Received: 29 October 2020

DOI: 10.1049/cit2.12052

ORIGINAL RESEARCH
- - -
Revised: 17 February 2021 Accepted: 18 May 2021 CAAI Transactions on Intelligence Technology

Stock values predictions using deep learning based hybrid

models

Konark Yadav1 | Milind Yadav2 | Sandeep Saini1

1
Department of Electronics and Communication Abstract
Engineering, The LNM Institute of Information
Technology, Jaipur, India
Predicting the correct values of stock prices in fast fluctuating high‐frequency financial
2
data is always a challenging task. A deep learning‐based model for live predictions of
Department of Computer Science and Engineering,
Rajasthan Technical University, Kota, Rajasthan, stock values is aimed to be developed here. The authors' have proposed two models for
India different applications. The first one is based on Fast Recurrent Neural Networks
(Fast RNNs). This model is used for stock price predictions for the first time in this work.
Correspondence The second model is a hybrid deep learning model developed by utilising the best features
Sandeep Saini, Department of Electronics and of FastRNNs, Convolutional Neural Networks, and Bi‐Directional Long Short Term
Communication Engineering, The LNM Institute of
Information Technology, Jaipur, India. Memory models to predict abrupt changes in the stock prices of a company. The 1‐min
Email: sandeep.saini@lnmiit.ac.in time interval stock data of four companies for a period of one and three days is
considered. Along with the lower Root Mean Squared Error (RMSE), the proposed
models have low computational complexity as well, so that they can also be used for live
predictions. The models' performance is measured by the RMSE along with computation
time. The model outperforms Auto Regressive Integrated Moving Average, FBProphet,
LSTM, and other proposed hybrid models on both RMSE and computation time for live
predictions of stock values.

1 | INTRODUCTION prediction. A lot of such systems have been developed with

high accuracies during the past decades as well [1]. Fore-
Economies around the globe are digital now and dependent casting problems can be further divided into three cate-
on each other. A multinational company is having its gories and listed as follows:
impact on a lot of companies. The fluctuations in the stock
values of such companies can change the economic sce- a. Short‐term forecasting, where prediction happens for the
narios for multiple stakeholders. Thus, forecasting the stock next second, minute, hour, day, or month
values is becoming more crucial now. Forecasting is the b. Medium‐term forecasting, where prediction happens for
process of predicting the future value of any series by the next 1 year, or 2 years.
considering the previous patterns or long historical data. c. Long‐term forecasting, where prediction happens beyond
For example, if the price of gold is increasing every year at 2 years.
Christmas time, then we can predict a similar trend for the
current year as well and plan the purchase well in advance
to avoid the high rates at Christmas time. Similarly, 1.1 | Motivation
computational models can help us in predicting the weather
for the next day, week, or month as well. With the high Stocks are not a simple time series with only one factor affecting
volume of money involved, the stock market values have the outcome. These can either be univariate or multivariate.
attracted the attention of computer scientists as well to Univariate stocks are rare and they are dependent on only one
design models and architectures for precise stock value factor or only one company's performance. With the emerging

This is an open access article under the terms of the Creative Commons Attribution‐NonCommercial License, which permits use, distribution and reproduction in any medium, provided

-
the original work is properly cited and is not used for commercial purposes.
© 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.

CAAI Trans. Intell. Technol. 2022;7:107–116. wileyonlinelibrary.com/journal/cit2 107

 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
108
- YADAV ET AL.

partnerships and dependence of every big company on its Convolutional Neural Network (CNN) and Bi‐Directional
partner's stocks, the second type of stock, that is, multivariate is Long Short Term Memory (Bi‐LSTM) networks. This model
more common now. So, in such cases, the prediction of exact not only provides the output with less execution time but also
future stock values can help a lot of investors and stakeholders. with improved RMSE values.
This is the motivation behind our proposed model.

1.3 | Organisation
1.2 | Stock market prediction
The rest of the paper is organised as follows. We have provided
The first model to predict the outcome of a time series was the literature review of the existing machine learning and deep
first proposed by Ahmad and Cook [2] in 1979. In this work, learning‐based models for stock market prediction in section 2.
Auto Regressive Integrated Moving Average (ARIMA) model We have focussed on single network‐based as well as hybrid
was introduced, which is one of the most trusted models for models in this section. The proposed models are described in
time series forecasting even now. This is also a reason behind complete detail in section 3. We have explained the mathe-
the fact that a lot of conventional models are based on Auto matical model as well as structural augmentations made to
Regression (AR) and Moving Average (MA) and Exponential implement these models in the same. Section 4 consists of
Smoothing [3, 4] and generalized autoregressive conditional details of our experimental setup, datasets, and simulation re-
heteroskedasticity models [5]. The existing stock market pre- sults for short‐term and long‐term stock price prediction. We
diction methods can be classified as follows. conclude our work in section 5.

1. Fundamental Analysis is concerned with the company that

underlies the stock itself. In such methods, the historical 2 | LITERATURE REVIEW
performance, as well as the credibility of the company's
accounts, is evaluated to predict future stock values. These We have described that there are three major methods for
methods are stochastic and have limited options to consider stock market prediction. We have focussed only on machine
unprecedented events to predict the stock in such cases. learning and more specifically on deep learning‐based ap-
2. Technical Analysis. These methods are not considering the proaches in our literature review. One of the earliest applica-
working principles of the company but mostly consider the tions of machine learning models for stock market forecasting
trends and historical stock values. AR, MA, ARIMA, and was from Tay and Cao [6] in 2001. They used Support Vector
similar methods are popular methods in this category. Machine (SVM) to verify the feasibility of SVMs for financial
These methods are more effective for short‐term pre- series forecasting. The promising results motivated a lot of
dictions rather than long‐term predictions. research groups to explore more machine learning algorithms
3. Machine Learning (Time Series Forecasting). Artificial In- for the same. In 2003, Egeli et al. [7] proposed a Multi‐Layer
telligence is emerging as a robust solution for time series Perceptron and Generalised Feed Forward network‐based ar-
forecasting. These methods include artificial neural net- chitecture for financial forecasting. These models were trained
works (ANNs) and Genetic Algorithms (GAs). The most and tested on MAs for 5 and 10‐day periods. The results
common form of ANN in use for stock market prediction outperformed conventional ARIMA‐based approaches. In
is the feed‐forward network utilising the backward propa- 2005, Enke and Thawornwong [8] combined conventional data
gation of errors algorithm to update the network weights. mining algorithms with neural networks to predict stock
In recent years, Recurrent Neural Network (RNN) has values. The authors used the Probabilistic Neural Network in
become the best ANN for time series forecasting. their model that was having three feed‐forward layers. In the
same year, Huang et al. used an SVM‐based model to predict
We have focussed on the Deep Learning‐based approaches the stock values and showed that the SVM‐based model out-
in this work. We have studied the latest models and architec- performs Linear Discriminant Analysis, Quadratic Discrimi-
tures proposed explicitly for stock market prediction. After nant Analysis, and Elman Backpropagation Neural Networks.
considering the shortcomings in terms of their computational With the encouraging results from single network‐based
time and the Root Mean Squared Error (RMSE) values, we models for time series forecasting, hybrid models were also
have proposed two different models in this work. We have explored for the purpose. In 2001, Abraham et al. [9] proposed
focussed on two performance parameters, that is, execution one of the first hybrid models that were based on neuro‐fuzzy
time and RMSE. A model can be very accurate in prediction logic and ANN. In 2005, Armano et al. [10] proposed the
but very slow in computing the output. Such a model cannot model with hybridisation of a genetic classifier designed to
be applied for live stock value prediction but very useful for control the activation of a feedforward ANN for performing a
long‐term predictions. On the other hand, a faster model with locally scoped forecasting activity. In 2007, Fu et al. [11] ana-
acceptable RMSE can be utilised for live predictions as well. lysed the problem based on a Fuzzy Cerebellar Model Artic-
The first model is based on FastRNN architecture. This model ulation Controller – Bayesian Ying Yang neural network. This
is designed to provide faster and accurate results. The sec- proposed model was motivated by the Chinese ancient Ying‐
ond model is proposed by hybridizing FastRNN with Yang philosophy that states that everything in the universe
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
YADAV ET AL.
- 109

can be viewed as a product of constant conflict between op- have considered four companies for our study and those are
posites, Ying, and Yang. In 2008, Choudhary and Garg [12] Facebook Inc., Uber Inc., Apple Inc., and Nike Inc. from the
proposed the GA‐SVM model for stock market prediction. New York Stock Exchange. We have obtained the stock values
The model is a hybrid version of the GA and SVMs. The from Yahoo finance.1 The dataset includes information about
system was tested on three of India's biggest companies, that is, day stamp, time stamp, transaction ID, the stock price (open and
TCS, Infosys, and RIL that had a trading data of 1386 days close), and volume of stock sold in each minute interval. For our
from 2002 to 2008. Further, it was tested on the data of 30 model, we have used the close price for each stock. Our work
companies. The hybrid model outperformed the SVM‐based also aims on creating a prototype for live prediction. We
models in terms of RMSE. In 2009, Tsai and Wang [13] also consider a working duration of 8 h and divide those hours into
performed the stock price forecasting by using a hybrid model training and testing time. We are predicting the future prices of
designed using a decision tree (DT) and ANNs. The proposed each minute for the next 50 min by keeping the initial 7 h 10 min
model, DT‐ANN, had an accuracy of 77% and was among the data in training. The best window length was identified by
most precise systems at that time. calculating the root mean square error for various window sizes.
In this decade, the focus has shifted to deep learning‐based We kept the size of the data the same for all the stocks, that
models. Ding et al. [14] developed an event‐driven deep is, each stock has 430 rows and the model was trained on 40
learning model. In this model, a variant of CNN, that is, Deep epochs. We have considered the error and computation time of
CNN was used to predict the stock values. The events are each model for our study. If the loss (mean squared error) for
extracted from the news articles and stored as dense vectors, the current epoch is less than the value obtained from the
trained using a novel neural tensor network. This dense network previous epoch, the weight matrix for that particular epoch is
was trained using the Deep‐CNN. Akita et al. [15] applied deep stored. After the completion of the training process, each of
learning models Paragraph Vector, and Long Short‐Term these models was tested on the remaining 50 values. In this
Memory (LSTM) to financial time series forecasting. The process, the model with the least RMSE is taken as the final
model also utilised the news article data and converted those into model for prediction.
the Paragraph vector, which was then fed to LSTM to predict the To compare our model with the existing similar models, we
stock prices. This model was tested on the Tokyo Stock Ex- have initially considered a few baseline models and then the
change. Fischer and Krauss [16] also proposed a similar LSTM‐ state‐of‐the‐art models. We have considered the following
based architecture. The model outperformed random forest, a baseline models for our study.
deep neural network (DNN), and LOG‐based models. CNN has
been employed by several researchers for the problem. 1. ARIMA is a widely known model for time series forecasting
Hoseinzade and Haratizadeh [17] proposed CNNpred, which that works on the concept of MA. It is based on the trend
was a CNN‐based model for establishing the relationships be- stationarity, and seasonality of the data. Here, we have to first
tween different stock markets across the globe and showed the make the input data stationary and then after finding their
positive correlation between the trends across the global stock auto‐correlation values and partial auto‐correlation, we
exchanges. Gudelek et al. [18] enhanced the regular CNN model would be able to forecast. The only glitch with this model is
to the two‐dimensional model and employed it for the task. The that it cannot be automated for all kinds of stocks, as we have
first convolutional layer had 32 filters of size 28 � 28 and the to set different p, d, q (number of auto‐regressive terms,
second one had 64 filters of the same sizes. This improved the number of non‐seasonal differences needed for stationarity,
RMSE values. Eapen et al. [19] presented a hybrid model that was number of lagged forecast errors) values for different stocks.
made using CNN and Bi‐LSTM. The proposed model was 9% 2. LSTM is a well‐established model used for time series
better than a single network‐based model. Selvin et al. [20] had forecasting and frequently applied for stock market price
proposed a similar architecture that was based on the hybrid- prediction [27–29]. It is a derived RNN with forget gate
isation of LSTM, RNN, and CNN. In the next section, we explain functionality. LSTMs are very effective for long as well as
our proposed models which aim to improve the RMSE values as short‐term predictions.
well as provide faster results. 3. FBProphet: is another time series forecasting model intro-
Significant work was carried out on stock by Nabipour et al. duced by Facebook Inc [30]. This model required very less
in the field of stock market predictions. The authors have car- computational time in comparison to other models
ried out Stock Market predictions Using Machine Learning and 4. A hybrid model of CNN and LSTM: This is a hybrid model
Deep Learning Algorithms via Continuous and Binary Data [21, designed recently and extensively used in stock price pre-
22]. Ecer et al. have used GAs and Particle Swarm Optimisation diction. The model consists of both LSTM and CNN layers
for Modelling Stock Price Index Prediction [23, 24]. These [31, 32].
techniques can be used in various other domains as well [25, 26].
These baseline models have their advantages and disad-
vantages. For example, FBProphet is the fastest model but
3 | PROPOSED MODEL underperforms in terms of error. Similarly, the hybrid model

In designing a stock market price prediction model, the fore-

1
most requirement is the availability of a suitable dataset. We https://finance.yahoo.com
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
110
- YADAV ET AL.

based on CNN and LSTM provides very little error but takes vector xt updates the hidden state. Also, FastRNN has two
more time in predictions. A single neural network‐based model more parameters than RNN and requires lesser computations,
performs well on one aspect of the problem while lags behind which is a very minute fraction of the per‐stride computational
on the other front. Thus, we decided to exploit the good complexity of the RNN. Unlike unitary methods [35, 36],
features of multiple networks. We have considered the FastRNN does not introduce any costly systemic constraints
following networks for their respective advantages in time on U and hence scales well to huge datasets with typical
series prediction. optimisation methods. The proposed FastRNN‐based model
for stock market predictions is shown in Figure 1.

3.1 | FastRNN‐based hybrid model

3.2 | FastRNN + CNN + Bi‐LSTM‐based
FastRNN was proposed by Microsoft in 2018 [33]. The model is
hybrid model
based on the Residual Network (ResNet) [34]. If we look at the
The first model proposed in this work is designed to provide a
structure of ResNet, then in this model, every network layer is not
faster computation time with reasonably good accuracy. The
stacked over each other, but they form a residual mapping. It is
second aim of this work is to maintain fast computation with
shown that when compared with unreferenced mapping, the
improved accuracy. In this direction, we propose a hybrid model
residual mapping is easier to optimise. The model is based on the
visual geometric group model and modified by adding residual that exploits the good features of multiple networks. We use
FastRNN as our first network that will provide faster results and
layers. ResNet is trained on 25.5 million parameters. This makes
augment it with CNN and Bi‐LSTM networks to improve the
ResNet a very heavy network as well. FastRNN exploits the
accuracy of prediction. Bi‐LSTM was proposed in 1997 [37].
learned residual connections of ResNet and provides stable
This model is designed to learn in both directions and is one of
training along with comparable training accuracies at a much
the best‐suited models for sequence to sequence learning. For
lower computational cost. The prediction accuracies of
time series forecasting as well, the model has been proven to be a
FastRNN are 19% higher than those of the regular RNN
good fit [38–40]. The encouraging results from these recent
structures and a little less than those of the gated RNN.
Let X = [x1, …, xT] be the input vector where xt ϵRD works have inspired us to consider Bi‐LSTM as our last network.
We have taken the conventional CNN in between FastRNN and
denotes the t‐th step feature vector. So, for a multi‐class RNN
Bi‐LSTM to stabilise the network. A CNN has similarities with
we aim to learn a function F:RDxT1, …, L that is going to
an ANN but it takes presumptions about the input data, which
predict any of these L classes for a given point of data that is X.
allows it to attain greater invariance when encoding the prop-
In a normal RNN architecture, hidden state vectors which are
erties of input data into the network. CNN needs to be trained on
represented by
huge amounts of training data to generate deep learning models
ht ¼ tanhðW xt þ Uht−1 þ bÞ ð1Þ that attain higher generalisation accuracy. The proposed design
of the hybrid model is shown in Figure 2.
In the proposed model, we have taken a three‐stage
Learning U, and W in the architecture given above is
pipeline consisting of FastRNN, CNN, and Bi‐LSTM. In the
arduous, as the gradient can have an exponentially large (in T)
first model, we have explained the added advantages of
condition number. Unitary network‐based ways can solve this
FastRNN for time series forecasting. The CNN layer is one‐
problem by expanding the network size and in this process,
dimensional with an ReLU activation function. The kernel
they become significantly large and the model accuracy may
size is kept as 3 and keeping padding as ‘same’ followed by
not remain the same.
Maxpooling. 1‐D CNN is used to extract the higher‐level
However, FastRNN utilises a simple weighted residual
features only. These extracted features are fed to the final
connection to stabilise the training by creating well‐
stage, that is, Bi‐LSTM network. Bi‐LSTM is followed by 2
conditioned gradients. In particular, FastRNN updates the
dense layers. In the proposed model, the Bi‐LSTM unit learns
hidden state ht as follows:

h0t ¼ σðW xt þ Uht−1 þ bÞ ð2Þ

0
ht ¼ α:ht þ β:ht−1 ð3Þ

Here, α and β are trainable weights, which are para-

meterised by sigmoid function and limited in the range 0 < α,
β < 1. σ is a non‐linear function such as hyperbolic tangent,
sigmoid, or Rectified Linear Unit (ReLU), and can vary across
datasets.
FastRNN reconditions hidden state in a calculated manner F I G U R E 1 Fast Recurrent Neural Networks‐based proposed model
with α, β limiting the extent to which the present attribute for faster stock values prediction
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
YADAV ET AL.
- 111

FIGURE 2 FastRNN, CNN, and bi‐directional long short term memory based hybrid model for higher accuracy stock market predictions

from both the backward and forward sequences of data and x0i ¼ CNNðxi Þ ð5Þ
uses concatenation to merge the two sequential outputs. There
was no overfitting observed, that is why we have not added a where xi is the data vector which is inputted to the CNN
dropout layer. Here, using Bi‐LSTM also prevented us from network and x0i is the CNN network output which is then
the vanishing gradient problem. Network parameters: As further passed to the Bi‐LSTM network. To get the idea of
mentioned before, we are using 430 stock values as input, and Bi‐LSTM, we add forget gate structure in the LSTM. The
the 1‐D CNN layer with the use of 3‐kernel windows to return equation is denoted by
another sequence of smaller size. The output from the 1‐D � ���
CNN is then fed into the Bi‐LSTM layer group and is giving it ¼ σ W i xt ; yt−1 ð6Þ
an output sequence of length 50. The output of this group is � ���
then fed to a dropout layer, which gives 25 units and 1 unit f t ¼ σ Wf xt ; yt−1 ð7Þ
output, respectively. As it can be considered as a standard
� ����
regression problem, we have used RMSE as the loss function ot ¼ σ W o xt ; yt−1 ð8Þ
to determine the error in the predicted versus actual data. We
have to keep input and output rates identical and are critical for � ���
gt ¼ tanh W g xt ; yt−1 ð9Þ
time‐series forecasting.
Let us assume that [x = x1, x2, x3, …., xn] is the one‐ ct ¼ f ⊙ ct−1 þ i ⊙ g ð10Þ
dimensional input for the 1‐D CNN layer. The equation
makes a feature map after it gets convoluted with the convo- yt ¼ o ⊙ tanhðct Þ ð11Þ
lution operator and is passed through a filter W ϵRf d , where f
signifies inherent attributes from the input data throwing out
where i is the input gate, f is the forget gate, o is the output
as output. A new features set fm from the new attributes f
gate, g, and c input modulation gate, respectively. Note that
represented in the equation below:
these are in n‐dimensional real vectors. In Equations (6)–(8),
fm
� � the σ is a sigmoid function and Wi, Wf, Wo, and Wg are fully
hl i ¼ tanh w f m xi:iþf −1 þ b ð4Þ connected neural networks for the input, forget, output, and
input modulation gates, respectively. The issue with the LSTM
Every set of features f uses the filter hl in the input defined model is that it only considers information from one direction
by [x1 − f, x2 − f + 1, …, xn − f + 1]. This operation generates a on a sequence which leads to an effective reduction of the
feature map denoted by [hl1, hl2, …, hln − f + 1]. LSTM model. Also, multi‐directional information in the
Convolution layer outputs are obtained as a sum of sequence can have valuable information data. Therefore,
weighted inputs after multiple linear transformations. For a Bi‐LSTM was developed and it combines backward and for-
non‐linear feature extraction problem, linear transformations ward directions in the sequence.
do not perform with satisfactory success and thus we have Order for the forward LSTM is [x1, x2, …, xn]and for the
to add non‐linear activation functions. In this model, we backward LSTM is [xn, xn−1, …., x1]. Post‐training, both the
have chosen the ReLU activation function which applies max forward and the backward LSTMs separately are integrated by
(0, x) on each input. The output is down‐sampled in the next combining their outputs in the previous step, which is denoted
step to reduce the information, so that the computation time in Equation (12) as,
can be improved. In our model, we have used max‐pooling
for that which is represented by hl = max (hl). Here, pooling yt ¼ yFðtÞyBðn − t þ 1Þ ð12Þ
helps the model to select the most relevant information and
the output of the max‐pooling layer can be denoted as where yF and yB are the outputs of the backward and forward
follows: LSTMs, respectively, while the notation denotes integration
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
112
- YADAV ET AL.

T A B L E 1 RMSE and computation time calculated for the state‐of‐ Apple, Facebook, Nike, and Uber are shown in Tables 1–4.
the‐art and the proposed models for Apple Inc. stock values These tables highlight the best values for each column.
Model name RMSE Time (in s) The training process is fast and could be carried out on a
CPU because the data‐size was not very high. The values of
ARIMA [41] 0.796109 1.63
loss functions for each of the 4 companies' data training are
BiLSTM_Attention_CNN_BiLSTM [42] 0.234644 25.72292113 shown in the graphical form in Figures 3 and 4.
CNN_LSTM_Attention_LSTM [43] 0.214821 16.00164294 From the mentioned tables, Tables 1–4, it can be observed
that the proposed models, not only have lesser RMSE but also
FBProphet [44] 0.935556 0.659962893
they perform better in terms of their computation speed,
LSTM [45] 0.228731 13.28157353 which gives a clear indication that these models can be useful
LSTM_Attention_CNN_BiLSTM [42] 0.263613 19.96186757 in making live next minute predictions of a stock price which
will help the investor to buy stocks more wisely as the market
LSTM_Attention_CNN_LSTM [45] 0.27994 17.32732081
can crash anytime due to any reason. The proposed models
LSTM_Attention_LSTM [45] 0.299334 19.28274226

LSTM_CNN_BiLSTM [46] 0.23489 16.76800251 T A B L E 2 RMSE and computation time calculated for the state‐of‐
the‐art models and the proposed models for Facebook Inc. stock values
FastRNN (proposed) 0.202456 3.337492943

FASTRNN_CNN_BiLSTM (proposed) 0.205647 13.49208355 Model name RMSE Time (in s)

Abbreviations: ARIMA, auto regressive integrated moving average; RMSE, root mean
ARIMA [41] 0.86664567 1.60001
squared error. BiLSTM_Attention_CNN_BiLSTM [42] 0.26834072 26.56901288

CNN_LSTM_Attention_LSTM [43] 0.15718991 16.61885238

operator such as a simple adder. The proposed model is
developed on Google Colab. Details of experimental setup and FBProphet [44] 1.59050836 0.405165672
results are provided in the next section. LSTM [45] 0.18714926 14.00746107

LSTM_Attention_CNN_BiLSTM [42] 0.22215487 20.65539312

4 | EXPERIMENTAL RESULTS AND LSTM_Attention_CNN_LSTM [45] 0.24017975 17.97418046

DISCUSSIONS LSTM_Attention_LSTM [45] 0.29138532 19.86988878

LSTM_CNN_BiLSTM [46] 0.19898068 17.52065849

4.1 | Experimental setup
FastRNN (proposed) 0.14932692 14.16171408
We have used a free version of Google Colab with AMD FASTRNN_CNN_BiLSTM (proposed) 0.15137204 3.447671652
EPYC 7B12 CPU as our execution environment with 12 GB
Abbreviations: ARIMA, auto regressive integrated moving average; RMSE, root mean
(adjustable) assigned RAM, with one socket, two threads per squared error.
core, 13,684K of L3 cache, CPU of around 2250 MHz, and
with No Power level. We have developed our model using the
Python programming language (Python 3.5). We used Keras
T A B L E 3 RMSE and computation time calculated for the state‐of‐
(with Tensorflow backend) library for creating our deep
the‐art models and the proposed models for Nike Inc. stock values
learning models. To compile the data into a suitable format, we
have used Pandas, and for dividing our data into test and train, Model name RMSE Time (in s)
we have used Numpy. For visualisation, we have the matplotlib ARIMA [41] 0.465822234 1.63001
library. The dataset and the corresponding codes are available
BiLSTM_Attention_CNN_BiLSTM [42] 0.106373725 26.56245756
at GitHub.2
CNN_LSTM_Attention_LSTM [43] 0.051930262 17.05163288

FBProphet [44] 0.550859082 0.509964705

4.2 | Results
LSTM [45] 0.052901 14.01486564

We have used 430 stock values for training and 70 for testing. LSTM_Attention_CNN_BiLSTM [42] 0.062675555 20.4820962
We have tested our models on the stock values of 4 companies, LSTM_Attention_CNN_LSTM [45] 0.061914832 17.52905965
that is, Apple, Facebook, Nike, and Uber. These proposed
LSTM_Attention_LSTM [45] 0.102347907 19.73444676
models are compared with nine other state‐of‐the‐art models.
We have trained the models for 40 epochs. The RMSE values LSTM_CNN_BiLSTM [46] 0.047966405 17.31208062
and the computation time for each of the nine models along FastRNN (proposed) 0.037727882 14.34483051
with the two proposed models for four companies, that is,
FASTRNN_CNN_BiLSTM (proposed) 0.039458341 3.807400703

Abbreviations: ARIMA, auto regressive integrated moving average; RMSE, root mean
2
https://github.com/MilindYadav‐97/Hybrid_FastRNN‐for‐stock‐predictions squared error.
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
YADAV ET AL.
- 113

T A B L E 4 RMSE and computation time calculated for the state‐of‐ work best on multiple stocks, which can be seen in the
the‐art models and the proposed models for Uber stock values mentioned tables that the proposed models have out-
Model name RMSE Time (in s) performed other classical and hybrid models in terms of both
RMSE and computation time. There is a visualised compara-
ARIMA [41] 0.163886332 1.610950985
tive study in Figure 3, which shows how the validation loss
BiLSTM_Attention_CNN_BiLSTM 0.025170157 24.91984868 (Mean Squared Error) on the validation dataset is improving
[42] after each epoch for all the studied stocks. Figure 4 shows how
CNN_LSTM_Attention_LSTM [43] 0.023596089 15.85278392 the model learning speed is increasing after each epoch, as we
FBProphet [44] 0.064013152 0.489181757
can see a horizontal line of training loss (Mean Squared Error)
after five epochs. Here, Figure 5 shows the comparative study
LSTM [45] 0.024660817 13.3820951 of the actual and predicted values of the baseline model and
LSTM_Attention_CNN_BiLSTM 0.031076263 19.3135848 our proposed models for the next 20 min and it can be
[42] observed that the proposed models' predictions were closer
LSTM_Attention_CNN_LSTM [45] 0.032733788 16.7529645 than the actual stock values for each of the visualised stock.
LSTM_Attention_LSTM [45] 0.027326071 18.67898536

LSTM_CNN_BiLSTM [46] 0.021864702 16.29291153 5 | CONCLUSIONS

FastRNN (proposed) 0.020682922 3.35441184
Financial data prediction can be very beneficial for companies
FASTRNN_CNN_BiLSTM 0.021809913 18.18150258
and investors. Deep learning‐based models have been quite
(proposed)
effective for such predictions in recent years. In this work, we
Abbreviations: ARIMA, auto regressive integrated moving average; RMSE, root mean have proposed two models for this purpose. The first model
squared error.
that is based on FastRNN, can provide faster predictions when

FIGURE 3 Comparison of plots of validation losses (mean squared error) of our proposed and baseline models at each epoch for all the studied stocks
24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ET AL.
YADAV

Comparative plots of training losses (mean square error) of our proposed and baseline models at each epoch for all the studied stocks

Predicted stock values with proposed model and baseline models

FIGURE 4

FIGURE 5
-114
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
YADAV ET AL.
- 115

compared with other state‐of‐the‐art models (except ARIMA 9. Abraham, A., Nath, B., Mahanti, P.K.: Hybrid intelligent systems for
and FBProphet). While improving the speed of prediction, it stock market analysis. In: International Conference on Computational
Science, pp. 337–345. Springer, Heidelberg, 28–30 May 2001. https://
also provides better or at par RMSE values as well. Thus, also
link.springer.com/content/pdf/10.1007/3‐540‐45718‐6_38.pdf
when compared with FBProphet and ARIMA, this is a better 10. Armano, G., Marchesi, M., Murru, A.: A hybrid genetic‐neural archi-
choice for a reliable model. The second model improves the tecture for stock indexes forecasting. Inf. Sci. 170(1), 3–33 (2005)
first proposed model while keeping the speed almost the same. 11. Fu, J., et al.: Stock prediction using FCMAC‐BYY. In: International
It compromises a bit on the speed of prediction but improves Symposium on Neural Networks, pp. 346–351. Springer, Heidelberg
the RMSE values. These models can help people who have (2007)
12. Choudhry, R., Garg, K.: A hybrid machine learning system for stock
invested in stocks. Using these models, if suppose the investor market forecasting. World Acad. Sci. Eng. Technol. 39(3), 315–318
gets to know that stock prices are crashing in the next minutes (2008)
or so, then the investor can sell the stocks at the right time 13. Tsai, C.F., Wang, S.P.: Stock price forecasting by hybrid machine learning
which can prevent the investor from heavy losses. Similarly, if techniques. In: Proceedings of the International MultiConference of
the investor knows that in the next few minutes, stock prices Engineers and Computer Scientists, vol. 1, pp. 60. IAENG, Hong Kong,
14‐16 March 2018. http://www.iaeng.org/publication/IMECS2018/
will rise, then accordingly the investor can make a move. The 14. Ding, X., et al.: Deep learning for event‐driven stock prediction. In:
only limitation with the proposed models is that we have not Twenty‐Fourth International Joint Conference on Artificial Intelli-
considered any other external factors on which stock prices gence, Buenos Aires, 25–31 July 2015
might be dependent on, such as their demand and supply, 15. Akita, R., et al.: Deep learning for stock prediction using numerical and
textual information. In: 2016 IEEE/ACIS 15th International Conference
geographical changes, company's profit/loss, etc. As we all
on Computer and Information Science (ICIS), Okayama, 26–29 June
know that stock prices are very volatile, it is better to take into 2016, pp. 1–6. IEEE, Piscataway, New Jersey (2016)
account other factors as well such as region, geographical 16. Fischer, T., Krauss, C.: Deep learning with long short‐term memory
seasons, inflation, cost of raw products, market competition, networks for financial market predictions. Eur. J. Oper. Res. 270(2),
demand, and supply, etc. on which stock prices are dependent. 654–669 (2018)
In the future, we will consider these other dependent attributes 17. Hoseinzade, E., Haratizadeh, S.: CNNpred: CNN‐based stock market
prediction using a diverse set of variables. Expert Syst. Appl. 129,
as well, which can affect the stock price values and will make 273–285 (2019)
our model more robust. 18. Gudelek, M.U., Boluk, S.A., Ozbayoglu, A.M.: A deep learning based
stock trading model with 2‐D CNN trend detection. In: 2017 IEEE
Symposium Series on computational Intelligence (SSCI), 27 November–
ACK NO W L ED GE MN TS 1 December 2017, Honolulu, pp. 1–8. IEEE, Piscataway, New Jersey
There is no funding source for this research. It is completed as (2017)
part of the individual level with the support of the host insti- 19. Eapen, J., Bein, D., Verma, A.: Novel deep learning model with CNN
tute, that is, the LNM Institute of Information Technology, and bi‐directional LSTM for improved stock market index prediction. In:
Jaipur, India. 2019 IEEE 9th Annual Computing and Communication Workshop and
Conference (CCWC), pp. 264–270. IEEE, New York, 6–8 January 2020
20. Selvin, S., et al.: Stock price prediction using LSTM, RNN and CNN‐
OR CID sliding window model. In: 2017 International Conference on Advances in
Sandeep Saini https://orcid.org/0000-0002-8906-8639 Computing, Communications and Informatics (ICACCI), 13–16
September 2017, Manipal, pp. 1643–1647. IEEE, Piscataway, New Jersey
(2017)
REFE R ENC ES 21. Nabipour, M., et al.: Predicting stock market trends using machine
1. Hsu, M.‐W., et al.: Bridging the divide in financial market forecasting: learning and deep learning algorithms via continuous and binary data: a
machine learners vs. financial economists. Expert Syst. Appl. 61, 215–234 comparative analysis. IEEE Access. 8, 150199–150212 (2020)
(2016) 22. Nabipour, M., et al.: Deep learning for stock market prediction. Entropy.
2. Ahmed, M.S., Cook, A.R.: Analysis of Freeway Traffic Time‐Series Data 22(8), 840 (2020)
by Using Box‐Jenkins Techniques, Vol. 722. Transportation Research 23. Ecer, F., et al.: Training multilayer perceptron with genetic algorithms and
Board, Washington, D.C. (1979) particle swarm optimization for modeling stock price index prediction.
3. Huang, S.‐J., Shih, K.‐R.: Short‐term load forecasting via ARMA model Entropy. 22(11), 1239 (2020)
identification including non‐Gaussian process considerations. IEEE 24. Nosratabadi, S., et al.: Data science in economics: comprehensive review
Trans. Power Syst. 18(2), 673–679 (2003) of advanced machine learning and deep learning methods. Mathematics.
4. Sharma, A., Bhuriya, D., Singh, U.: Survey of stock market prediction 8(10), 1799 (2020)
using machine learning approach. In: 2017 International Conference of 25. Shamshirband, S., Rabczuk, T., Chau, K.‐W.: A survey of deep learning
Electronics, Communication and Aerospace Technology (ICECA), vol. techniques: application in wind and solar energy resources. IEEE Access.
2, pp. 506–509. IEEE, New York, 20–22 April 2017. https://ieeexplore. 7, 164650–164666 (2019)
ieee.org/document/8212715 26. Shamshirband, S., et al.: A review on deep learning approaches in
5. Franses, P.H., Van Dijk, D.: Forecasting stock market volatility using healthcare systems: taxonomies, challenges, and open issues. J. Biomed.
(non‐linear) GARCH models. J. Forecast. 15(3), 229–235 (1996) Inform, 113, 103627 (2020)
6. Tay, F.E.H., Cao, L.: Application of support vector machines in financial 27. Gers, F.A., Eck, D., Schmidhuber, J.: Applying LSTM to time series
time series forecasting. Omega. 29(4), 309–317 (2001) predictable through time‐window approaches. In: Neural Nets WIRN
7. Egeli, B., Ozturan, M., Badur, B.: Stock market prediction using artificial Vietri‐01, pp. 193–200. Springer, Heidelberg, 21–25 August 2001
neural networks. Decis. Support Syst. 22, 171–185 (2003) 28. Yadav, K., et al.: Bi‐LSTM and ensemble based bilingual sentiment
8. Enke, D., Thawornwong, S.: The use of data mining and neural networks analysis for a code‐mixed Hindi‐English social media text. In: 2020 IEEE
for forecasting stock market returns. Expert Syst. Appl. 29(4), 927–940 17th India Council International Conference (INDICON), 11–13
(2005) December 2020, Delhi, pp. 1–6. IEEE, Piscataway, New Jersey (2020)
 24682322, 2022, 1, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/cit2.12052, Wiley Online Library on [12/05/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
116
- YADAV ET AL.

29. Saini, S., Sahula, V.: Neural machine translation for English to Hindi. In: 39. Kim, J., Moon, N.: BiLSTM model based on multivariate time series data
2018 Fourth International Conference on Information Retrieval and in multiple field for forecasting trading area. J. Ambient Intell. Humaniz.
Knowledge Management (CAMP), Kota kinabalu, Malaysia, pp. 1–6. Comput. 10(4), 1–10 (2019)
IEEE, New York, 26–28 March 2018. https://ieeexplore.ieee.org/ 40. Long, J., et al.: An integrated framework of deep learning and knowledge
document/8464781 graph for prediction of stock price trend: an application in Chinese stock
30. Taylor, S.J., Letham, B.: Forecasting at scale. Am. Stat. 72(1), 37–45 (2018) exchange market. Appl Soft Comput. 91, 106205 (2020)
31. Kim, T., Kim, H.Y.: Forecasting stock prices with a feature fusion LSTM‐ 41. Ariyo, A.A., Adewumi, A.O., Ayo, C.K.: Stock price prediction using the
CNN model using different representations of the same data. PLoS One. ARIMA model. In: 2014 UKSim‐AMSS 16th International Conference
14(2), e0212320 (2019) on Computer Modelling and Simulation, pp. 106–112. IEEE, New York,
32. Yadav, K., et al.: Bilingual sentiment analysis for a code‐mixed Punjabi 26–28 March 2014. https://ieeexplore.ieee.org/document/7046047
English social media text. In: 2020 5th International Conference on 42. Wang, M., Cheng, J., Zhai, H.: Life prediction for machinery components
Computing, Communication and Security (ICCCS), 14–16 October 2020, based on CNN‐BiLSTM network and attention model. In: 2020 IEEE
Patna, pp. 1–5. IEEE, Piscataway, New Jersey (2020) 5th Information Technology and Mechatronics Engineering conference
33. Kusupati, A., et al.: FastGRNN: a fast, accurate, stable and tiny (ITOEC), 12–14 June 2020, Chongqing, pp. 851–855. IEEE, Piscataway,
kilobyte sized gated recurrent neural network. In: NIPS’18: Pro- New Jersey (2020)
ceedings of the 32nd International Conference on Neural Informa- 43. Zhang, Y., et al.: A text sentiment classification modeling method based
tion Processing Systems, San Diego, pp. 9017–9028 Montréal, 8 on coordinated CNN‐LSTM‐attention model. Chinese J. Electron. 28(1),
December 2018. https://proceedings.neurips.cc/paper/2018/hash/ 120–126 (2019)
ab013ca67cf2d50796b0c11d1b8bc95d-Abstract.html 44. Chikkakrishna, N.K., et al.: Short‐term traffic prediction using SARIMA
34. He, K., et al.: Deep residual learning for image recognition. In: Pro- and FbPROPHET. In: 2019 IEEE 16th India Council International
ceedings of the IEEE Conference on Computer Vision and Pattern conference (INDICON), 13–15 December 2019, Rajkot, pp. 1–4. IEEE,
Recognition, 27–30 June 2016, Las Vegas, pp. 770–778. IEEE, Piscat- Piscataway, New Jersey (2019)
away, New Jersey (2016) 45. Qiu, J., Wang, B., Zhou, C.: Forecasting stock prices with long‐short term
35. Arjovsky, M., Shah, A., Bengio, Y.: Unitary evolution recurrent neural memory neural network based on attention mechanism. PLoS One.
networks. In: International Conference on Machine Learning, pp. 15(1), e0227222 (2020)
1120–1128 (2016) 46. Le, T., et al.: Improving electric energy consumption prediction using
36. Zhang, J., Lei, Q., Dhillon, I.S.: Stabilizing gradients for deep neural net- CNN and Bi‐LSTM. Appl. Sci. 9(20), 4237 (2019)
works via efficient SVD parameterization. arXiv Prepr. arXiv1803.09327
(2018)
37. Schuster, M., Paliwal, K.K.: Bidirectional recurrent neural networks. How to cite this article: Yadav, K., Yadav, M., Saini,
IEEE Trans. Signal Process. 45(11), 2673–2681 (1997)
S.: Stock values predictions using deep learning based
38. Siami‐Namini, S., Tavakoli, N., Namin, A.S.: The performance
of LSTM and BiLSTM in forecasting time series. In: 2019 IEEE hybrid models. CAAI Trans. Intell. Technol. 7(1),
International Conference on Big Data (Big Data), pp. 3285–3292 107–116 (2022). https://doi.org/10.1049/cit2.12052
(2019)