 August 2020 is taken into ac-
the filter. Stride is count from the same source as a case study. The working
the number of pixels that can be shifted over the input dataset has seven features in it, out of which five features
image. Figure 4 shows the convolution operation on the have been taken into consideration for this project. The
image features that have been taken into account are ‘Date,’
�with the 7 7 dimension with filter
size 3 3 and stride length = 2. Pooling layers are introduced ‘Open,’ ‘Close,’ ‘High,‘, ‘Low’. A sample from the original
to extract essential features or downsize the extensive features dataset used in the work has been provided below in Table
extracted by the convolution layers. In most cases, max‐ 1. Initially, all the features are plotted to generate a pattern
pooling or average‐pooling is used. The fully connected on
network then works on the reduced number of features how the stock data moves along with time. The plots of each
hence producing efficient predictions quickly. Figure 5 feature are shown above in Figure 2. The approach using 2‐
describes the max pooling operation with D histograms has not been implemented earlier; hence, an
filter size 3 � 3 and stride length = 2. endeavour has been taken to predict a day's stock market
This study focusses on generating synthetic images in 2‐D prices based on the recent past data. These approaches
histograms [35] from stock market data spanning over a few proved to work well with the prediction of time‐series data.
Based on this movement pattern, two deep learning models
have been devised to predict future data values considering the
present values. The models that have been used in this study
are discussed in the following section.

4 | PROPOSED MODEL
4.1 | Using the ANN with backpropagation
algorithm

This study implements the backpropagation algorithm on a

deep feed‐forward neural network to predict stock market data.
The model involves predicting the ‘Open,’ ‘Close,’ ‘High,’
and ‘Low’ values of NIFTY of the nth day based on these four
F I G U R E 4 Convolution operation with kernel size = (3 � 3)
and stride length = 2
TABL E 1 Raw data sample

F I G U R E 5 Max pooling with kernel size = (2 � 2) and

stride length = 2
Date Open High Low Close
03‐Jan‐08 6184.25 6230.15 6126.4 6178.55
8 ■
04‐Jan‐08
MUKHERJEE ET

6179.1 6300.05 6179.1 6274.3

07‐Jan‐08 6271 6289.8 6193.35 6279.1
08‐Jan‐08 6282.45 6357.1 6221.6 6287.85
09‐Jan‐08 6287.55 6338.3 6231.25 6272
10‐Jan‐08 6278.1 6347 6142.9 6156.95
 MUKHERJEE ET
■ 8
features of the n − 1 th and n − 2 th days. The model is
designed to haveð eight Þinput neurons
ð to accept the values of
4.2 | Using the convolutional neural
Þ network with 2‐D histograms
each of the four features of the previous two days, as shown
in Figure 6. The input layer is followed by two hidden layers
The study was further taken to the implementation of deep
con- sisting of 2 neurons, each, in turn, followed by the output
convolution neural networks in the field of stock market
layer with four nodes to predict the following day's values.
prediction.
The number of hidden layers and the number of neurons in
each of the two hidden layers were fixed at two after
careful experi-
mentation with higher values, which often led to over‐fitting
4.2.1 | Network Model
and
decreased prediction accuracy. The activation function that was The network model designed for the prediction is shown in
Figure 7. The model takes the 20 20
� matrix generated from a 2‐
used in the particular model is the sigmoid function. The model
D histogram as input and generates a fraction as output. The
was trained with 2000 input combinations for more than 10,000 predicted fraction multiplied with the nð − 1 th day gives the
epochs with a learning rate of 0.1. The model was then predicted data value of the nth day.
tested with more than 2500 input combinations of stock price
data.
The accuracy of each prediction was calculated using the 4.2.2 | Training
following formula:
� � This CNN model's training process was much faster and
|tar geti − outputi |
accuracyi ðin %Þ ¼ 1 − ∗ 100 efficient than the previously discussed ANN model. Only
targeti 1000 input combinations and just about 150 epochs were
ð7Þ
sufficient for the desired training, which is way less than
what was required for the ANN model. The Adam opti-
miser was used for the training, and a learning rate of 0.01
proved to be good enough for proper convergence of the
cost function. The cost function that was used is the Mean
Squared Error (MSE). Both MSE and Mean Absolute Error
(MAE) were used to judge the model's performance during
the training.

1 X 2
MSE ¼ ðy − ^yÞ
n
1X
MAE ¼ |y − ^y|
n
FIGURE 6 ANN model used in this prediction

FIGURE 7 CNN model used in this study

 9 ■ MUKHERJEE ET

Here n is the number of training examples, y is the actual

output value, and ^y is the predicted output value.

5 | EXPERIMENTAL SETUP AND

SIMULATION

Themodels were trained and tested using the Keras API on top
of Theano in Python version 3.6.8. A machine with the
Windows 7 Operating Systemwasusedfor the entire study.
Thesystemuseda dual‐core Intel(R) Core(TM) 2
Duoprocessorcoupledwith 2 GB RAM and 829 MB of
Graphics memory. The Graphics chipset used was an Intel(R) 4
Series Express Chipset. This hardware and software
configuration took around 3 h to train the Artificial
Neural Network model for 10,000 epochs, and it took about 1 h
30 min to train the CNN model for 150 epochs. The results ob-
tained after implementing the model on this system is discussed
below.

5.1 | Image Preparation

A 2‐dimensional histogram also known as a density heat‐
map is generated by grouping into bins multiple points
according to their x and y coordinates. In a 1‐dimensional
histogram, the frequency of points belonging to a bin is
characterised by the bar's height corresponding to that bin.
Whereas in a 2‐D histogram, the colour corresponding to
a bin signifies the frequency of points mapped to that
bin. Thus, 2‐D histograms are used to visualise the den-
sity of points in a 2‐D plane. In this study, a 2‐D his-
togram has been generated from the stock market data.
Here, a point in the 2‐D plane represents the stock price
index of two consecutive days. The x‐coordinate gives the
price index of day one, and the y‐coordinate gives the F I G U R E 8 (a) and (b). Sample 2‐D histograms generated from
price index of day two. Fifty such sets of points are the time‐series data
considered for generating a single 2‐D histogram. A grey‐
scale 2‐D histogram has been used in this study. Hence,
the whiter a bin, the denser it is. A 2‐D histogram ex- 6 | RESULTS AND ANALYSIS
plains how the price index varied over consecutive days
for the last 50 days. This insight helps to predict the 6.1 | Using the ANN with backpropagation
stock price index of the next day. algorithm
The model that was designed can predict the ‘Open,’
‘Close,’ ‘High,’ and ‘Low’ values of NIFTY, or in general, any The model was tested with more than 2500 input combinations
time‐series data value of the next day based on the corre- of stock price data, and it predicted the outputs with an ac-
sponding values of the last 51 days. curacy of 97.66%.
In our work, a 2‐D greyscale histogram is generated by The time versus feature plot shows how the data moves
plotting any of the ‘Open,’ ‘Close,’ ‘High,’ and ‘Low’ values according to the prescribed model. It also correlates the model
of the n − 51 th day to n − 2 th day on the x‐axis and the with the expected output against the received output. Its ac-
ð Þ ð
values of the corresponding succeeding day n − 50 th curacy layup structure has accompanied each graph.
Þ ðð
day to thn − 1 day in the y‐axis. Each 2‐D histogram has 20
ðbins along
Þ the two axes, as shown in Figure 8a and b. Each The time versus feature graphs based on the ANN model
for the features ‘Open,’ ‘Close,’ ‘High,’ and ‘Low’ is shown in
such 2‐Þ D histogram gives rise to a 20 20 matrix where the (i;
� values in the ith and jth bins'
j)th cell corresponds to the pixel Figure 9a–d respectively.
inter- section along the x and y‐axis, respectively. Each such Figure 10a–d shows the time versus prediction accuracy
matrix is fed to the deep convolution network to predict the layup of the features ‘Open,’ ‘Close,’ ‘High,’ and ‘Low,’
nth day's respectively, based on the ANN model.
corresponding feature. The above graphs clearly show that the model's pre-
dictions were very close to the actual values. The ‘Open’
MUKHERJEE ET
■ 9

FIGURE 9 (a–d). The prediction of all the features based on the ANN
F I G U R E 10 (a–d). The accuracy layup of all the features based on the
model
ANN model

feature was predicted with an average accuracy of respectively. Next, the same predictions were carried out
97.87%, followed by the ‘High,’ ‘Low,’ and ‘Close’ fea- using Convolutional Neural Networks to obtain even
tures with accuracies of 97.77%, 97.53%, and 97.47%, better results.
 9 ■ MUKHERJEE ET

6.2 | Using the convolutional neural

network with 2‐D histograms 6.3 | Quantisation table

The MSE and MAE metrics were plotted against the Table 2 shows that a clear picture is obtained on how the
number of epochs, and they converged to sufficiently small accuracy of the CNN model on the given dataset is ob-
values, as shown in Figure 11 below. tained. The entire dataset is broken down into 15 segments
randomly. The CNN model is implemented on each
segment to determine the highest accuracy in that particular
6.2.1 | Testing segment. In each segment, the maximum and minimum
value of the feature (‘High’ and ‘Low’) is taken into
consideration within that time frame, which is then
The model was tested with more than 1600 input combi-
compared with the corresponding predicted maximum and
nations or matrices. The training data of 1000 data points
minimum value. Based on this, the accuracy is generated.
was also put to test for analysing the performance of the
Taking the ‘High’ and ‘Low’ features into consideration, the
model based on known and unknown data points. The ac-
CNN model's accuracy was found to be 99.34%. The total
curacy was calculated in the same way as it was done for the
accuracy is calculated by only taking the arithmetic mean of
ANN model. Accuracy on the training set and the testing
the accuracy of all segments.
set was found to be close. The ‘Low’ feature was predicted
Statistically, overfitting can be defined as producing an
most accurately with a prediction accuracy of 99.06%, fol-
analysis that corresponds precisely to a particular dataset and
lowed by ‘Close’ at 99.03%, ‘High’ at 98.94%, and ‘Open’ at
may therefore fail to fit additional data or fail to predict
98.63%.
future observations on unknown data reliably. An overfitted
The results that were obtained after the prediction of the
model is a statistical model that contains more parameters than
features ‘Open,’ ‘Close,’ ‘High,’ and ‘Low’ are depicted in
that can be justified by the data and hence has higher model
Figure 12a–d, respectively.
complexity. The essence of overfitting is to have
Figure 13a–d show time versus prediction accuracy layup
unknowingly extracted some of the residual variations as if
of the features ‘Open,’ ‘Close,’ ‘High,’ and ‘Low,’
that variation represented the underlying model structure.
respectively, based on the CNN model.
Regularisation is a process in which extra information is
The prediction graphs show that the CNN model used
the 2‐D histograms as input predicted even better than the added to the current analysis to solve an ill proposed (incorrect
ANN model. The ‘target’ and ‘output’ plots seem to almost model usage) problem or prevent overfitting. The Regularisa-
coincide with each other. The accuracy graphs validate the tion method is used extensively in this study to address the
same as we see that the accuracy remained between 98% problem of overfitting. The model is experimented with a
and 100% for most of the time. range of λ values, which is known as the regularisation
parameter. An optimal λ value was chosen based on the

FIGURE 11 The convergence of Mean Squared Error (MSE) and Mean Absolute Error (MAE)
 MUKHERJEE ET
■ 9

FIGURE 12 (a–d). The prediction of all features based on the CNN

F I G U R E 13 (a), (b), (c), (d). The accuracy layup of all the features
model
based on the CNN model

performance of the model on the cross‐validation set. This

difference and were found to be sufficiently high. This finding
eliminated the chances of the model having high variance. The
validates the fact that the model neither suffered from over-
model's training accuracy and testing accuracy had a negligible
fitting nor underfitting.
 9 ■ MUKHERJEE ET

TABLE 2 Accuracy determination by quantisation of the dataset into 15

segments

‘High’ price index (Rs) ‘Low’ price index (Rs)

Stock market data Actual value Predicted value Accuracy (%) Actual value Predicted value Accuracy (%)
(Apr‐2008)–(Nov‐2008) 5298.85 5297.73 99.97886334 2252.75 2251.84 99.95960493
(Dec‐2008)–(Jun‐2009) 4693.2 4666.45 99.43002642 2539.45 2541.03 99.93778180
(Jul‐2009)–(Dec‐2009) 5221.85 5183.20 99.25984086 3918.75 3886.11 99.16708134
(Jan‐2010)–(Jul‐2010) 5447.5 5477.13 99.99324509 4675.4 4658.97 99.64858622
(Aug‐2010)–(Apr‐2011) 6338.5 6436.02 98.46146565 5177.7 5176.29 99.79894548
(May‐2011)–(Dec‐2011) 5775.25 5845.08 98.79087485 4531.15 4498.00 99.26839765
(Jan‐2012)–(Aug‐2012) 5629.95 5639.80 99.82504285 4675.8 4610.96 98.61328543
(Sep‐2012)–(May‐2013) 6229.45 6377.19 97.62836205 4888.2 4870.71 99.64219958
(Jun‐2013)–(Jan‐2014) 6415.25 6476.07 99.05194653 5118.85 5114.17 99.90857321
(Feb‐2014)–(Dec‐14) 8626.95 8713.91 98.99199601 5933.3 5922.29 99.81443716
(Jan‐2015)–(Oct‐2015) 9119.2 9130.92 99.87147995 7539.5 7524.16 99.79653823
(Nov‐2015)–(Sep‐2016) 8968.7 9059.88 98.98335322 6825.8 6828.34 99.96278824
(Oct‐2016)–(Apr‐2017) 9367.15 9623.66 97.26160038 7893.8 7737.58 98.02097849
(May‐2017)–(Feb‐2018) 11,171.55 11,180.32 99.92149702 9269.9 9243.23 99.71229463
(Mar‐2018)–(Dec‐2018) 11,760.2 11,758.02 99.98146290 9951.9 9904.56 99.52431194

The same model has been implemented for predicting

time frame of 15th January to 5th August. The sudden
the stock market values during a situation of extreme market
downfall in the accuracy occurred because of the heavy
fluctuation, that is, during the peak phase of the COVID‐19
market fluctu- ations during late March. However, the
pandemic in March and April 2020. During that time, the
model adjusted with the fluctuations and provided mostly
stock market was extraordinarily motile, and severe fluctuation accurate predictions, resulting in an aggregate of 91%
was observed. The study shown below provides an insight into accuracy.
how a situation of this sort can be predicted.

6.5 | Comparison table

6.4 | Case study
From the comparison Table 3 shown above, a clear picture
To test the effectiveness and accuracy of the model in unex- is received on the level of accuracy obtained from the model
pected situations, it was made to predict every day ‘Open,’
used in this study compared to the studies performed earlier
‘Close,’ ‘High,’ ‘Low’ features of NIFTY during the Covid‐19
on similar topics. The boosted tree and the ELSTM model
pandemic in March and April 2020. NIFTY plunged drastically
pro- vide valuable insight on the predictions, but the
during this period, and thus it would be a difficult test for
accuracy is considerably less compared to that of the
the model. The model performed pretty well in this case
proposed model in this study. The model used in this study
study.
obtained a significantly higher level of accuracy. The
Although the model's average prediction accuracy decreased to
Convolutional Neural Network model proved to give a
around 91%–94% during the first few days of the downfall,
it soon adapted to the situation. The model adjusted quickly, significantly good result even during
and the average prediction accuracy increased to a range of extreme market fluctuations like in the case of the COVID‐19
95%‐ 98%. Therefore, even though the model suffered from pandemic.
an un- expected scenario, it did well to cope with it
efficiently. The graphs showing the prediction value and
actual value of NIFTY 7 | DISCUSSION
from 15 January 2020 to 5 August 2020 for the features
‘Open’, ‘Close’, ‘High’, ‘Low’ are given in Figure 14a–d, Artificial Neural Networks which have been used before for
respectively. stock market prediction gives about 97.66% accuracy, whereas
The changes in the prediction accuracy of the four features the Convolutional Neural Network model gives 98.92% ac-
and the average prediction accuracy over the period are curacy on the dataset. The working of CNN model is clearly
given in Figure 15 below. better than that of the ANN model. This is because the ANN
The graph shown in Figure 15 depicts the movement of model loses context from the time series data before the
prediction accuracy for the prediction that took place in the CNN model. The dense connections with regularisation
are not
 MUKHERJEE ET
■ 9

F I G U R E 15 Accuracy in predicting the features

T A B L E 3 Comparison of accuracy of our proposed model with two

recent works
Works Accuracy in (%)
Aparna Nayak et al.
Work done: Boosted decision tree 76.90

Xiongwen Pang et al.

Work done: Neural network with 57.20
embedded layer (ELSTM)
Proposed work:
(I) Artificial neural network (with 2 hidden layers) 97.66 (ANN)
(II)CNN (with 2‐D histograms generated from
stock market data) 98.92 (CNN)

(III) Case study on severe market fluctuations due

to COVID‐19 91 (under COVID
situation)

8 | CONCLUSION

This study proposes two approaches to predict stock market

indices and stock prices. This study first uses a Feed‐
forward Neural Network and performs the backpropagation
algo- rithm for the training process. This model gave a
fundamental insight into the prediction trend and provided a
graphical result on how the prediction should look. This
model gave satis-
factory results with an average prediction accuracy of 97.66%,
but it required many training data and epochs to get to the
above accuracy. Moreover, it also suffered from overfitting to
some extent, which was handled using the regularisation
method. Thus, to overcome the cost of extensive training,
this study further gives another cleaner approach. The
Convolution Neural Network model proved to give better
F I G U R E 14 (a–d). The predicted value versus actual value of NIFTY results on a given
dataset from 15 January 2020 to 5 August 2020 for all the features dataset. The study finds a different approach to analyse
time‐ series data. It utilises greyscale 2‐D histograms
generated from time‐series data for prediction. The entire
capable of retaining the time series context. This leads to the dataset is
CNN model performing better than the ANN model, as it broken down into 15 segments, every segment is fed into the
has more connection to get context from the dataset. CNN model, and maximum accuracy is obtained for each of
them. This increase in input to the model makes a vast dif-
ference in the training time and prediction accuracy. This
 9 ■ MUKHERJEE ET

model required way less training data and time than the pre-
14. Das, N., Ghosh, P., Roy, D.: Effect of Demonetization on stock
vious model. After the training phase and quantisation, the market correlated with geo twitter sentiment analysis. In: Learning and
average prediction accuracy of 98.92% is obtained, which is Analytics in Intelligent Systems. Springer, Siliguri (2020).
also better than that of the previous model. https://doi.org/10.1007/ 978‐3‐030‐42363‐6_92
On the other hand, the generation of synthetic images in 2‐ 15. Li, W., et al.: User reviews: Sentiment analysis using lexicon integrated
two‐channel CNN–LSTM family models. Appl. Soft Comput. 94, 106435
D histograms adds an overhead that is not present in the first (2020). https://doi.org/10.1016/j.asoc.2020.106435
approach. Both the methods have their advantages and dis- 16. Nayak, A., Pai, M.M.M., Pai, R.M.: Prediction models for Indian stock
advantages, but both can provide nearly accurate predictions market, In Procedia Comput. Sci. 89, 441–449 (2016). https://doi.org/
for a dataset of this sort. Such an efficient prediction of the 10.1016/j.procs.2016.06.096
stock market is essential in today's world. It would be of 17. Drucker, H., Cortes, C.: Boosting decision trees. Adv Neural Inf Process
Syst. 0, 6 (1995). https://dl.acm.org/doi/10.5555/2998828.2998896
great use to stock market analysts who would be able to 18. Pang, X., et al.: An innovative neural network approach for stock market
develop solutions to help companies and economies by prediction. J. Supercomput. 76(3), 2098–2118 (2020). https://doi.org/10.
predicting future patterns in the stock market. 1007/s11227-017-2228-y
19. Lien, M.D.: Deep learning approach for short‐term stock trends pre-
diction based on two‐stream gated recurrent unit network. IEEE Access.
ACKNOWLEDGEMENTS 6, 55392‐55404 (2018). https://doi.org/10.1109/access.2018.2868970
None. 20. Chen, J.‐F., et al.: Financial time‐series data analysis using deep convolu-
tional neural networks. In International conference of cloud
DATA AVAILABILITY STATEMENT Computing and Big Data (2016).
https://doi.org/10.1109/CCBD.2016.027
The data that support the findings of this study are available 21. Naeini, M.P., Taremian, H.: Stock Market Value Prediction Using Neural
from the corresponding author upon reasonable request. Networks. IEEE (2010). https://doi.org/10.1109/CISIM.2010.5643675
22. Chen, S., He, H.: Stock prediction using convolutional neural network. In
ORCID IOP conference Series: Materials Science and Engineering (2018)
Somenath Mukherjee https://orcid.org/0000-0003-1729- 23. Du, X., et al.: Overview of deep learning. In Youth Academic Annual
conference of Chinese Association of Automation, Wuhan (2016).
0502 https://doi.org/10.1109/YAC.2016.7804882
Bikash Sadhukhan https://orcid.org/0000-0001-5469-0711 24. N.S Exchange: NSE India, [Online]. https://www.nseindia.com/
25. Hiransha, M., et al.: NSE stock market prediction using deep‐learning
models. Procedia Comput. Sci. 132, 1351–1362 (2018).
REFERENCES https://doi.org/ 10.1016/j.procs.2018.05.050
1. Statman, M., Fisher, K.L.: Consumer Confidence and Stock returns. 26. Ding, G.J., et al.: Hierarchical multi‐scale Gaussian transformer for stock
SSRN Journal. https://doi.org/10.2139/ssrn.317304 movement prediction. International Joint Conference on Artificial In-
2. Wilcox, J.: Consumer Sentiment and Consumer Spending. FRBSF Eco- telligence Organization (2020). https://doi.org/10.24963/ijcai.2020/640
nomic Letter (2008) 27. Li, S.Q., et al.: Modeling the stock relation with graph network for
3. Kaul, V. https://www.deccanherald.com/specials/sunday‐spotlight/ overnight stock movement prediction. International Joint Conference on
what‐s‐in‐store‐for‐the‐indian‐economy‐in‐2020‐789545.html Artificial Intelligence Organization (2020). https://doi.org/10.24963/
4. Sharma, A., Bhuriya, D., Singh, U.: Survey of stock market prediction ijcai.2020/626
using machine learning approach. In: International conference of 28. Lu, W., et al.: A CNN‐BiLSTM‐AM method for stock price prediction.
Elec- tronics, Communication and Aerospace Technology (ICECA). Neural. Comput. Appl. 33(10), 4741–4753 (2021). https://doi.org/10.
(2017). https://doi.org/10.1109/ICECA.2017.8212715 1007/s00521-020-05532-z
5. Nivetha, R.Y., Dhaya, C.: Developing a prediction model for stock 29. Liu, W.X., et al.: Multi‐scale two way deep neural network for stock trend
analysis. In: 2017 International Conference on Technical Advancements prediction. International Joint Conference on Artificial Intelligence Or-
in Computers and Communications (ICTACC), IEEE (2017). https:// ganization (2020). https://doi.org/10.24963/ijcai.2020/628
doi.org/10.1109/ICTACC.2017.11 30. Nayak, A., Pai, M.M.M., Pai, R.M.: Prediction models for Indian stock
6. Jayanth Balaji, A., Harish Ram, D.S., Nair, B.B.: Applicability of deep Market. Procedia Comput. Sci. 89, 441–449 (2016). https://doi.org/10.
learning models for stock price forecasting an empirical study on 1016/j.procs.2016.06.096
BANKEX data. Procedia Comput. Sci. 143, 947–953 (2018). https://doi. 31. Pang, X., et al.: An innovative neural network approach for stock market
org/10.1016/j.procs.2018.10.340 prediction. J. Supercomput. 22 (2018)
7. Lin, T.‐W., Yu, C.‐C.: Forecasting stock market with neural networks. 32. B.S Exchange: NSE India, national stock exchange. [Online]. https://
SSRN Journal. https://doi.org/10.2139/ssrn.1327544 www1.nseindia.com/
8. Shah, V.S., et al.: Stock market prediction using neural networks. Int. 33. Simard, P.Y., Steinkraus, D., Platt, J.C.: Best Practices for Convolutional
J. Soft Comput. Eng. 6(1), (2016) Neural Networks.MicrosoftRedmond, WA. (2003). https://doi.org/10.
9. Kalyvas, E.: Using Neural Networks and Genetic Algorithms to 1109/ICDAR.2003.1227801
Predict Stock Market Returns. University of Manchester (2001) 34. Krizhevsky, A., Sutskever, I., Hinton, G.E.: ImageNet classification with
10. Girija, V.A., et al.: Stock Market Prediction: A Big Data Approach. IEEE deep convolutional. Adv. Neural Inf. Process. Syst. 25(NIPS), 9 (2012).
(2015). https://doi.org/10.1109/TENCON.2015.7373006 https://dl.acm.org/doi/10.5555/2999134.2999257
11. Ertuna, L.: Stock Market Prediction Using Neural Network and Time 35. 2D Histograms in Python, Plotly, [Online]. Available: https://plotly.
com/python/2D‐Histogram/
Series Forecasting. IEEE (2016). https://doi.org/10.13140/RG.2.1.1954.
1368
12. Wen, M., et al.: Stock market trend prediction using high‐order infor-
mation of time series. IEEE Access. 7, 28299–28308 (2019). https://doi. How to cite this article: Mukherjee, S., et al.: Stock
org/10.1109/access.2019.2901842 market prediction using deep learning algorithms. CAAI
13. Usmaini, M., et al.: Stock Market Prediction Using Machine Learning Trans. Intell. Technol. 8(1), 82–94 (2023). https://doi.
Techniques. ICCOINS (2016). https://doi.org/10.1109/ICCOINS.2016.
org/10.1049/cit2.12059
7783235