This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI

An On-Chain Analysis-Based Approach

to Predict Ethereum Prices
Nishant Jagannath1 , (Member, IEEE), Tudor Barbulescu1 , Karam Sallam1 ,(Member, IEEE) ,
Ibrahim Elgendi1 , (Member, IEEE), Braden McGrath2 , Abbas Jamalipour3 , (Fellow, IEEE),
Mohamed Abdel-Basset 4 , (Senior Member, IEEE), Kumudu Munasinghe1 , (Senior Member,
IEEE)
1
School of IT and Systems, University of Canberra, ACT 2601, Australia
2
Faculty of Human Factors and Behavioral Neurobiology, Embry-Riddle Aeronautical University, United States of America
3
School of Electrical and Information Engineering, University of Sydney, NSW 2006, Australia
4
Faculty of Computers and Informatics, Zagazig University, Zagazig 44519, Egypt
Corresponding author:Nishant Jagannath (e-mail: [email protected]).

ABSTRACT The Ethereum blockchain generates a significant amount of data due to its intrinsic
transparency and decentralized nature. It is also referred to as on-chain data and is openly accessible to
the world. Moreover, the on-chain data is timestamped, integrated, and validated into an open ledger. This
important blockchain feature enables us to assess the network’s health and usage. It serves as a massive
data warehouse for complex prediction algorithms that can effectively detect systemic trends and forecast
future behavior. We adopt a quantitative approach using a subset of these metrics to determine the network’s
true monetary value by developing a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN)
with the metrics most closely associated with the price as inputs. Since several hyperparameters regulate the
learning process in an RNN, they are highly sensitive to their values. It is thus critical, to select optimal
hyperparameters so that the training is quick and effective. Determining the optimal parameters of an
RNN model is a tedious and complex process. Hence, previous studies have developed several self-adaptive
approaches to determine the optimal values for various parameters effectively. However, none of the prior
studies explore self-adaptive algorithms in deep learning models in conjunction with on-chain data to predict
cryptocurrency prices. In this paper, we propose three self-adaptive techniques, each of which converges on a
set of optimal parameters to predict the price of Ethereum accurately. We compare our results to a traditional
LSTM model. Our approach exhibits 86.94% accuracy while maintaining a minimum error rate.

INDEX TERMS Blockchain, Deep neural network, Differential evolution, Ethereum, Evolutionary
algorithms, jSO, LSTM

I. INTRODUCTION of conducting transactions. Beyond cryptocurrencies, several

applications use blockchain technology, such as electronic
Blockchain is proving to be one of the most significant health care and identity management systems. [2].
technological advancements in the public and private sectors.
Bitcoin [3], the most well-known use of blockchain tech-
According to Australia’s National Blockchain Roadmap pro-
nology to date, was introduced in 2008 through a paper
posed by the Department of Industry, Science, Energy and
written under the pseudonym, Satoshi Nakamoto. Bitcoin’s
Resources, by 2025, blockchain technology is expected to
success stems from the fact that it was the first digital cur-
generate over US 175 billion in annual business value and
rency to overcome the double-spending and the "Byzantine
further grow to over US 3 trillion in 2030 [1]. To this end,
General" problems, allowing for a secure way of transferring
the concept of blockchain technology has emerged to address
value online without relying on a trusted third party. The
the limitations of a centralized architecture, primarily the
advent of other cryptocurrencies has given rise to a new asset
single point of failure alongside addressing security issues.
class. Bitcoin provided a safe and inexpensive way to move
The ability of blockchain to securely store and transfer digital
currency across a decentralized, peer-to-peer network.
assets in a decentralized, distributed, and tamper-proof way
is a quality that many researchers attribute to it as a means Though Bitcoin was the first currency to gain traction,

VOLUME 4, 2016 1

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

the concept of digital currency has been around for decades. overview of relevant work, emphasising current research on
Since the 1990’s Cypher-punks have been experimenting Ethereum price prediction. Section III covers the research’s
with the notion of using an alternative digital currency en- proposed methodology and is divided into four subsections,
tirely independent of any financial institutions and govern- including subsection A, B, C and D. Subsection A describes
ment organisations issuing it. Some of the initial projects data collection methods, data characteristics, and analysis
that solved cryptographic puzzles to create a concept for measures. We study on-chain events and their effects on
cryptocurrencies were b-money [4], bit gold [5] and reusable price, correlation analysis with the price to understand price
proof of works [6]. Nonetheless, these projects failed primar- patterns better. Subsection B describes the architecture of an
ily as a result of their reliance on a centralized intermediary LSTM model, while subsection C introduces differential evo-
[7]. lution algorithms, covers the three self-adaptive techniques
Bitcoin is the world’s first decentralized cryptocurrency, and justifies their use. Subsection D describes the proposed
established in 2009 with the creation of the genesis block. system model. Section IV describes the experimental setup,
Alternative cryptocurrencies (also known as altcoins) were whereas Section V presents the results and their interpre-
developed since 2011, with several of them continuing in use tation. Finally, in Section VI, we summarise our findings,
today. The majority of cryptocurrencies are forks of Bitcoin, conclusions and outline areas for further research.
meaning they were developed using the same codebase as
Bitcoin, hence the name "alternative coin". Ethereum is the II. RELATED WORK
second most popular blockchain platform after Bitcoin, with Technical Analysis is a topic that has been extensively ex-
a market cap of 267 billion dollars [8]. It permits decentral- plored in academic literature. Although simple to implement
ized money transmission and the design and participation in and interpret, traditional statistical methods require many
rule-based smart contracts (code) that run on a decentralized statistical assumptions that could be unrealistic, leaving ma-
infrastructure comparable to Bitcoin. In 2013, virtually every chine learning as the best technology capable of predicting
week saw the birth of new altcoins, resulting in the growth of price based on historical data, among others [16]. Numerous
an extensive, highly speculative trading market [9]. studies have examined the effectiveness of various machine
Investors often use one of three approaches to evaluate a learning techniques and technical indicators in predicting bit-
crypto asset: a technical, fundamental, or quantitative analy- coin values [11], [17]. The problem with cryptocurrencies is
sis. Technical analysis is the most frequently utilized method that their value is dependent on unpredictable market move-
as a directional hypothesis in bitcoin trading. However, de- ments and social sentiment. Moreover, cryptocurrencies, pri-
spite claims that profits from the technical analysis have marily Bitcoin, have a low correlation with major financial
declined over time, research reveals that profits in the for- assets as seen in [18], [19], indicating that conventional
eign currency markets have grown considerably [10]. Even economic theories and models are insufficient for forecasting
today, several investors still base their investing decisions the volatility connected with the price of cryptocurrencies.
on technical indicators since they focus on profitability via There has been much research interest in Bitcoin in the
statistics and data volume. It is difficult to anticipate the price last decade, whereas only a few studies have investigated
of a crypto asset using technical analysis since the indicators Ethereum. Ethereum is one of the most popular blockchain
change rapidly in the cryptocurrency market. Researchers platforms after Bitcoin. It enables users to use blockchain-
have used machine learning algorithms combined with tech- based applications with a smart contract to be easily deployed
nical indications to predict the price of crypto assets, partic- in a completely decentralized way. In 2016, a Decentral-
ularly Bitcoin, to solve these difficulties [11]. A recent study ized Autonomous Organization (DAO) attack on Ethereum
discovered a link between the price of Bitcoin, Google, and resulted in a loss of millions of dollars, more here [20]. Smart
Wikipedia [12]. Several studies used blockchain data such contract analysis can help detect the similarity of codes and
as transaction volume, hash rate, and difficulty to forecast calls, vulnerability, and fraud detection in contracts. Securify
bitcoin prices [13], [14], [15]. [21] was released publicly and has analyzed over eighteen
The research above has focused on modelling machine thousand contracts for vulnerabilities to address this issue.
learning algorithms to predict Ethereum pricing without fully Recent studies such as [22], [23] have proposed classification
contextualizing the characteristics of the on-chain data and methods to detect frauds on smart contracts by analysing the
their potential impact on Ethereum prices. Furthermore, un- data available on Ethereum, primarily Ponzi schemes.
predictable swings in cryptocurrency prices due to various Since Ethereum is a permission-less blockchain, it uses
factors such as supply and demand, inflation, and political an open data structure. Therefore, data analysis of the data
factors are significant barriers to accurate price prediction. on the blockchain can provide tools that can further en-
This presents many opportunities to realise the value and able understanding of the underlying network and its us-
information of on-chain data related to Ethereum that has age, providing valuable information on the user, miner be-
been overlooked while utilising more advanced optimization havior and other activities on the blockchain [24]. A few
algorithms to better adapt to the volatility associated with studies explored the blockchain information to understand
Ethereum prices. the characteristics and inherent usage of the network for
This article is organised as follows. Section II provides an Ethereum. For example, authors in [25] investigate the trans-
2 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

TABLE 1. Correlation coefficient values between Bitcoin and other major for Ethereum with 50% test data. A recent study proposed
financial assets between 2016 to 2020 [18]
a stochastic neural network model utilising a perturbation
factor (gamma) to introduce randomness into the model
Asset Correlation with Bitcoin
to simulate market volatility [15]. It achieved an average
Gold 0.0459 improvement of 4.15% on mean absolute percentage error
Silver 0.0071 (MAPE) for Ethereum compared to Multi-Layer Perceptron
WTI (Crude Oil) 0.0158 (MLP) and LSTM models.
S&P 500 0.0491 Evolutionary Algorithms (EA) have been widely used to
MSCI World 0.0457
solve optimization problems in several fields and real-world
applications. The term "optimization" refers to the process
MSCI EM50 0.0042
of identifying the most appropriate solution to a problem
given some constraints. Portfolio optimization has tradition-
ally relied on metaheuristics. The utilisation of EA for Multi-
actions data on the blockchain to understand the behavioral Objective Portfolio Management is reviewed in [34]. Evo-
traits of addresses on the Ethereum Network. A study into lutionary optimisers for continuous parameter spaces, such
the statistical characteristics of Ethereum pricing utilising as Differential Evolution (DE) [35], [36], [37], [38], [39],
blockchain information, mainly using Ethereum network fea- developed by Storn and Price, are becoming increasingly
tures, determined that hash rate, difficulty, and transaction powerful and flexible. As a result, DE and its variations have
cost are strongly associated with price [14]. Nevertheless, the established themselves as one of the most competitive and
public Ethereum blockchain data is vast and has expanded flexible groups of evolutionary computing algorithms. They
exponentially over time, reaching 248.93GB in 2021 [26]. have been effectively used to address a wide range of real-
Parsing this large volume of data is time-consuming. To world issues in science and industry [40], [41]. However,
address this issue, several researchers have proposed open- none of the above studies investigates the application of DE
source frameworks to extract crucial information from popu- algorithms for hyperparameter optimization in deep learning
lar permissionless blockchain platforms such as Bitcoin and models in conjunction with on-chain data to predict cryp-
Ethereum [27], [28], [29]. A recent study proposed a frame- tocurrency prices.
work that focuses on collecting and analysing datasets from The above studies use a few metrics, such as hash rate,
Ethereum and released a series of datasets for developers and difficulty, transaction count, among other metrics, to identify
researchers to analyze user behavior and other blockchain price trends in Ethereum. This paper adopts a comprehensive
system operations [30]. approach by using on-chain datasets to help us better un-
Smuts [31] adopted a LSTM model to predict short- derstand core activities related to blockchain. In comparison,
term trends in cryptocurrency prices by investigating the researchers have used LSTM, Random Forest and Stochastic
influence of Telegram and Google Trends search analysis models, among other techniques, to predict Ethereum price.
on the price of Ethereum. The above study achieved an In our previous work, we proposed a self-adaptive technique
accuracy rate of 56% for Ethereum and 63% for Bitcoin and for predicting bitcoin prices and achieved a higher accuracy
concluded that Telegram proved to be a better indicator of rate compared to a traditional LSTM model [42]. This study
the price of bitcoin and Ethereum. The machine learning further explores advanced algorithms (self-adaptive tech-
model trained using historical data from non-blockchain niques) in conjunction with deep learning models that predict
sources, such as social media feeds and publicly accessible price trends better and adapt to the volatility associated with
websites, may cause the model to make incorrect assump- the Ethereum price.
tions due to the datasets inherent bias [32], [33], particularly
when the model is converged with a blockchain system. III. PROPOSED DESIGN AND METHODOLOGY
Since the data on the blockchain is timestamped, embed- This section explains the meaning of on-chain metrics and
ded, and vetted into an open ledger, this creates a massive their purpose in the context of blockchain in general and
data warehouse for advanced predictive models to identify Ethereum in particular. Additionally, this section discusses
trends and fraud in the system accurately. This is possible the methodologies utilized to analyze the information to
as the data on the blockchain is immutable. This important understand better the blockchain network’s activity, which
characteristic of blockchain has gained tremendous interest enables us to develop a prediction model.
among researchers in using advanced predictive models with
blockchain (on-chain) data. A. ON-CHAIN ANALYSIS
Conversely, only a few studies have adopted various 1) WHAT ARE ON-CHAIN METRICS?
machine learning approaches using blockchain information On-chain metrics are data points derived from information
to predict the price trends for Ethereum. Authors in [14] generated by the blockchain network, such as the size of
adopted a multivariate regression model using features of the blockchain, the number of blocks attached to it, or
the Ethereum network to explain the Ethereum dynamics the difficulty of mining blocks. Their purpose is to inform
further, achieving a mean absolute error (MAE) of 0.0563 interested parties about the state of a blockchain network;
VOLUME 4, 2016 3

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

this data inherits the properties of these networks, such as count, hash rate, transactions gas limit, difficulty, transfer
their inherent transparency, tamper-resistant and decentral- rate, total gas used, wallets address > 1, >10, >100 coins, ex-
ized nature. Since on-chain data is often recorded in a time- change deposits, exchange withdrawals, external and internal
series manner, each metric offers insight into the historical contract calls, total addresses.
activity of a blockchain. As a result, everyone involved with
law enforcement, such as agents tracking illegal activity [43] 4) DATA CHARACTERISTICS
or those in the financial industry determining the viability Cryptocurrencies use more accurate valuation methods as
of a proposed investment in enhanced decision-making, will opposed to stocks. While some metrics are derived from data
benefit from this knowledge. generated within the blockchain like block size, others are
derived from external and on-chain information - such as the
2) ON-CHAIN METRICS AS FINANCIAL TOOLS market capitalisation.
The financial sector has already been noted as a potential Market Cap = Circulating supply × Last market price
use case for on-chain data. Comparing blockchain networks
For Bitcoin, the price changes every day as it is traded
to companies, there are various similarities to be observed
on various exchanges, and the supply also increases as more
- most networks have a coin, equivalent to the stock of a
coins are mined in circulation. At the time of writing this,
company, and both have a price and are traded on exchanges.
the price of BTC is $49,027.71, and the circulating supply
There are fundamental differences: blockchain networks are
is 18,792,243.00 BTC [48]. The supply of Bitcoin increases
not structured like companies, there is no need for executives,
and it is capped at 21 million. While not all coins are in
and the economic parameters are written in code. The most
circulation (some have yet to be mined), there will never
recent story about Bitcoin is that it is an improved version
be more than 21 million bitcoins in circulation, with the
of Gold [44] for its store of value and non-inflationary
mining reward halving roughly every four years. Considering
properties, coupled with its lower friction of moving assets
that one satoshi is 1 × 10−8 bitcoin, this choice seems
around.
arbitrary. For Ethereum, on the other hand, the supply is not
However, seeing as their popularity as financial instru-
yet fixed. As it stands today, there is a circulating supply
ments are increasing and ever more credible investors are
of 117,189,429 ETH [47], one of which has a price of
adding them to their portfolios [44], we can ascertain that
$3,254.38.
cryptocurrencies have gained their category as crypto-assets.
Market capitalisation is not unique to blockchains, and it
They are traded on dedicated exchanges and are becoming
is a suboptimal method of valuing a blockchain network. For
significant parts of investors’ portfolios around the world.
one, it does not account for the fact that only a few coins
Hence, the need arises for tools and protocols to analyze these
have been traded at the latest spot prices, and it also ignores
crypto-assets as financial instruments and decide whether to
the significant amount of coins irrevocably lost due to lost
make them part of one’s portfolio or not. On-chain metrics
private keys [49]. Enter realised capitalisation, developed by
provide the best basis for such tools, as decentralized and
Carter et al. [50], uses the extra information provided by
tamper-proof data that enables anyone access to in-depth
blockchains (coinage) to give a more accurate valuation. The
information about a network’s usage, with no entity having
formula for the realised cap is:
the ability to censor it. We view this as akin to everybody
having access to insider information on any publicly traded Realised Cap = Σ Coin amount × Price last transacted
company. We consider this could act as a way of levelling the In Ethereum, a block contains a block number, difficulty,
financial playing field and getting us closer to the efficient gas limit, transaction list and the most recent state. A block
market hypothesis [45] which states that all the publicly and also holds the answer to a unique, difficult-to-solve mathe-
privately accessible information available to market partici- matical problem; miners gather all the information included
pants is already reflected in the price. in that block before the block is released into circulation. The
first miner to solve the puzzle created by this information
3) DATA COLLECTION wins a block reward. Unlike Bitcoin, whose block rewards
In this paper, the data has been gathered from the public are halved every 210,000 blocks (about four years), there is
Ethereum blockchain and application programming inter- no established number of ethers in circulation. The block
faces (APIs) of online resources [46], [47] from 2016 through rewards were adjusted from 3 ETH to 2 ETH after the
2021. We analyze metrics with the Ethereum price using Constantinople hard fork in 2018 [51].
on-chain data from these resources. This research aims to Ethereum employs an account-based model to keep track
provide a broader overview by incorporating on-chain met- of the state of the transactions in the network, unlike Bitcoin,
rics relating to miners, users, and exchange activity and their which uses the Unspent Transactions Output (UTXO) model
possible impact on Ethereum pricing. The following metrics to calculate the sum of the unspent transactions. The account-
are obtained from the on-chain data; block size, block height, based model is similar to bank accounts; the balance is the
transaction count, daily active addresses, miners revenue, number of coins left in the account, using which the user can
miner fees, miner to exchanges, total new addresses, sup- send or receive transactions. Using the UTXO model, it is
ply in smart contracts, gas price, transactions rate, transfers easier to extract valuable on-chain information such as the
4 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

normalized transaction count graph looks very similar to the

normalized active addresses.
On-chain metrics such as active addresses, total addresses
and transaction volume indicate the usage and adoption of the
network. In comparison, the hash rate indicates the number
of computational resources dedicated to the network, which
indicates the number of miners committed to this network.
Miners are generally financially committed, having invested
in equipment, infrastructure and ongoing power consump-
tion, none of which have an instant return on investment.
More miners committed to a network means that more people
FIGURE 1. State Transition Function in Ethereum. truly believe in the network (by having invested in it long
term), but it also means the security of the network increases,
making it more valuable. Hence, metrics such as hash rate
age of the coins in circulation, and therefore there are fewer and mining rate can help us understand the network’s overall
unique on-chain metrics available for Ethereum. However, security.
effective ways have been developed for mapping between the The above relationships are straightforward to conceptu-
account-based and UTXO models [52] which makes it easier alize, but they are by far the only ones. A recent article [54]
to gain similar insights into the Ethereum network. outlined several less obvious relationships. We can determine
Exchange operations must be open and transparent in light the age of each currency traded on the Bitcoin network by us-
of the numerous fraudulent activities that have been investi- ing a UTXO-based ledger. For Ethereum, we do not have this
gated in the past [53]. Analysis of exchange on-chain activity, information directly as it has an account-based accounting
such as deposits and withdrawals, provides a verifiable and model but can estimate this information accurately enough
immutable data source that might help researchers better for our purposes [52]. The ability to extract a coin’s age
understand the influence of exchange operations on pricing. allows us to group the coins of the entire supply based on the
last time they were spent - thus creating a group of coins that
5) ON-CHAIN METRICS AND THEIR EFFECT ON THE were spent less than one day ago, a group last spent between
PRICE 1 day and one week, one week and one month, 1-3 months,
We examine a few metrics and critical aspects connected 3-6 months.
to pricing to illustrate the parallels we observed between Given the number of fraudulent activities reported in the
metrics and price that led us to investigate this relationship past, exchange operations must be transparent. Analysis of
further. exchange on-chain activity, such as deposits and withdrawals,
Perhaps the most prominent relationship between an on- provides a verifiable and immutable data source that might
chain metric and the pricing can be illustrated by referring to help researchers better understand the influence of exchange
the daily transaction volume. On Ethereum, a transaction is operations on pricing. Supply distribution indicates the dis-
defined as any interaction with the blockchain that changes tribution of economic resources across the network. A more
the blockchain’s state. Ethereum uses transaction fees to centralized distribution means that fewer entities have to
minimize network overload and a possible attack vector, gain from an increase in the coin’s value and that a small
Bitcoin, and other blockchains. As a result, the transaction number of people have much influence on the price, which
count accurately represents the network’s underlying activity might decrease the overall attractiveness of the network as an
and demand for Ether, which is also used to pay transaction investment.
fees. From 2017 through 2021, we can detect similar patterns Following our exploration of the effects of on-chain met-
in these graphs by plotting the normalized price against the rics on the price of Ethereum, on-chain activity adds value
normalized transaction count. to portfolio managers looking to invest in cryptocurrencies.
The number of active addresses is another simple metric Nonetheless, we are interested in determining if quantitative
to conceptualize and extract since all blockchain transactions traders could incorporate some of these metrics into their
are timestamped. We can easily find all the transactions trading algorithms. We devise a simple method for choosing a
that occurred on a specific day, out of which we count subset of metrics that may have a more significant influence
the number of individual addresses which initiated and sent in forecasting price: calculate the price-metric correlations
transactions. The resulting number approximates the number and choose the most strongly associated metrics.
of people that transacted on the network on a specific day For selecting the metrics, we chose two correlation meth-
(some participants have several wallets), with more active ods - Pearson, which is the commonly used method for
addresses indicating increased demand for the network’s financial markets because of the linear nature of the data, and
native coin. According to supply and demand theory, periods Spearman’s, a rank-based correlation method, to introduce
when more addresses are active than normal should result in new metrics potentially. We also attempted to use Kendall’s
a higher price, as seen in Fig. 3. We can also observe that the Tau, but our trials did not introduce any new metrics besides
VOLUME 4, 2016 5

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

1.0 Price
transaction_count
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 2. Correlation between normalized prices and transaction count.

1.0 Price
active_addresses
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 3. Correlation between normalized prices and active addresses.

1.0 Price
block_size_mean
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 4. Correlation between normalized prices and block size.

6 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

1.0 Price
gas_used_total
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 5. Correlation between normalized prices and total gas used.

1.0 Price
miner_revenue_total
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 6. Correlation between normalized prices and miner revenue.

1.0 Price
hash_rate
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 7. Correlation between normalized prices and hash rate.

VOLUME 4, 2016 7

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

1.0 Price
supply_in_smart_contracts
0.8
Normalised Values

0.6

0.4

0.2

0.0
0 200 400 600 800 1000 1200 1400
Days

FIGURE 8. Correlation between normalized prices and supply in smart contracts.

what the Spearman method had already provided.

TABLE 2. Mean correlation value for every on-chain metric between 2016 and
6) ANALYSIS METRICS AND APPROACH 2021

In this study, each significant Ethereum on-chain metric’s

Pearson Correla- Spearman Corre-
linear impact on the price is investigated using Pearson’s and Data Category
tion lation
Spearman correlation coefficients. Pearson coefficients vary
Block Size 0.7465 0.5494
from -1 to 1, with 1 indicating perfect positive correlation,
-1 indicating perfect negative correlation, and 0 indicating Block Height 0.4483 0.1640
no correlation at all. Equation (1) represents the Pearson Hash rate 0.8791 0.5681
correlation coefficient. Difficulty 0.8865 0.5663
Pn Transactions Volume 0.3892 0.4213
(xi − x)(yi − y)
r = pPn i=1 Pn (1) Transactions rate 0.7651 0.7723
2 2
i=1 (xi − x̄) i=1 (yi − ȳ) Gas price 0.4420 0.6019
The Spearman rank correlation coefficient [55] is a non- Total Gas used 0.6038 0.3875
parametric measurement correlation used to evaluate the Supply in Smart Contracts 0.8725 0.6153
monotonic relationship between two variables.
Internal Contract Calls 0.6218 0.3549

6 di 2
P
External Contract Calls 0.5128 0.3549
ρ=1− (2)
n(n2 − 1) Miners Revenue 0.5924 0.4440
Miners Inflow 0.1335 0.3191
Pearson correlation coefficients are used to quantify the
linear connection between variables. In contrast, Spearman Miners Outflow 0.0612 0.2136
correlation coefficients are only applicable to monotonic Miners to exchanges 0.0376 0.1904
connections, in which variables tend to move in the same or Exchange Withdrawls 0.3943 0.5902
opposite direction but not necessarily at the same rate. In a Exchange Deposits 0.1778 0.3161
linear relationship, the rate is constant. Exchange Inflow 0.1589 0.3427
Correlated data cannot be displayed linearly to represent
Exchange Outflow 0.0156 0.1139
the scaling of negative to positive values accurately. We use
Exchange NetFlow 0.4292 0.2105
the normalisation value to re-scale the data between 0 and 1.
The normalising value is determined by using (3). Daily Active Address 0.7889 0.7277
Total Address 0.5297 0.1640
xi − min(x)
z= (3) Total New Address 0.6509 0.6856
max(x) − min(x)
Addresses (> 1 coin) 0.5872 0.5365
This research aims to examine a wide range of on-chain Addresses (> 10 coins) 0.6215 0.5912
metrics to determine the metrics that can be used in advanced Addresses (> 100 coins) 0.4545 0.4163
prediction algorithms and those that should be excluded.
Although a highly correlated metric seems significant, it
8 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

cannot be used as a sophisticated measure to determine cell, respectively. σ, g, and h are activation functions for
whether it is favourable. Hence, it is crucial to understand the gates, inputs, and outputs, respectively. Weight coefficients
characteristics and the effects of on-chain metrics on price are denoted by W .
and perform correlation analysis to determine if a metric
requires further investigation. C. EVOLUTIONARY ALGORITHMS
Evolutionary Algorithms (EAs) have been around since the
B. LONG SHORT-TERM MEMORY MODEL early 1950s. It is comprised of Genetic Algorithms (GA),
Long short-term memory (LSTM) is a specific type of recur- Evolutionary programming (EP), Evolution Strategies (EP).
rent neural network (RNN) capable of solving both long-term For many years, EAs have been widely used to solve op-
and short-term dependence issues in a resilient and efficient timization problems in diverse fields and the real world.
manner. The memory cell is the backbone of the LSTM Applications began to arise to the optimization issues that
network, which takes the role of the traditional neuron’s stem from various research and engineering fields in the
hidden layers. It contains three gates (input, output, and late 1990s. The Differential Evolution (DE) was developed
forget); it can add or remove information from the cell’s state. by Price and Storn [35], [36], [37], [38], [39] as a highly
competitive optimization algorithm. Compared to other EA’s,
DE is an appealing optimization algorithm as it is easy to
code and implement, and it uses minimal control settings
(crossover rate (Cr), scaling parameter (F ), and population
size (NP)). The performance of DE has been extensively
examined with regard to these factors [56], [57], [58].

DIFFERENTIAL EVOLUTION ALGORITHMS (DE)

Differential Evolution is a real-parameter optimization al-
gorithm. Its operation process consists of four basic steps:
initialisation, mutation, crossover and selection operations. It
also contains three main parameters: population size (NP),
scaling factor (F ) and crossover rate (Cr). It is complex and
time-consuming to determine the optimal combination of NP,
F and Cr by trial and error. However, various self-adaptive
FIGURE 9. LSTM Memory Cell. approaches have evolved over the years to aid in finding
suitable parameter values. Three of these self-adaptive tech-
Updating the cell state and computing the LSTM model’s niques will be employed in this study [59], [56], [60], [61].
output is done by: An in-depth explanation of the DE algorithm follows in the
next subsections.

it = σ(Wix xt + Wim mt−1 + Wic Ct−1 + bi ) (4) a: Initialization phase

In this phase, a random initial population of size NP, p0 =
(x1,0 , x2,0 , ...xNP,0 ) dis generated using (10).
ft = σ(Wf x xt + Wf m mt−1 + Wf c Ct−1 + bf ) (5)
xi,j = xmin max min
i,j + rand × (xi,j − xi,j )
(10)
i = 1, 2, ..., NP and j = 1, 2, ..., D
ct = f ⊙ ct−1 + it ⊙ g(Wcx xt + Wcm mt−1 + bc ) (6)
where D refers to the problem dimension (number of deci-
sion variables), and xmax min
i,j , xi,j the upper and lower bounds of

Ot = σ(Wox xt + Wom mt−1 + Woc Ct−1 + bo ) (7) j th decision variable, NP is the population size or the number
of solutions evolved by the DE algorithm.

mt = Ot ⊙ h(ct ) (8) b: Mutation Operator

Mutation operators are employed after the initialisation phase
yt = Wym mt + by (9) to create a mutant solution. According to the literature,
mutations may be categorised into a variety of methods
where xt represents the input data, yt denotes the output with varying capacities and features [59]. While some are
data at time t. it denotes the input, ot the output, and ft useful for investigation, others are capable of being exploited.
forget gates at time t. mt refers to the activation vector for The following techniques are the most often used mutation
each memory block, and ct the activation vector for each approaches.
VOLUME 4, 2016 9

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

• DE/best/1
vgi = xgbest + F (xgr1 − wxgr2 ) (11)
• DE/best/2
vgi = xgbest + F (xgr1 − xgr2 ) + F (wxgr3 − xgr4 ) (12)
• DE/rand/1
vgi = xgr1 + F (xgr2 − xgr3 ) (13)
• DE/rand/2
FIGURE 10. Taxonomy of parameter setting.
vgi = xgr1 + F (xgr2 − xgr3 ) + F (xgr4 − xgr5 ) (14)
• DE/current-to-pbest/1 However, the main disadvantage of parameter tuning is its
vgi = xgi + F (xgbest − xgi ) + F (xgr1 − xgr2 ) (15) lack of flexibility during this process. While in the parameter
control based method, the values of different parameters can
• DE/current-to-best/1 be changed during the evolutionary process, and as per Fig.
10, it can be classified as deterministic, adaptive and self-
vgi = xgi + F (xgbest − xgi ) + F (xgr1 − xgr2 ) (16)
adaptive. As per the literature [60]–[62], DE algorithms with
where the indexes r1 ̸= r2 ̸= r3 ̸= r4 ̸= r5 ̸= i and adaptive and self-adaptive techniques have often been more
r1 , r2 , r3 , r4 , r5 are randomly generated integers in [1, NP]. successful than classical ones. Thus, all selected algorithms
The current individual or solution is represented using i. The applied in this study used a self-adaptive technique to set the
current generation g, the best solution xbest in the current parameters [56], [61].
iteration and xpbest is an individual that is randomly chosen
from the top p% solutions. d: Selection Operator
The parent (xgi ) and trial (ugi ) populations are subjected to
c: Crossover Operator a greedy selection to determine which solution from each
After mutating, every mutant vector vi is converted to a group (xgi ) and (ugi ) enters the next generation. As long as the
trial/offspring vector ui using a crossover operator. Exponen- trial solution f(ugi ) has a greater fitness function value than
tial and binomial crossover operators [56], [60] are the two the parent solution f(xgi ), the trial solution is included in the
primary crossover operators. new population; else, the parent solution will be included.
The first crossover is the binomial crossover operator, This procedure is performed mathematically in the following
which is performed using (17). way:
(
(
g g+1 ugi , if f (ugi ) ≤ f (xgi ),
vi,j , if rand(0, 1) ≤ Cr or j = jrand , xi = (18)
ugi,j = g (17) xgi , Otherwise.
xi,j , Otherwise.
1) L-SHADE OPTIMIZATION ALGORITHM
where jrand ∈ [1, 2, ..., D] and rand is a random number
belongs to [0, 1]. The values of jrand and rand are chosen Researchers in [59] recently proposed a DE-based evolu-
randomly to ensure that at least one variable from the trial tionary algorithm known as L-SHADE. It is an improved
vector is picked. The crossover rate (Cr) is used to determine variant of the SHADE algorithm that was developed by [63].
how many variables are inherited from the donor vector. In L-SHADE, the DE/current-to-pbest/1 mutation operator
As per the literature review, the performance of the DE described in the following equation has been used to generate
algorithm is mainly dependent on its parameter values [56], new solutions.
[60]. The balance between exploration and exploitation is
achieved by using F . Smaller values of F indicate an in- vgi = xgi + F (xgbest − xgi ) + F (xgr1 − xgr2 ) (19)
crease in the convergence rate, while larger values indicate Instead of using fixed population size, a linear population
maintaining population diversity. The rate of change in an size mechanism is used, in which the population size is
individual of the population is determined using Cr. Setting linearly reduced from a high value to a small value [59].
the control parameter values in DE can be split into two For each generation, the population size is set to one Ninit ,
processes, parameter tuning and parameter control, as is and the population after the run is set to one Nmin . After
shown in Fig. 10. each generation, the next generation’s population number is
In parameter tuning, good parameter values can be deter- calculated as follows:
mined before an algorithm is executed, and then these values
are not changed during the optimization process; that is, all NPmin − NPinit
the parameter values are static during the search process. NPg+1 = round[( ) × fes + NPinit ] (20)
maxfes
10 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

Algorithm 1 L-SHADE algorithm Algorithm 2 jSO algorithm

1: g = 1, Ng = N init , Archive A = ∅; 1: Define g ← 0, A ← {};
2: Randomly, generate an initial population P0 = 2: Generate an initial random population (P0 ) of size NP,
(xg1 , ..., xgN ); that represents the initial structure of the LSTM;
3: initialize MCr , MF , and set their values to 0.5; 3: Evaluate f (P0 ).
4: while The termination condition is not satisfied do 4: Update fes; fes ← fes + NP;
5: SCr = {}, SF = {}; 5: while fes ≤ maxfes do
6: for i = 1 to N do 6: g ← g + 1;
7: ri = Select from [1, H] randomly; 7: for i = 1 : NP do
8: If MCr,ri =⊥, Cri,g = 0. Otherwise 8: Use (21) to produce mutant vector (vgi );
9: Cri,g = randni (MCr,ri , 0.1); 9: Use (17) (crossover operation) to produce new so-
10: Fi,g = randci (MF,ri , 0.1); lution (ugi ), which represents a newly generated
11: Generate trial vector ugi according to current-to- LSTM structure;
pbest/1/bin; 10: Evaluate f (ugi )
12: end for 11: end for
13: for i = 1 : N do 12: for i = 1 : NP do
14: if f (ugi ) ≤ f (xgi ) then 13: if f (ugi ) ≤ f (xgi ) then
15: xg+1
i ← ugi ; 14: xg+1
i ← ugi ;
16: else 15: else
17: xg+1
i ← xgi ; 16: xg+1
i ← xgi ;
18: end if 17: end if
19: if f (ugi ) < f (xgi ) then 18: if f (ugi ) < f (xgi ) then
20: xgi → A; 19: xgi → A;
21: Cri,g → SCr , Fi,g → SF ; 20: Update F and Cr as in [59], and Fw as in (22);
22: end if 21: end if
23: end for 22: If needed, update the archive A;
24: If necessary, delete randomly selected individuals 23: Use (20) to linearly update the population size;
from the archive such that the archive size is |A|. 24: end for
25: Update memories MCr and MF (Algorithm 1); 25: end while
26: Calculate Ng+1 according to (10);
27: if Ng < Ng+1 then
28: Sort individuals in P based on their fitness values where Fw is a weighted version of the scaling factor F
and delete lowest Ng − Ng+1 members; calculated by:
29: Resize archive size |A| according to new |P |; 
30: end if 0.7 × F, nfes < 0.2 × maxfes ,

31: g + +; Fw = 0.8 × F, nfes < 0.4 × maxfes , (22)
32: end while

1.2 × F, otherwise.


where nfes is the current number of function evaluations and

min maxfes is the maximum number of function evaluations. The
where NP is the minimum number of individuals the algo-
F values are computed using the same approach as Tanabe
rithm can use, fes the current number of function evaluations
et al. [59]. These strategies employ a lower factor Fw to
(fes), maxfes the largest number of fes.
multiply the difference of vectors which appears early in the
optimization process, while a larger factor F is used later in
2) jSO OPTIMIZATION ALGORITHM the process. The main steps of the jSO algorithm are pre-
As mentioned in the above section, the L-SHADE algo- sented in Algorithm 2. The jSO method, like the L-SHADE
rithm is an enhanced variant of a DE-based algorithm that algorithm, employs a linear population size mechanism to
employs a technique to reduce the population linearly. Later, update the total number of individuals in a population every
an enhancement to the L-SHADE algorithm was proposed, generation [59].
named iL-SHADE, to solve optimization problems [64].
Brest et al. [56] introduced a more powerful version of iL- 3) MPEDE OPTIMIZATION ALGORITHM
SHADE, named jSO, that employs a novel weighted muta- The multi-population based ensemble of mutation strategies
tion operator that is a modified version of the DE/current-to- (MPEDE) is a multi-population based-DE algorithm [65],
pbest/1 mutation operator as in which the whole population is dynamically divided into
several indicator subpopulations (equal and relatively lesser
vgi = xgi + Fw (xgpbest − xgi ) + F (xgr1 − xgr2 ) (21) sized) and one relatively large reward subpopulation for each
VOLUME 4, 2016 11

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

generation to materialize the effective ensemble of various • DE/pbad-to-pbest/1

mutation methods into one DE variant. A mutation strategy
is allocated to each indicator subpopulation. The reward vgi = xgi + F (xgpbest − xgpbad ) (25)
subpopulation is allocated to the mutation approach that
where xgpbad is a solution selected randomly from the bad top
performs best in the most recent generations based on the
100 p% of solutions in the current population with p ∈ (0, 1]
objective function values.
As a result, the top-performing mutation method can acquire
Algorithm 3 MPEDE algorithm additional computing resources.
1: Set µCrj = 0.5, µFj = 0.5, △fj = 0 and △fesj = 0 for
each j = 1, ...., 4; D. PROPOSED SYSTEM MODEL
2: Initialize, NP, ng for each j = 1,....4; The proposed system model is represented in Fig. 11, where
3: Initialize the pop randomly distributed in the solution we discuss the two modules employed in this study, namely
space; on-chain analysis and deep learning framework. In the on-
4: Initial λj and set NPj = λj × NP; chain analysis module, we collect on-chain metrics derived
5: Randomly partition pop into pop1 , pop2 , pop3 and pop4 from data provided by the blockchain network, such as the
with respect to their sizes; size of the blockchain, the number of blocks linked to it,
6: Randomly select a subpopulation popj (j = 1, 2, 3) and or the difficulty of mining blocks from the Ethereum pub-
combine popj with pop4 . Let popj = popj ∪ pop4 and lic blockchain and API of online resources [46], [47]. We
NPj = NPj + NP4 ; normalise the on-chain metrics to appropriately display data
7: Set g = 0; in graph form and analyze the data features of the obtained
8: while g ≤ maxg do on-chain metrics. The effects of on-chain metrics on price
9: g = g + 1; are analyzed by plotting a graph of each normalised on-chain
10: for j = 1 → 3 do metric against the price. Correlation analysis enables us to
11: Calculate µCrj and µFj ; choose a subset of metrics that may significantly affect price
12: Calculate Cri,j and Fi,j for each individual xi in forecasting: we analyze the price-metric correlations and
popj ; select the most strongly linked metrics. Pearson’s correlation
13: Perform the jth mutation strategy and related coefficient determines the linear relationship between two
crossover operators over subpopulation popj ; variables, while the Spearman Correlation coefficient deter-
14: Set SCr,j = ∅ and SF,j = ∅; mines the monotonic relationship between two variables. The
15: end for highly correlated datasets are split into train and test datasets.
16: for i = 1 → NP do In the deep learning framework module, self-adaptive
17: if f (xgi ) ≤ f (ugi ) then algorithm-based deep learning models use differential evo-
18: xgi+1 = xgi ; lution algorithms to assist us in determining the optimal
19: else values for the deep-learning model’s hyperparameters since
20: xgi+1 = ugi ; △fj = △fj + f (xgi ) − f (ugi ); determining the optimal values for hyperparameters is a
21: Cri,j → SCr,j ; Fi,j → SF,j time-consuming process. We use three distinct self-adaptive
22: end if methods, namely L-SHADE, jSO, and MPEDE, to ensure
23: end for effective and efficient training. Each self-adaptive approach
24: pop = ∪j=1,....3 popj returns the optimal values for the hyperparameters deter-
25: if mod(g, ng) == 0 then mined after the deep learning model’s training. We develop
△f a deep-learning model using the optimal hyperparameter val-
26: k = arg(max1≤j≤3 =( ng.NPj ))
j
27: △fj = 0; ues acquired from each self-adaptive method. We utilize test
28: end if datasets to test self-adaptive algorithms-based deep learning
29: Randomly partition pop into pop1 , pop2 , pop3 and models and conduct MAE, mean squared error (MSE) and
pop4 ; MAPE analysis to assess the accuracy of self-adaptive algo-
30: Let popk = popk ∪ pop4 and NPk = NPk + NP4 ; rithms.
31: end while
IV. EXPERIMENTAL SETUP
In MPEDE, the following mutation strategies have been The proposed optimization methods were implemented in
utilized. Python 3.7.6 and TensorFlow (TF-API). The training was
• DE/current-to-pbest/1
done on Google Colab using the Tensor Processing Unit
(TPU). We used three unique self-adaptive techniques to
vgi = xgi + F (xgpbest − xgi ) + F (xgr1 − xgr2 ) (23) adjust the 17 hyperparameters involved in the RNN architec-
ture. The proposed algorithms produced 150 distinct models,
• DE/current-to-rand/1
each with its own set of parameters. The optimal parameter
vgi = xgi + F (xgi − xgi ) + F (xgr1 − xgr2 ) (24) combination with the most negligible loss is selected to de-
12 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

ON-CHAIN ANALYSIS

Train Test
Dataset
Dataset

DEEP LEARNING FRAMEWORK

Input layer

Hidden layer Output layer

Price Predic on
for (N+1)th day
Input Cell

Memory Cell

Output Cell

Input layer

Hidden layer Output layer

Evaluation
Metrics

FIGURE 11. Proposed System Model.

velop a prediction model for Ethereum. The hyperparameters three self-adaptive techniques: L-SHADE, jSO, and MPEDE.
are listed in the Table 3. We used optimization algorithms due to their unparalleled
speed in generating ideal model parameters to precisely
calculate and closely monitor the current Ethereum price. The
V. RESULTS AND ANALYSIS
proposed algorithms time performance is evaluated based on
This section summarises and discusses the findings of our re- optimization, training, and testing times. It is worth empha-
search. The correlation analysis results were used to develop sising that optimization time refers to the time it takes for
a deep learning model based on the data characteristics of algorithms that use the optimization operator to determine the
on-chain data and their influence on the Ethereum price. The best parameter combination for each algorithm with the most
deep learning model’s hyperparameters were optimised using
VOLUME 4, 2016 13

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

400 Real Price

Predicted Price(LSTM)
350
Ethereum Price (USD)

300

250

200

150

100
0 50 100 150 200 250 300 350
Days

FIGURE 12. Real Price vs Predicted Price using LSTM.

Real Price
350 Predicted Price(L-SHADE)

300
Ethereum Price (USD)

250

200

150

100

0 50 100 150 200 250 300 350

Days

FIGURE 13. Real Price vs Predicted Price using L-SHADE-LSTM optimization algorithm.

Real Price
350 Predicted Price(jSO)

300
Ethereum Price (USD)

250

200

150

100

0 50 100 150 200 250 300 350

Days

FIGURE 14. Real Price vs Predicted Price using jSO-LSTM optimization algorithm.

14 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

350
Real Price
Predicted Price(MPEDE)
300
Ethereum Price (USD)

250

200

150

100

0 50 100 150 200 250 300 350

Days

FIGURE 15. Real Price vs Predicted Price using MPEDE-LSTM optimization algorithm.

negligible loss, as shown in Table 4. The optimization time TABLE 4. The best Hyperparameters values obtained by L-SHADE, jSO &
MPEDE algorithms
of the proposed algorithms L-SHADE-LSTM, jSO-LSTM
and MPEDE-LSTM to achieve the optimal parameters are Hyperparameters L-SHADE jSO MPEDE
23077.59s, 24547.86s and 23872.21s, respectively. As shown No. of LSTM layers 3 2 3
in Table 5, our proposed approach also takes the least training LSTM Layer 1 42 18 58
and testing time compared to a traditional LSTM model. Recurrent Dropout in Layer 1 37 37 37
LSTM Layer 2 48 48 72
In Fig. 12 -15, a plot of the predicted price for Ethereum Recurrent Dropout in Layer 2 34 35 40
was obtained by using the traditional LSTM model and LSTM Layer 3 233 109 236
self-adaptive techniques- L-SHADE, jSO and MPEDE for Recurrent Dropout in Layer 3 37 34 30
LSTM Layer 4 143 201 131
60% training size compared to the actual Ethereum price is Recurrent Dropout in Layer 4 43 31 34
depicted. Our model adopts a holistic approach by analysing Dropout between LSTM and
33 35 43
the influence of core components of Ethereum, including Dense layers
Number of Dense layers 2 2 3
miners, nodes, and exchanges activities, that helps us identify Dense Layer 1 435 510 437
price trends better. Comparing Fig. 12-15, it is evident that Dense Layer 2 142 239 214
the predicted price obtained using the L-SHADE, jSO, and Dense Layer 3 82 79 120
MPEDE algorithms is relatively closer to the actual price of Dense Layer 4 38 32 32
Dense Layer 5 18 24 30
Ethereum when compared to the traditional LSTM model.
We conducted a mean absolute error (MAE), mean square
error (MSE) and mean absolute percentage error (MAPE) TABLE 5. Comparison of computational overhead times for training and
testing between LSTM and different optimization algorithms.

TABLE 3. Lower and upper bounds of Hyperparameters Model Training (s) Testing (s)
LSTM 551.40 4.43
Hyperparameters Min Max
No. of LSTM layers 2 4 L-SHADE-LSTM 527.86 3.29
LSTM Layer 1 8 64 jSO-LSTM 538.16 4.15
Recurrent Dropout in Layer 1 30 40
LSTM Layer 2 32 128 MPEDE-LSTM 497.22 2.84
Recurrent Dropout in Layer 2 30 50
LSTM Layer 3 86 256
Recurrent Dropout in Layer 3 30 50
LSTM Layer 4 86 256 comparison between the different self-adaptive techniques
Recurrent Dropout in Layer 4 30 50 and the traditional LSTM model to validate this finding.
Dropout between LSTM and Dense layers 30 50
Number of Dense layers 0 5 The MAE, MSE and MAPE values are calculated using the
Number of neurons in each dense Layer following formula:
Dense Layer 1 256 512
Dense Layer 2 128 256
Dense Layer 3 64 128 n
Dense Layer 4 32 64 1X
Dense Layer 5 16 32 MAE = |yk − yˆk | (26)
n
k=1
VOLUME 4, 2016 15

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

n
and health. In future work, we aim to extend this research
1X by including the attention mechanism into LSTM framework
MSE = (yk − yˆk )2 (27)
n [66] and applying it for on-chain data for predicting cryp-
k=1
tocurrency prices. In addition, it will be interesting to explore
n
1 X |yk − yˆk | on-chain and off-chain variables such as Reddit, Twitter, and
MAPE = × 100 (28) Google Trends, among others, in conjunction with a self-
n |yk |
k=1
adaptive algorithm-based LSTM model to attain near-perfect
where n is the number of data points and yk and yˆk are the price prediction accuracy.
real and predicted prices of sth point.
REFERENCES
TABLE 6. MAE, MSE & MAPE values for LSTM and different optimization [1] Department of Industry, Science, Energy and Resources, “National
algorithms at 60% training size
blockchain roadmap,” Department of Industry, Science, Energy and
Resources. [Online]. Available: https://www.industry.gov.au/data-and-
MAE MSE MAPE publications/national-blockchain-roadmap (accessed Jan. 29, 2021).
LSTM 26.41243 936.22145 43.64065 [2] X. Xiang, M. Wang, and W. Fan, “A permissioned blockchain-based iden-
L-SHADE-LSTM 16.87477 559.22696 36.02115 tity management and user authentication scheme for E-health systems,”
jSO-LSTM 21.40334 735.06323 39.93681 IEEE Access, vol. 8, pp. 171 771–171 783, 2020.
MPEDE-LSTM 19.82180 597.57913 37.84366 [3] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system.” [Online].
Available: https://bitcoin.org/bitcoin.pdf (accessed Jan. 3, 2021).
[4] W. Dai, “B-Money.” Satoshi Nakamoto Institute. [Online]. Available:
https://nakamotoinstitute.org/b-money (accessed Dec. 29, 2020).
TABLE 7. MAE, MSE & MAPE values for LSTM and different optimization [5] N. Szabo, “Bit Gold.” Satoshi Nakamoto Institute. [Online]. Available:
algorithms at 90% training size https://nakamotoinstitute.org/bit-gold (accessed Nov. 15, 2020).
[6] H. Finney, “RPOW - reusable proofs of work.” Satoshi Nakamoto
MAE MSE MAPE Institute. [Online]. Available: https://nakamotoinstitute.org/finney/rpow
LSTM 19.87341 591.55436 29.06433 (accessed Nov. 12, 2020).
L-SHADE-LSTM 13.62016 286.46870 23.84312 [7] V. Buterin, “Ethereum whitepaper.” Ethereum.org. [Online]. Available:
jSO-LSTM 15.31869 381.55121 25.02581 https://ethereum.org/en/whitepaper (accessed Mar. 6, 2021).
MPEDE-LSTM 14.89678 383.43369 24.52417 [8] CoinMarketCap, “Cryptocurrency prices, charts and market capitaliza-
tions.” CoinMarketCap. [Online]. Available: https://coinmarketcap.com/
(accessed Apr. 18, 2021).
The results of the MAE, MSE and MAPE analysis are [9] A. Eross, F. McGroarty, A. Urquhart, and S. Wolfe, “The intraday dynam-
presented in Table 6 and Table 7. It is observed that for train- ics of Bitcoin,” Research in International Business and Finance, vol. 49,
ing sizes of 60% and 90%, the self-adaptive algorithm-based pp. 71–81, 2019.
[10] R. Hudson and A. Urquhart, “Technical trading and cryptocurrencies,”
LSTM models, namely L-SHADE-LSTM, jSO-LSTM, and Annals of Operations Research, vol. 297, no. 1, pp. 191–220, 2021.
MPEDE-LSTM, yield lower MAE, MSE and MAPE values. [11] J.-Z. Huang, W. Huang, and J. Ni, “Predicting Bitcoin returns using
Correspondingly, our proposed model achieved the following high-dimensional technical indicators,” The Journal of Finance and Data
Science, vol. 5, no. 3, pp. 140–155, 2019.
improvements, 33.79%, 20.93% and 24.99% improvement [12] L. Kristoufek, “BitCoin meets Google Trends and Wikipedia: Quantify-
for MAE, 45.91%, 28.49% and 35.67% improvement for ing the relationship between phenomena of the Internet era,” Scientific
MSE and 17.70%, 12.02% and 14.45% for MAPE over Reports, vol. 3, no. 1, pp. 1–7, 2013.
[13] H. Jang and J. Lee, “An empirical study on modeling and prediction
its LSTM model counterpart. The L-SHADE-LSTM model of bitcoin prices with Bayesian neural networks based on blockchain
performed the best, yielding the lowest error rate compared to information,” IEEE Access, vol. 6, pp. 5427–5437, 2017.
jSO and MPEDE algorithms. These results suggest that self- [14] M. Saad, J. Choi, D. Nyang, J. Kim, and A. Mohaisen, “Toward character-
izing blockchain-based cryptocurrencies for highly accurate predictions,”
adaptive algorithm-based LSTM models provide a quicker IEEE Systems Journal, vol. 14, no. 1, pp. 321–332, 2019.
and more accurate prediction of Ethereum prices. [15] P. Jay, V. Kalariya, P. Parmar, S. Tanwar, N. Kumar, and M. Alazab,
“Stochastic neural networks for cryptocurrency price prediction,” IEEE
Access, vol. 8, pp. 82 804–82 818, 2020.
VI. CONCLUSION AND FUTURE WORKS
[16] A. M. Khedr, I. Arif, P. R. P V, M. El-Bannany, S. M. Alhashmi, and
In this article, we analyze factors that impact the price of M. Sreedharan, “Cryptocurrency price prediction using traditional statis-
Ethereum using a profusion of on-chain metrics relating to tical and machine-learning techniques: A survey,” Intelligent Systems in
Accounting, Finance and Management, vol. 28, no. 1, pp. 3–34, 2021.
the user, miner, and exchange activity. Using these findings,
[17] S. McNally, J. Roche, and S. Caton, “Predicting the price of bitcoin using
we develop an LSTM model that employs three distinct self- machine learning,” in 2018 26th Euromicro International Conference on
adaptive methods to identify the best hyperparameter values Parallel, Distributed and Network-based Processing (PDP), 2018, pp.
339–343.
for predicting the price of Ethereum. We compare each self-
[18] T. Klein, H. P. Thu, and T. Walther, “Bitcoin is not the New Gold–A com-
adaptive technique to one another as well as to a traditional parison of volatility, correlation, and portfolio performance,” International
LSTM model. In comparison, self-adaptive algorithm based Review of Financial Analysis, vol. 59, pp. 105–116, 2018.
LSTM models provide a quick and more accurate price [19] G. Gajardo, W. D. Kristjanpoller, and M. Minutolo, “Does Bitcoin exhibit
the same asymmetric multifractal cross-correlations with crude oil, gold
prediction for Ethereum. and DJIA as the Euro, Great British Pound and Yen?” Chaos, Solitons &
As demonstrated in this article, on-chain metrics can be Fractals, vol. 109, pp. 195–205, 2018.
utilized as a supplementary tool in combination with existing [20] M. I. Mehar, C. L. Shier, A. Giambattista, E. Gong, G. Fletcher, R. Sanay-
hie, H. M. Kim, and M. Laskowski, “Understanding a revolutionary and
techniques to predict prices more accurately. They provide flawed grand experiment in blockchain: The DAO attack,” Journal of
important insights into the blockchain network’s utilisation Cases on Information Technology (JCIT), vol. 21, no. 1, pp. 19–32, 2019.

16 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

[21] P. Tsankov, A. Dan, D. Drachsler-Cohen, A. Gervais, F. Buenzli, and [43] B. Wolf, “US law enforcers partner with cryptocurrency tracking
M. Vechev, “Securify: Practical security analysis of smart contracts,” in firm to fight financial crime.” Thomson Reuters. [Online]. Avail-
Proceedings of the 2018 ACM SIGSAC Conference on Computer and able: https://www.thomsonreuters.com/en-us/posts/investigation-fraud-
Communications Security, 2018, pp. 67–82. and-risk/cryptocurrency-financial-crime (accessed Jun. 24, 2021).
[22] W. Chen, Z. Zheng, E. C.-H. Ngai, P. Zheng, and Y. Zhou, “Exploiting [44] A. Nathan, G. L. Galbraith, and J. Grimberg, “Crypto:
blockchain data to detect smart Ponzi schemes on Ethereum,” IEEE A new asset class?” Goldman Sachs. [Online]. Avail-
Access, vol. 7, pp. 37 575–37 586, 2019. able: https://www.goldmansachs.com/insights/pages/crypto-a-new-asset-
[23] M. Bartoletti, S. Carta, T. Cimoli, and R. Saia, “Dissecting Ponzi schemes class.html (accessed Jun. 21, 2021).
on Ethereum: Identification, analysis, and impact,” Future Generation [45] B. G. Malkiel, “Efficient market hypothesis,” in The New Palgrave:
Computer Systems, vol. 102, pp. 259–277, 2020. Finance. London, UK: Palgrave Macmillan, 1989, pp. 127–134.
[24] H.-N. Dai, Z. Zheng, and Y. Zhang, “Blockchain for Internet of Things: [46] Glassnode Studio, “Glassnode studio - on-chain market intelligence.”
A survey,” IEEE Internet of Things Journal, vol. 6, no. 5, pp. 8076–8094, Glassnode Studio. [Online]. Available: https://studio.glassnode.com
2019. (accessed Aug. 15, 2021).
[25] M. S. Bhargavi, S. M. Katti, M. Shilpa, V. P. Kulkarni, and S. Prasad, [47] The Block, “Ethereum archives.” The Block. [Online]. Avail-
“Transactional data analytics for inferring behavioural traits in Ethereum able: https://www.theblockcrypto.com/data/on-chain-metrics/ethereum
blockchain network,” in 2020 IEEE 16th International Conference on (accessed Aug. 6, 2021).
Intelligent Computer Communication and Processing (ICCP), 2020, pp. [48] Bitcoin.com, “Bitcoin.com markets: Price, charts, news.” Bitcoin.com.
485–490. [Online]. Available: https://markets.bitcoin.com/crypto/BTC (accessed
[26] Blockchair, “Ethereum blockchain size chart.” Blockchair. [On- Aug. 18, 2021).
line]. Available: https://blockchair.com/ethereum/charts/blockchain-size [49] E. Krause, “A fifth of all Bitcoin is missing. These
(accessed Apr. 21, 2021). crypto hunters can help.” The Wall Street Journal. [Online].
[27] M. Bartoletti, S. Lande, L. Pompianu, and A. Bracciali, “A general Available: https://www.wsj.com/articles/a-fifth-of-all-bitcoin-is-missing-
framework for blockchain analytics,” in Proceedings of the 1st Workshop these-crypto-hunters-can-help-1530798731 (accessed Jan. 12, 2021).
on Scalable and Resilient Infrastructures for Distributed Ledgers, 2017. [50] N. Carter and A. L. Calvez, “Introducing realized capitalization.” Coin
[28] H. Kalodner, M. Möser, K. Lee, S. Goldfeder, M. Plattner, A. Chator, and Metrics. [Online]. Available: https://coinmetrics.io/realized-capitalization
A. Narayanan, “Blocksci: Design and applications of a blockchain analysis (accessed Oct. 5, 2020).
platform,” in 29th {USENIX} Security Symposium ({USENIX} Security [51] A. Schoedon, “EIP-1234: Constantinople difficulty bomb delay and
20), 2020, pp. 2721–2738. block reward adjustment.” Ethereum Improvement Proposals. [Online].
[29] Y. Li, K. Zheng, Y. Yan, Q. Liu, and X. Zhou, “EtherQL: A query layer for Available: https://eips.ethereum.org/EIPS/eip-1234 (accessed Dec. 3,
blockchain system,” in International Conference on Database Systems for 2020).
Advanced Applications, 2017, pp. 556–567. [52] K. Heeg and R. Schultze-Kraft, “Introducing account-based on-
[30] P. Zheng, Z. Zheng, J. Wu, and H.-N. Dai, “Xblock-ETH: Extracting and chain metrics for Bitcoin and Ethereum.” Glassnode Insights.
exploring blockchain data from Ethereum,” IEEE Open Journal of the [Online]. Available: https://insights.glassnode.com/account-based-metrics
Computer Society, vol. 1, pp. 95–106, 2020. (accessed Mar. 25, 2021).
[31] N. Smuts, “What drives cryptocurrency prices? An investigation of Google [53] P. Xia, H. Wang, B. Zhang, R. Ji, B. Gao, L. Wu, X. Luo, and G. Xu,
Trends and Telegram sentiment,” ACM SIGMETRICS Performance Eval- “Characterizing cryptocurrency exchange scams,” Computers & Security,
uation Review, vol. 46, no. 3, pp. 131–134, 2019. vol. 98, p. 101993, 2020.
[32] L. Muñoz-González, B. Biggio, A. Demontis, A. Paudice, V. Wongras- [54] Glassnode, “Bitcoin’s on-chain market cycles.” Glassnode Insights. [On-
samee, E. C. Lupu, and F. Roli, “Towards poisoning of deep learning line]. Available: https://insights.glassnode.com/bitcoin-onchain-market-
algorithms with back-gradient optimization,” in Proceedings of the 10th cycles (accessed Apr. 12, 2021).
ACM Workshop on Artificial Intelligence and Security, 2017, pp. 27–38. [55] C. Spearman, “The proof and measurement of association between two
[33] M. Jagielski, A. Oprea, B. Biggio, C. Liu, C. Nita-Rotaru, and B. Li, things,” The American Journal of Psychology, vol. 100, no. 3/4, pp. 441–
“Manipulating machine learning: Poisoning attacks and countermeasures 471, 1987.
for regression learning,” 2018 IEEE Symposium on Security and Privacy [56] J. Brest, S. Greiner, B. Boskovic, M. Mernik, and V. Zumer, “Self-
(SP), pp. 19–35, 2018. adapting control parameters in differential evolution: A comparative study
[34] K. Metaxiotis and K. Liagkouras, “Multiobjective evolutionary algorithms on numerical benchmark problems,” IEEE Transactions on Evolutionary
for portfolio management: A comprehensive literature review,” Expert Computation, vol. 10, no. 6, pp. 646–657, 2006.
Systems with Applications, vol. 39, no. 14, pp. 11 685–11 698, 2012. [57] A. K. Qin, V. L. Huang, and P. N. Suganthan, “Differential evolution
[35] K. V. Price, “Differential evolution vs. the functions of the 2/sup nd/ algorithm with strategy adaptation for global numerical optimization,”
ICEO,” in Proceedings of 1997 IEEE International Conference on Evo- IEEE Transactions on Evolutionary Computation, vol. 13, no. 2, pp. 398–
lutionary Computation (ICEC ’97), 1997, pp. 153–157. 417, 2009.
[36] K. V. Price, “Differential evolution: A fast and simple numerical opti- [58] J. Zhang and A. C. Sanderson, “JADE: Adaptive differential evolution with
mizer,” in Proceedings of North American Fuzzy Information Processing, optional external archive,” IEEE Transactions on Evolutionary Computa-
1996, pp. 524–527. tion, vol. 13, no. 5, pp. 945–958, 2009.
[37] R. Storn and K. Price, “Differential evolution–A simple and efficient [59] R. Tanabe and A. S. Fukunaga, “Improving the search performance of
heuristic for global optimization over continuous spaces,” Journal of SHADE using linear population size reduction,” in 2014 IEEE congress
Global Optimization, vol. 11, no. 4, pp. 341–359, 1997. on evolutionary computation (CEC), 2014, pp. 1658–1665.
[38] R. Storn and K. Price, “Minimizing the real functions of the ICEC’96 [60] K. M. Sallam, S. M. Elsayed, R. K. Chakrabortty, and M. J. Ryan, “Im-
contest by differential evolution,” in Proceedings of IEEE International proved multi-operator differential evolution algorithm for solving uncon-
Conference on Evolutionary Computation, 1996, pp. 842–844. strained problems,” in 2020 IEEE Congress on Evolutionary Computation
[39] R. Storn, “On the usage of differential evolution for function optimization,” (CEC), 2020.
in Proceedings of North American Fuzzy Information Processing, 1996, [61] K. M. Sallam, S. M. Elsayed, R. K. Chakrabortty, and M. J. Ryan,
pp. 519–523. “Evolutionary framework with reinforcement learning-based mutation
[40] F. Neri and V. Tirronen, “Recent advances in differential evolution: A adaptation,” IEEE Access, vol. 8, pp. 194 045–194 071, 2020.
survey and experimental analysis,” Artificial Intelligence Review, vol. 33, [62] K. M. Sallam, S. M. Elsayed, R. A. Sarker, and D. L. Essam, “Landscape-
no. 1, pp. 61–106, 2010. assisted multi-operator differential evolution for solving constrained op-
[41] S. Das and P. N. Suganthan, “Differential evolution: A survey of the state- timization problems,” Expert Systems with Applications, vol. 162, p.
of-the-art,” IEEE Transactions on Evolutionary Computation, vol. 15, 113033, 2020.
no. 1, pp. 4–31, 2011. [63] R. Tanabe and A. Fukunaga, “Success-history based parameter adaptation
[42] N. Jagannath, T. Barbulescu, K. M. Sallam, I. Elgendi, A. A. Okon, for differential evolution,” in 2013 IEEE Congress on Evolutionary Com-
B. Mcgrath, A. Jamalipour, and K. Munasinghe, “A self-adaptive deep putation, 2013, pp. 71–78.
learning-based algorithm for predictive analysis of bitcoin price,” IEEE [64] J. Brest, M. S. Maučec, and B. Bošković, “iL-SHADE: Improved L-
Access, vol. 9, pp. 34 054–34 066, 2021. SHADE algorithm for single objective real-parameter optimization,” in

VOLUME 4, 2016 17

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.:An On-Chain Analysis-Based Approach to Predict Ethereum Prices

2016 IEEE Congress on Evolutionary Computation (CEC), 2016, pp. IBRAHIM ELGENDI (M’16)holds a PhD in In-
1188–1195. formation Technology from University of Can-
[65] G. Wu, R. Mallipeddi, P. N. Suganthan, R. Wang, and H. Chen, “Dif- berra, Australia. He is currently the lecturer in
ferential evolution with multi-population based ensemble of mutation Networking and Cybersecurity. He is a member
strategies,” Information Sciences, vol. 329, pp. 329–345, 2016. of the Institution of Engineers Australia, His re-
[66] W. Zheng, P. Zhao, K. Huang, and G. Chen, “Understanding the prop- search focuses on Mobile and Wireless Networks,
erty of long term memory for the LSTM with attention mechanism,” in Internet-of-Things, Machine Learning, and Cyber-
Proceedings of the 30th ACM International Conference on Information &
Physical-Security. Dr. Elgendi has over 14 refer-
Knowledge Management, 2021, pp. 2708–2717.
eed publications with over 63 citations (H-index
5) in highly prestigious journals, and conference
proceedings. He has served as a reviewer for a number of journals such
as IEEE Wireless Communications magazine, IEEE Transaction on Mobile
Computing, and IEEE/ACM Transactions on Networking. He has 17 years’
experience from industry in the field of Automation and AI.

NISHANT JAGANNATH is currently pursuing a

PhD degree at the University of Canberra, Aus-
tralia. His doctoral research investigates the cur-
rent issues in adopting blockchain technology,
given the diverse nature of its implementations
from financial markets to supply chain. He holds BRADEN MCGRATH is a Research Professor
a Master’s degree in Network Engineering and is at Embry-Riddle Aeronautical University and the
also a part-time faculty member in the Science and CEO of Cann Pharmaceutical Australia. Braden is
Technology department at the University of Can- a Chartered Engineer in the Engineers Australia
berra. His research interests include integrating Colleges of Biomedical Engineering and Lead-
blockchain technology with the Internet of things (IoT) and cyber-physical ership and Management; a Fellow of Engineers
systems. Australia; a QinetiQ Fellow; a member of the
Society of Flight Test Engineers; and a certified
Sports Medicine Trainer. Braden holds a Ph.D. in
Aeronautical Engineering from the University of
Sydney and a Master’s in Aeronautics and Astronautics from MIT and
his current research activities include the development of secure supply
chain technology for pharmaceutical-quality medicinal cannabinoid drug
development on the human neurological system.

TUDOR BARBULESCU holds a BIT from the

Australian National University and is currently
pursuing a Masters in Computing / Business Infor-
matics from the University of Canberra. Recently
he has been working as a software developer on
personal projects and for clients and his interest
lies at the intersection of algorithmic trading and ABBAS JAMALIPOUR (S’86–M’91–SM’00–
blockchain. Before this he co founded an aug- F’07) is the Professor of Ubiquitous Mobile Net-
mented reality startup. working at the University of Sydney, Australia,
and holds a PhD in Electrical Engineering from
Nagoya University, Japan. He is a Fellow of the
Institute of Electrical, Information, and Commu-
nication Engineers (IEICE) and the Institution of
Engineers Australia, an ACM Professional Mem-
ber, and an IEEE Distinguished Lecturer. He has
authored nine technical books, eleven book chap-
ters, over 450 technical papers, and five patents, all in the area of wireless
KARAM M. SALLAM received the Ph.D. de- communications. Dr. Jamalipour is the President and an elected member
gree in computer science from the University of of the Board of Governors of the IEEE Vehicular Technology Society. He
New South Wales at Canberra, Australian Force was the Editor-in-Chief IEEE Wireless Communications, Vice President-
Academy, Canberra, Australia, in 2018. He is Conferences and a member of Board of Governors of the IEEE Communi-
currently a Lecturer at University of Canberra, cations Society, and serves as an editor of IEEE Access, IEEE Transactions
Canberra, Australia. His current research interests on Vehicular Technology, and several other journals. He has been a General
include evolutionary algorithms and optimization, Chair or Technical Program Chair for a number of conferences, including
constrained-handling techniques for evolutionary IEEE ICC, GLOBECOM, WCNC and PIMRC. He is the recipient of a
algorithms, operation research, machine learning, number of prestigious awards such as the 2019 IEEE ComSoc Distinguished
deep learning, Cybersecurity and IoT. He was the Technical Achievement Award in Green Communications, the 2016 IEEE
winner of IEEE-CEC2020 and IEEE-CEC2021 competitions. He serves as ComSoc Distinguished Technical Achievement Award in Communications
an organizing committee member of different conferences in the evolution- Switching and Routing, the 2010 IEEE ComSoc Harold Sobol Award, the
ary computation field and a reviewer for several international journals. 2006 IEEE ComSoc Best Tutorial Paper Award, as well as 15 Best Paper
Awards.

18 VOLUME 4, 2016

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2021.3135620, IEEE Access

N. Jagannath et al.: An On-Chain Analysis-Based Approach to Predict Ethereum Prices

MOHAMED ABDEL-BASSET received the

B.Sc., M.Sc., and Ph.D. degrees in operations
research from the Faculty of Computers and In-
formatics, Zagazig University, Egypt. He is an
Associate Professor with the Faculty of Comput-
ers and Informatics, Zagazig University. He has
published more than 200 articles in international
journals and conference proceedings. His current
research interests are optimization, operations re-
search, data mining, computational intelligence,
applied statistics, decision support systems, robust optimization, engineering
optimization, multiobjective optimization, swarm intelligence, evolutionary
algorithms, and artificial neural networks. He is working on the application
of multiobjective and robust metaheuristic optimization techniques. He is
also an/a Editor/Reviewer in different international journals and conferences.

KUMUDU MUNASINGHE holds a PhD in

Telecommunications Engineering from University
of Sydney. He is currently the Associate Profes-
sor in Network Engineering, leader of the IoT
Research Group at the Human Centred Research
Centre, University of Canberra. His research fo-
cuses on Next Generation Mobile and Wireless
Networks, Internet-of-Things, Green Communi-
cation, Smart Grid Communications, and Cyber-
Physical-Security. A/Prof Munasinghe has over
100 refereed publications with over 900 citations (H-index 17) in highly
prestigious journals, conference proceedings and two books to his credit.
He has secured over $ 1.6 Million dollars in competitive research fund-
ing by winning grants from the Australian Research Council (ARC), the
Commonwealth and State Governments, Department of Defence, and the
industry. He has also won the highly prestigious ARC Australian Postdoc-
toral Fellowship, served as a co-chair for many international conferences,
served as an editorial board member for a number of journals. A/Prof
Munasinghe’s research has been highly commended through many research
awards including two VC’s Research Awards and three IEEE Best Paper
Awards. He is currently a Member of the IEEE, a Charted Professional
Engineer, an Engineering Executive and a Companion (Fellow Status) of
Engineers Australia.

VOLUME 4, 2016 19

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/