Available online at [Link].

com
ScienceDirect
ScienceDirect
Procedia online
Available Computer
at Science 00 (2019) 000–000
[Link]
Procedia Computer Science 00 (2019) 000–000 [Link]/locate/procedia
[Link]/locate/procedia
ScienceDirect
Procedia Computer Science 167 (2020) 2091–2100

International Conference on Computational Intelligence and Data Science (ICCIDS 2019)

International Conference on Computational Intelligence and Data Science (ICCIDS 2019)
Optimizing LSTM for time series prediction in Indian stock market
Optimizing LSTM for time series prediction in Indian stock market
Anita Yadava,*
a,*
, C K Jhaaa, Aditi Sharanbb
a
Anita Yadav , C K Jha , Aditi Sharan
Department of Computer Science, AIM & ACT, Banasthali Vidyapith, Rajasthan- 304022, India
ba
School of Computer
Department & Systems
of Computer Sciences,
Science, AIM &Jawaharlal Nehru University,
ACT, Banasthali New Delhi -304022,
Vidyapith, Rajasthan- 110067,India
India
b
School of Computer & Systems Sciences, Jawaharlal Nehru University, New Delhi - 110067, India

Abstract
Abstract
Long Short Term Memory (LSTM) is among the most popular deep learning models used today. It is also being applied to time
series Short
Long prediction
Term which
Memory is a(LSTM)
particularly hardthe
is among problem to solvedeep
most popular due learning
to the presence of long
models used term
today. trend,
It is also seasonal and cyclical
being applied to time
fluctuations and random
series prediction which isnoise. The performance
a particularly of LSTM
hard problem to issolve
highly
duedependent on choice
to the presence of of
longseveral
termhyper-parameters which
trend, seasonal and need
cyclical
to be chosenand
fluctuations very carefully,
random [Link]
order to get goodofresults.
performance LSTM Being a relatively
is highly dependentnew model,ofthere
on choice are hyper-parameters
several no established guidelines
which needfor
configuring
to be chosenLSTM. In this paper
very carefully, thistoresearch
in order gapresults.
get good was addressed. A datasetnew
Being a relatively wasmodel,
createdthere
fromare
thenoIndian stock market
established and for
guidelines an
LSTM model
configuring was developed
LSTM. for this
In this paper it. Itresearch
was thengapoptimized by comparing
was addressed. stateless
A dataset and stateful
was created models
from the Indianand by market
stock tuning for
andthe
an
numbermodel
LSTM of hidden
waslayers.
developed for it. It was then optimized by comparing stateless and stateful models and by tuning for the
number of hidden layers.
© 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
© 2020 The Authors. Published by Elsevier B.V.
([Link]
© 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
This is an open access article under the CC BY-NC-ND license ([Link]
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and
([Link]
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data
Data Science
Peer-review
Science (ICCIDS
under
(ICCIDS 2019)
responsibility
2019). of the scientific committee of the International Conference on Computational Intelligence and
Data Science (ICCIDS 2019)
Keywords: LSTM; Hyperparameters; Stateful Stateless; Hidden layers; Time series prediction
Keywords: LSTM; Hyperparameters; Stateful Stateless; Hidden layers; Time series prediction

1. Introduction
1. Introduction
Stock price prediction is the process by which future stock prices are forecast on the basis of past prices. Stock price
prediction
Stock priceisprediction
useful forisinvestors to increase
the process by whichthe profits
future from
stock stock
prices aretrading.
forecastThere
on theare twooftraditional
basis approaches
past prices. Stock priceto
prediction:
prediction istechnical andinvestors
useful for fundamental analysis.
to increase theTechnical analysis
profits from stocktries to identify
trading. There some patterns
are two from historical
traditional approaches data
to
while fundamental
prediction: technicalanalysis focuses onanalysis.
and fundamental overall economy,
Technicalthe company’s
analysis tries tofinancial
identify condition and its
some patterns management
from [1].
historical data
Stock fundamental
while price prediction is a very
analysis challenging
focuses task
on overall as it is highly
economy, volatile, financial
the company’s non-linearcondition
and dynamic and
and its is impacted[1].
management by
variousprice
Stock factors like political
prediction conditions,
is a very economic
challenging task assituation, trend,
it is highly seasonality,
volatile, investor
non-linear and psychology
dynamic and etc.
is impacted by
various factors like political conditions, economic situation, trend, seasonality, investor psychology etc.
* Corresponding author. Tel.: +91-9899782208.
* Corresponding anitayadav_07@[Link]
E-mail address:author. Tel.: +91-9899782208.
E-mail address: anitayadav_07@[Link]
1877-0509 © 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
([Link]
1877-0509 © 2019 The Authors. Published by Elsevier B.V.. This is an open access article under the CC BY-NC-ND license
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science
([Link]
(ICCIDS
Peer-review2019)
under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data Science
(ICCIDS 2019)

1877-0509 © 2020 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license ([Link]
Peer-review under responsibility of the scientific committee of the International Conference on Computational Intelligence and Data
Science (ICCIDS 2019).
10.1016/[Link].2020.03.257
2092 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100
2 Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000

Nomenclature

ANN Artificial Neural Network

DL Deep Learning
KNN K-nearest neighbors
LSTM Long Short Term Memory
ML Machine Learning
MLP Multilayer Perceptron
NB Naive Bayes
NLP Natural Language Processing
NSE National Stock Exchange
RAF Random Forest
RMSE Root Mean Square Error
RNN Recurrent Neural Network
SVM Support Vector Machine

Due to information explosion in recent years, a large amount of data has become available online and manually
analyzing this data is not feasible. Various supervised and unsupervised machine learning techniques have been
developed to automatically gather and analyze data and make predictions. Deep learning techniques have become
the hottest tool delivering state-of-the art results in areas as diverse as computer vision, speech recognition and NLP
problems. But deep learning techniques also present a unique set of challenges.
The performance of deep learning models depends upon a number of parameters called hyper-parameters. Hyper-
parameters are manually set training variables which are determined before training the model. Some automatic
techniques for tuning hyper-parameters have been proposed but these methods don’t provide insight to the end-users
about the interactions between different hyper-parameters and their relative importance [2]. For the deep learning
model used in this paper, namely LSTM, the important hyper-parameters are [3] activation function (sigmoid, tanh,
softmax etc.), optimizer (Adam, Adadelta, RMSprop etc.), batch size, number of epochs, number of hidden layers
etc. LSTM can be stateful or stateless. By creating a dataset of stock prices from the Indian stock market and
building an LSTM model to predict stock prices, various configurations of LSTM can be tested and compared to
optimize the model.

2. Literature Review

Neural networks have been widely used for forecasting time series. Foster et al. [4] compared the forecasting of
noisy time series using neural networks with linear regression and exponential smoothing methods. Initially the
forecast accuracy of linear regression was better. The forecast accuracy of neural networks improved after removing
seasonal variation from the series. [5] proposed the use of independent component analysis and support vector
regression using two stage modeling in forecasting financial time series. It was opined that the most important
challenge of working with financial time series was the presence of noise and the need to make them stationary.
Hamzacebi et al. [6] used artificial neural network for multi periodic forecasting using iterative and direct
methods and compared their results using grey relational analysis. [7] applied deep neural networks (DNNs) for
prediction of financial market. The trading strategy was back-tested on commodity and futures markets. [8] used
data from France metropolitan’s electricity consumption data for electric load forecasting. The number of hidden
neural layers and optimal time lags were tuned for LSTM using genetic algorithm. It was found to give better
performance as compared to machine learning models.
[9] used genetic algorithms to optimize technical analysis parameters. These are then passed on to a deep neural
network for predictions. Enhanced stock trading performance was reported. [10] used a special type of artificial
neural network called CEFLANN. Stock trading prediction was modeled as a classification problem and model
performance was compared with other ML techniques like SVM, NB, KNN and DT. [11] used LSTM for out-of-
 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100 2093
Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000 3

sample prediction of financial time series. It was found to outperform logistic regression, DNN and RAF. A short
term trading strategy was devised based on the LSTM outputs.
[12] compared LSTM, RNN, CNN and MLP with linear and non-linear regression models. The networks were
trained with one company from NSE and then tested on five companies from NSE and NYSE. CNN was found to
outperform other models. [13] was one of the early papers that addressed the question of how many hidden layers
should be there in a neural network. Using remote sensing data, a near optimal solution was discovered with genetic
algorithm. [14] evaluated eight LSTM variants on three tasks: music modeling, handwriting recognition and speech
recognition. It was found that none of the variants significantly outperformed the basic LSTM architecture and that
the hyper-parameters were virtually independent. [15] compared LSTM with SVM, backpropagation and Kalman
filter for stock market for different number of epochs varying from 10 to 100. LSTM was found to have high
accuracy and low variance. [16] compared manual search and grid search techniques of hyper-parameter
optimization for neural networks and deep belief networks. It was found that the former is able to find as good or
even better models than the latter within lesser computation time. It was suggested that manual search be used as a
baseline for judging hyper-parameter optimization algorithms.

3. Long Short Term Memory (LSTM)

LSTM is a type of recurrent neural network (RNN). RNNs are a powerful type of artificial neural network that
can internally maintain memory of the input. This makes them particularly suited for solving problems involving
sequential data like a time series. However RNNs frequently suffer from a problem called vanishing gradient which
leads to the model learning becoming too slow or stopping altogether. LSTMs were created in the 1990’s to solve
this problem. LSTMs have longer memories and can learn from inputs that are separated from each other by long
time lags.
An LSTM has three gates: an input gate which determines whether or not to let the new input in, a forget gate
which deletes information that is not important and an output gate which decides what information to output. These
three gates are analog gates based on the sigmoid function which works on the range 0 to 1. These three sigmoid
gates can be seen in Fig. 1 below. A horizontal line that can be seen running through the cell represents the cell
state.

Fig. 1. LSTM Architecture [17]

2094 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100
4 Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000

3.1. Proposed System

Fig. 2. Main steps of the proposed model

3.2. Tuning Hyper-parameters

Fig. 3. Hyper-parameter tuning process

Fig. 3 shows a typical development cycle for any neural network. For large neural nets, training times can be long so
repeatedly running the network and evaluating the results can take days or weeks. For faster development of neural
network models, it is important to explore hyper-parameter tuning and develop recommendations that can provide a
good initial starting point for the kind of network being developed.
 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100 2095
Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000 5

4. Experimental Setup

4.1. Selection of companies

The most widely used benchmark for stocks in India is the NIFTY 50. This is provided by the National Stock
Exchange (NSE) which is the largest Indian stock exchange. As the name signifies, NIFTY 50 is made up of fifty
stocks. For this paper, stocks were picked from different sectors so that LSTM performance could be viewed on time
series that were influenced by different factors. The sector with the highest weightage of 37.94% in NIFTY is
financial services. Out of this ICICI Bank, which is the largest private sector bank in India, was selected. The sector
with the next highest weightage of 15.31% is Energy. Here Reliance, which is the first Indian company to reach
$100 billion in market capitalization, was chosen. The next sector is IT with a weightage of 14.46%. The biggest
company here is Tata Consultancy services Limited (TCS) which was chosen. The fourth biggest sector is
Automobiles where again the biggest company, Maruti, was chosen. The stock price data for these companies was
downloaded from the World Wide Web as shown in Fig. 2.
Keeping in mind easier allocation of batch size later while configuring LSTM, data was extracted for these four
companies for 2560 days for the period 29-12-2008 to 24-05-2019 which is over ten years of data. An 80:20 ratio
was implemented for training and testing data, yielding 512 days of testing.

4.2. Adjustment of closing price

The last quoted price for a stock during a day is called its closing price. This is the standard price considered
during financial time series analysis. This, however, needs to be adjusted for corporate actions before doing any
analysis on historical data. These actions could be stock splits, dividends or rights offerings. All of these affect the
“nominal” price of the stock though none of them affects its “true” price. For proper training of the LSTM, the price
effect of these actions needs to be removed from the nominal stock price. The closing price after these adjustments
is called the adjusted close price and has been used in our analysis.

4.3. Setting up the environment

Python 3.6.8 was used as the programming environment. A virtual environment was created to run the
experiments for this paper. In this virtual environment, the following packages were installed:
• Tensorflow 1.13.1
• Keras 2.2.4-tf
• Pandas 0.24.2
• Sklearn 0.21.1
• Numpy 1.16.3
• Matplotlib 3.0.3

4.4. Data Preparation

Before applying time series techniques, the data series needs to be made stationary. This means that it should be a
flat looking series without any long term trend or seasonality. A common method to do this is differencing.
Y=
i Z i − Z i −1 (1)

The resulting series has one element less because the first item has nothing to difference it from and has to be
skipped. For this reason 2560+1 data points were included in the dataset. After forecasting through LSTM this
difference has to be added back to get the original series.
The next step is to add an output component to the data. LSTM assumes that there are input values (time series)
which are to be used to predict an output value. Since the time series data only had an input series, the stock price
value from time t-1 was used as input for predicting the stock price value from time t as the output. A new input
2096 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100
6 Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000

series was created by pushing the values on the original series back by one time period. The original times series
contains the output variables.
For the LSTM default activation function, hyperbolic tangent (tanh) was used, whose output values range
between -1 and 1. The input data was scaled down to fit it within the range of the activation function. Again, once
the forecasts have been made, the forecasts must be scaled back up to the original scale.

4.5. LSTM Model

Experiment 1: Data is fed to LSTM in batches. Batch size is decided by the user (64 in our case) and defines the
number of rows from the dataset processed by the model before it updates its weights. While a batch is being
processed, LSTM maintains its state. Between batches, the state can be maintained or cleared. By default, the state is
cleared. This is called “Stateless.” In our first experiment this was compared with the other option of maintaining
state called “Stateful” to see which one yields better results.

Experiment 2: An important hyper-parameter in configuring LSTM is the number of neural layers. There are
three types of these layers in any LSTM – input layer, output layer and hidden layer. Out of these, input and output
have a single layer and this number is not configurable. The number of hidden layers can be configured. In the
second experiment LSTM was run with hidden layers varying from one to seven and the performance was
compared. This has been done for two of the companies – ICICI and TCS.
In both experiments, “mean_squared_error” was used as the LSTM loss function and ADAM as the optimization
algorithm. Root Mean Squared Error (RMSE) was used to evaluate the performance of the various models. A batch
size of 64 was used for training and 1 for testing. The number of epochs was fixed at 30 and the tests were repeated
30 times to ensure consistency.

5. Results and Analysis

The input data was visualized by plotting a combined line plot for the four companies. It is shown in Fig. 4.

Fig. 4. Combined plot of input data for four companies

 Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000 7
Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100 2097

Experiment 1: The input data was run through the LSTM and predictions were made on the test dataset. This
was done for Stateless and Stateful LSTM. The results of 30 repetitions were then statistically analyzed. These are
presented below where results for stateful are labelled A and those for stateless are labelled B.

Table 1. Comparison of Stateful (A) and Stateless (B) LSTM

ICICI-A ICICI-B TCS-A TCS-B RELIANCE-A RELIANCE-B MARUTI-A MARUTI-B
count 30 30 30 30 30 30 30 30
mean 5.851305 5.852057 25.596914 25.531364 17.00645 16.789178 112.526895 112.428441
std 0.026454 0.024332 0.247032 0.162271 0.298829 0.147629 0.383043 0.347485
min 5.824255 5.818488 25.374967 25.396562 16.715145 16.624479 111.932105 111.976718
25% 5.830218 5.829164 25.411047 25.412839 16.842389 16.668147 112.299 112.201765
50% 5.843647 5.849717 25.463858 25.457608 16.883783 16.752088 112.496621 112.320372
75% 5.86248 5.868858 25.78003 25.623765 17.104022 16.859496 112.682002 112.569707
max 5.920165 5.897414 26.257963 26.102652 18.08476 17.147931 113.44813 113.425736

Fig. 5. Box plots for Stateful (A) and Stateless (B) LSTM

From table 1 and Fig. 5, no major difference is found between the performance of stateful and stateless LSTM.
Mean RMSE for ICICI is lower for stateful as compared to stateless but for TCS, RELIANCE and MARUTI
stateless gives a lower mean RMSE. To find out if these differences are meaningful, a P-value test for statistical
significance was applied for a 5% significance level (alpha). The P-values were as follows: ICICI 0.898, TCS 0.216,
RELIANCE 0.002 and MARUTI 0.263. The values for three companies out of four (except RELIANCE) are greater
than the significance level (0.05) so it can be concluded that the differences in prediction accuracy between stateful
and stateless LSTM are not statistically significant.
8 Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000
2098 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100

Experiment 2: Other hyper-parameters were kept constant while the number of hidden layers (n) was varied. The
experiment was conducted with n values from 1 to 7. The results are given in table 2.
Table 2. Comparison of number of hidden layers for ICICI

One Two Three Four Five Six Seven

count 30 30 30 30 30 30 30
mean 5.851305 5.860783 5.867778 5.880775 5.881493 5.88441 5.892269
std 0.026454 0.015801 0.019696 0.011612 0.011708 0.011299 0.008682
min 5.824255 5.827461 5.832743 5.855847 5.860987 5.854128 5.87459
25% 5.830218 5.851372 5.85345 5.873032 5.871773 5.878954 5.886606
50% 5.843647 5.858564 5.864112 5.881874 5.88012 5.885006 5.890901
75% 5.86248 5.871088 5.879801 5.887195 5.889274 5.891017 5.896691
max 5.920165 5.890493 5.908064 5.902414 5.903165 5.903842 5.909898

Fig. 6. Box plots for number of hidden layers for ICICI

Table 3. Comparison of number of hidden layers for TCS

One Two Three Four Five Six Seven
count 30 30 30 30 30 30 30
mean 25.59691 25.47586 25.49538 25.45115 25.48475 25.48042 25.47779
std 0.247032 0.106443 0.101739 0.073064 0.059278 0.062683 0.046589
min 25.37497 25.3578 25.36995 25.35989 25.38053 25.39786 25.39795
25% 25.41105 25.4084 25.40937 25.38487 25.44425 25.42941 25.44357
50% 25.46386 25.45043 25.46922 25.43292 25.47954 25.46087 25.48252
75% 25.78003 25.50188 25.57136 25.50593 25.51148 25.52651 25.50313
max 26.25796 25.8193 25.72406 25.61197 25.62057 25.61386 25.57116
 Anita
Anita Yadav
Yadav et al.
et al./ / Procedia
Procedia Computer
Computer Science
Science 167 (2020)
00 (2019) 2091–2100
000–000 20999

Fig. 7. Box plots for number of hidden layers for TCS

Fig. 6 for ICICI shows very clearly that n = 1 is the best configuration as far as the mean RMSE is concerned.
This is less clear from Fig. 7 for TCS where there is not much to differentiate the various configurations. To validate
this statistically, one-way ANOVA test was applied. It was found that for ICICI there was a statistically significant
difference between groups as determined by one-way ANOVA (F(6,203) = 24.80, p = 5.54x10-22). For TCS, again,
there was a statistically significant difference between groups as determined by one-way ANOVA (F(6,203) = 4.76,
p = 1.45x10-4). Another thing that stands out clearly from all the figures is that the standard deviation decreases
drastically as n increases from 1 to 7.

6. Conclusion

In Experiment 1, stateful and stateless LSTMs were compared for four companies - ICICI, TCS, Reliance and
Maruti. The P-values showed that the differences between stateful and stateless LSTM for the stock price prediction
problem chosen for the experiment, are statistically insignificant. Most of the difference in values can be accounted
for by the random seeding that occurs for every LSTM run that produces small variations in output. The standard
deviation values and the spread in the box and whisker plot diagram shows that the deviation or spread for stateless
is less than that for stateful. This suggests that stateless LSTM is more stable compared to stateful LSTM. It can be
concluded that for time series prediction problems, a stateless LSTM model is preferable due to higher stability. If
one is doing language modelling where successive batches are successive chunks of text with possible
interconnections, it makes sense to go for stateful LSTM.
In Experiment 2, the number of hidden layers was varied from one to seven. The results show that n = 1 appears
to be the best configuration as far as mean RMSE is concerned. This was also validated by the one-way ANOVA
test. The recommendation here would be to go with one hidden layer for most problems due to better accuracy,
easier training and less risk of over-fitting. If there is a problem that is very complex and cannot be efficiently
modelled with one layer, one may want to go with higher number of layers but such cases are expected to be rare.
2100 Anita Yadav et al. / Procedia Computer Science 167 (2020) 2091–2100
10 Anita Yadav et al./ Procedia Computer Science 00 (2019) 000–000

An advantage of increasing the number of hidden layers is that the LSTM becomes more stable as revealed by the
decreasing standard deviation values and the spread in the box and whisker plot diagram.

7. Future Work and limitations

Deep learning techniques are highly computation intensive. The present work has been performed on Google
Colab which provides eight Tensor processing units (TPUs). If dedicated computational resources are available,
more experiments can be run with more data and higher epochs. This may provide more clarity on hyper-parameter
tuning.
The present work is limited to tuning the basic LSTM architecture. Different types of LSTM models can be
similarly tuned.

References

[1] Lam, Monica. (2004) “Neural network techniques for financial performance prediction: integrating fundamental and technical analysis.”
Decision Support Systems 37 (4): 567-581.
[2] Hutter, Frank, Holger Hoos, Kevin Leyton-Brown. (2014) “An Efficient Approach for Assessing Hyperparameter Importance.” Proceedings
of the 31st International Conference on Machine Learning (PMLR) 32 (1): 754-762.
[3] Persio, Luca Di, and Oleksandr Honchar. (2016) “Artificial Neural Networks architectures for stock price prediction: comparisons and
applications.” International Journal of Circuits, Systems and Signal Processing 10: 403-413.
[4] Foster, W. R., F. Collopy, and L. H. Ungar. (1992) “Neural network forecasting of short, noisy time series.” Computers & Chemical
Engineering 16 (4): 293-297.
[5] Lu, Chi-Jie, Tian-Shyug Lee, and Chih-Chou Chiu. (2009) “Financial time series forecasting using independent component analysis and
support vector regression.” Decision Support Systems 47 (2): 115–125.
[6] Hamzaebi, Coskun, Akay Diyar, and Kutay Fevzi. (2009) “Comparison of direct and iterative artificial neural network forecast approaches in
multi-periodic time series forecasting.” Expert Systems with Applications 36 (2): 3839-3844.
[7] Dixon, Matthew, Diego Klabjan, and Jin Hoon Bang. (2017) “Classification-based financial markets prediction using deep neural networks.”
Algorithmic Finance 6: 67–77.
[8] Bouktif, Salah, Ali Fiaz, Ali Ouni, and Mohamed Adel Serhani. (2018) “Optimal Deep Learning LSTM Model for Electric Load Forecasting
using Feature Selection and Genetic Algorithm: Comparison with Machine Learning Approaches.” Energies 11 (7): 1636.
[9] Sezer, Omer Berat, Murat Ozbayoglu, and Erdogan Dogdu. (2017) “A Deep Neural-Network Based Stock Trading System Based on
Evolutionary Optimized Technical Analysis Parameters.” Procedia Computer Science 114: 473-480.
[10] Dash, Rajashree, and Pradipta Kishore Dash. (2016) “A hybrid stock trading framework integrating technical analysis with machine learning
techniques.” The Journal of Finance and Data Science 2 (1): 42-57.
[11] Fischer, Thomas, and Christopher Krauss. (2018) “Deep learning with long short-term memory networks for financial market predictions.”
European Journal of Operational Research 270 (2): 654-669.
[12] Hiransha, M, Gopalakrishnan E. A., Vijay Krishna Menon, and Soman K. P. (2018) “NSE Stock Market Prediction Using Deep-Learning
Models.” Procedia Computer Science 132: 1351-1362.
[13] Stathakis, D. (2009) “How many hidden layers and nodes?” International Journal of Remote Sensing 30 (8): 2133–2147.
[14] Greff, Klaus, Rupesh Kumar Srivastava, Jan Koutník, Bas R. Steunebrink, and Jürgen Schmidhuber. (2017) “LSTM: A Search Space
Odyssey.” IEEE Transactions on Neural Networks and Learning Systems 28 (10): 2222 - 2232.
[15] Karmiani, D., R. Kazi, A. Nambisan, A. Shah, and V. Kamble. (2019) “Comparison of Predictive Algorithms: Backpropagation, SVM,
LSTM and Kalman Filter for Stock Market.” Amity International Conference on Artificial Intelligence (AICAI), Dubai, United Arab
Emirates: 228-234.
[16] Bergstra, James, and Yoshua Bengio. (2012) “Random Search for Hyper-Parameter Optimization.” Journal of Machine Learning Research
13: 281-305.
[17] Olah, Christopher. (2015) Understanding LSTM Networks [Blog post]. Retrieved from [Link]
LSTMs/