Received 14 June 2024, accepted 13 August 2024, date of publication 22 August 2024, date of current version 2 September 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3448234

Sell or HODL Cryptos: Cryptocurrency

Short-to-Long Term Projection Using
Simultaneous Classification-Regression
Deep Learning Framework
MUHAMMAD IQBAL 1 , ARSHAD IQBAL 1 , (Member, IEEE),
ABDULLAH ALSHAMMARI2 , (Member, IEEE), IHTISHAM ALI 1 ,
LOUAI A. MAGHRABI 3 , (Member, IEEE), AND NIGHAT USMAN4
1 Sino-Pak Center for Artificial Intelligence (SPCAI), Pak-Austria Fachhochschule: Institute of Applied Sciences and Technology (PAF-IAST), Harı̄pur 22620,
Pakistan
2 College of Computer Science and Engineering, University of Hafr Al Batin, Hafar Al Batin 31991, Saudi Arabia
3 Department of Software Engineering, College of Engineering, University of Business and Technology, Jeddah 21361, Saudi Arabia
4 Department of Computer Science, COMSATS University Islamabad, Abbottabad 22060, Pakistan

Corresponding author: Arshad Iqbal ([email protected])

ABSTRACT Decentralized cryptocurrencies like Bitcoin are digital assets with a price volatility nature, that
allow for blockchain-based, peer-to-peer monetary transactions. Due to the price volatility problem with
decentralized cryptocurrencies, research into the underlying pricing mechanism is required. Additionally,
the behavior of Bitcoin prices is non-stationary, meaning that the statistical distribution of data varies over
time. The proposed framework demonstrates the use of sophisticated machine learning models in predicting
the short and medium-term trends and actual values of Bitcoin prices. This research goes beyond previous
work that has only looked at machine learning-based categorization for a single day by instead using such
models to forecast price changes seven, thirty, and ninety days into the future. The generated models are
useful and work admirably, with the classification models reaching a maximum accuracy enhancement up
to 31.48% for a 90-day prediction and a 11.76% F1-score for a forecast extending to the thirtieth day. In the
case of regression the margin of error shifts from the one-time horizon for price projections to the next.
A significant notable downfall occurs in different error metrics. These findings suggest that the models
given here outperform those already found in the literature.

INDEX TERMS Bitcoin, machine learning, support vector machine (SVM), deep learning, long short-term
memory (LSTM), artificial neural network (ANN).

I. INTRODUCTION dependent on the ruling government of a state [2]. Negative

Amove from traditional payment methods like checks and consequences might result from the government meddling
money orders to more modern digital payment systems in the financial sector, potentially catastrophic outcomes
reflects the ever-changing nature of the technological that were not envisaged. However, digital currencies’ worth
world [1]. Money’s utility, whether as a means of exchange depends on the reliability and safety of the network on which
or a store of value, depends on a business arrangement they are used [3]. When using traditional forms of digital
or investing to estimate its future worth. When dealing currency, it is possible to spend the same amount twice.
with monetary systems, the value and stability are largely The most significant change affecting economic and
banking institutions currently is the widespread use of
The associate editor coordinating the review of this manuscript and digital technologies [4]. Economies and financial institutions
approving it for publication was Daniel Augusto Ribeiro Chaves . throughout the globe seem to have become digital at a
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/ 118169
 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

FIGURE 1. Cost of BTC between September 2014 and November 2022.

rate that has never been seen before. The analysis predicts decentralized currency system needs confirmation from most
that by 2025, the combined value of all digital assets, of the connected digital nodes in the blockchain network
both physical and invisible or ethereal, would account for to mitigate dual-spending issues [14]. These decentralized
25% of the global economy, or $23 trillion [5]. Another network confirmations make the network more secure. The
survey states that by 2024, the cryptocurrency industry blockchain ledger’s financial transactions are incorruptible
is expected to increase from its current value of $266 since it would be computationally difficult to change a record
billion by a compound annual growth rate (CAGR) of across the channel’s connected nodes [15]. As a result,
11.9% [6]. Distributed ledger technology (DLT), with Bitcoin the successful operation also known as transaction makes
being it is most well-known implementation is the leading the record unalterable [16]. The widespread availability
technology for creating and exchanging digital assets. and features of cryptocurrencies enable it as a means of
These advancements have led to blockchain technology’s transactions and savings [17].
acceptance at the crossroads of Financial Inclusion and Next However, cryptocurrency value is still very susceptible to
Generation Networks [7]. Attempts on the security of online market dynamics volatility and social impacts [18]. Further-
digital currency exchanges may compromise the integrity more, crypto assets tend to be uncorrelated with traditional
of customers’ financial information [8]. More and more of ones. This renders obsolete the time-tested practices of
these problems make conventional currencies vulnerable to the financial sector. The fluctuation of intangible digital
volatility and deterioration. assets, particularly cryptocurrencies, is a major concern.
Using digital currencies based on a distributed ledger In FIGURE 1, we analyze the market worth of one Bitcoin
like Bitcoin might be a workable answer to the mentioned (BTC) for eight consecutive years. The price of BTC has been
challenges [9]. Blockchain is a new technology that keeps quite volatile recently. Although the pricing value climbed by
data permanently throughout a network, offering safety, 1900% during 2017, it has fallen by 72% during 2018 [19].
decentralization, and visibility that are essential for a Moreover, 80% of all cryptocurrency enterprise value was
functional currency. Cryptocurrencies are digital payment held by Bitcoin alone during 2017, but this percentage has
systems that use cryptographic hash functions rather than a dropped below 50% by 2022 [20]. The value of BTC, its use
central bank or central clearing house [10]. The last decade in online transactions, and the amount of attention it received
has seen unprecedented expansion in the domain of digital from the general public were all very unremarkable before
currency, with cryptocurrencies leading the way in terms of that.
innovation. Cryptocurrency’s primary characteristic is that Bitcoin’s price may be wildly unpredictable, but the
they are resistant to corruption because of their decentralized, digital currency itself is surprisingly robust, as seen by its
blockchain-based design [11]. ability to recover its value from deep declines, and changes
Decentralized characteristics of cryptocurrency consist in the business environment, such as that which occurred
of transparent public ledgers, reduced transaction costs, even during the COVID-19 outbreak [21]. Considering the
improved accessibility, strong security measures, financial volatility of the BTC rate has been the subject of several
inclusivity, resistance to censorship and inflation, promotion attempts at forecasting. There has been a lot of research that
of innovation, and price volatility [12]. The decentralization investigates the possibility of predicting BTC price utilizing
of cryptocurrencies is inherently resistant to the devalua- technical indicators and these analyses have revealed the
tions that might occur through government fiat [13]. The presence of important patterns, and gain in foresight [22].

118170 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

In other latest research, a variety of artificial intelligence- • We propose cryptocurrency short-to-long-term projec-
related techniques were used to predict prices after each tion using a joint classification-regression deep learning
day and whether or not they would go up or down [23]. framework to learn categorical price increases and
Given the scale of the BTC price fluctuations following these decreases, and perennial price fluctuations.
times, understanding fluctuation patterns is a challenging • The proposed learning framework employs SVM, ANN,
task. Furthermore, the public will understand the instant and LSTM to predict the categorical increase and
currency trends, the anticipation of daily closing prices, and decrease, and perennial price fluctuation using their
daily price increases or decreases. derived algorithms for classification and regression.
Bitcoin prices display non-stationary behaviour with • We propose enhanced deep learning architecture for the
trends, volatility clustering, and structural breakdowns, mak- learning framework and identify optimized hyperparam-
ing standard analytical methods less useful [24]. The data’s eters aiming to achieve accurately predicted short-to-
non-stationary behaviour in Bitcoin prices means that the long-term price projections.
statistical characteristics of Bitcoin’s price data fluctuate over • The proposed framework is evaluated individually for
time, which goes against the premise of stationarity [25]. analyzing the performance of classification and regres-
Statistical features like mean, variance, and autocorrelation sion. Metrics of Mean Absolute Error (MAE), Mean
stay consistent across time in a stationary time series, Absolute Percentage Error (MAPE), and Root Means
enabling accurate forecasting and analysis. Non-stationarity Square Error (RMSE) are used for regression analysis
may result from reasons such as changes in market attitude, to identify the model enhancement for perennial price
the regulatory environment, technical improvements, and fluctuations. The area under the curve, accuracy, and F1-
adjustments in supply and demand dynamics [26]. Analysts score metrics are used to identify the enhancement of the
need to consider the non-stationarity of Bitcoin prices when learning algorithms for categorical price prediction.
creating models. They use approaches like time-varying
parameter models, regime-switching models, or detrending II. LITERATURE REVIEW
techniques to correctly depict the changing character of Research on cryptocurrency price forecasting has surged due
Bitcoin’s price dynamics. The difference between long- to the growing public investment interest prior to buying
term and short-term viewpoints is based on the period digital coins. TABLE 1. shows recent research on different
over which events, trends, and choices are assessed [27]. cryptocurrency’s price forecasting. Guo et al. proposed
Both viewpoints are crucial for thorough examination and the Transform and Casual Multi-Head Attention Temporal
decision-making in many domains, such as finance and Convolutional Network (WT-CATCN) algorithm where the
economics. associations between multiple data aspects and catch crucial
The existing schemes are based on bitcoin price prediction points in input vectors are discussed [31]. Zhang et al.
in the short term using machine learning models achieving presented a Weighted and Attentive Memory Channels
considerable accuracy. The proposed mechanism stems (WAMC) model forecasting the daily closing price as well
from the lack of an enhanced cryptocurrency prediction as the volatility of cryptocurrencies. The proposed model has
mechanism to forecast the short-to-medium-to-long-term shown robust modelling capabilities through deep learning
price considering instant price volatility. A simultaneous methods i.e., GRU and CNN, and the robust correlations
classification and regression-deep learning framework are across different cryptocurrencies [32]. However, the proposed
proposed to assist in the Sell-or-HODL of Bitcoin. Cryp- algorithm’s usability is limited to the cryptocurrency market
tocurrency investors and influencers signals to hold on for trends. Ammer and Aldhyani proposed all cryptocurrency
dear life (HODL) referring to retain their digital coins for price predictions made by the LSTM algorithm closely
an extended period, foreseeing a significant price increase related to the natural trends limited to six-months data [33].
[28], [29], [30]. Tanwar et al. trained and validated the models on industry-
The proposed method focuses on the categorical price standard datasets to foresee the impact of leading coins on
movement predictions, whether they increase or decrease. the altcoins [34]. Multiple cryptocurrencies are considered
The dataset encloses both price increases and decreases for the LSTM and GRU model to investigate individual
reflecting the actual price movements of the cryptocurrency. impacts [35], [36]. As bitcoin contains time series data, the
Therefore, the models learn price movements in positive author in [37] claims that there are three distinct model-
and negative directions. To achieve this, the data is labelled based methods of time-series forecasting, pure algorithms
considering the price change in preprocessing to let the are the simplest method of prediction since they rely just
models learn the trends in classification. This work covers on past records of the target variable. ARIMA [38] and
both long-term price predictions (up to 90 days out), we also Generalized Autoregressive Conditional Heteroskedasticity
provide medium-term BTC price projections (30 days out) (GARCH) [39] respectively are two examples of periodic
and short-term price predictions (one week out) and price prediction methods. Furthermore, the suggested machine-
hike predictions using a variety of different models. Our learning approach tackles autocorrelation, periodicity, and
contributions are listed as follows; pattern effects, although the learning of pure time-series

VOLUME 12, 2024 118171

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

TABLE 1. A relevant analysis of significant research in the area of cryptocurrency price forecasting.

models requires human adjustment to address such impli- periodic models display changing statistical distributions
cations because of the vast variety of characteristics in the over time, leading to shifting behaviour in the dependencies
data. between input and output variables. The nonlinear and
When trying to foretell the future value of a target non-static properties of the data are explored by machine
parameter, explanatory models instead rely on a set of learning-based methods [41]. When they take into account
predictor variables. The author in [40] carried out empirical the fundamental elements that have an effect on the
research on modelling and forecasting the BTC price using outcome variables, they may also benefit from the descriptive
daily BTC data to evaluate the performance of the Bayesian characteristics.
neural network (BNN) in comparison to existing linear and Numerous studies have been conducted to model and
nonlinear standard algorithms. It is proposed in [40] article, predict BTC’s worth via ML algorithms, the suggested
that BNN is effective at forecasting the log-transformed value study in [42] focuses on normalizing a specific dataset in
of BTC, which helps to account for its unpredictability. order to better predict the price of Bitcoin. The suggested
Nevertheless, the aforementioned studies have been deceptive stacking ensemble approach makes use of feature selection
since they have reported performance measurements based and feature extraction to lower the forecaster’s characteristics
on log-transformed cost values, which are often lower prior to meta-learning, the postulated stacking ensemble
than performance metrics calculated using actual costs. incorporating the Least Absolute Shrinkage and Selection
We calculated performance indicators using log-normalized Operator (LASSO) significantly outperforms the standard
values and compared them to those derived from non-log- stacking ensemble and practically all hybrid models in
normalized data in our study. terms of prediction accuracy [43]. In this research [44],
Our research shows that the MAPE score may be the author uses BTC, Litecoin (LTC), and Ethereum (ETH)
drastically lowered by using log-normalized market prices. price predictions derived from three distinct recurrent neural
Given that the dynamics of bitcoin rates are irregular and non- networks (RNN) methods. The results show that when it
stationary, the prediction effectiveness may suffer if certain comes to forecasting the price of different cryptocurrencies,
data distribution presuppositions are used. Non-stationary the gated recurrent unit (GRU) is superior to both the

118172 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

FIGURE 2. An example of the proposed short-to-long term cryptocurrency projection framework using simultaneous classification-regression and deep
learning.

long short-term memory (LSTM) and bidirectional LSTM The Rectified Linear Unit (ReLu) is an activation function
(bi-LSTM) models. used to clip negative values. Mathematically it is represented
as follows,
III. LEARNING MODELS ARCHITECTURE AND
f(x) = max(0, x) (2)
OPERATIONS
Machine learning models such as ANN, LSTM networks, where x is the input to this function which maps input to a
and SVM could be employed for price projections of various positive output value where a greater value between and zero
cryptocurrencies. The detailed architectural structure and is chosen. It Introduces non-linearity to let the model learn
operational procedures of each model are presented as complex patterns. Specifically, the ReLu activation function
follows; solves vanishing gradient issues that occur due to low
step size. The ReLU function exhibits enhanced computing
A. ARTIFICIAL NEURAL NETWORK LEARNING MODEL efficiency over sigmoid and hyperbolic tangent activation
It is a type of neural network organized into a layered functions. The ReLU activation function promotes sparsity
network structure where each layer is composed of artificial in the network by producing an output of 0 for negative
neural nodes. The layered network structure is distributed input values. Sparse representations improve the efficiency
into the input layer, hidden layer, and output layer. In the of the deep learning network and reduce the likelihood of
network structure, each neuron in the next layer gets weighted overfitting.
inputs from its previous one, along with a bias as inputs.
Mathematically, it can be represented as, B. SVM MODEL
Nj−1
The SVM model is a supervised learning technique used for
X j−1 classification and regression analysis. It distinctly separates
Zi = ( Xk Wk,i − bk ) (1)
the data points of two different classes with a hyperplane
k=1
in a training split of the data. The points adjacent to the
j−1
where Xk represents the input data at the k-th layer out from hyperplane are referred to as the support vector. For the case
j-th node, The value of Wk,j is the weight of the connection of the limited dataset, the SVM performs well by distinctively
between node k and the preceding layer’s nodes, bk is the separating the classes in a dataset as it requires minimal
node’s bias and Nj−1 denotes the number of nodes. Activation training data. The separation of margin is resolved using
functions are used to control the output of a node through a soft margin and a hard margin. The Soft-Margin (SVM)
a specific function. The linear activation maps the input algorithm aims to identify a hyperplane that optimizes the
through the linear function, while the nonlinear activation margin between the two classes. The Sell or HODL SVM
function maps it to a non-linear output. Model employs the conventional SVM model, with enhanced

VOLUME 12, 2024 118173

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

parameters like the Radial Basis Function (RBF) kernel C. LONG SHORT-TERM MODEL LEARNING MODEL
function to handle the relationship of non-linearity in the data. This model is a form of recurrent neural network that is
A high regularization parameter C = 105 aims to fit the used to learn immediate as well as long-term dependencies.
training data with generalized outcomes on unseen data. Therefore, LSTM is used to predict temporal dependency in
Importantly, it allows a certain amount of misclassification decentralized time-dependent cryptocurrencies. The architec-
so that the algorithm works well on unseen data. The objective ture of LSTM is shown in FIGURE 3.
function
n
1 X
||w||2 + C ζi (3)
2
i=1

is Pused to represent soft margin optimization. The term

C ni=1 ζi represents the regularisation or penalty linked
with the weight vector w. The primary purpose is to
minimize the objective function to get a hyperplane that
has a substantial margin. C ni=1 ζi implement a penalty for
P
incorrect categorizations. The symbol ζi is the slack variable
for each training sample, whereas C is the regularization
parameter that determines the balance between minimizing
training error and maximizing margin. A greater value of C
imposes higher penalties for misclassifications, resulting in a FIGURE 3. LSTM principal working flowchart.
narrower margin. Conversely, a lower value of C permits a
wider margin but may lead to more misclassification.
The constraints guarantee that the Support Vector Machine Cryptocurrency price projection stems from long-to-sort
accurately categorizes the training samples, while still time-series data e.g., the Bitcoin price changes momentarily.
permitting a certain level of misclassification. The limitations The LSTM network is composed of input i, output o, and
are mathematically expressed as forget f gates which are used to manage the information
inflow and outflow. Cell state controls the state of the memory
yi (w · xi + b) ≥ 1 − ζi (4) cell at each time step and forwards the memory information
to the next LSTM cell. Mathematically, it can be represented
where the term yi (w · xi + b) denotes the output of the as
decision function, while ζi introduces a margin of error
Ct = ft ⊗ Ct−1 + it ⊗ C̃t (6)
in case the example falls inside the margin or on the
incorrect side of the hyperplane. The inequality helps in where Ct denotes the cellular state at a time t, ft is the
accurately categorizing every training case xi with a minimum forget gate which is used to determine information from
margin of 1. Kernel functions are used to handle complex the previous cell state Ct−1 to forget the memory, and ⊗ is
data distribution and robust to outliers accurately. A kernel used for element-wise multiplication. Similarly, Ct−1 retains
function, in its simplest form is a mathematical metric for the past information at time t − 1. The LSTM framework
determining the degree to which two data points xi and xj are solves the vanishing gradient problem and learns long-term
similar. A similarity score is derived from these two pieces of relationships in a sequential data distribution. The forget
information. gate in the LSTM design allows the network to selectively
The goal is to determine the degree of similarity between retain or discard information from the previous hidden state,
xi and xj in a more complex feature space where Gaussian depending on the current input. The selective memory process
kernel functions. plays a vital role in tackling the difficulties associated with
||xi −xj ||2
learning long-term dependencies in sequential traits. Forget
(− ) gate,
k(xi , xj ) = e 2σ 2 (5)
ft = σg (Wf xt + Uf ht−1 + bf ) (7)
where ||xi − xj||2 gives the Euclidean distance between xi and
xj , in the original input space. The parameter is the bandwidth It operates to forget the information received from the
or the scale of the kernel, a smoothing controlling factor, previous cell state Ct−1 . The sigmoid activation function,
which indicates how rapidly the similarity score drops as the denoted as g, compresses the output to a range of values
distance between xi and xj . The kernel becomes smoother from 0 to 1. Higher values i.e., closer to 1, signify that
as increases, whereas it becomes more distance-sensitive as all information should be maintained, while a lower value
it decreases. The similarity score increases as the distance i.e., closer to 0, implies that all information should be
between xi and xj decreases, therefore the closer the two discarded. The weight matrix Wf is the weight of the
numbers are, the higher the score. corresponding input xt indicates the contribution of the

118174 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

present input to the forget gate. Similarly, Uf is the weight of must first create a dataset. After this step, Machine Learning
the preceding hidden state ht1 which shows the contribution models are trained, and subsequent forecasts are made using
of the preceding hidden state ht1 to the forget gate, and bf is these models for varying time horizons. It is challenging to
the bias. The input gate controls the amount of information identify and analyze the core inter-dependencies in a time-
from the current input xt that is permitted to enter the cell series forecast of bitcoin prices.
state Ct. The input gate at time step t is mathematically Variance/mean squared error and root-mean-square error
represented as are two measures of statistical dispersion that evolve over
time. Throughout our research, the selected features of the
it = σg (Wi xt + Ui ht−1 + bi ) (8) BTC prices were gathered using web scraping. In the pre-
where Wi is the weight of the input xt , Ui is weight of the processing phase, information is compiled, scrubbed, and
preceding hidden state ht−1 , and bi is the bias of the input standardized/scaled. Three time periods are created from
gate. The output gate ot regulates the amount of information the processed data on Bitcoin transactions. It is indeed
from the current cell state Ct that is sent as the output or possible to concentrate on the right features by employing
hidden state ht . Mathematically, it can be expressed as, feature extraction. We generate numerous datasets for three
distinct time frames and perform feature extraction inde-
ot = σg (Wo xt + Uo ht−1 + bo ) (9) pendently for every dataset based on the predicted horizon.
The selected features are revealing features derived from
where Wo is the weight of input xt , Uo is weight for the
several blockchain-specific statistics and technical available
preceding hidden state ht−1 , bo is the bias of this layer. As part
analyses. Training and validation sets are constructed from
of its capacity to learn and retain important details from
datasets. The foundational classification and regression
sequential input, the LSTM architecture’s candidate cell state
models of the ML system are trained using the training
determines what new information is added to the cell state at
set. In contrast, the validation/testing set is used to check
a certain time step. Mathematical expression is
the accuracy of the ML models. We employ the following
C̃t = tanh(Wc xt + Uc ht−1 + bc ) (10) machine learning models for regression and classification:
Artificial ANN, LSTM, and SVM.
where Ct represents the status of the candidate cells at time
t, tanh is the tangent hyperbolic activation function that A. DATASET SOURCE
considers the input values within the range of −1 to 1. The There is no cost to view or use any of the information on
network can record both good and negative information. Bitcoin that can be found online. The information used in this
The weight matrix Wc corresponds to the input matrix xt , analysis was gathered from [45]. A vast number of technical
indicating the impact of input matters for the candidate indicator-based characteristics were gathered. This massive
cell’s state. The weight matrix Uc represents the impact of feature collection was then pruned using a feature selection
the previous hidden state ht−1 on the current hidden state. technique to provide a more manageable set of features.
It demonstrates how much the prior hidden state contributed The technical indicators include the variance, rate of change,
to the current candidate cell state. Similarly, the LSTM and triple-moving exponential. Other indicators include the
architecture’s hidden state ht can be calculated as, relative strength index and weighted moving average. The
ht = ot ⊗ tanh(ct ) (11) 7-day, 30-day, and 90-day duration are only some of the
time frames that may be used to determine these analytical
where ht represents the concealed state at time t. Both the indications. Bitcoin price time-series characteristics are used
output gate ot and the hyperbolic tangent of the cell state ct in the calculation of these technical analyses. For instance,
affect it, and it is the output of the LSTM cell. The tangent instead of only reporting raw transaction and hash-rate data,
hyperbolic of the cell state ct is denoted as tanh(ct ). they illustrate the relationship between the Bitcoin prices and
the root-mean-square error.
IV. PROPOSED SELL-OR-HODL LEARNING FRAMEWORK The dataset used in this study covers the period from
A variety of ANN, SVM, and LSTM-based machine learning April 2013 to December 2019, including over seven years
regression and classification models are employed to project of historical Bitcoin price data. The dataset consists of daily
future Bitcoin prices. In this article, we focus on the time- price observations. It is comprised of 737 features of various
series estimate of BTC costs utilizing AI. In FIGURE 2, properties ranging from statistical variations to social media
we can see a visual representation of the paper’s ML-based impacts on price projections. Considering these variations,
technique. As the name implies, a time series is a collection 26 different features are used to analyze the impact of dif-
of data collected at various intervals in time. Predictions of ferent models in accurately predicting the short to long-term
future values in a time series are the outcomes of such an price projections due to the statistical variation’s impacts on
analysis, which is what time-series forecasting is all about. the price projections. Furthermore, the proposed framework
Our range of forecasting covers not just the next seven uses 26 key features, and statistical variations, impacting
days, but also the following thirty and ninety. To use the the price projections of cryptocurrencies. In comparison,
Machine Learning-based time-series forecasting method, one the conventional method elaborated 18 listed features, and

VOLUME 12, 2024 118175

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

TABLE 2. The underlying characteristics/features used to construct technical analysis.

only 5 common features including top100cap, transac- while the remaining 80% were kept for training purposes. The
tion_fee30trxUSD, hashrate90var, mining_profitability, and dataset was linearly divided into two distinct groups for any
fee_to_reward_USD are used. Features details are given in and all classification techniques: training and testing. Eighty
TABLE 2. per cent of the information was used as training data, while
the remaining twenty per cent was set aside for testing. When
B. DATASET PRE-PROCESSING training ANN and LSTM algorithms, we used a mix of long-
When feasible, absent instances were substituted in the term scaling and min-max scaling to make adjustments to the
preprocessing phase using the linear interpolation approach. features.
If that’s not possible, the feature’s most common value is Min-max scaling is a method for re-scaling features
imputed. The dataset was mixed up and divided into two for from 0 to 1 while maintaining their original magnitude
every single regression model: There are two sets of data used and significance. Rubustscaler and Min-Max scaler are
in training and validating. Only 20% were utilized for testing, sequentially used where the RobustScaler technique is

118176 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

employed to effectively manage outliers by using robust them in getting the predicted outcome of price hikes. The
statistical measures such as the median and interquartile usage of binary representation streamlines the prediction
range (IQR). Similarly, the min-max scaler standardizes data work by emphasizing the fundamental direction of price
to the range of 0 and 1. After the scaling values have been movement instead of attempting to forecast exact price
optimized with the help of the training/learning setting, they levels. Moreover, binary classification enables the use of
are applied to the testing set. conventional assessment metrics and approaches to evaluate
The characteristics used for training SVM were scaled to the performance of the model and make well-informed
the standard scale since, according to our data, this resulted trading decisions.
in the best prediction accuracy. For 7-day projections, the ReLU activation functions have been shown to work well
cost column is moved ahead by 7 days. In this way, in LSTM networks in the past, especially when working with
the regression algorithm may be trained to discover the sequential data and long-range dependencies. This is because
connection between the characteristics and the predicted they can fix the vanishing gradient problem and speed up
pricing. For classification, a price rise or holdover from the convergence during training. When it comes to robustness and
last day is translated into a value of 1 and a price decrease is generalisation, Softmax is suitable for classification tasks.
given a value of 0. If not specified, a value of 0 is used. However, using ReLU activations in LSTM layers provides
A similar method is utilized for N-day price prediction. advantages in terms of model expressiveness, computational
In order to predict prices 30 days in the future, for instance, efficiency, and resistance to overfitting. This ultimately leads
the 30-day price is compared to the current price, and a value to superior generalisation performance.
of 0 or 1 is given depending on the result. Price outliers were RBF is used as a kernel for SVM to map non-linear higher
shown to have a negative impact on prediction accuracy, and dimensional data patterns to a linear classification problem to
their removal led to better overall results. It is constructed identify the target hyperplane. In this case, the regularization
on the premise that anomalies are uncommon and exhibit parameter C = 1 × 104 has a smooth decision boundary to
characteristics different from the norm. If the price of Bitcoin distinctively categorize the classes.
suddenly spikes by a huge amount, for no apparent reason,
this would be an anomaly. C. EVALUATION METRICS
Regression model effectiveness is measured by the following
TABLE 3. Neural networks hyperparameter tuning. indicators: Mean Absolute Error (MAE), Mean Absolute
Percentage Error (MAPE), and Root Mean Squared Error
(RMSE).
m
1X
MMAE = |xi − x̂i | (12)
m
i=1

The MAE Equation (12) is often used in regression analysis

to quantify the precision of prediction models. The symbol
The Sell-or-HODL framework integrates classification- xi represents the actual values inside a dataset, whereas x̂i
regression and deep learning algorithms, such as ANN, denotes the model predicted values. The accuracy of the
LSTM, and SVM, for short-to-long-term price projections. model’s predictions is assessed by calculating the average
Hyperparameter tuning optimizes model performance as absolute difference between the actual and projected values
shown in TABLE 3. for the proposed Sell-Or-HODL method across all observations. Low Mean absolute Errors (MAEs)
and conventional methods. Logcosh serves as the loss indicate enhanced performance in regression analysis. The
function due to its nature of smoothness, differentiable low error metric in forecasting reflects the close alignment of
and robustness to outliers. The Adam optimizer adapts the the predicted data with real-world data. The MAPE Equation
learning rate to recent changes efficiently. Furthermore,
m
the architectural details of the proposed models and the 100 X |xi − x̂i |
MMAPE = (13)
conventional model are depicted in TABLE 4. Softmax is m xi
i=1
used in the classification output layer due to the required
probability distribution for class labels. ReLu is used for all is a performance metric used in regression analysis for
models due to its continuous nature required in Regression a considerable size of data points where xi is the actual
leading to clipping negative values. Similarly, LeakyReLu is values, and the x̂i values represent the estimated values
used in the middle layers to avoid dying neurons and let each predicted by the model. The formula is used to calculate
node contribute to learning features. the average percentage difference between the observed and
In the proposed model binary classification model is projected values. The MAPE provides a measure of the
used where the classes are (i) price increase and (ii) prices model’s predictive performance when proportional errors
decrease labelled as 1 and 0 respectively. Considering the are considered. For regression applications like demand
interest of crypto investors, the proposed method would assist prediction and forecasting, a lower MAPE suggests a more

VOLUME 12, 2024 118177

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

TABLE 4. Neural networks hyperparameter tuning.

accurate model. V. FRAMEWORK PERFORMANCE EVALUATION

v The performance improvements achieved through the indi-
u m
u1 X vidual models up to 18%, 25% and 31.48% for LSTM, ANN
RRMSE =t |xi − x̂i |2 (14)
m and SVC classification models respectively as compared
i=1
to the conventional schemes. Similarly, the performance
The RMSE is a popular statistic Equation (14) for evaluating enhancement achieved by LSTM, ANN and SVR regres-
the quality of prediction models in the context of regression. sion models at most are 42.88%, 23.06%, and 12.45%
The xi , in this case, refers to true values in the dataset, whereas respectively. The paper explores the anticipated results of
the x̂i reflects the AI model’s prediction. The formula uses the regression and classification. Results from machine learning
square root of the mean squared deviation between observed and deep learning-based regression and classification models
and forecasted values to convert back to the original scale. are presented in this section.
Lower RMSE values indicate higher model performance, To analyze and predict the connection across a variable that
which is a measure of how well the predicted values match is dependent and a set of independent variables, statisticians
the actual values. Regression applications in AI, such as utilize a technique called regression. A regression model
forecasting housing or stock market prices, may benefit makes predictions about the dependent component based on
greatly from this measure. the values of the independent variables, and the predicted
Classification model effectiveness may be gauged using a result is a consistent measurement in numbers. In contrast,
variety of assessment indicators. Measures of accuracy are classification is a method that divides information into
shown in Equation (15). The accuracy of a classification distinct classes according to predetermined criteria. A classi-
system is an often reported measure since it is easily fication model’s predicted result is a categorically classified
understood and interpreted. A superior-performing model is value that places a data point in one of many predetermined
associated with greater accuracy. classes. The predicted results of regression and classification
models allow analysts to acquire knowledge about the
TP + TN connections between variables and anticipate future instances
AAccuracy = (15)
TP + TN + FP + FN of patterns and indicators.
The model is doing well if its F-1 score Equation (16), which
is a combination of the precision P and recall R metrics, A. PRICE PREDICTIONS USING REGRESSION MODELS
is growing. Using metrics like the F1-score may shed light Bitcoin price predictions may be made using regression
on situations where the distribution of reported classes is models, a common statistical method. Regression models
unequal, for as when greater numbers of instances of reduced can shed light on foreseeable patterns of Bitcoin price
price are reported than raised ones. fluctuations by analyzing the previous data of BTC prices
2×P×R and finding significant variables affecting its price. Here,
FF1−score = (16) we identify using regression models to foretell future BTC
P+R
prices. In TABLE 5, from April 2013 to December 2019,
The Area Under the Curve (AUC) Equation (17), of the model we can see the whole conclusion of the regression models’
in distinguishing between true positive and true negative, predictions for the nth-day BTC prices.
predictions is quantified by the AUC score. Scores closer to A significant decrease of 29.07% in MAE, 42.88%
one indicate greater classification ability, while scores closer in MAPE, and 24.27% in RMSE as compared to the
to zero indicate that no differentiation can be made between conventional scheme. This demonstrates how well the LSTM
classes and the receiver operating characteristic (ROC) curve. model works at producing more precise forecasts. ANNs
1

TP · (TP − 1)
 show promising outcomes by reducing the MAE, MAPE,
AAUC = TP · (FP) + (17) and RMSE by 1.98%, 15.99%, and 6.31% respectively.
2 2
Furthermore, when we apply the SVR mode the MAE is
The AUC is used to evaluate the performance of classifiers reduced by 3.46% while MAPE and RMSE are increased
to reflect the higher impact of true instances over negative by 20.96% and 9.18% respectively. Thus, the proposed
instances [47], where the TP and FP represent the count of framework shows a significant enhancement over the com-
true positives and number of false positives respectively. parative study. Specifically, the LSTM model in the proposed

118178 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

TABLE 5. Regression models to forecast the price of Bitcoin on the N-th Day.

TABLE 6. Regression models confidence intervals.

framework significantly outperforms the baseline study for displays greater performance and becomes more worthy,
forecasting using the last 90 days. on the other hand. Decreases of 11.95% points may be seen
Compared to the base paper, the performance of the built in the MAE, 10.94% points in the MAPE, and 12.45% points
LSTM model significantly improves after 30 days. The MAE in the RMSE. This suggests that the SVR model is more
dropped by 17.73%, the MAPE went down by 9.39%, and robust in terms of retaining its predicted accuracy during the
the RMSE reduced by 18.16%, all of which are considerable condensed time frame.
improvements. This suggests that the prediction accuracy of The outcomes show diverse patterns in the confidence
the LSTM model suffers inside this window of time. The intervals regarding LSTM, ANN, and SVR for varying
performance of an ANN improves as well. There is a sizeable forecast horizons, as shown in TABLE 6. Comparatively,
drop of 7.40% in the MAE, a decline of 23.60% in the MAPE, the proposed models show diverse regression outcomes on
and a modest decrease of 3.29% in the RMSE. According to the datasets, which is also reflected in their confidence
these findings, the ANN model flares up some of its efficacy intervals. The proposed ANN model shows smaller confi-
after 30 days. In the speed boost shown with the SVR model, dence intervals as compared to SVR and LSTM signifying
there is a tiny dip in two measures, with the MAE declining its outcomes consistency and uniformity over time due to
by 1.94%, the MAPE increasing by 2.90%, and the RMSE optimal architecture and hyperparameters. On the other side,
falling by 1.18%. Based on these results, it seems that the LSTM and SVR show a range of margin errors assisting in
SVR model can more reliably maintain its accuracy within averaging the variability in regression performance over the
the specified time frame. different prediction prospects.
The applied LSTM model shows a significant improve-
ment in performance after 7 days compared to the founda-
tional research paper. The MAE drops by 40.81% points, B. PRICE PREDICTION USING CLASSIFICATION MODELS
the MAPE falls by 42.03% points, and the RMSE drops By analyzing past data and determining the most important
by 38.70% points. According to these findings, the LSTM aspects influencing BTC’s price, classification models,
model’s predicted accuracy rises dramatically over such a a common artificial intelligence approach, may be used to
short period of time. The ANN model, quite similar to the make price predictions. Here, we will get into the topic of
ML model, also shows performance improvement. There is using classification models to forecast Bitcoin values and
a significant drop in all three metrics: 4.79% for the MAE, talk about the results. TABLE 7 shows the expected periodic
4.85% for the MAPE, and 17.14% for the RMSE. These shift (increase or reduction) in Bitcoin price. The constructed
results imply that the ANN model, like the LSTM model, LSTM classification model outperforms the referenced
is influenced by the 7-day timescale, but to a lesser level. SVR research paper by a large margin over 90 days. There is a

VOLUME 12, 2024 118179

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

TABLE 7. Price prediction using classification models on N-th Day.

significant rise in Accuracy of 15.62%, an increase in F1- Moreover, variations in assessment techniques, such as
score 4.22%, and an increase in AUC 13.63%. These findings inconsistencies in train-test divisions or cross-validation
demonstrate that the LSTM model may improve accuracy methods, exacerbate these constraints. AUC, on the other
and overall performance measures over such a long time hand, shows a significant rise of 10.71%. These results
period. On the other hand, the ANN model shows substantial suggest that the LSTM model’s accuracy and selected
improvement. There is a considerable rise in all three metrics: metrics improve throughout the 7-day timespan. Similarly,
Accuracy by 12.90%, F1-score by 25.00%, and AUC by the performance of the ANN model improves. There was a
18.64%. These findings suggest that both the performance considerable increase in Accuracy of 15.68% and no change
and accuracy metrics improve over a period of 90 days as a in the F1-score of 0.00%. The AUC increases significantly
consequence of the reduced effectiveness of the ANN model. by 11.32%. This indicates that there is a favourable effect on
The performance of the SVC model exhibits comparable the AUC and the ANN model’s accuracy and performance
improvements. The accuracy experiences a 31.48% rise, the during the 7-day period. Additionally, the effectiveness of the
F1-score shows a 19.69% increase and the AUC demonstrates SVC model is improved. There is a little increase in Accuracy
a 15.78% increase. This implies that during the 90-day period, by 1.61%, a moderate rise in F1-score 1.72%, and a sizable
there is a noticeable impact on the projected precision of the increase in AUC 5.00%. These findings suggest that the SVC
SVC model, leading to enhanced performance metrics. model’s overall performance and predicted accuracy improve
The proposed LSTM model demonstrates substantial over the course of a week.
performance improvements compared to the cited research Similarly, when considering the classification of price
article within a 30-day timeframe. The accuracy has risen decrease or increase the confidence interval and margin errors
by 15.38%, the F1-score has grown by 1.47%, and the show distinct and static outcomes for the proposed LSTM,
AUC has increased by 18.00%. The data illustrate the ANN, and SVC. The variations would affect the F1-score
efficacy of the LSTM model in enhancing accuracy and leading to false positives and false negatives. Comparatively,
other crucial performance metrics over the span of 30 days. the proposed models achieve distinct outcomes reflecting the
The ANN model, which is comparable, also demonstrates model’s architecture and hyperparameters tuning as shown
enhanced performance. While the Accuracy shows a notable in TABLE 8. The proposed SVC model shows a lower
increase of 7.69%, it shows a huge loss of 19.11% in the error margin as compared to LSTM and ANN showing
F1-score, suggesting a decline in performance. There is a performance reliability and consistency due to its decision
significant 18.00% rise in AUC. These results demonstrate boundary separating the classes. The proposed LSTM and
that the performance of the ANN model fluctuates across ANN error margin and confidence intervals indicate more
several criteria throughout the 30 days. Additionally, the variability in its classification findings across the different
effectiveness of the SVC model is improving. A significant classes due to feature complexities.
improvement of 21.15% is shown in Accuracy, a rise of Both the ups and downs of the price movement and the
11.76% in the F1-score, and a small gain of 6.00% in AUC. actual price need to be modeled when attempting to predict
According to these findings, the SVC model’s prediction the future value of Bitcoin. In the proposed framework,
accuracy and overall performance improved over a month. we demonstrate that the latter is possible with almost no
The deployed LSTM model shows a significant per- inaccuracies. The earlier, however, remains a mystery that
formance increase after 7 days compared to the baseline scientists have yet to solve. Researchers have classified
study paper. Although the Accuracy improves by 12.72%, the rise and fall of Bitcoin’s price using both internal and
the F1-score drops by 12.30%, indicating that performance external causes, as described in the aforementioned works.
is worsening in this area. The key factor to examine The price of Bitcoin fluctuates randomly, and it is impossible
this significant downfall is the variability in model archi- to predict its future using a static set of characteristics. Yet,
tecture, hyperparameters, and optimization methodologies. academics have had mixed results when trying to estimate

118180 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

TABLE 8. Classification models confidence intervals.

BTC values using a variety of extracted features. In this medium-, and long-term price variations [49]. The enhanced
article, we have implemented certain blockchain-related price predictions of decentralized cryptocurrencies by the
functionality. For instance, the hashrate and difficulty will proposed framework would help in evaluating and minimiz-
be high if there is significant interest from miners in Bitcoin ing risks related to price variations [50]. This would enable
mining. The same holds true for transactional characteristics regulators, lawmakers, and industry stakeholders to make
like the number of active addresses and the volume of informed policy decisions considering future price projec-
transactions, they will both increase as user adoption grows. tions and assessing social impacts [51]. Enhanced accuracy
All of these characteristics meet the criteria for time-series and F1-score in cryptocurrency applications improve transac-
characteristics. tion verification and fraud detection, which are essential for
Statistical measures are straightforward statistical tech- upholding confidence and security in financial transactions,
niques for transforming this fast raw information into ultimately promoting broader acceptance and stability in the
refined time-series characteristics for use in establishing cryptocurrency market [52].
estimations of the mean. In order to produce a huge chunk The limitation of the proposed models is the identification
of features that are amenable to machine learning, it is of optimal hyperparameters, which is reflected in the
helpful to combine the technical indications across many time downfall of different error metrics. LSTM learns long-term
frames. Finding the best features requires a strong feature dependencies of time series data and has the potential for
decision-making model. There is no uniformity between improvement for accurate prediction on the instant price
the chosen characteristics for the various periods in time impacts. Also, training LSTM on a large dataset requires
and prediction ranges. The effectiveness of the algorithms more computational resources. Similarly, training ANN
provides proof that the proposed method for choosing models on a dataset with many features or insufficient data
features may be utilized to make excellent judgments. lacks generalization capability, giving increased error metrics
Likewise, the principal component analysis technique was on the test dataset. It requires an optimal architecture for
tried out for low-dimensional embedding. Performance- defining hidden layers and the optimal number of nodes in a
based principal component analysis (PCA)-based regression layer leading to an impactful gap in error metrics. While SVM
models lagged behind those trained on carefully chosen regulates optimum decision boundaries to separate various
characteristics. Our method is effective because, in addition classes in a dataset. In addition, it requires feature selection,
to extracting the feature significance, the approaches utilized scaling, and normalization to achieve optimal performance.
in our feature selection process also deal with problems
like multi-collinearity and cross-correlations. Our addition VI. CONCLUSION AND FUTURE WORK
of predictive value with random forest makes it possible to We apply ML models to the problem of forecasting the
identify specific features with high importance, which wasn’t short- and medium-term value of Bitcoin. This research is
achievable with PCA, even though PCA has a significant the first of its kind to take into account all price indications
impact on constructing innovative parameters that are linearly up to December 31, 2019, and utilizes machine learning to
independent. provide surprisingly accurate short-term (within 7 days) and
In the trending cryptocurrency market, this proposed mid-term (30 and 90-day) forecasts of BTC price. We have
framework assists in making informed investment decisions, used various machine learning models, including LSTM,
mitigating, or minimizing the potential risks, and strategizing SVM, and ANN. Compared to the conventional schemes,
the investment plans [48]. To better aid investors in making the results have shown an increase in the accuracy with
educated decisions about entry points and exit strategies which daily closing prices and price changes can be predicted.
and optimizing their portfolios, the proposed framework Acceptable findings were achieved, showing promise for
would assist in accurately finding the projections of short-, applications in fields as diverse as financial technology,

VOLUME 12, 2024 118181

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

blockchain technology, and artificial intelligence research. generalization capability. Studying the impacts of social
Our research shows that although it is possible to accurately media and news articles would be an interesting future
estimate the present BTC price, it is far more difficult to direction to analyze the possible market manipulation using
predict its swings, in particular its increase and decrease. natural language processing (NLP) techniques. The ensemble
The findings of this investigation are unparalleled. Despite or hybrid approach can produce a more robust and accurate
the claims made, further study is needed to improve the performance than individual models.
classification models of Bitcoin. It is suggested that future
studies include hourly Bitcoin prices and technical indicators, REFERENCES
and investigate ensemble models that combine several kinds [1] M. Javaid, A. Haleem, R. P. Singh, R. Suman, and S. Khan, ‘‘A review of
of models to enhance forecast accuracy. blockchain technology applications for financial services,’’ BenchCouncil
Trans. Benchmarks, Standards Evaluations, vol. 2, no. 3, Jul. 2022,
To evaluate the potential financial risks involved with using Art. no. 100073.
blockchain technology, this article asserts that future research [2] F. Allen, X. Gu, and J. Jagtiani, ‘‘Fintech, cryptocurrencies, and CBDC:
might investigate the use of artificial intelligence in modeling Financial structural transformation in China,’’ J. Int. Money Finance,
bitcoin pricing. Possible uses for the model include spotting vol. 124, Jun. 2022, Art. no. 102625.
[3] H. Song, N. Zhu, R. Xue, J. He, K. Zhang, and J. Wang, ‘‘Proof-of-
instances of fraud and other forms of anomalous behavior. contribution consensus mechanism for blockchain and its application in
The presence of external influences, such as important intellectual property protection,’’ Inf. Process. Manage., vol. 58, no. 3,
world events or fraudulent actions like fake pumps and May 2021, Art. no. 102507.
[4] Y. Chen, E. K. Kumara, and V. Sivakumar, ‘‘Investigation of finance
dumps, may be indicated by a large discrepancy between industry on risk awareness model and digital economic growth,’’ Ann.
the real and modeled behavior of the price. Price modeling Oper. Res., vol. 326, pp. 1–22, Oct. 2021.
and forecasting aren’t the only ways to uncover external [5] X. W. Huawei. Digital Spillover. Accessed: Nov. 3, 2022.
[Online]. Available: https://www.huawei.com/minisite/gci/en/digital-
influences, but they have the potential to be used in the spillover/files/gci_digital_spillover.pdf
fight against fraud. To figure out how stable cryptocurrencies [6] CoinMarketcap. Digital Currency. Accessed: Nov. 3, 2022. [Online].
are, a mix of machine learning-based pricing models and Available: https://coinmarketcap.com/charts/
techniques for finding outliers could be used. This would [7] T. Renduchintala, H. Alfauri, Z. Yang, R. D. Pietro, and R. Jain, ‘‘A survey
of blockchain applications in the FinTech sector,’’ J. Open Innov., Technol.,
include external data sources that show global events and Market, Complex., vol. 8, no. 4, p. 185, Dec. 2022.
financial risk. [8] R. Awadallah, A. Samsudin, J. S. Teh, and M. Almazrooie, ‘‘An
Future research directions include enhancing the resilience integrated architecture for maintaining security in cloud comput-
ing based on blockchain,’’ IEEE Access, vol. 9, pp. 69513–69526,
and reliability of blockchain technology and cryptocur- 2021.
rency price trends. The model efficiency can be improved [9] E. J. Scheid, B. B. Rodrigues, C. Killer, M. F. Franco, S. Rafati, and
by identifying optimal hyperparameters, transfer learning B. Stiller, ‘‘Blockchains and distributed ledgers uncovered: Clarifications,
achievements, and open issues,’’ in Advancing Research in Information
techniques to utilize the performance of existing complex and Communication Technology: IFIP’s Exciting First 60+ Years, Views
architecture, and investigating time series RNN variants From the Technical Committees and Working Groups, M. Goedicke,
such as GRUs. Also, regularization approaches, feature E. Neuhold, and K. Rannenberg, Eds., Cham, Switzerland: Springer, 2021,
pp. 289–317, doi: 10.1007/978-3-030-81701-5_12.
engineering, and cross-validation may enhance the models’
[10] P. R. Cunha, P. Melo, and H. Sebastião, ‘‘From Bitcoin to central bank
generalization capability. Nevertheless, research directions digital currencies: Making sense of the digital money revolution,’’ Future
include enhancing the resilience and reliability of blockchain Internet, vol. 13, no. 7, p. 165, Jun. 2021.
technology and cryptocurrency price trends. The ensemble [11] M. A. Alrowaily, M. Alghamdi, I. Alkhazi, A. B. Hassanat,
M. M. S. Arbab, and C. Z. Liu, ‘‘Modeling and analysis of proof-based
or hybrid approach can produce more robust and accurate strategies for distributed consensus in blockchain-based peer-to-peer
performance than individual models. Given the significance networks,’’ Sustainability, vol. 15, no. 2, p. 1478, Jan. 2023.
of the detailed features used in this study, the current data has [12] R. Ahuja, J. Khandelwal, and J. Anjali, ‘‘Challenges, opportunities and risk
analysis of adoption of decentralized finance applications,’’ ICTACT J. Soft
been constrained by the available data, which extends only Comput., vol. 14, no. 1, p. 3114, 2023.
until 2019. However, as a future direction, we plan to study [13] J. Vaz and K. Brown, ‘‘Sustainable development and cryptocurrencies
the impact of the data post-2019 with the most recent bitcoin as private money,’’ J. Ind. Bus. Econ., vol. 47, no. 1, pp. 163–184,
Mar. 2020.
price using the proposed framework along with additional
[14] M. N. M. Bhutta, A. A. Khwaja, A. Nadeem, H. F. Ahmad, M. K. Khan,
features of social networking apps and news influencing price M. A. Hanif, H. Song, M. Alshamari, and Y. Cao, ‘‘A survey on blockchain
manipulation. External factors such as political events and technology: Evolution, architecture and security,’’ IEEE Access, vol. 9,
influential figures can have a significant influence on price pp. 61048–61073, 2021.
[15] E. N. Tas, D. Tse, F. Gai, S. Kannan, M. A. Maddah-Ali, and F. Yu,
fluctuation. For this purpose, the price of bitcoins is to be ‘‘Bitcoin-enhanced proof-of-stake security: Possibilities and impossibili-
studied with the news and social media interaction with a ties,’’ in Proc. IEEE Symp. Secur. Privacy (SP), May 2023, pp. 126–145.
historical price timeline considering the NLP techniques. [16] E. Zaghloul, T. Li, M. W. Mutka, and J. Ren, ‘‘Bitcoin and blockchain:
Security and privacy,’’ IEEE Internet Things J., vol. 7, no. 10,
Furthermore, the model efficiency can be improved pp. 10288–10313, Oct. 2020.
by identifying optimal hyperparameters, transfer learning [17] CNBC. Digital Coin. Accessed: Nov. 11, 2022. [Online]. Available:
techniques to utilize the performance of existing complex https://www.cnbc.com/2022/05/20/crypto-as-currency-managing-your-
financial-life-using-digital-coins
architecture, and investigating time series RNN variants
[18] T. Klein, P. T. Hien, and T. Walther, ‘‘Bitcoin is not the new gold: A
such as GRUs. Also, regularization approaches, feature comparison of volatility, correlation, and portfolio performance,’’ SSRN
engineering, and cross-validation may enhance the models’ Electron. J., vol. 59, pp. 105–116, Oct. 2018.

118182 VOLUME 12, 2024

M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

[19] M. Irfan, M. A. Rehman, S. Nawazish, and Y. Hao, ‘‘Performance [41] J.-M. Kim, C. Cho, and C. Jun, ‘‘Forecasting the price of the cryptocur-
analysis of gold- and fiat-backed cryptocurrencies: Risk-based choice for rency using linear and nonlinear error correction model,’’ J. Risk Financial
a portfolio,’’ J. Risk Financial Manage., vol. 16, no. 2, p. 99, Feb. 2023. Manage., vol. 15, no. 2, p. 74, Feb. 2022.
[20] A. Bloomenthal. (2022). What Determines Bitcoin’s Price? Accessed: [42] H. K. Andi, ‘‘An accurate Bitcoin price prediction using logistic regression
Nov. 8, 2022. [Online]. Available: https://www.investopedia. with LSTM machine learning model,’’ J. Soft Comput. Paradigm, vol. 3,
com/tech/what-determines-value-1-bitcoin/#:~:text=Bitcoin’s%20price% no. 3, pp. 205–217, Sep. 2021.
20is%20primarily%20affected,to%20be%20mined%20in%202140 [43] S. Aras, ‘‘Stacking hybrid GARCH models for forecasting Bitcoin
[21] M. B. Arouxet, A. F. Bariviera, V. E. Pastor, and V. Vampa, ‘‘COVID-19 volatility,’’ Expert Syst. Appl., vol. 174, Jul. 2021, Art. no. 114747.
impact on cryptocurrencies: Evidence from a wavelet-based Hurst expo- [44] M. J. Hamayel and A. Y. Owda, ‘‘A novel cryptocurrency price prediction
nent,’’ Phys. A, Stat. Mech. Appl., vol. 596, Jun. 2022, Art. no. 127170. model using GRU, LSTM and bi-LSTM machine learning algorithms,’’ AI,
[22] W. Chen, H. Xu, L. Jia, and Y. Gao, ‘‘Machine learning model for Bitcoin vol. 2, no. 4, pp. 477–496, Oct. 2021.
exchange rate prediction using economic and technology determinants,’’ [45] BitInfoCharts. Accessed: May 7, 2022. [Online]. Available:
Int. J. Forecasting, vol. 37, no. 1, pp. 28–43, Jan. 2021. https://bitinfocharts.com/
[23] S. Ziweritin, ‘‘Height-end multi-layer perceptron and machine learning [46] M. Mudassir, S. Bennbaia, D. Unal, and M. Hammoudeh, ‘‘Time-series
methods of forecasting Bitcoin price time series,’’ TechRxiv, pp. 1–17, forecasting of Bitcoin prices using high-dimensional features: A machine
2022, doi: 10.36227/techrxiv.20891200.v1. learning approach,’’ Neural Comput. Appl., pp. 1–15, Jul. 2020.
[47] D. V. Kute, B. Pradhan, N. Shukla, and A. Alamri, ‘‘Deep learning
[24] O. Özdemir, ‘‘Cue the volatility spillover in the cryptocurrency markets
and explainable artificial intelligence techniques applied for detect-
during the COVID-19 pandemic: Evidence from DCC-GARCH and
ing money laundering—A critical review,’’ IEEE Access, vol. 9,
wavelet analysis,’’ Financial Innov., vol. 8, no. 1, p. 12, Dec. 2022.
pp. 82300–82317, 2021.
[25] A. Kumar, N. Sharma, R. Chauhan, and M. Sharma, ‘‘Cryptocurrency price
[48] T. Choithani, A. Chowdhury, S. Patel, P. Patel, D. Patel, and M. Shah,
forecasting in a volatile landscape: SARIMAX modeling and short-term
‘‘A comprehensive study of artificial intelligence and cybersecurity on
strategies,’’ in Proc. 3rd Int. Conf. Smart Gener. Comput., Commun. Netw.
Bitcoin, crypto currency and banking system,’’ Ann. Data Sci., vol. 11,
(SMART GENCON), Dec. 2023, pp. 1–6.
no. 1, pp. 103–135, Feb. 2024.
[26] X. Ren, J. Li, F. He, and B. Lucey, ‘‘Impact of climate policy uncertainty on [49] A. Ashta and H. Herrmann, ‘‘Artificial intelligence and fintech: An
traditional energy and green markets: Evidence from time-varying Granger overview of opportunities and risks for banking, investments, and
tests,’’ Renew. Sustain. Energy Rev., vol. 173, Mar. 2023, Art. no. 113058. microfinance,’’ Strategic Change, vol. 30, no. 3, pp. 211–222, May 2021.
[27] A. Antenucci, P. Crespo del Granado, B. Gjorgiev, and G. Sansavini, [50] S. Syed, A. Iqbal, W. Mehmood, Z. Syed, M. Khan, and G. Pau,
‘‘Can models for long-term decarbonization policies guarantee security of ‘‘Split-second cryptocurrency forecast using prognostic deep learning
power supply? A perspective from gas and power sector coupling,’’ Energy algorithms: Data curation by deephaven,’’ IEEE Access, vol. 11,
Strategy Rev., vol. 26, Nov. 2019, Art. no. 100410. pp. 128644–128654, 2023.
[28] A. Urquhart, ‘‘Under the hood of the Ethereum blockchain,’’ Finance Res. [51] M. A. Rudd, L. Bratcher, S. Collins, D. Branscum, M. Carson,
Lett., vol. 47, Jun. 2022, Art. no. 102628. S. Connell, E. David, M. Gronowska, S. Hess, A. Mitchell, M. Prusak, K.
[29] T. Bakhtiar, X. Luo, and I. Adelopo, ‘‘Network effects and store-of-value Schneps, M. Serezhin, S. A. Wolfe, and D. Porter, ‘‘Bitcoin and its energy,
features in the cryptocurrency market,’’ Technol. Soc., vol. 74, Aug. 2023, environmental, and social impacts: An assessment of key research needs
Art. no. 102320. in the mining sector,’’ Challenges, vol. 14, no. 4, p. 47, Nov. 2023.
[30] Y. Yogarajah, ‘‘‘Hodling’ on: Memetic storytelling and digital folklore [52] X. Wang, W. Lin, W. Zhang, Y. Huang, Z. Li, Q. Liu, X. Yang, Y. Yao,
within a cryptocurrency world,’’ Economy Soc., vol. 51, no. 3, pp. 467–488, and C. Lv, ‘‘Integrating Merkle trees with transformer networks for secure
Jul. 2022. financial computation,’’ Appl. Sci., vol. 14, no. 4, p. 1386, Feb. 2024.
[31] H. Guo, D. Zhang, S. Liu, L. Wang, and Y. Ding, ‘‘Bitcoin price
forecasting: A perspective of underlying blockchain transactions,’’ Decis. MUHAMMAD IQBAL received the B.S. degree in
Support Syst., vol. 151, Dec. 2021, Art. no. 113650. computer science from Abdul Wali Khan Univer-
[32] Z. Zhang, H.-N. Dai, J. Zhou, S. K. Mondal, M. M. García, and H. Wang, sity Mardan, Pakistan, in 2020. He is currently pur-
‘‘Forecasting cryptocurrency price using convolutional neural networks suing the Master of Science degree in the artificial
with weighted and attentive memory channels,’’ Expert Syst. Appl., intelligence with Pak-Austria Fachhochschule:
vol. 183, Nov. 2021, Art. no. 115378. Institute of Applied Sciences and Technology,
[33] M. A. Ammer and T. H. H. Aldhyani, ‘‘Deep learning algorithm to predict Pakistan. His research interests include blockchain
cryptocurrency fluctuation prices: Increasing investment awareness,’’ technology and bioinformatics.
Electronics, vol. 11, no. 15, p. 2349, Jul. 2022.
[34] S. Tanwar, N. P. Patel, S. N. Patel, J. R. Patel, G. Sharma, and
I. E. Davidson, ‘‘Deep learning-based cryptocurrency price prediction
scheme with inter-dependent relations,’’ IEEE Access, vol. 9, ARSHAD IQBAL (Member, IEEE) received the
pp. 138633–138646, 2021. B.S. degree in electrical and computer engineering
[35] I. E. Livieris, N. Kiriakidou, S. Stavroyiannis, and P. Pintelas, ‘‘An from the COMSATS Institute of Information Tech-
advanced CNN-LSTM model for cryptocurrency forecasting,’’ Electron- nology (CIIT), Abbottabad, Pakistan, in 2013,
ics, vol. 10, no. 3, p. 287, Jan. 2021. and the M.S. and Ph.D. degrees in electrical and
[36] G. R. Patra and M. N. Mohanty, ‘‘Price prediction of cryptocurrency using computer engineering from Sungkyunkwan Uni-
a multi-layer gated recurrent unit network with multi features,’’ Comput. versity, Suwon, South Korea, in 2020. Since 2021,
Econ., vol. 62, no. 4, pp. 1525–1544, Dec. 2023. he has been an Assistant Professor with the Sino-
[37] R. Hyndman and G. Athanasopoulos, Forecasting: Principles and Pak Center for Artificial Intelligence (SPCAI),
Practice, 3rd ed., Australia: OTexts, 2021. Pak-Austria Fachhochschule: Institute of Applied
[38] A. Ibrahim, R. Kashef, and L. Corrigan, ‘‘Predicting market movement Sciences and Technology (PAF-IAST), Harı̄pur, Pakistan. His research
direction for Bitcoin: A comparison of time series modeling methods,’’
interests include medium access control, resource allocation, the Internet
Comput. Electr. Eng., vol. 89, Jan. 2021, Art. no. 106905.
of Things, applied artificial intelligence, WLAN, sensor networks, energy
[39] L. Flach, R. de Oliveira, J. D. Sallaberry, and L. K. de Mattos, ‘‘Blockchain
harvesting networks, backscatter communication networks, power saving,
and new digital technologies: Explaining the Bitcoin volatility with
a generalized autoregressive conditional heteroskedasticity model,’’ in distributed communication networks, and next-generation communication
Digital Technologies and Transformation in Business, Industry and networks. He is an Associate Member and the Team Lead of European
Organizations, R. Pereira, I. Bianchi, and Á. Rocha, Eds., Cham, Organization for Nuclear Research (CERN) PAF-IAST Team. He was a
Switzerland: Springer, 2022, pp. 167–180, doi: 10.1007/978-3-031-07626- recipient of the fully-funded ICT research and development scholarship for
8_8. undergraduates by the Ministry of Information Technology (IT), Pakistan,
[40] H. Jang and J. Lee, ‘‘An empirical study on modeling and prediction and the HEC scholarship under the Human Resource Development (HRD)
of Bitcoin prices with Bayesian neural networks based on blockchain initiative M.S. leading to the Ph.D. program of faculty development for
information,’’ IEEE Access, vol. 6, pp. 5427–5437, 2018. UESTPS, Phase-1 Batch-IV.

VOLUME 12, 2024 118183

 M. Iqbal et al.: Sell or HODL Cryptos: Cryptocurrency Short-to-Long Term Projection

ABDULLAH ALSHAMMARI (Member, IEEE) LOUAI A. MAGHRABI (Member, IEEE) received

received the bachelor’s degree from Tennessee B.Sc. degree in computer science from Lebanese
State University, Nashville, TN, USA, the master’s American University, Beirut, Lebanon, the M.Sc.
degree from Howard University, Washington, degree in information technology from the Uni-
DC, USA, and the Ph.D. degree from Howard versity of the West of England, Bristol, U.K.,
University, in 2020, under the supervision of Prof. and the Ph.D. degree in cybersecurity from
Danda B. Rawat. He was a Researcher with the Kingston University, London, U.K. He is currently
Cybersecurity and Wireless Networking Innova- an Assistant Professor with the Department of
tions (CWiNs) Laboratory, Howard University, Software Engineering, College of Engineering,
from 2016 to 2020. He was a Graduate Researcher University of Business and Technology, Jeddah,
with the Data Science and Cybersecurity Center (DSC2), Howard University, Saudi Arabia. His research interests include cybersecurity, risk assessment,
during the Ph.D. degree. Formerly, he led Cybersecurity Innovation, cryptography, artificial intelligence, machine learning, the IoT, blockchain,
Research, and Development, NEOM, and NEOM AUTHORITY. In addition, drones, metaverse, quantum computing, and game theory. He received the
he was an Assistant Professor of cybersecurity and artificial intelligence with Best Research Paper Award in 2021.
the Royal Commission for Jubail and Yanbu. He is currently an Assistant
Professor with the Department of Electrical Engineering and Computer
Science, University of Hafr Al Batin, Hafr Al Batin, Saudi Arabia. He is
engaged in research and teaching in the areas of cybersecurity, machine
learning, big data analytics, and wireless networking for emerging networked
systems, including cyberphysical systems, the Internet of Things, smart
cities, edge computing, cognitive city, mobile computing, network security,
and artificial intelligence. He was a recipient of the SACM Award in 2012,
2013, 2014, 2015, and 2016. He has been serving as a Reviewer for over
five international journals, including Sustainability (MDPI) and The Journal
of Supercomputing (Spring Nature). He has been in organizing committees
for several conferences. He has delivered keynotes and invited speeches at
international conferences and workshops.
NIGHAT USMAN received the master’s degree in
information security from COMSATS University
IHTISHAM ALI received the Ph.D. degree Islamabad (CUI), Islamabad Campus, Abbot-
(Hons.) in distributed systems from Gyeongsang tabad. She is currently a Lecturer with the Depart-
National University, Jinju, South Korea, in 2020. ment of Computer Science, CUI. She has gained
He has completed his postdoctoral research at pretty good knowledge regarding IBM QRadar,
the Sino-Pakistan Centre for Artificial Intelligence OSSIM, and elasticsearch. She has published
(SPCAI), Pak-Austria Fachhochschule: Institute several research articles in the area of Internet-
of Applied Sciences and Technology (PAF-IAST). of-Things, cybersecurity, and data security and
He currently holds the position of an Assis- privacy. Furthermore, she was rewarded with the
tant Professor of computer science and infor- Award by Trillium Pakistan for her research contribution to the rule mining
mation technology with the Machine Learning technique for proling log behavior.
Department, PAF-IAST.

118184 VOLUME 12, 2024