Received 13 March 2024, accepted 28 May 2024, date of publication 3 June 2024, date of current version 18 June 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3408222

Predicting Market Performance Using Machine

and Deep Learning Techniques
MOHAMED EL MAHJOUBY 1 , MOHAMED TAJ BENNANI 1 , MOHAMED LAMRINI1 ,
BADRE BOSSOUFI 2 , THAMER A. H. ALGHAMDI 3,4 , AND MOHAMED EL FAR1
1 Department of Computer Science, LPAIS Laboratory, Faculty of Science Dhar El Mahraz, Sidi Mohamed Ben Abdellah University, Fez 30003, Morocco
2 LIMAS Laboratory, Faculty of Sciences Dhar El Mahraz, Sidi Mohamed Ben Abdellah University, Fez 30003, Morocco
3 Wolfson Centre for Magnetics, School of Engineering, Cardiff University, CF24 3AA Cardiff, U.K.
4 Electrical Engineering Department, Faulty of Engineering, Al-Baha University, Al-Baha 65779, Saudi Arabia

Corresponding authors: Mohamed El Mahjouby ([email protected]) and Badre Bossoufi

([email protected])

ABSTRACT Today, forecasting the stock market has been one of the most challenging issues for the
‘‘artificial intelligence’’ AI research community. Stock market investment methods are sophisticated and
rely on analyzing massive volumes of data. In recent years, machine-learning techniques have come under
increasing scrutiny to assess and improve market predictions over traditional approaches. The observation
in time is due to their dependence. Their predictions are crucial tasks in data mining and have attracted
great interest and considerable effort over the past decades. Tackling this challenge remains difficult due to
the inherent characteristics of time series data, including its high dimensionality, large volume of data, and
constant updates. Exploration of Machine Learning and Deep Learning methods undertaken to enhance the
effectiveness of conventional approaches. In this document, we aim precisely to forecast the performance of
the stock market at the close of the day by applying various machine-learning algorithms on the two data
sets ‘‘CoinMarketCap, CryptoCurrency’’ and thus analyze the predictions of the architectures.

INDEX TERMS Machine learning, deep learning, LSTM, ARIMA, linear regression.

I. INTRODUCTION is not viable since investors would quickly uncover such

The stock market is one of the indicators of a country’s strategies, and good forecasting rules will self-destruct. The
economy. Few people excel at correctly comprehending the stock market is both a difficulty and a chance for investors
evolution of the stock trend, so many people are afraid to benefit. However, investors are aware of the ‘‘High Risk,
to invest in stocks. By its forecast scope, data science High Return’’ philosophy and are looking for ways to reduce
(data analytics) has shattered the assumed paradigm of risk while increasing earnings. Fundamental analysis and
the only engagement of economics and finance in stocks. technical analysis are the two methods of stock forecast-
In the financial domain, stock price is a type of time ing. Fundamental analysis forecasts prices using income
series. Financial time series variations are dynamic, selective, statements, balance sheets, and cash flows to find chances
nonlinear, nonstationary, and noisy, making forecasting for growth, financial health, and competitive advantages.
challenging. One of the most pressing concerns is predicting It also examines the business and national environment,
stock prices effectively utilizing data mining or machine the country’s economy, and business possibilities in certain
learning approaches. However, because the stock price will industrial groups. Technical analysis, on the other hand,
follow a random work pattern, it is challenging to make forecasts stock values based on price history.
forecasts based on the premise of the efficient market Stock forecasting has long been a challenging topic for
hypothesis. Furthermore, a stationary prediction approach statisticians and finance experts. The reason for this forecast
is to acquire stocks expected to gain in value and then sell
The associate editor coordinating the review of this manuscript and stocks expected to fall in value. In general, there are two
approving it for publication was Ahmed Farouk . approaches to stock market forecasting. Two approaches
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
VOLUME 12, 2024 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ 82033
 M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

are utilized for analyzing stock markets. The initial method Researchers typically use the Decision Tree algorithm to
is fundamental analysis, which depends on a company’s predict financial time series because it generates rules to
strategy and essential data, including its market position, understand and analyze [5]. McNally et al. [6] estimated the
expenditures, and yearly growth rates. The second method coin’s price with 52% accuracy using a Bayesian-optimized
involves technical analysis, relying on historical stock prices recurrent neural network. Rekha et al. [7] studied the use
and values. This study forecasts future prices using past of RNN and CNN algorithms. These models’ precision
charts and trends. Financial experts used to be able to foretell is assessed by comparing their accuracy to actual stock
stock market movements. However, developments in machine market values in real-world settings. Another study in [8]
learning techniques have enabled data scientists to overcome used autoregressive (AR) properties in an LSTM network
prediction difficulties. In addition, computer scientists have to forecast daily BTC values. Their suggested LSTM-AR
begun to use machine-learning algorithms to increase the model outperformed a regular LSTM model in terms of
performance and accuracy of forecasting models. MSE) RMSE, MAPE, and MAE. Using the hourly values
The next step was to employ deep learning to improve the for BTC, ETH, and XRP, Livieris et al. [9] employed deep
prediction models’ performance. Stock market forecasting learning methods to predict patterns and values of specific
is fraught with difficulties, and data scientists frequently cryptocurrencies. They suggested an ensemble learning
encounter challenges when attempting to produce a predictive system that incorporated Bi-LSTM, LSTM, and CNN and
model. The overarching question is how to anticipate whether discovered that it was capable of making accurate and
the price of given stocks will rise or fall over the next year. dependable predictions.
It is critical to address this issue. In this paper, we employed In a different investigation [10], a combined approach that
methodologies from Machine Learning and Deep Learning. integrated LSTM and GRU networks was applied to predict
In general, different academics utilized time series to the prices of LTC and Monero. This hybrid method contrasted
forecast prices using deep learning and machine learning with a sole LSTM approach, and the results indicated that
methodologies. This study provides recommendations to help the combined model demonstrated superior accuracy in
investors make informed choices about closing prices. forecasting the prices of the specified cryptocurrencies. In a
The organization of our paper is as follows: Initially, study by Chowdhury et al. [11], the application of ensemble
we offer an introduction, and after an exploration of the approaches based on machine learning, such as KNN,
literature review and various machine learning algorithms. ANN, gradient-boosted trees, and a combined ensemble
Next, we feature an interview focused on deep learning long model, was examined to predict the prices of 9 prominent
short-term memory. After that, we present and describe our cryptocurrencies. The findings indicated that the ensemble-
methodologies. Subsequently, we present and analyze our learning model exhibited the lowest prediction error.
findings. Finally, we conclude with final remarks.
III. MACHINE LEARNING
II. LITERATURE REVIEW Machine learning is based on experience, and is a system
Sun et al. [1] identified daily patterns in the Bitcoin that works from algorithms that, fed with data, tend to learn
market while examining semantics and insights connected and improve automatically. The continuous improvement
to elements impacting Bitcoin’s price. As part of their processes are based on experience and not through pro-
research, they anticipated changes in Bitcoin’s daily price gramming. Thus, learning consists of processing observations
using both Random Forest and Bayesian regression models. or data (examples, experience, or instructions) to find
Pirani et al. [2] introduced a comparative study into existing models that allow the implementation of predictions and
literature, examining the performance of the BiLSTM model decision-making.
in forecasting time series data. Their analysis primarily
centered on evaluating the behaviors of ARIMA, BiLSTM, A. LINEAR REGRESSION
GRU, and LSTM models. Makala and Li [3] investigated We use linear regression (LR) to forecast cryptocurrency.
gold commodity forecasting using ARIMA and a Support In this scenario, the linear regression mathematical equation,
Vector Machine, and three types of kernel were utilized in this you can find the formula in Equation (1).
publication. Lee et al. [4] used three models to forecast four
w = bc + y (1)
key currency pairs. This research compares the performance
of the CNN, LSTM, and ARIMA models in forecasting In this case, the dependent variable is symbolized by W,
currency pair prices. Each model uses EUR/USD closing while the independent variable is marked by y. The slope is
price data from 1, 3, 5, and 7 years. The forecasting findings denoted by b, and the interceptor by c.
revealed that ARIMA is the best for data with fewer than
seven years of closing prices, while LSTM and CNN are B. SGDREGRESSOR
better suited for training with data with seven years of closing SGD Regressor is a machine-learning algorithm that
prices. Because of its success in processing vast volumes of makes predictions using stochastic gradient descent. The
data, LSTM is the better of the two deep learning models for SGDRegressor is a technique that iteratively optimizes an
currency forecasts. objective function chosen for its smoothness properties.

82034 VOLUME 12, 2024

M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

Its stochastic nature stems from the random selection or components: input, output, and one or more hidden layers
shuffling of samples to assess gradients. This randomness of hidden units [18]. Deep learning handles vast amounts
characterizes SGD as a stochastic approximation of gradient of data by organizing intricate data structures in manners
descent optimization. SGDRegressor proves to be a straight- beyond the capability of an artificial neural network (ANN)
forward yet highly effective method for discriminative [19]. A prototype of data processing that works in the same
learning commonly useful for convex loss functions like way as the human nervous system, it analyzes and processes
those found in linear Support Vector Machines and Logistic information in a similar way to our brain this prototype
Regression [12]. contains a lot of elements connected and working to solve
different problems the main reason to build a neural network
C. RANDOM FOREST is its similarity with the human brain, they wanted to test
Random Forest (RF) [13] is a versatile ensemble learning the different possible manipulations that can happen to our
approach for classification and regression tasks. Employing brain, so they presented it as a group of connected elements
bootstrapped aggregation and random feature selection that create a neural network or as called neural. We should
constructs decision trees within a forest. RF amalgamates know that our brain has over a billion neurons and each one
the simplicity of single decision trees, yielding class mode is connected to a million cells.
for classification and mean prediction for regression by Deep learning [20], a subset of machine learning tech-
leveraging multiple trees. Renowned for its favorable traits niques, focuses on discerning distinctions within data through
such as robust generalization, simplicity, resilience, and various architectural approaches. Artificial neural networks
reduced variance, RF finds extensive application. serve as the foundation for deep learning methods, executing
the learning process through multi-layered neural networks.
D. BAGGING The objective is to create abstract models that effec-
Bagging [14] is an abbreviation for Bootstrap Aggregation, tively handle data, encompassing diverse deep learning
one of the most well-known and successful ensemble learning architectures such as Recurrent Neural Networks, Forward
approaches. Breiman popularized bagging. The basic concept Neural Networks, and Convolutional Neural Networks. This
of bagging [15] is straightforward. Bagging intends to deep learning has made a revolution in the field of voice
generate multiple classifiers in parallel and then aggregate and image recognition, financial prediction, handwritten
them together. As a result, it chooses a basic classifier character recognition, and data compression, and also in the
algorithm to train the base classifiers on randomly distributed field of auto-driving such as the automatic detection of traffic
training datasets. The primary concept of bagging [16] is lights or cuts, without forgetting the medical field which
to randomly sample from a training dataset using a weak allows the scientist to automatically analyze cancer cells, also
learning algorithm. This sampling process involves selecting in the field of safety, especially in factories, he can detect the
n samples from the original training set, with some samples safety distance between employers and machines.
potentially being chosen multiple times and others not being
selected at all. Each subset of the training data is then
assigned a base classifier. When faced with an unknown
dataset, the findings of each base classifier are considered as A. LSTM
votes, and the final classification outcomes are established by Long short-term memories, commonly referred to as
aggregating the votes. ‘‘LSTM’’ [21], are a type of recurrent neural network [22]
that excels at capturing long-term dependencies. Initially
E. ADABOOST proposed by Hochreiter and Schmidhuber in 1997, these
This gives more weight [17] to the observation that was concepts have undergone continuous refinement and gained
misclassified to the last weak learner usually it is a decision widespread attention through subsequent research by various
stub which means a level one decision tree, which also means scholars. They are currently frequently utilized and function
the tree is based on decision variables and contains a root quite well in a wide range of situations. Long Short-Term
and the child connected to it. At each iteration, a model is Memory (LSTM) [23] represents a specialized form of
learned and the samples are re-weighted so that subsequent Recurrent Neural Network (RNN) tailored to efficiently
classifiers focus on samples that were incorrectly predicted handle and forecast data characterized by recurrent time
by the previous classifier. Correctly predicted samples have series. Its utility extends across diverse fields like natural
their weights reduced. The weights of samples that were language processing, speech recognition, and time series
inaccurately predicted are increased. The final prediction is forecasting. Unlike conventional RNN architectures, LSTM
a weighted average of model projections. demonstrates superior performance particularly with lengthy
data sequences, mitigating challenges such as vanishing
IV. DEEP LEARNING and exploding gradients. Additionally, LSTM stands out in
A deep neural network is a parallel processing architecture capturing extended patterns and dependencies within the data
composed of linked neurons arranged in layers. A deep neural over time. LSTMs (Long Short-Term Memory networks)
network’s fundamental structure is made up of three crucial [24] also possess a similar chain-like structure, but the

VOLUME 12, 2024 82035

 M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

repeating module differs. Instead of employing a solitary ARIMA (Autoregressive integrated moving average)
neural network layer, four layers interact distinctively. to predict the closing prices for CoinMarketCap and
CryptoCurrency.
V. MATERIALS AND METHODS 3) Predict closing prices: In this part, we have forecasted
A. DATA closing prices for CoinMarketCap and CryptoCur-
Our dataset includes the following attributes: Open, High, rency.
Low, and Close, representing the daily stock prices of 4) Comparing different prices: We compared different
cryptocurrency and CoinMarketCap. The cryptocurrency forecasted closing prices.
data spans from September 17, 2014, to December 8,
2023, while the CoinMarketCap data covers the period from C. EVALUATION MEASURE
December 1, 2013, to December 11, 2023, without any One metric, including R2 , is used for comparing various
information gaps. machine learning and deep learning approaches mentioned
above. You can find the formula in Equation (2). y Is the mean
B. METHODOLOGY of the metric, and b
yi is the forecasted outcome.
After downloading the datasets, we used different machine n
yi )2
P
learning and deep learning techniques such as linear regres- (yi − b
i=1
sion, SGDRegressor (stochastic gradient descent regressor), R2 = 1 − n (2)
random forest regressor, adaptive boosting regressor, LSTM P
yi )2
(yi − b
(Long short-term memory), and ARIMA (Autoregressive i=1
integrated moving average) to predict the closing prices for
CoinMarketCap and CryptoCurrency. After that, we com- VI. RESULTS AND DISCUSSION
pared the results obtained from various algorithms used in A. ANALYSIS OF RESULTS
this study, which provides recommendations to help investors This part contains an analysis of predictions of the following
make informed choices about closing prices. We utilized one architectures: LSTM, ARIMA, and some machine learning
measure of accuracy for comparing different results. Figure 1 methods applied to the two databases Coin_Market_cap
decribes our methodology. and Crypto Currency. For our study, we will apply the
architectures on the two databases and proceed by analyzing
the results of the predictions of each architecture and then be
able to compare them.
In our application, we apply several models that will
analyze the stock market quotes in real-time and be able to
make a prediction of the future and assess the future risk
through artificial intelligence. Thus, the system will be able
to make the appropriate decision at a given moment to invest
or to withdraw from the market. This paragraph presents
the different architectures LSTM and ARIMA [25] and
the different Machine Learning algorithms applied to these
databases. Finally, we will analyze the predictions and the
results obtained to compare the LSTM, Linear Regression,
SGDRegressor, and ARIMA approaches. This research
allows investors to make informed decisions on closing
prices. Figures 2 through 11 demonstrate the application of
diverse machine learning and deep learning algorithms.
As you can see in Figure 2, and Figure 3, we have showed
the forecasted price with actual price using ARIMA model.
As you can see, in Figure 4, the accuracy of LSTM reaches
FIGURE 1. Our methodologies. 98.49% for CoinMarketCap, and in Figure 5, the accuracy of
LSTM is 98.35% for CryptoCurrency.
1) Dataset: We have collected our datasets consisting of In Figure 6, the adaptive accuracy stands at 98.35%,
daily prices for CoinMarketCap and CryptoCurrency. bagging achieves an accuracy of 99.25%, and Random Forest
2) Applied various machine learning and deep learning demonstrates an accuracy rate of 99.17%.
techniques: In this phase, we used different machine In Figure 7, the adaptive accuracy stands at 97.62%,
learning and deep learning techniques such as linear bagging achieves an accuracy of 98.68%, and Random Forest
regression, SGDRegressor (stochastic gradient descent demonstrates an accuracy of 98.51%.
regressor), random forest regressor, adaptive boosting Figure 8 displays the performance of Linear Regression on
regressor, LSTM (Long short-term memory), and CoinmarketCap, achieving an accuracy of 98.81%. Predicted

82036 VOLUME 12, 2024

M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

FIGURE 5. LSTM Cryptocurrency.

FIGURE 2. ARIMA CoinMarketCap.

FIGURE 6. Several models for Cryptocurrency.

FIGURE 3. ARIMA CryptoCurrency.

FIGURE 4. LSTM CoinMarketCap.

FIGURE 7. Several models for CoinMarketCap.

values are depicted by the orange line, and the actual trends
are represented by the blue line. In Figure 11, we employed the SGDRegressor model
In Figure 9, the SGDRegressor model applied to CoinMar- for Cryptocurrency analysis achieving a precision rate of
ketCap demonstrates a precision rate of 98.82%. The orange 99.72%. We represented the anticipated values by the orange
line illustrates the projected values, while the blue line depicts line and the observed trends by the blue line. Figures 2
the observed trends. through 11 demonstrate the application of diverse machine
In Figure 10, Linear Regression achieves an accuracy learning and deep learning algorithms. The previous results
of 99.82% in predicting cryptocurrency values. The orange show that the forecasts of our models are well predicted by
line illustrates the predicted values, while the blue line the stock market actions and follow the market trend whether
corresponds to the actual trends. it is bullish or bearish.
VOLUME 12, 2024 82037
 M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

FIGURE 8. Lineare regression CoinMarketCap.

FIGURE 11. SGDRegressor Cryptocurrency.

us the best results is the fact that it makes trees

grow randomly, and alternately of looking for the
best characteristics when a division of nodes occurs,
it looks for the best advantage from a random subset
of forest characteristics, which gives a great diversity.
More neural networks have become increasingly used
in forecasting the time series because they correct some
defects of ‘classical’ methods such as ARIMA.
3) ARIMA also gives a good result, called robust and effi-
cient in forecasting financial time series. A statistical
technique that uses historical market data to predict its
performance. The results show a performance compar-
FIGURE 9. SGDRegressor CoinMarketCap.
ison of various machine-learning methods. Experiment
analysis shows that linear regression outperforms
SGDRegressor, LSTM, ARIMA (Autoregressive Inte-
grated Moving Average) in words of accuracy.
Table 1 illustrates our results, which show that the accuracy
of the linear regression model outperforms other algorithms.
The findings in Table 1 indicate that the linear regres-
sion model achieves a prediction accuracy of 99.82% for
the closing price of the cryptocurrency, followed by the
SGDRegressor algorithm in second place with an accuracy of
99.72%, and bagging regressor algorithm in third place with
an accuracy of 99.25%.

TABLE 1. Comparison between different algorithms.

FIGURE 10. Lineare regression Cryptocurrency.

1) As you see, the results are different by contributions

to the base of données.la coin market cap database and
more voluminous than the crypto_currency impacts the
results of no models ‘a good result by contribution
to the data set coin market cap’. A voluminous data
set helps our models to train well and gives good
predictions. VII. CONCLUSION
2) Random forest gives the good results among the other This thesis contributes to the research theme by comparing
models because The main reason random Forest gives different Machine Learning and Deep Learning architectures

82038 VOLUME 12, 2024

M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

for predicting stock market trends. The main targeted applica- [15] W. Wu and M. Peng, ‘‘A data mining approach combining K -means
tion of this work is to study the existing state of the art to apply clustering with bagging neural network for short-term wind power
forecasting,’’ IEEE Internet Things J., vol. 4, no. 4, pp. 979–986,
it to a database and thus make predictions and analyze them. Aug. 2017, doi: 10.1109/JIOT.2017.2677578.
The latter allowed us to evaluate our models to know their [16] X.-D. Zeng, S. Chao, and F. Wong, ‘‘Optimization of bagging classifiers
strong and weak points and thus to extract the advantages based on SBCB algorithm,’’ in Proc. Int. Conf. Mach. Learn. Cybern.,
Jul. 2010, vol. 1, pp. 262–267, doi: 10.1109/ICMLC.2010.5581054.
of each to make a better prediction. We have proposed [17] Y.-P. Wu, J.-M. Mao, and W.-F. Li, ‘‘Predication of futures market by
to improve prediction by a method that joins the trend using boosting algorithm,’’ in Proc. Int. Conf. Wireless Commun., Signal
predicted by different algorithms. To understand the different Process. Netw. (WiSPNET), Mar. 2018, pp. 1–4, doi: 10.1109/WiSP-
NET.2018.8538586.
aspects of the architectures, we applied modifications to the [18] A. Baharin, A. Abdullah, and S. N. M. Yousoff, ‘‘Prediction of bioprocess
architectures and examined the change in the predictions to production using deep neural network method,’’ Telkomnika Telecommun.
understand the effect of the last on the result. In this article, Comput. Electron. Control, vol. 15, no. 2, p. 805, Mar. 2017, doi:
10.12928/telkomnika.v15i2.6124.
we have used distinct algorithms to predict closing prices [19] H. H. Martínez, H. M. Ariza, and L. A. G. Roa, ‘‘Strategy to determine the
for CoinMarketCap and cryptocurrency. In our future work, foot plantar center of pressure of a person through deep learning neural
we recommend using algorithms on other stock markets. networks,’’ Telkomnika Telecommunication Comput. Electron. Control,
vol. 18, no. 4, p. 1954, Aug. 2020, doi: 10.12928/telkomnika.v18i4.15200.
[20] S. Uzun, S. Kaçar, and B. Arıcıoğlu, ‘‘Deep learning based classification
REFERENCES of time series of chaotic systems over graphic images,’’ Multimedia Tools
Appl., vol. 83, no. 3, pp. 8413–8437, Jan. 2024, doi: 10.1007/s11042-023-
[1] X. Sun, M. Liu, and Z. Sima, ‘‘A novel cryptocurrency price trend 15944-3.
forecasting model based on LightGBM,’’ Finance Res. Lett., vol. 32, [21] W. Kong, Z. Y. Dong, Y. Jia, D. J. Hill, Y. Xu, and Y. Zhang, ‘‘Short-term
Jan. 2020, Art. no. 101084, doi: 10.1016/j.frl.2018.12.032. residential load forecasting based on LSTM recurrent neural network,’’
[2] M. Pirani, P. Thakkar, P. Jivrani, M. H. Bohara, and D. Garg, ‘‘A IEEE Trans. Smart Grid, vol. 10, no. 1, pp. 841–851, Jan. 2019, doi:
comparative analysis of ARIMA, GRU, LSTM and BiLSTM on financial 10.1109/TSG.2017.2753802.
time series forecasting,’’ in Proc. IEEE Int. Conf. Distrib. Comput. Electr. [22] W. Chen, Y. Zhang, C. K. Yeo, C. T. Lau, and B. S. Lee, ‘‘Stock
Circuits Electron. (ICDCECE), Apr. 2022, pp. 1–6, doi: 10.1109/ICD- market prediction using neural network through news on online social
CECE53908.2022.9793213. networks,’’ in Proc. Int. Smart Cities Conf. (ISC2), Sep. 2017, pp. 1–6,
[3] D. Makala and Z. Li, ‘‘Prediction of gold price with ARIMA and SVM,’’ doi: 10.1109/ISC2.2017.8090834.
J. Phys., Conf. Ser., vol. 1767, no. 1, Feb. 2021, Art. no. 012022, doi: [23] H. Zhang, L. Wang, and W. Shi, ‘‘Seismic control of adaptive variable
10.1088/1742-6596/1767/1/012022. stiffness intelligent structures using fuzzy control strategy combined with
[4] J.-S. Lee, Y.-L. Yong, M.-H. Hoo, and K.-C. Khor, ‘‘A comparison of LSTM,’’ J. Building Eng., vol. 78, Nov. 2023, Art. no. 107549, doi:
machine learning models for currency pair forecasting,’’ in Intelligent 10.1016/j.jobe.2023.107549.
Computing & Optimization (Lecture Notes in Networks and Systems). [24] S. Hochreiter and J. Schmidhuber, ‘‘Long short-term memory,’’
Algeria: Springer, 2023, pp. 499–506, doi: 10.1007/978-3-031-19958- Neural Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 1997, doi:
5_47. 10.1162/neco.1997.9.8.1735.
[5] C.-F. Tsai, Y.-C. Lin, D. C. Yen, and Y.-M. Chen, ‘‘Predicting stock [25] R. Y. C. Tse, ‘‘An application of the ARIMA model to real-estate prices in
returns by classifier ensembles,’’ Appl. Soft Comput., vol. 11, no. 2, Hong Kong,’’ J. Property Finance, vol. 8, no. 2, pp. 152–163, Jun. 1997,
pp. 2452–2459, Mar. 2011, doi: 10.1016/j.asoc.2010.10.001. doi: 10.1108/09588689710167843.
[6] S. McNally, J. Roche, and S. Caton, ‘‘Predicting the price of Bitcoin
using machine learning,’’ in Proc. 26th Euromicro Int. Conf. Parallel,
Distrib. Network-Based Process. (PDP), Mar. 2018, pp. 339–343, doi:
10.1109/PDP2018.2018.00060.
[7] G. Rekha, D. B. Sravanthi, S. Ramasubbareddy, and K. Govinda,
‘‘Prediction of stock market using neural network strategies,’’ J. Comput.
Theor. Nanoscience, vol. 16, no. 5, pp. 2333–2336, May 2019, doi: MOHAMED EL MAHJOUBY received the mas-
10.1166/jctn.2019.7895.
ter’s degree in information systems engineering
[8] C.-H. Wu, C.-C. Lu, Y.-F. Ma, and R.-S. Lu, ‘‘A new forecasting (MSID) from Sidi Mohamed Ben Abdellah Uni-
framework for Bitcoin price with LSTM,’’ in Proc. IEEE Int. Conf.
versity, Fez, Morocco, in 2020. His thesis was
Data Mining Workshops (ICDMW), Nov. 2018, pp. 168–175, doi:
on ‘‘stock market predictive system analysis and
10.1109/ICDMW.2018.00032.
modeling.’’ His research interests include artificial
[9] I. E. Livieris, E. Pintelas, S. Stavroyiannis, and P. Pintelas, ‘‘Ensemble deep
intelligence and machine learning.
learning models for forecasting cryptocurrency time-series,’’ Algorithms,
vol. 13, no. 5, p. 121, May 2020, doi: 10.3390/a13050121.
[10] M. M. Patel, S. Tanwar, R. Gupta, and N. Kumar, ‘‘A deep learning-
based cryptocurrency price prediction scheme for financial institu-
tions,’’ J. Inf. Secur. Appl., vol. 55, Dec. 2020, Art. no. 102583, doi:
10.1016/j.jisa.2020.102583.
[11] R. Chowdhury, M. A. Rahman, M. S. Rahman, and M. R. C. Mahdy,
‘‘An approach to predict and forecast the price of constituents and index
of cryptocurrency using machine learning,’’ Phys. A, Stat. Mech. Appl.,
vol. 551, Aug. 2020, Art. no. 124569, doi: 10.1016/j.physa.2020.124569. MOHAMED TAJ BENNANI received the mas-
[12] A. Gonzalez-Briones, G. Hernandez, T. Pinto, Z. Vale, and J. M. Corchado, ter’s and Ph.D. degrees in computer science and
‘‘A review of the main machine learning methods for predicting residential networking from the Faculty of Science and Tech-
energy consumption,’’ in Proc. Int. Conf. Eur. Energy Mark. (EEM), nology of Tangier, Tangier, in 2011 and 2019,
Sep. 2019, pp. 1–6, doi: 10.1109/EEM.2019.8916406. respectively. At present, he is working as a
[13] H. H. Htun, M. Biehl, and N. Petkov, ‘‘Survey of feature selection and Professor with the Faculty of Science Dhar El
extraction techniques for stock market prediction,’’ Financial Innov., vol. 9, Mahraz, Fez, since 2019. He has been a qualified
no. 1, pp. 83–96, Jan. 2023, doi: 10.1186/s40854-022-00441-7. university professor since June 2023.
[14] R. Richman and M. V. Wüthrich, ‘‘Nagging predictors,’’ Risks, vol. 8,
no. 3, p. 83, Aug. 2020, doi: 10.3390/risks8030083.

VOLUME 12, 2024 82039

 M. El Mahjouby et al.: Predicting Market Performance Using Machine and Deep Learning Techniques

MOHAMED LAMRINI received the Ph.D. degree THAMER A. H. ALGHAMDI received the
from the Claude Bernard University Lyon 1, B.Sc. degree from Al-Baha University, Al-Baha,
in 1993. He is currently a Professor of computer Saudi Arabia, in 2012, the M.Sc. degree from
science with USMBA, Fez, where he is also a Northumbria University, Newcastle upon Tyne,
member of the LPAIS Laboratory. His research U.K., in 2016, and the Ph.D. degree from Cardiff
interests include software quality (CMMI, Six University, Cardiff, U.K., in 2023. He was a Power
Sigma, ISO 9001,. . . ) and industrial engineering Distribution Engineer at Saudi Electricity Com-
(methods and statistical tools). pany (SEC), in 2013. He was a Lecturer Assistant
at Al-Baha University, from 2016 to 2018, where
he is currently an Assistant Professor of electrical
power engineering. His main research interests include power systems,
power quality, the integration of renewables, and AI applications in electrical
engineering.

BADRE BOSSOUFI was born in Fez, Morocco,

in 1985. He received the Ph.D. degree in electrical
engineering from the Faculty of Sciences, Sidi
Mohammed Ben Abdellah University, Morocco, MOHAMED EL FAR is currently a Professor
and the joint Ph.D. degree from the Faculty of with Sidi Mohamed Ben Abdellah University,
Electronics and Computer, University of Pitesti, Fez, Morocco. He teaches computer science. His
Romania, and the Montefiore Institute of Electri- research interests include data science, big data,
cal Engineering, Liège, Belgium, in 2013. He was and artificial intelligence. He has published many
an Assistant Professor of electrical engineering researches.
at the Higher School of Technology, Oujda,
Morocco. His research interests include static converters, electrical motor
drives, power electronics, smart grids, renewable energy, and artificial
intelligence.

82040 VOLUME 12, 2024