algorithms

Article
Applying Artificial Intelligence in Cryptocurrency Markets:
A Survey
Rasoul Amirzadeh * , Asef Nazari and Dhananjay Thiruvady

School of Information Technology, Deakin University, Geelong, VIC 3216, Australia

* Correspondence: [email protected]

Abstract: The total capital in cryptocurrency markets is around two trillion dollars in 2022, which is
almost the same as Apple’s market capitalisation at the same time. Increasingly, cryptocurrencies
have become established in financial markets with an enormous number of transactions and trades
happening every day. Similar to other financial systems, price prediction is one of the main challenges
in cryptocurrency trading. Therefore, the application of artificial intelligence, as one of the tools of
prediction, has emerged as a recently popular subject of investigation in the cryptocurrency domain.
Since machine learning models, as opposed to traditional financial models, demonstrate satisfactory
performance in quantitative finance, they seem ideal for coping with the price prediction problem in
the complex and volatile cryptocurrency market. There have been several studies that have focused
on applying machine learning for price and movement prediction and portfolio management in
cryptocurrency markets, though these methods and models are in their early stages. This survey
paper aims to review the current research trends in applications of supervised and reinforcement
learning models in cryptocurrency price prediction. This study also highlights potential research gaps
and possible areas for improvement. In addition, it emphasises potential challenges and research
directions that will be of interest in the artificial intelligence and machine learning communities
focusing on cryptocurrencies.
Citation: Amirzadeh, R.; Nazari, A.;
Thiruvady, D. Applying Artificial Keywords: cryptocurrency markets; artificial intelligence; price prediction; FinTech; reinforcement
Intelligence in Cryptocurrency learning
Markets: A Survey. Algorithms 2022,
15, 428. https://doi.org/10.3390/
a15110428

Academic Editors: Danilo Pelusi and

1. Introduction
Asit Kumar Das Cryptocurrency markets have experienced a remarkable transformation recently, both
in statistical terms and public acknowledgement. The total market capitalisation of cryp-
Received: 9 October 2022
tocurrencies rose from 267.8 billion to 1.664 trillion between November 2017 and March
Accepted: 9 November 2022
2022 (based on data on www.tradingview.com, accessed on 1 November 2022). Excluding
Published: 14 November 2022
Bitcoin as the most recognised cryptocurrency with the highest market capital, the capi-
Publisher’s Note: MDPI stays neutral talisation of “alternate coins”, or altcoins, grew from 86.31 billion to 1.007 trillion in the
with regard to jurisdictional claims in same period. Moreover, there is an increasing trend among international companies that
published maps and institutional affil- have invested in cryptocurrency-based business solutions. For example, the possibility of
iations. purchasing Tesla merchandise using Dogecoin is a revealing example of the phenomenon
in the cryptocurrency world. Thus, cryptocurrency markets have become more popular,
established, and their integration with other financial assets appears inevitable [1,2].
Cryptocurrency markets have been a prominent subject for investors. Many small in-
Copyright: © 2022 by the authors.
vestors have made a small fortune by speculating on cryptocurrencies. They follow trending
Licensee MDPI, Basel, Switzerland.
This article is an open access article
news on social media or waves of excitement in cryptocurrencies. On the other hand, there
distributed under the terms and
is also a risk, and investors are prone to losing money because of the volatile nature of cryp-
conditions of the Creative Commons tocurrency markets. For example, based on the analysis of data on (www.coinmarketcap.
Attribution (CC BY) license (https:// com, accessed on 1 November 2022), the Bitcoin market capitalisation dropped from 1.18 tril-
creativecommons.org/licenses/by/ lion to 935 billion USD in only ten days in April 2021, and it almost halved to 602 billion
4.0/). USD in the following three months up to July 2021 (www.coinmarketcap.com). Thus,

Algorithms 2022, 15, 428. https://doi.org/10.3390/a15110428 https://www.mdpi.com/journal/algorithms

Algorithms 2022, 15, 428 2 of 26

this provided a trading system to predict market directions, and eventually mitigating
investment risks is an open problem in cryptocurrency markets. To this end, there is an
increasing number of promising methods to deal with cryptocurrency analysis and predict
inherent trends. However, high volatility and the fact that cryptocurrencies do not behave
like fiat currencies make them uncertain for investors [3] and require them to have their
own specific trading strategies. The uncertainty in predictability suggests that several
price formation elements of cryptocurrencies still have not been thoroughly examined.
Hence, these challenges have led to the necessity of further investigation to gain a better
understanding of cryptocurrency markets [4].
Asset trading has been experiencing substantial changes with the rapid enhancements
in computing and telecommunication infrastructures that support a highly productive
improvement in quantitative trading [5]. Besides, the growing number of investors and,
consequently, transactions, in addition to multiple sources of alternative data such as social
media hashtags, tweets, and feeds, have built massive blocks of big data around cryptocur-
rency markets. As a result, market participants look beyond traditional approaches to build
automated-profitable trading models dealing with this data. In contrast to quantitative
trading approaches utilised for decades, financial technology (FinTech) companies have
significantly changed these trading strategies to handle big data appropriately. FinTechs
are mainly established around integrating artificial intelligence (AI) and finance [6]. As one
of the primary tools of FinTechs, AI has become a popular research subject in quantita-
tive trading and presents excellent efficiency in discovering profitable trading rules [7,8].
In contrast to traditional statistical models, AI techniques as a ubiquitous analysis tool
can deal with high-dimensional data and complex environments, such as cryptocurrency
markets. Further references on FinTech companies and their approaches to dealing with
the challenges and complexities of cryptocurrencies can be found in [9–11].
The number of academic publications is rapidly growing in the field of cryptocur-
rencies in connection to various domains, such as economics, finance, and AI. To further
support this claim, a python package called “academic-keyword-occurrence” [12] is used to
extract the historic word occurrences of a set of keywords in academic papers from Google
Scholar. Figure 1 demonstrates the exponential increase in the number of publications based
on four different keywords, namely altcoin, altcoins, cryptocurrency, and cryptocurrencies,
from 2015 until 2021, based on the package results.

Figure 1. Publications associated with cryptocurrencies between 2015–2021.

Alternatively, another possible categorisation of literature in the realm of cryptocur-

rency trading is about trading software systems, systematic trading, emergent trading
Algorithms 2022, 15, 428 3 of 26

technologies, crypto-asset portfolio research, market condition, and so forth [13]. In this
paper, we mainly focus on the third category (emergent trading technologies), including
econometric methods, machine learning technology, etc. As evidence of growing literature
on the application of AI models in predicting cryptocurrency prices, several surveys and
reviews of recent publications on this topic have been published recently. For example,
Mosavi et al. [14] review deep learning methods in various finance and economic sectors
such as insurance, auction mechanisms, and banking. Additionally, Sabry et al. [15] present
a survey on current challenges and opportunities of AI applications in several cryptocur-
rency domains, such as volatility prediction, cryptocurrency mining, and fraud detection.
Murat Ozbayoglu et al. [16] provide a state-of-the-art snapshot of deep learning models for
a range of financial applications, including algorithmic trading, risk management, fraud
detection, and behavioural finance. The primary purpose of this paper is to review recent
studies on AI applications in cryptocurrencies as well. However, unlike other studies, it
only focuses on cryptocurrency price prediction by machine learning (ML) models and
provides an in-depth review of the challenge of predicting the price of cryptocurrencies.
Furthermore, this work investigates the financial aspects underlying the price prediction
presented in recent studies, which are examples of studies conducted for the first time on
this topic. Therefore, to achieve the main purpose of our study, this survey aims to answer
the following questions to fill the gap in the existing surveys of recent publications in AI
and cryptocurrency price prediction.
1. What factors influence the price of cryptocurrencies?
2. What is the state-of-the-art in AI research in the domain of cryptocurrency price
prediction?
3. What are current gaps in the literature that may be addressed by conducting future
research?
The rest of the paper is structured as follows. Section 2 primarily focuses on the domain
knowledge of the cryptocurrency sector from a financial and economic perspective. Then,
ML approaches used in predicting cryptocurrency prices are divided into two subsections,
namely supervised and reinforcement learning, and analysed in Section 3. Subsequently,
Section 4 discusses current research gaps and recommends potential future research. Finally,
Section 5 concludes the paper.

2. Cryptocurrency Markets
In this section, a short introduction to cryptocurrencies and blockchain technology,
as the digital infrastructure underpinning the financial transactions of cryptocurrencies,
is provided. Then, the fundamental drivers of price formation on cryptocurrencies are
discussed from the perspective of a famous framework for the life cycle of data science
projects, namely, the Cross-Industry Standard Process for Data Mining (CRISP-DM). In ad-
dition to supply-demand equilibrium in deciding cryptocurrency prices, it is observed that
several other drivers influence price fluctuations in cryptocurrency markets.

2.1. A Short Background on Cryptocurrencies

Traditionally, economic systems typically use third-party financial institutions such
as banks to process payments. These entities operate as mediators among parties for ex-
changing funds, and they have complete control over all transactions. Although traditional
financial systems have effectively executed financial transactions, there are a few drawbacks
to such systems. For example, limits on the amount of money transacted, transaction costs,
a lack of trust, security problems, transparency, and flexibility are among them. There were
unsuccessful attempts to develop decentralised unregulated virtual currencies to tackle
these problems. However, it was the invention of blockchain by Stuart Haber and W. Scott
Stornett in 1991 that addressed some of these difficulties [17].
As a distributed database solution, blockchain technology maintains an expanding list
of data records confirmed by participant nodes. Each transaction information is recorded in
a public ledger that is available to all nodes. Thus, a blockchain network is more transparent
Algorithms 2022, 15, 428 4 of 26

than a centralised transaction network involving a third party. Furthermore, as participant

nodes are all anonymous in a blockchain network, it is more secure for other nodes to verify
the transactions [18]. The term blockchain is derived from the fact that it is incrementally
built up by blocks of data or chains of data where each of those chains contains three main
elements [19] as follows:
• Block-data is a set of messages or transactions;
• Chaining-hash is a copy of the hash value of the immediately preceding block; and
• Block-hash is the calculated value of the hash of the data block.
Since the invention of the technology, various blockchain projects have been developed
in different sectors, from international payments and healthcare to music royalties tracking.
Moreover, attempts have been made to create a decentralised virtual currency, such as Bit
Gold in 1998. However, the advent of Nakamoto’s Bitcoin paper [20] in 2008 was a historic
milestone for both blockchain and digital currencies.
Cryptocurrencies rely on a decentralised peer-to-peer network, using blockchain tech-
nology to store all transactions in a decentralised public ledger. In addition, blockchain
technology is responsible for verifying transactions and synchronising nodes in a network
of participants. It provides a digital platform for fast, efficient, and secure cryptocur-
rency transactions.
In a blockchain system for cryptocurrencies, a block header consists of the main
metadata in cryptocurrencies, such as the previous block, block version, hash, timestamp,
nonce, and transaction details. It is used to identify a particular block in a blockchain system,
and each block has a unique header. Table 1 presents information attributes included in a
cryptocurrency blockchain, and a short description of each item is presented.

Table 1. Information attributes of a blockchain system for cryptocurrencies.

Headers Attributes and Definitions

Total circulation of crypto: the total number of mined cryptocurrency
coins
Crypto statistical info
Crypto price: the price of the coin
Market capitalisation: the total value of cryptocurrency in circulation
Blockchain size: total size of the blockchain
Avg. block size: average block size for the past 24 h
Avg. trans per block: average number of transactions per block for
the last 24 h
Block info Avg. payments per block: the average number of payments per block
for last 24 h
Total no. of trans: the total number of transactions on blockchain
Median (avg.) confirmation time: the median (avg.) time for a mined
block to be added to the public ledger
Total hash rate: the estimated number of terahashes per second
Hash rate distribution: an estimation of hash rate distribution
amongst the largest mining pools
Mining info Network difficulty: the difficulty of mining a new block
Miners revenue: total value of cryptocurrency block rewards and
transaction fees paid to miners
Total transaction fee
Fees per transaction: average transaction fees per transaction

Traditional financial systems and cryptocurrencies can be compared from various

perspectives, such as their monetary, regulatory, and decentralised functionality. From a
monetary function view, three conventional money functions are the medium of exchange,
store of value, and unit of account. In economics, a medium of exchange is a transitional
instrument widely accepted for exchanging products and services, such as fiat money.
In the monetary economy, a store of value means any form of financial assets that can be
used to save and eventually be exchanged in the future. For example, gold and silver are
Algorithms 2022, 15, 428 5 of 26

two popular commodities that have been saved throughout history due to their ability
to store value. A unit of account is a monetary unit that can measure the value of other
assets or products. Although cryptocurrencies can technically fulfil the monetary role of
medium of exchange, they are not currently capable of performing two other monetary
roles [21]. From a regulatory standpoint, the regulation should not target cryptocurrencies,
as there are several practical constraints for regulating this decentralised open-source
ecosystem. In other words, the centralised command-and-control approach of traditional
regulation may not succeed in decentralised cryptocurrencies. Instead, regulations seem to
aim at the interface between financial associations and cryptocurrencies [22]. Regarding the
decentralised function, financial institutions are the essential intermediaries that control
and facilitate financial transactions in a centralised economic system. However, these
institutions can dominate financial activities when they grow and impose disproportionate
market power. On the contrary, decentralised platforms reduce transaction costs and
eliminate the monopoly power of any entity by creating effective decentralised peer-to-peer
networks [23].
Moreover, the uniqueness of cryptocurrencies is based on three distinct features,
namely, anonymity, decentralisation (no central authority), and double-spending attack
protection [24]. However, even with these different characteristics, there is still ambiguity
between electronic money and cryptocurrency. The European Central Bank (ECB) has
provided a clear distinction between virtual and electronic currencies, presented in Table 2
based on [25]. In particular, virtual money is different from electronic money considering
its format, acceptance, legal status, issuer, and so forth.

Table 2. Comparison of electronic money and virtual currency based on the definition of the European
Central Bank.

Money
Electronic Virtual
Attributes
Money format Digital Digital
By undertakings other than Usually within a specific
Acceptance
the issuer virtual community
Legal status Regulated Unregulated
Legally established electronic Non-financial private
Issuer
money institution company
Not fixed (depends on issuer’s
Supply of money Fixed
decisions)
Supervision Yes No
Legal, credit, liquidity,
Type of risk Mainly operational
and operational

Bitcoin is the most popular and well-established cryptocurrency and a de facto stan-
dard in cryptocurrency markets. However, there are nearly 17,000 other cryptocurrencies
called altcoins, which have around 50% of the total market capitalisation. Figure 2 shows
the market capitalisation of the top ten cryptocurrencies in January 2022 based on Coin-
MarketCap (www.coinmarketcap.com). This figure shows BTC, ETH, USDT, BNB, USDC,
SOL, ADA, XRP, LUNA, and DOT, which stand for Bitcoin, Ethereum, Tether, Binance coin,
USD coin, Solana, Cardano, Ripple, Terra, and Polkadot.
At a technical level, altcoins use almost the same or similar blockchain technology
used by Bitcoin. Nevertheless, altcoins fundamentally seek to introduce some new alterna-
tive features compared to Bitcoin to increase their market share. For instance, Ethereum
incorporates almost all of Bitcoin’s attributes with additional features such as a fee limit
or a digital platform to run smart contracts (SC) [26]. Litecoin, as another example, is
designed to cope with the computational requirements for mining cryptocurrency coins.
Furthermore, the aim of creating Dash coin is to have a faster transactions process and
Algorithms 2022, 15, 428 6 of 26

extend privacy protection [27]. Aside from the technical differences, cryptocurrencies can
generally be categorised as follows.
• Mining-based altcoins: they have similar characteristics to Bitcoin, and as the name
implies, they use the typical mining process for generating new coins. One of the most
famous leading altcoins belonging to this category is Ethereum.
• Stablecoins: One of the main issues of mining-based cryptocurrencies is high volatility
and fluctuation in their prices, making their trading complicated. Hence, stablecoins
were introduced to address this challenge, which is valued based on stable existing
currencies such as fiat currencies. Additionally, stable assets behind stablecoins secure
and support their value. For example, Diem (previously Libra), developed by Face-
book, and Tether, with the highest capital among stablecoins, are two famous coins in
this cryptocurrency category.
• Utility tokens: this type of cryptocurrency can give value to its investors by providing
access to a future product or service. For example, Filecoin is a famous open-source
cryptocurrency that aims to store data on hard drive storage spaces compared to cloud
storage companies such as Amazon.

Figure 2. The Market capitalisation of different cryptocurrencies.

Cryptocurrencies can also be categorised based on the currency domain, platform

category facilities, and the domain of application [28].
The cryptocurrency ecosystem mainly relies on many exchanges as they provide trad-
ing tools for investors. Exchanges facilitate cryptocurrency trades and allow traders to
sell or buy digital assets for other assets, primarily fiat currencies. Centralised exchanges
(CEX), decentralised exchanges (DEX), and hybrid exchanges are three types of cryptocur-
rency exchanges. CEX is governed by a corporation, such as Binance and Coinbase. DEX
provides an automated process for peer-to-peer trades. For example, DODO and Uniswap
are part of this group. Finally, hybrid exchanges, as the name implies, are a combination
of both centralised and decentralised exchanges [29]. There are numerous cryptocurrency
exchanges worldwide. For example, Bitcoin can be traded in various currency cross pairs in
at least 387 markets [30]. Payment methods, supported coins, transaction fees, transaction
speed, and trading volume are a few factors of an exchange’s popularity. Table 3 presents
a number of the most known exchanges along with their specifications. The number of
supported coins in the second column of the table shows the number of cryptocurrencies
available in the exchanges for trade. Furthermore, their transaction fees, the location of the
Algorithms 2022, 15, 428 7 of 26

headquarters, and the year in which the exchange was founded are represented in columns
three, four, and five in Table 3, respectively. It is observed that transaction fees can vary for
each exchange and are between 0% and 4% per trade.

Table 3. Information about a number of most known cryptocurrency exchanges.

Exchange No. Supported Coins Transaction Fee (%) Headquarter Location Founded
Binance 320+ 0.100 Malta 2017
Coinbase 40+ 0.500 San Francisco, US 2012
BitMex 160+ 0.075 Eden Island, Seychelles 2014
Okex 230 0.150 Malta 2017
Huobi 310+ 0.200 Seychelles 2013
Bitfinex 30+ 0.200 Hong Hong 2012
Kraken 60 0.260 San Fransisco, US 2011
Bitterx 320+ 0.350 Seattle, US 2014
BitStamp 10+ 0.500 Luxembourg 2011
KuCoin 270+ 0.100 Mahe, Seychelles 2017

Cryptocurrencies, similar to most new technologies, come with their disadvantages.

Decentralisation supports cryptocurrencies to be less prone to government interventions,
and blockchain technology provides a medium with the semi-anonymity of owners. There-
fore, these lead to involvement in illegal activities as one of the main drawbacks [15].
For instance, the famous story of the Silk Road website portrayed a negative image of
Bitcoin, which was a dark web black market for trading illegal substances [31]. It is worth
mentioning that illicit internet activities are not only limited to cryptocurrencies and
blockchain technologies. For example, it is estimated that 10 million Americans are victims
of online identity theft each year [32].

2.2. Price Determinants of Cryptocurrencies

The growing importance of cryptocurrency markets in changing the worldwide econ-
omy is not negligible. It is necessary to understand the underlying features of financial
aspects of cryptocurrencies before taking further action in devising an ML model with
high prediction capability in forecasting their behaviours. Moreover, there are motives
for both governmental bodies and individual investors to expand their knowledge about
the factors behind the price formation of cryptocurrencies. For instance, cryptocurrency
markets can be a potential source of instability in economic systems. Consequently, they
may influence policymakers’ decisions and limit the authority of governments all over the
world. Moreover, an investor’s asset portfolio allocation and risk management strategy
can be influenced by fluctuations in cryptocurrencies [2]. To manage the consequences,
the econometrics of cryptocurrencies seek interdisciplinary research to explain how traders’
actions inform price discovery [33].
For the purpose of interdisciplinary research, the Cross-Industry Standard Process
for Data Mining is considered a framework for analysing the problem of price predic-
tion. It is one of the most common methodologies used to describe data science projects
(Figure 3) [34]. CRISP-DM is a process model that naturally describes the data science
life cycle and contains six phases. First, the business understanding phase is about the
objectives and requirements of a particular project. Next, the data understanding phase
focuses on identifying, collecting, and analysing datasets to achieve the project goals. Then,
the data preparation phase processes the final dataset to be used in the modelling phase,
which builds and assesses several models to find the best model(s). Finally, the evalua-
tion and deployment phases are self-descriptive, which are about evaluating results and
deploying the model into action. A significant observation and motivation behind incor-
Algorithms 2022, 15, 428 8 of 26

porating CRISP-DM are that the study of current literature in the cross-disciplinary field
of AI and cryptocurrencies demonstrates that AI community researchers mainly focus
on the modelling phase of CRISP-DM, and the business understanding phase is mainly
neglected. Thus, this section provides information to understand the factors behind the
price formation in cryptocurrency markets by focusing on business, finance, and economic
perspectives by means of CRISP-DM.

Figure 3. Six phases of CRISP-DM for data analysis projects.

Factors that drive the price of cryptocurrencies can be divided into internal factors and
external drivers, including marco-financial and attractiveness for investors [35]. The inter-
nal factors are about the generation of the data blocks via the mining process in a blockchain
network, and they directly impact the supply and demand of each cryptocurrency [36].
Furthermore, similar to a fiat currency, which is valued based on the trust in which it will be
accepted as an exchange medium, cryptocurrencies establish their trustworthiness among
market participants by creating trust and credibility [1]. In other words, the attractiveness
of cryptocurrencies for investors is considered an external factor driving price fluctuations.
Despite the general supply-demand equilibrium concept that stabilises the price of
commodities, the price of Bitcoin seems to follow different paradigms. Ciaian et al. [1]
expand a gold standard model to investigate factors in the formation of Bitcoin price
variation. This study considers three main factors including Bitcoin supply and demand
market forces, its attractiveness for investors, and global macro-financial development.
The results demonstrate that the impact of attractiveness has changed over time, which
means there are periods when investors are interested in investing in Bitcoin regardless of
the supply-demand equilibrium, and the macro-financial development factor has a minor
influence on Bitcoin price in the short run. The key finding is that the demand-side pressure
substantially affects the Bitcoin price in comparison to supply-side factors.
Another fundamental question in the price prediction of a cryptocurrency is whether
other financial assets impact the change in price value or price fluctuations. Therefore,
the degree to which cryptocurrency markets integrate with other financial assets and their
own interconnections is important to be investigated by investors for building better price-
prediction strategies. Ji et al. [2] use a directed acyclic graph (DAG) and Granger causality
test to uncover the causal structure between three categories of assets, including interna-
tional equities and currency assets. An energy index is also incorporated as a commodity
component due to the fact that electricity price is important in Bitcoin mining. The re-
sults demonstrate that Bitcoin is an isolated market in the contemporary causal structure.
However, there are time-variant causal relationships, especially in bearish market trends, ac-
cording to the time-lagged causality structure. In another study, Zeng et al. [37] use a vector
autoregression (VAR) model to examine the connectedness and volatility spillover effects
Algorithms 2022, 15, 428 9 of 26

relationship between Bitcoin and conventional financial assets. There is a weak relationship
between Bitcoin and financial assets; however, the correlation among cryptocurrencies is
comparably stronger. Moreover, there is a stronger connectedness via negative returns than
positive returns. The study also demonstrates that the connectedness between Bitcoin and
other financial assets varies over time. Corbet et al. [38] also use a spillover approach to
explore relationships between Bitcoin, Ripple, and Litecoin, as well as relationships with a
combination of financial assets. Volatility spillover results are categorised based on low and
high frequencies. There is little evidence of volatility spillovers between cryptocurrencies
and financial assets at short frequencies. However, there are obvious connections between
cryptocurrencies in high and low frequencies.
Furthermore, other studies investigate the correlation between the prices of cryptocur-
rencies to build better price prediction models. The study by Gkillas et al. [39] does a
pairwise comparison of ten different cryptocurrencies considering ninety combinations
of them. They apply multivariate extreme value theory and estimate a bias-corrected
extreme correlation coefficient. They find that the extreme correlation coefficient is more
observable in bear markets than bull markets among different pairs. In a negative re-
turn state, the extreme correlation coefficient generally has the highest value where one
side of the cryptocurrency pair is either Bitcoin or Litecoin. On the other hand, for pos-
itive return exceeding, either Bitcoin or Ethereum is one side of the cryptocurrency pair.
The results also demonstrates pairs including Dash on one side, which has a weak depen-
dency level with other cryptocurrencies. Stosic et al. [40] also analyse cross-correlations
between 119 cryptocurrencies. This study utilised Random matrix theory and minimum
spanning trees (MST) to develop their methods. In contrast to the common belief that
Bitcoin has a global influence on all cryptocurrencies, five communities of cryptocurrencies
are discovered by MST, indicating the existence of diverse collective behaviour between
cryptocurrencies. Bitcoin and Ethereum communities are extremely close to each other
among all communities. The Synereo community, including cryptocurrencies such as
Tether, has a strong anti-correlation. This result is aligned with the findings of [39].
Additionally, due to the high volatility in cryptocurrency prices, it is reasonable to
investigate whether volatilities are just the results of bubbles. A study by [41] examines
possible pricing bubbles in Bitcoin and Ethereum as two leading significant cryptocurren-
cies. Their findings reveal that the mining difficulty, hash rate, and cryptocurrency liquidity
of the coins are fundamental drivers of their price structures. In their results, there is no
obvious indication of bubble-based volatility in cryptocurrency prices. However, there are
bubbles related to incidents such as a Bitcoin seizure from the Silk Road website or the
declaration from the US court that Bitcoin meets an investment contract definition.

3. Artificial Intelligence and Cryptocurrencies

The term “artificial intelligence” was first used in a summer research project at Dart-
mouth College in 1956 [42]. It was originally founded as a research discipline for building
a machine to simulate every aspect of learning or any other feature of intelligence that can
be described in principle [43]. Although there is still not a universally accepted definition
for AI, as a general definition, AI leverages computers to solve problems and make de-
cisions by mimicking the human brain’s thinking ability and intelligence. AI empowers
machines to exhibit human-like behaviours, such as driving a car autonomously, improving
corporate productivity, or completing dangerous tasks [44]. Despite having several winters
as a seasonal metaphor, when technology, business, and the media paid less attention
to AI, and recent predictions about another possible winter [45], major tech companies
still prioritise AI over other IT initiatives. As a consequence, the implementation of AI
systems is expanding rapidly in a wide spectrum of domains, from health, criminal justice,
welfare, and stream history-influenced video viewing suggestions to real-time evaluation
of enormous data sets (big data) and fraud detection [46,47].
There is no clear border to distinguish different topics in AI and ML. However, it
is essential to differentiate these relevant concepts from each other for an adequate un-
Algorithms 2022, 15, 428 10 of 26

derstanding of AI. We follow the frequently used framework in categorising AI and ML

models, which considers ML as a subset of AI. Therefore, we have adopted a plausible
categorization from [48,49] to assist in explaining the models used for cryptocurrency price
prediction as shown in Figure 4. As one of the subsets of AI, ML is an umbrella term for
methods and algorithms enabling machines to discover patterns without explicit program-
ming instructions [8]. ML methods perform the experiential “learning” associated with
human intelligence and have the capability to improve their analyses via computational al-
gorithms [47]. Different ML models learn from data in different ways including supervised,
unsupervised, semi-supervised, reinforcement learning, and so forth. Supervised learning
aims to use labelled data to train algorithms for prediction or classification. On the other
hand, the purpose of unsupervised learning is to organise datasets into similar groups or
clusters. There are no labels associated with data points in unsupervised learning models.
A combination of supervised and unsupervised learning when there are only partially
labeled data, which leads to semi-supervised learning. As the next category of ML models,
RL algorithms choose an action based on previously received rewards in response to past
actions and the environmental setting, intending to find decision policies to maximise the
total reward.

Figure 4. A generic categorisation of AI concepts and their subcategories.

3.1. Application of Machine Learning in Cryptocurrency

Computer scientists, mathematicians, statisticians, and data scientists have been
developing and refining various ML algorithms to extract high-quality knowledge from
data to develop trustable and accurate price and movement predictors to perform profitable
trading in cryptocurrencies. In this section, we focus on the literature on supervised and
unsupervised learning applications in cryptocurrencies. RL in cryptocurrencies will be
discussed later in Section 3.2.2. Portfolio performance in financial markets is affected by
various factors, including the quality of input data, the data granularity, market maturity,
and forecasting models [50]. Thus, in the following, the papers reviewed in supervised and
unsupervised learning models are catalogued and summarised based on these metrics in
addition to features including baseline methods, trade frequencies, coin types, and data
sources (eight attributes in total) in Tables 4 and 5. In addition, Table 4 has been created for
the purpose of quick access to the references of the papers in Table 5.
Algorithms 2022, 15, 428 11 of 26

Table 4. Key to accessing the information about each publication in Table 5.

No Reference No Reference No Reference

1 Mittal et al. [4] 6 Chowdhury et al. [51] 11 Chen et al. [52]
2 Poongodi et al. [53] 7 McNally [54] 12 Kim et al. [55]
3 Patel et al. [28] 8 McNally [54] 13 Derbentsev et al. [34]
4 Alessandretti et al. [56] 9 Peng et al. [57] 14 Lamon et al. [58]
5 Sun et al. [59] 10 Jang and Lee [36] 15 Lahmiri and Bekiros [60]

Hybrid ML models are widely used in the cryptocurrency domain. Hybrid ML models
are constructed by combining ML methods that make use of the relative advantages of
each model on its own to enhance the overall model’s performance [61]. Patel et al. [28]
propose a hybrid model of long short-term memory (LSTM) and gated recurrent unit
(GRU) networks to predict the price of two less common altcoins, namely, Litecoin and
Monero. Although the proposed model effectively predicts daily prices, it is not effective
for longer periods, for example, a seven-day prediction. They also demonstrate that LSTM-
based methods are more suitable for short-term prediction. Furthermore, ref. [54] uses a
combination of recurrent neural networks (RNN) and LSTM methods to forecast the price of
Bitcoin, and a random forest (RF) for feature engineering. Since different exchanges provide
different prices of cryptocurrencies based on their supply and demand factors, the study
considers the average closing price of five major exchanges instead of one specific exchange.
Thus, the data are less noisy, and they lead to a more generalised model. As a result,
the LSTM performance is considerably better for long periods, such as 100 days, while it is
not the case for RNN. In another study incorporating LSTM, Alessandretti et al. [56] build an
investment portfolio system to predict the return on investment by using three supervised
learning methods, two gradient-based boosting decision trees, and one LSTM. First, two
regression models relying on XGBoost are applied to all cryptocurrencies, and then different
models separately are checked on each currency. LSTM performs best; however, gradient-
boosting decision trees allow better interpretation. Interestingly, based on their findings,
the models have a better performance when the coins’ prices are fed based on Bitcoin price
rather than their worth converted into the US dollar value.
Although cryptocurrencies are theoretically recognised as tools for economic decen-
tralisation, they rely on fiat currencies such as the US dollar to be traded. Thus, whether
traditional financial assets affect the price of cryptocurrencies is an open question. Jang
and Lee [36] implement Bayesian networks (BNs) to predict the price of Bitcoin. As an
advantage of the proposed model, blockchain information and microeconomic factors are
incorporated to forecast the price. Blockchain-related variables, ten global macroeconomic
indices, and five international exchange rates are the inputs of the BN model. Log price and
volatility of Bitcoin are compared to linear regression (LR) and support vector regression
(SVR) as baseline models. The Bayesian network model demonstrates better predictive
performance. Furthermore, their investigation shows that the price of Bitcoin is correlated
to factors including macroeconomic variables such as stock indices, exchange rates, and oil
prices. In addition, another study by Sun et al. [59] obtains forty features from around
ten indices to explain the changes in cryptocurrency markets. In the paper, 42 different
types of cryptocurrencies are tested by a novel algorithm called a light gradient boosting
machine (LightGBM). The results demonstrate that the model is more suitable for a two-
week prediction. Furthermore, the algorithm demonstrates improved forecasting for the
top 10 cryptocurrencies in more mature markets. However, they analyse a relatively small
amount of data, accounting for seven months, though a dataset with a more extended
period is required for generalising the model.
Algorithms 2022, 15, 428 12 of 26

Table 5. Detailed information on supervised and unsupervised ML models used in the literature covered in this survey. For each study, the method used for the
price prediction task, the baseline model for comparing approaches (if exists), in addition to the target feature, and the specification of data is provided.

No Method Baseline(s) Prediction Feature Frequency Prediction Performance Metric(s) Data Period Crypto(s) Data Source(s)
1 Multivariate LR - Highest price 1D F-score - 10 coins kaggle.com
Cost function accuracy
2 LR and SVM - Price 1H - Ethereum etherchain.org
score
MAE, MSE, MAPE,
3 LSTM and GRU LSTM Price 1D, 3D, 7D 2016–2020 Litecoin Monero investing.com
and RMSE
Simple moving Geometric mean return
4 LSTM, regression Cumulative Return 3D, 5D, 7D, and 10D 2015–2018 10 coins coinmarketcap.com
average strategy Sharpe ratio
5 LightGBM SVM and RF Price 2D, 2W, and 2M AUC indicator 2018 42 coins investing.com
Gboosted trees, NNs, RMSE and squared coinmarketcap.com
6 LSTM, RNN, and ARIMA Closing price 1D 2016–2019 9 coins and cci30
and K-NN correlation cci30.com
coinmarketcap.com
LDA, LR, RF, XGB, QDA, Precision, accuracy, recall,
7 - Price 1D and 5 min 2017–2019 Bitcoin Bitcoinity.org
SVM, and LSTM and F1-score
blockchain.com
Sensitivity, specificity,
8 RF, LSTM, and RNN ARIMA Price 1D precision, accuracy, 2013–2016 Bitcoin Coindesk Blockchain.info
and RMSE
Bitcoin, Ethereum, alt19.com
9 SVR-GARCH, and SVR Price 1H and 1D RMSE and MAE 2016–2017
and DashCoin fxhistoricaldata.com
10 BNN SVR, and LR Price 1D RMSE and MAPE 2011–2017 Bitcoin Bitcoincharts.com
Twitter’s API,
11 VADER - Price 1D Pearson R and p-value 2018 Bitcoin and Ethereum
Google Trends
Accuracy rate, F-measure, Coindesk
Bitcoin, Ethereum,
12 AODE - Price - and Matthews correlation 2013–2015 CoinMarketCap
and Ripple
coefficient Etherscan RippleCharts
5D, 10D, 14D, 21D, Bitcoin, Ethereum,
13 BART ARIMA and ARFIMA Price RMSE 2017–2019 Yahoo Finance
and 30D and Ripple
Logistic Regression, SVM, Confusion matrix Bitcoin, Ethereum, Kaggle.com,
14 - Price 1D 2017
and Naive Bayes accuracy and Litecoin Twitter’s API
SVR, GRP, RT, kNN,
15 FFNN, BRNN, - Price 5 min RMSE 2016–2018 Bitcoin -
and RBFNN
Algorithms 2022, 15, 428 13 of 26

Linear regression models are widely used as an approximation for real-world models
and have commonly been the most popular in the past. However, real-world phenomena
tend to have quite a lot of nonlinear components, and these methods fail to efficiently fit
complex datasets where other models tend to be more effective. Mittal et al. [4] implement
a multivariate linear regression to forecast the prices of ten different cryptocurrencies.
The high price on a given day is predicted based on the open price, close price, and low
price values of previous days. In the same stream of research, Poongodi et al [53] obtain
the prediction of the price of Ethereum by combining Linear regression and support vector
machine (SVM) models. The results reveal that SVM without additional features has about
ten percent higher accuracy than linear regression.
Lahmiri and Bekiros [60] examine seven ML models for Bitcoin intra-day price predic-
tion. The models are categorised into three distinct groups of algorithms, including a sta-
tistical ML category such as support vector regression, algorithmic techniques (regression
trees), and AI-based neural network topologies. Aside from radial basis function networks
(RBFNN), all neural networks outperform other algorithms based on the RMSE metric.
Due to the nature of cryptocurrencies, deciding the best approach to model their behaviour
is an open question. Thus, in another SVR-based study, Peng et al. [57] compare mod-
els in two distinct financial markets of cryptocurrencies and three major fiat currencies.
The results are presented as model-based comparisons and financial market-based com-
parisons. It is understood from the findings that forecasting errors of traditional GARCH
family models and SVR models are higher for cryptocurrencies than fiat currencies. More-
over, low-frequency windows have higher error metrics than high-frequency windows in
both markets.
Since a large part of the behavioural finance research emphasises that investors’ emo-
tions have significant effects on financial decisions, researchers pay more attention to
analysing investor sentiments [62]. Like traditional financial markets, there appears to be
a correlation between media sentiment and cryptocurrency price movements. Therefore,
one research trend is to employ sentiment analysis in cryptocurrency price prediction.
Abraham et al. [3] examine three sources of data to perform sentiment analysis for pre-
dicting the price of Bitcoin and Ethereum. Based on the results, although Google trends
and tweet volumes are highly correlated with the price of cryptocurrencies, the sentiment
of tweets is determined as an unreliable source of prediction. The sentiment of tweets
is inclined to keep staying positive even when the market trend is falling. In another
study on sentiment analysis, Amon et al. [58] use daily news headlines and tweets for
price fluctuations and implement four classification algorithms for a binary classification
problem. While logistic regression is the best classifier for Bitcoin and Litecoin, a version
of the Naive Bayes classifier performs the best prediction for Ethereum. Although the
proposed model identifies the general trends, it cannot anticipate the price fluctuations
when they are not in line with the general market trend. Kim et al. [55] study the transaction
number and information of online communities of cryptocurrencies by using averaged
one-dependence (AODE) and user comments and replies for price predictions. The results
generally vary across different cryptocurrencies. Both the Bitcoin price and the transaction
number demonstrate a significant positive association with user replies. It is consistent
with the outcomes of [52], where there are tweets with positive sentiment. However, unlike
Bitcoin, Ethereum, and Ripple have a negative correlation with very negative comments.
A unique aspect of financial assets is that several external factors may seriously influ-
ence their price movements. These factors may include macroeconomic factors, a trader’s
psychology, fundamental elements, market sentiment, and so forth [63,64]. It is hence a
considerable challenge to discover all influencing factors and extract intelligent trading
rules from the dynamic financial markets environment [65,66]. Therefore, features con-
taining valuable information are necessary to develop decision-support trading systems
for cryptocurrency price prediction in financial markets. This highlights the vital role
of feature engineering and selection in cryptocurrency markets [63]. There is a trade-off
between a higher number of features, which increases the training time, and having a
Algorithms 2022, 15, 428 14 of 26

lower number of features, which decreases model performance. Chen et al. [52] introduce
a different perspective regarding data granularity and feature engineering. Features of
low and high frequencies are separately selected, and distinct methods are implemented
on low and high frequencies data. The results suggest statistical methods such as logistic
regression perform better on low-frequency data while other ML approaches including
linear discriminant analysis and SVM accomplish better outcomes for high-frequency data.
In addition, the study introduces two sentimental features, Baidu media search volume,
and Google trend search, combined with other features as another novelty.
Further ML models have been used to investigate different aspects of cryptocurrencies.
Chowdhury et al. [51] apply four ML models for forecasting the closing price of nine differ-
ent cryptocurrencies. The prediction of cryptocurrencies index 30 (cci30) (www.cci30.com,
accessed on 1 November 2022), which measures the overall growth of cryptocurrency
markets, and the implementation of models on the RapidMiner platform, are two distinct
characteristics of this research. A comparison of algorithms shows that a K-nearest neigh-
bours (K-NN) model performs poorly in prediction. In another study, Erbentsev et al. [34]
adopt a modified model of a binary auto-regressive tree (BART) from standard models
of regression trees. The proposed model combines different components of classification,
regression trees, and autoregressive models ARIMA. For three major cryptocurrency lead-
ers, the algorithm is tested based on the dynamics of the cryptocurrencies, namely, stable
period, falling trend, transition dynamics (change of trend), and rising trend. The model
demonstrates a better performance than the baseline models in both falling and rising
trends; however, its forecast is worse in periods of rapid trend changes.

3.2. Reinforcement Learning

One initial goal of AI is to create a fully autonomous agent capable of interacting with
an environment for learning optimal behaviours that improve over time considering a
particular objective [67]. To respond to this goal, an RL agent interacts with an environment
over time by following a procedure. At each time step, the agent is in a state (from a state
space) and then selects an action (from an action space belonging to the state). The agent
follows a policy (the agent’s behaviour) for all these state-action-state procedures and
receives a scalar reward [68]. This configuration of moving from one state to another
by selecting actions and receiving rewards over time makes RL particularly effective in
achieving the initial goal of the AI field. Figure 5 shows a generic framework for an
agent–environment interactions. At any time t, an agent is in a state St , chooses an action
At , and receives an reward of Rt . This action takes the agent to another state St+1 . All these
states, actions, and rewards are defined inside an environment.

Figure 5. Classic RL cycle.

RL follows a different paradigm in terms of learning in comparison to supervised

and unsupervised learning in ML. Basically, it is a principled mathematical framework
of experience-driven autonomous learning [69]. RL cannot be classified as a supervised
learning approach since labeled data are not provided to an agent. Furthermore, RL
methods are suitable for problems that have sequential dynamics and involve optimisation
of a scalar performance objective, while supervised learning methods are usually applied
to problems that involve static input-output mappings and minimise a mismatch between
the data and the model [70]. On the other hand, it is not also an unsupervised learning
Algorithms 2022, 15, 428 15 of 26

approach, since the agent is provided with information about a rewarding scheme that
guides the algorithm through each state-action-state iteration [71].
RL has been around for more than two decades; however, two recent breakthroughs in
the last seven years have significantly changed the course of this approach by going beyond
theoretical concepts via successful implementation of RL algorithms across a range of
problems [72]. First, in 2015, an agent on classic Atari 2600 games surpassed a professional
human game tester and all previous algorithms [73]. Later, in 2016, AlphaGo achieved a
99.8% winning rate against all other computer programs in the game Go and could defeat
a human professional player for the first time [74]. Despite these successful instances of
incorporating RL in problem-solving, there still are several unexplored areas that warrant
investigation in using this framework to solve challenging real-world problems.

3.2.1. An Overview on Reinforcement Learning Overview

The fundamental nature of RL is to learn through a process of trial and error. This
means that an RL agent has continuous interactions with its environment, and it learns to
modify its behaviour based on rewards received after taking actions [67]. RL can be consid-
ered as creating an environment for a learner to learn by interacting with the environment
and choosing actions without being told. The learner chooses actions in an environment
based on the consequences of prior actions referred to as rewards. The agent modifies its
behaviour to choose better actions in order to maximise cumulative rewards. To this end,
RL has two distinctive features. One is the trial-and-error search to optimise cumulative
rewards. The second is the delayed reward which implies that actions may not only have
an immediate reward, but they may affect all subsequent rewards or may have rewards in
the later stages of learning [14,75].
An RL problem can be modeled and described as a Markov decision process (MDP)
by a five-tuple (S , A, π, R, λ). An RL algorithm is composed of the following:
States and observations S : A state is a complete description of the circumstances of
an environment, and an agent obtains all the information regarding the environment
through a state. Additionally, observation is considered in case some information
might be omitted due to a partial representation of a state.
Action spaces A: Each environment allows performing several actions. The set of
all legitimate actions in a given environment is called an action space. Unlike a
continuous action space, an agent has a finite number of available actions in a discrete
action space.
Policy π: A policy π is the way an agent behaves at a given time. It determines the
action that has to be chosen by the agent when it is in a particular state. In a proba-
bilistic setting, it maps the current state of the environment into a set of probabilities
for taking actions from the action space.
Rewards R: A reward component is an important concept in RL. It is an immediate
or instantaneous gain that an agent receives when choosing an action in the current
state to move to the next state.
Discount factor λ: The quantity λ ∈ [0, 1) is the discount factor and generates dis-
counted rewards to prevent infinite cumulative rewards when running for a long
period. As a general intuition, discounted rewards mean rewards today are worth
more than rewards tomorrow. If it is zero, an agent considers only immediate rewards;
while λ closer to one means that the agent evaluates its actions based on cumulative
rewards in the future.

Figure 6 is one possible taxonomy of RL algorithms produced partially based on work

by Zhang and Yu [76]. A short description of each component is provided in the following.
Algorithms 2022, 15, 428 16 of 26

Figure 6. The taxonomy of RL algorithms.

Model-Free and Model-Based Approaches

Based on the proposed taxonomy in Figure 6, two computational strategies consti-
tute reinforcement learning, namely, model-free (MF) and model-based (MB) approaches.
In a model-based approach, rewards for taking actions come from internal models of
the environment, whereas in a model-free approach a value is associated with actions
to maximise the rewards. Additionally, model-free approaches require access to a large
number of interactions for the purposes of training, while model-based methods are ca-
pable of achieving an optimal policy quickly based on what they have learned about the
environment [77]. A model of environment allows an agent to predict future states, and a
model-based strategy helps to enhance the efficiency of RL substantially [78]. Additionally,
model-free approaches are more suitable for learning complex tasks, such as games.

Value-Based Methods
In an RL setting, an agent receives a reward signal based on choosing an action to move
from one state to another. The agent maximises the total received reward (or discounted
total reward in an RL setting with an infinite horizon) [75]. The discounted total reward,
also called the value, in a state that is gathered through state-action-state iterations in RL,
plays a central role in guiding the algorithm towards an optimal policy. These values are
obtained dynamically using the Bellman equation. Moreover, a key concept in RL is the
trade-off between exploration and exploitation performed by an agent in searching among
states and actions considering a non-optimal policy learned so far (on-policy scheme) or
without a particular policy (off-policy scheme). An agent should not only learn from
previous experiences in earning rewards, but it has to also explore new unattended states
and actions to ensure improvements in selecting actions in the future. To this end, on-policy
methods seek to improve the policy previously used in decision making while off-policy
methods evaluate a policy different from what was used to generate data [75]. In addition,
off-policy methods are more sample efficient, meaning that they need comparably lees
samples to reach a certain level of performance than on-policy methods because they can
learn from any trajectory sampled from the same environment. However, not having
stable interactions with the value functions or their approximations is a fundamental
shortcoming [79]. Moreover, Off-policy methods learn from previously collected data with
no interaction. Thus, it is possible in principle to leverage massive data. Nevertheless,
Off-policy RL algorithms have significant technical obstacles in practice, arising from
Algorithms 2022, 15, 428 17 of 26

the distributional shift between the policy deducted from collected data and the learned
policy [80]. SARSA and Q-Learning are the two most common value-based methods, that
use off-policy and on-policy schemes, respectively.

Policy-Based Methods
In a value-based approach, an agent learns an optimal policy based on its interac-
tions with the environment in terms of the estimates of an optimal action-value function.
In contrast, a policy-based method learns an optimal policy without using a value function
and is only based on its interactions with the environment. In particular, policy-based
methods incorporating gradients (PG) compete with value-based techniques in a discrete
environment and are successfully implemented in continuous control (as opposed to a
discrete action space, in which agents decide which distinct action to select from a set of
actions, in continuous action space, actions are defined as real-valued vectors [81]). The
efficiency of value-based methods is limited in continuous control settings due to the dimen-
sionality problem; however, policy-based methods are more effective in continuous space
by directly learning a policy distribution [82,83]. Instability and requiring small learning
rates for training are two problems of standard PG methods, given that the strategy of
policy update is simple. Trust-region policy-based (TRPO) schemes address the problems
by bounding policy updates to a trust region to optimise standard PG algorithms [83,84].
Deep deterministic policy gradients (DDPG) and proximal policy optimisation (PPO) are
other popular PG algorithms.

3.2.2. Reinforcement Learning Applications in Cryptocurrency Markets

There are a wide range of applications of RL in different industries from games [85]
and robotics [86] to natural language processing [87] and computer vision [88]. RL is also
a natural solution for several problems in finance and economics such as option pricing,
multi-period portfolio optimisation, and risk management due to their dynamic nature.
The focus of this section is on the application of RL in dealing with better investment
policies in cryptocurrency markets.
There are two main streams of literature for using RL models in cryptocurrency
markets. Firstly, there is growing interest in explaining and predicting cryptocurrency
price behaviours and trading strategies [89]. This stream describes that cryptocurrencies
consist of distinct governing factors which distinguish them from traditional financial
assets. However, it focuses on the price volatility of the cryptocurrency assets, and there is
no systematic analysis of cryptocurrency trading [90]. The second line of research covers
economic applications for developing automated trading systems [89]. Since it is in its
early stages of development, most studies may not provide practical solutions and a
competitive edge for investors [90,91]. In addition, creating better market simulations by
including transaction fees and applying feature engineering to cryptocurrency trading data
are investigated in the literature.
To mimic a real cryptocurrency market, Sadighian [92] designs a new framework,
called deep RL applied to market-making, that uses advantage actor-critic and proximal
policy optimisation RL algorithms. These two policy-based methods constructed upon
gradient are trained by Bitcoin, Ether, and Litecoin data. Positional profit-and-loss and the
trade completion ratio are defined as the reward function, and the average daily returns
are compared between currencies considering total accumulated rewards. BTC has the
highest return on investment among all cryptocurrencies investigated in the study. In this
work, limit order books, order flow imbalances, and trade flow imbalances are used as the
environment state space, which is the novelty of the study. However, the short time-frame,
two days, could be increased to achieve improved generalisation of the proposed approach.
In addition, Koker and Koutmos [89] employ an active trading strategy using deep RL to
achieve the greatest risk-adjusted returns. To this end, a gradient ascent algorithm optimises
the Sortino ratio as the reward function. Five portfolio performance metrics, including
cumulative returns, Sharpe and Sortino ratios, maximum drawdown, and value-at-risk
Algorithms 2022, 15, 428 18 of 26

are compared with the buy-and-hold trading strategy as the baseline model. The results
indicate that the proposed model outperforms the buy-hold strategy in all tested five
cryptocurrencies. On a similar topic of extracting trade ideas, Sattarov et al. [93] propose
a four-layer architecture deep RL for recommending trading suggestions to maximise
short-term profit. The reward function is based on the difference between the selling price
and the purchase price. Additionally, as its unique feature, the agent receives negative
rewards if the number of sequential purchases is more than a limited number (20 in the
study). Thus, it prevents the agent from many open positions to increase the reward.
Double-cross strategy, swing trading, and scalping trading strategies are compared with
the proposed model based on the number of trading actions and trading quality. Deep
RL finds more trading opportunities and hence performs a higher number of trades than
other trading strategies. In addition, the net return is 14.4% for Bitcoin within one month of
simulated trading, while the scalping strategy has only 6.1% growth as the best performer
of traditional strategies.
The study by Lee et al. [94] proposes the integration of inverse RL and agent-based
modelling for Bitcoin market movements. Inverse RL assumes that the reward of an agent
is unknown and tries to infer the reward function in the presence of optimal policy or
observed behaviour. The RL model proposed by the study has an environment whose
states are defined by price, value possessed, the value realised of Bitcoin, and the difference
between its price and a moving average of the price. A main advantage of the proposed
model is that it considers interactions between market participants. Another variant of
deep RL, called deep double Q-networks, and proximal policy optimisation are utilised to
solve the problem of optimising the placement of limit orders by [95]. Designing a relatively
comprehensive state-space RL environment, by including a wide range of financial aspects,
is the main advantage of the study. To accomplish the task, market states are divided into
four main sections, including transaction imbalances, best-order volumes and imbalances,
volatility and current price drift, and current cost of liquidity. The results indicate that the
agent guided by the proximal policy optimisation for order placement is more aggressive
against unfavourable price moves. The study also concludes that queue imbalances and
volumes are the most important features impacting the behaviour of the agents.
A multi-agent RL method is proposed by Lucarelli and Borrotti [96] to develop a
deep Q-learning portfolio management framework. A local agent guided by three learning
techniques, namely, deep Q-Networks, double deep Q-Networks, and duelling double deep
Q-Networks is evaluated based on two global reward functions. The reward functions are
the sum of the local rewards, the weighted sum of the Sharpe ratio of the portfolio, and the
net portfolio return. As a result, the framework outperforms weighted portfolio and portfo-
lio selection using a genetic algorithm as the baseline. In another study, a state-augmented
RL framework for managing portfolios on two well-known datasets, namely, the Bitcoin
market and the HighTech stock market, is proposed by Ye et al. [97]. The proposed port-
folio management algorithm performs better in terms of accumulated and risk-adjusted
profits in comparison to the standard RL-based portfolio management. Jiang and Liang [98]
implement a deterministic deep RL method and convolutional neural network (CNN)
for portfolio management. Weight allocation and adjustment are performed by using a
Monte Carlo policy gradient method for CNN training. The CNN agent outperforms three
benchmark strategies and two out of three portfolio management algorithms.
One unrealistic assumption in many trading algorithms is that they neglect trading
fees. In particular, in automated systems that perform many trades in a short amount of
time, the accumulated fees, even negligible for a single trade, are considerable and can affect
the optimal policy over a long time horizon. Hegazy and Mumford [99] examine seven
algorithms, including an altered recurrent RL (RRL), for studying Bitcoin trading strategies.
To evaluate the performance of algorithms, they define several metrics including cumulative
weighted confidence without a trading fee, profit considering only 0.25% of transaction
fees, and finally, incorporating the correct transaction rate for the task of predicting the
Bitcoin price. RRL generates the best results for both with or without fee considerations. It
Algorithms 2022, 15, 428 19 of 26

likely demonstrates the way RL algorithms incorporate fees into the process of optimising
trading strategies rather than adding only during trading. In addition, Betancourt and
Chen [100] propose an RL model that takes into account the same assumptions as [99],
which are similar to real markets. Their RL model utilises proximal policy optimisation
that considers several assets in cryptocurrency markets and adopts the number of assets
that may vary over time. In this study, the two agents evaluate the total return and the
differential Sharpe ratio metrics. The model performs competitively in comparison to
temporal-difference learning, convolutional neural network, and double Q-networks.
The majority of existing RL studies either use all the features in historical trade data,
including open price, high price, low price, close price, volume, and so forth, or a row
combination of them, while proper feature engineering is not often emphasised in those
studies. Feature engineering focuses on obtaining significant factors by using feature
selection methods [101]. In some studies, feature engineering is treated separately from
the then task of learning. For example, in [95], a neural network architecture is used to
directly learn significant features for cybersecurity attack detection, instead of handcrafted
approaches. In another study for dealing with feature engineering in the realm of cryp-
tocurrencies, Weng et al. [102] propose a portfolio trading system with a deep RL model
using a multidimensional attention-gating mechanism for twenty cryptocurrencies. In this
study, XGBoost performs for quantifying the importance of features before feeding data
into the RL network. Among all available features, close, high, and low prices have the
highest importance, respectively, as determined by XGBoost.

4. Discussion and Potential Future Research

As evidenced in the discussion so far, AI studies in cryptocurrency markets have been
growing, especially during the last few years. However, with all advancements in both
science and technology, several unexplored topics still warrant further investigation in the
cryptocurrency domain. For this reason, this section discusses potential research directions
based on the literature reviews in Sections 3.1 and 3.2.2.

4.1. Integration within Cryptocurrencies or with Other Financial Assets

Recent fundamental changes in cryptocurrency markets indicate that they are in-
creasingly integrated with other traditional financial and economic systems. For instance,
Coinbase, one of the most popular cryptocurrency exchanges, went public on the NASDAQ
exchange, and there are an increasing number of ways in which cryptocurrencies can be
used to pay for services and products. This evidence strongly indicates the acceptance of
cryptocurrencies by existing financial bodies as well as the general public. Since financial
and economic systems continuously and dynamically evolve, investigating the integration
of cryptocurrency markets with other financial assets can be of significant benefit. Inter-
nationally, it is important to study the financial impacts of the price of traditional assets
and commodities on the price of cryptocurrencies, and vice versa. In a similar fashion,
the question of whether or not the prices of cryptocurrencies can be predicted based on
changes in the prices of other financial assets (e.g., oil and the US dollar). Investigating
the influence of financial markets and political sentiments on the price of cryptocurrencies
and whether or not cryptocurrencies can influence traditional assets such as oil prices, is of
significant interest.
In addition, the influence of different cryptocurrencies on each other warrants further
attention and investigation. In particular, the emergence of government-based cryptocur-
rencies, governmental restrictions on trading and using cryptocurrencies in some countries
(e.g., China), and preferences regarding certain coins based on their liquidity ease, may
create an influence from one coin to another. Furthermore, creating a profitable portfolio of
different cryptocurrencies needs a better understanding of the interactions and influences
between the coins. The necessity of attending to these gaps is evident in the recent changes
in the cryptocurrency markets. For example, the market capitalisation of altcoins has in-
creased substantially in the last three years, where it has surged from 500 million dollars in
Algorithms 2022, 15, 428 20 of 26

January 2019 to 1.5 trillion dollars in December 2021 (www.coinmarketcap.com). Therefore,

the question of whether leading cryptocurrencies such as Bitcoin and Ethereum influence
the prices of other altcoins needs to be studied. Modelling approaches such as Bayesian
networks [103] can be used to gain insights by modelling causal relationships between
cryptocurrencies and other financial systems, and among cryptocurrencies themselves.
Although studies have been undertaken to investigate integrating cryptocurrencies
within other markets, they are mainly based on financial and economic models, such as
a gold standard model applied in [1] to study the price of Bitcoin, or in another study
GARCH models (statistical models) are widely used in the financial sector. Moreover,
there are other gaps in the literature that need further attention. One impending issue is
that of dealing with big data, where a sound approach from AI can be of great benefit.
Additionally, statistical models used to date place a strong emphasis on explainability
rather than accuracy. Therefore, selecting an appropriate AI framework that is accurate
and explainable considering feature engineering, fitting, resampling, and testing is crucial.
Furthermore, designing trustable AI tools that are capable of predicting market movements,
price changes, and profitable trade strategies is vitally important. In particular, such a tool
should be able to take as input different types of data including price histories, information
from social media, news items from international political interactions, governmental
decisions from economically influential countries, and expert opinion.

4.2. Macroeconomic Factors as the States in RL

Reinforcement learning appears to be an appropriate tool founded within AI to deal
with open problems within cryptocurrency and its price prediction. In particular, RL can
be a useful tool in dealing with unknown factors of cryptocurrency price fluctuations. This
capability of RL eventually leads to detecting optimal policy for designing automated
trading systems. The tool needs to be able to deal with feature engineering, analysing
correlational and causal relationships among historical price data of different assets, reliable
price prediction capability, social media analysis, and expert opinion. The main challenge
in developing such a tool is the definition of the fundamental elements of an RL system
(discussed in Section 3.2.1). For example, the connection between macroeconomic factors
and cryptocurrencies is studied in supervised and unsupervised learning techniques used in
prediction tasks. However, it is not straightforward to use these factors in an RL framework.
Thus, designing an RL environment whose states contain macroeconomic information can
be beneficial as a potential future research direction. In addition, the definition of action
space, reward function, and the structure of a policy, and the discounting factor in dealing
with challenges in cryptocurrency markets needs further attention and investigation.

4.3. Minor Challenges Related to Cryptocurrencies

There are further research directions that can be considered in the area of using AI
techniques in predicting the behaviour of cryptocurrencies. For example, the COVID-19
pandemic has influenced our daily lives in many ways, from working habits to public
transport, and the economy. There are studies that have been conducted concerning
the effects of the COVID-19 outbreak on the cryptocurrency market, such as [104,105],
which mainly focus on financial models. It is worthwhile to investigate the changes in the
parameters of AI models that are fitted using the data before and after the pandemic due
to the different characteristics in data including changes in the mean and variance. This
comparison and analysis can be used to develop strategies and create policies is situations
in the future, when events similar to the COVID-19 pandemic may occur. Moreover, even
though cryptocurrencies are designed as a tool of decentralisation, their price historically
has demonstrated sensitivity toward governmental laws such as banning cryptocurrency
mining in some countries such as China. On the other hand, ML techniques are increasingly
being used across different fields, for example, in law and the legal domain in general [106].
Thus, the application of ML models to investigate the effects of laws and regulations on
cryptocurrency price prediction can be an area of potential future research. In addition,
Algorithms 2022, 15, 428 21 of 26

studying the time granularity of time series data used in training models can impact the
accuracy of predictions, and hence, methods for data prepossessing need to be considered
carefully to determine an appropriate time granularity.

4.4. Sentiment Analysis

Behavioural finance research increasingly emphasises the point that investors’ emo-
tions significantly affect financial decisions. A number of researchers have devoted their
attention to investor sentiment [62] by considering their activities in social media. There
are several supervised and unsupervised ML studies that combine sentiment indicators
with some financial features. However, RL studies mainly focus on sentiment analysis as a
standalone market input to predict prices. Hence, it is worthwhile studying sentimental
factors along with other financial inputs in an RL environment. As an example of the
impact of sentiment analysis and social influencers, Figure 7 shows the volatility in Bitcoin
prices due to two instances of Elon Musk’s comments regarding Bitcoin. It is clear in the
figure that large green (increase in price) and large red (decrease in price) candlesticks,
for positive and negative comments of Elon Musk on Twitter, respectively, are due to
the influence traders accept based on the comments of influencers. Similar diagrams can
be extracted after major geo-political events and the emotional influence they can create
on the trading behaviour of investors. Thus, investigating the effects of social media on
overall market sentiment combined with other data sources can reveal important drivers in
cryptocurrency price formation and prediction.

Figure 7. A chart of Bitcoin focusing on volatilities impacted by social media. Green and red candles in
the chart occurred directly after Elon Musk’s positive and negative tweets about Bitcoin, respectively.

4.5. Further Attention to Altcoins

Bitcoin, as the first cryptocurrency, is studied and referenced more than other coins.
For example, Figure 8 presents Google books Ngram viewer for comparing Bitcoin (in
blue) and altcoins (in green) between 2010 and 2019. In addition, the red curve shows
the number of appearances of the word cryptocurrencies in the Google books Ngram
viewer to provide a better context for comparison. As observed in the figure, Bitcoin has
generally been referenced more than altcoins, demonstrating the potential unexplored areas
in altcoins. Additionally, altcoins with enhanced features might be good candidates to
diversify investment portfolios and minimise the total risk of investment [107]. Moreover,
with the increasing share of the market by altcoins, there is an opportunity to investigate
influential factors in the price prediction of the cryptocurrencies among themselves by
Algorithms 2022, 15, 428 22 of 26

removing the volatility pertaining to Bitcoin. It might assist in understanding the whole
cryptocurrency ecosystem.

Figure 8. Altcoin and Bitcoin Ngram comparison.

4.6. Extreme Condition Detection in Cryptocurrencies

The past financial market turmoil, such as the 2007–2008 Global Financial Crisis or
March 2020 market turmoil caused by a global health crisis, has shown that rare extreme
market events can have severe consequences and spillover effects for world economies [108].
Therefore, it is important for financial risk management portfolio diversification to control
and monitor extreme downside market risk [109]. As a relatively new financial sector,
the cryptocurrency market witnessed various extreme market movements, leading to sub-
stantial capital losses for investors. Hence, detecting extreme market conditions in the
cryptocurrency market can prevent participants from investment loss. Although consider-
able research has been conducted to deal with these events, anomaly detection in economic
data has been widely ignored [110]. Therefore, detecting anomalies in the cryptocurrency
market by implementing AI methods is another area of future research that will eventually
assist investors in making informed decisions that minimise risk. The methods, such as the
local outlier factor, autoencoders, and Bayesian networks, can serve as anomaly detection
techniques for this purpose.

5. Conclusions
This paper first uses a cross-disciplinary approach to discuss the price determinants
of cryptocurrencies from a financial and economic perspective. Then, recent studies on
the use of various AI models in cryptocurrency price prediction are reviewed through a
comparative survey. Recent studies on the use of various AI models in cryptocurrency
price prediction are reviewed through a comparative survey. As a relatively new field
in finance and economics, many open issues still warrant further investigation in the
area of cryptocurrencies, and in particular the use of advanced AI methods to achieve
accurate prediction of their prices. More specifically, further attention is needed concerning
integrating cryptocurrencies with traditional markets and analysing two-way influence in
terms of correlational and causal relationships. Meanwhile, AI is capable of addressing
these open problems, and future work in this field has to consider different components and
tools to cope with the data engineering aspects of a prediction tool for cryptocurrencies. It
is expected that this survey serves as a guideline for researchers to comprehend the current
state of the literature, and we discuss open challenges that may be considered as part of
future research. The potential of AI in the cryptocurrency domain is evident, and research
in this direction can lead to great benefits for cryptocurrency market participants and policy-
and decision-makers.
Algorithms 2022, 15, 428 23 of 26

Author Contributions: Conceptualization, Investigation, Visualization, Writing—original draft, and

Writing—review & editing, R.A.; Conceptualization, Investigation, Visualization, Writing—original
draft, Writing—review & editing, and supervision, A.N. and D.T. All authors have read and agreed
to the published version of the manuscript.
Acknowledgments: The authors would like to acknowledge Mong Shan Ee (Deakin Business School,
Deakin University) for her guidance regarding cryptocurrencies, and for reviewing the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Ciaian, P.; Rajcaniova, M.; Kancs, d. The economics of BitCoin price formation. Appl. Econ. 2016, 48, 1799–1815. [CrossRef]
2. Ji, Q.; Bouri, E.; Gupta, R.; Roubaud, D. Network causality structures among Bitcoin and other financial assets: A directed acyclic
graph approach. Q. Rev. Econ. Financ. 2018, 70, 203–213. [CrossRef]
3. Abraham, J.; Higdon, D.; Nelson, J.; Ibarra, J. Cryptocurrency price prediction using tweet volumes and sentiment analysis. SMU
Data Sci. Rev. 2018, 1, 1.
4. Mittal, R.; Arora, S.; Bhatia, M. Automated cryptocurrencies prices prediction using machine learning. Div. Comput. Eng. Netaji
Subhas Inst. Technol. India 2018, 8, 2229–6956.
5. Wu, X.; Chen, H.; Wang, J.; Troiano, L.; Loia, V.; Fujita, H. Adaptive stock trading strategies with deep reinforcement learning
methods. Inf. Sci. 2020, 538, 142–158. [CrossRef]
6. Dai, S.; Wu, X.; Pei, M.; Du, Z. Big data framework for quantitative trading system. J. Shanghai Jiaotong Univ. (Sci.) 2017,
22, 193–197. [CrossRef]
7. Huang, B.; Huan, Y.; Xu, L.D.; Zheng, L.; Zou, Z. Automated trading systems statistical and machine learning methods and
hardware implementation: A survey. Enterp. Inf. Syst. 2019, 13, 132–144. [CrossRef]
8. Rasekhschaffe, K.C.; Jones, R.C. Machine learning for stock selection. Financ. Anal. J. 2019, 75, 70–88. [CrossRef]
9. Lee, D.; Deng, R.H. Handbook of Blockchain, Digital Finance, and Inclusion: Cryptocurrency, FinTech, InsurTech, Regulation, ChinaTech,
Mobile Security, and Distributed Ledger; Academic Press: Cambridge, MA, USA, 2017.
10. Leong, K.; Sung, A. FinTech (Financial Technology): What is it and how to use technologies to create business value in fintech
way? Int. J. Innov. Manag. Technol. 2018, 9, 74–78. [CrossRef]
11. Stulz, R.M. Fintech, bigtech, and the future of banks. J. Appl. Corp. Financ. 2019, 31, 86–97. [CrossRef]
12. Strobel, V. Pold87/academic-keyword-occurrence: First release. Zenodo 2018. doi: 10.5281/zenodo.1218409. [CrossRef]
13. Fang, F.; Ventre, C.; Basios, M.; Kanthan, L.; Martinez-Rego, D.; Wu, F.; Li, L. Cryptocurrency trading: A comprehensive survey.
Financ. Innov. 2022, 8, 13. [CrossRef]
14. Mosavi, A.; Faghan, Y.; Ghamisi, P.; Duan, P.; Ardabili, S.F.; Salwana, E.; Band, S.S. Comprehensive review of deep reinforcement
learning methods and applications in economics. Mathematics 2020, 8, 1640. [CrossRef]
15. Sabry, F.; Labda, W.; Erbad, A.; Malluhi, Q. Cryptocurrencies and Artificial Intelligence: Challenges and Opportunities. IEEE
Access 2020, 8, 175840–175858. [CrossRef]
16. Murat Ozbayoglu, A.; Ugur Gudelek, M.; Berat Sezer, O. Deep Learning for Financial Applications: A Survey. arXiv 2020,
arXiv:2002.05786.
17. Al-Jaroodi, J.; Mohamed, N. Blockchain in industries: A survey. IEEE Access 2019, 7, 36500–36515. [CrossRef]
18. Yli-Huumo, J.; Ko, D.; Choi, S.; Park, S.; Smolander, K. Where is current research on blockchain technology?—A systematic
review. PLoS ONE 2016, 11, e0163477. [CrossRef]
19. de Leon, D.C.; Stalick, A.Q.; Jillepalli, A.A.; Haney, M.A.; Sheldon, F.T. Blockchain: Properties and misconceptions. Asia Pac. J.
Innov. Entrep. 2017, 11, 286–300. [CrossRef]
20. Nakamoto, S.; Bitcoin, A. A peer-to-peer electronic cash system. Bitcoin 2008, 4. Available online: https://bitcoin.org/bitcoin.pdf
(accessed on 1 November 2022).
21. Ammous, S. Can cryptocurrencies fulfil the functions of money? Q. Rev. Econ. Financ. 2018, 70, 38–51. [CrossRef]
22. Nabilou, H. How to regulate bitcoin? Decentralized regulation for a decentralized cryptocurrency. Int. J. Law Inf. Technol. 2019,
27, 266–291. [CrossRef]
23. Chen, Y.; Bellavitis, C. Decentralized finance: Blockchain technology and the quest for an open financial system. SSRN J. 2019.
[CrossRef]
24. Lansky, J. Possible state approaches to cryptocurrencies. J. Syst. Integr. 2018, 9, 19–31. [CrossRef]
25. Bank, E.C. Virtual Currency Schemes; EC Bank, Virtual Currency Schemes (h. 13–14); European Central Bank: Frankfurt,
Germany, 2012.
26. Monrat, A.A.; Schelén, O.; Andersson, K. A survey of blockchain from the perspectives of applications, challenges, and
opportunities. IEEE Access 2019, 7, 117134–117151. [CrossRef]
27. Ciaian, P.; Rajcaniova, M. Virtual relationships: Short-and long-run evidence from BitCoin and altcoin markets. J. Int. Financ.
Mark. Inst. Money 2018, 52, 173–195. [CrossRef]
28. Patel, M.M.; Tanwar, S.; Gupta, R.; Kumar, N. A Deep Learning-based Cryptocurrency Price Prediction Scheme for Financial
Institutions. J. Inf. Secur. Appl. 2020, 55, 102583. [CrossRef]
Algorithms 2022, 15, 428 24 of 26

29. Xia, P.; Wang, H.; Zhang, B.; Ji, R.; Gao, B.; Wu, L.; Luo, X.; Xu, G. Characterizing cryptocurrency exchange scams. Comput. Secur.
2020, 98, 101993. [CrossRef]
30. Wei, W.C. Liquidity and market efficiency in cryptocurrencies. Econ. Lett. 2018, 168, 21–24. [CrossRef]
31. DeVries, P.D. An analysis of cryptocurrency, bitcoin, and the future. Int. J. Bus. Manag. Commer. 2016, 1, 1293–1302.
32. Irshad, S.; Soomro, T.R. Identity theft and social media. Int. J. Comput. Sci. Netw. Secur. 2018, 18, 43–55.
33. Akcora, C.G.; Dixon, M.F.; Gel, Y.R.; Kantarcioglu, M. Bitcoin risk modeling with blockchain graphs. Econ. Lett. 2018, 173, 138–142.
[CrossRef]
34. Derbentsev, V.; Datsenko, N.; Stepanenko, O.; Bezkorovainyi, V. Forecasting cryptocurrency prices time series using machine
learning approach. In Proceedings of the SHS Web of Conferences, the 8th International Conference on Monitoring, Modeling
& Management of Emergent Economy (M3E2 2019), Odessa, Ukraine, 22–24 May 2019; EDP Sciences: Les Ulis, France, 2019;
Volume 65, p. 02001.
35. Poyser, O. Exploring the dynamics of Bitcoin’s price: A Bayesian structural time series approach. Eurasian Econ. Rev. 2019,
9, 29–60. [CrossRef]
36. Jang, H.; Lee, J. An empirical study on modeling and prediction of bitcoin prices with bayesian neural networks based on
blockchain information. IEEE Access 2017, 6, 5427–5437. [CrossRef]
37. Zeng, T.; Yang, M.; Shen, Y. Fancy Bitcoin and conventional financial assets: Measuring market integration based on connectedness
networks. Econ. Model. 2020, 90, 209–220. [CrossRef]
38. Corbet, S.; Meegan, A.; Larkin, C.; Lucey, B.; Yarovaya, L. Exploring the dynamic relationships between cryptocurrencies and
other financial assets. Econ. Lett. 2018, 165, 28–34. [CrossRef]
39. Gkillas, K.; Bekiros, S.; Siriopoulos, C. Extreme correlation in cryptocurrency markets. SSRN 2018. [CrossRef]
40. Stosic, D.; Stosic, D.; Ludermir, T.B.; Stosic, T. Collective behavior of cryptocurrency price changes. Phys. A Stat. Mech. Its Appl.
2018, 507, 499–509. [CrossRef]
41. Corbet, S.; Lucey, B.; Yarovaya, L. Datestamping the Bitcoin and Ethereum bubbles. Financ. Res. Lett. 2018, 26, 81–88. [CrossRef]
42. McCarthy, J.; Minsky, M.L.; Rochester, N.; Shannon, C.E. A proposal for the dartmouth summer research project on artificial
intelligence, August 31, 1955. AI Mag. 2006, 27, 12.
43. Dick, S. Artificial Intelligence. HDSR, 1 July 2019. Available online: https://hdsr.duqduq.org/pub/0aytgrau (accessed on 1
November 2022).
44. Artificial Intelligence the Next Digital Frontier. Available online: https://apo.org.au/node/210501 (accessed on 1 Novem-
ber 2022).
45. Floridi, L. AI and its new winter: From myths to realities. Philos. Technol. 2020, 33, 1–3. [CrossRef]
46. Whittaker, M.; Crawford, K.; Dobbe, R.; Fried, G.; Kaziunas, E.; Mathur, V.; West, S.M.; Richardson, R.; Schultz, J.; Schwartz, O. AI
Now Report 2018; AI Now Institute at New York University: New York, NY, USA, 2018.
47. Helm, J.M.; Swiergosz, A.M.; Haeberle, H.S.; Karnuta, J.M.; Schaffer, J.L.; Krebs, V.E.; Spitzer, A.I.; Ramkumar, P.N. Machine
learning and artificial intelligence: Definitions, applications, and future directions. Curr. Rev. Musculoskelet. Med. 2020, 13, 69–76.
[CrossRef]
48. Janiesch, C.; Zschech, P.; Heinrich, K. Machine learning and deep learning. Electron. Mark. 2021, 31, 685–695. [CrossRef]
49. Zhang, W.; Li, H.; Li, Y.; Liu, H.; Chen, Y.; Ding, X. Application of deep learning algorithms in geotechnical engineering: A short
critical review. Artif. Intell. Rev. 2021, 54, 5633–5673. [CrossRef]
50. Pang, X.; Zhou, Y.; Wang, P.; Lin, W.; Chang, V. An innovative neural network approach for stock market prediction. J. Supercomput.
2020, 76, 2098–2118. [CrossRef]
51. Chowdhury, R.; Rahman, M.A.; Rahman, M.S.; Mahdy, M. An approach to predict and forecast the price of constituents and
index of cryptocurrency using machine learning. Phys. A Stat. Mech. Its Appl. 2020, 551, 124569. [CrossRef]
52. Chen, Z.; Li, C.; Sun, W. Bitcoin price prediction using machine learning: An approach to sample dimension engineering.
J. Comput. Appl. Math. 2020, 365, 112395. [CrossRef]
53. Poongodi, M.; Sharma, A.; Vijayakumar, V.; Bhardwaj, V.; Sharma, A.P.; Iqbal, R.; Kumar, R. Prediction of the price of Ethereum
blockchain cryptocurrency in an industrial finance system. Comput. Electr. Eng. 2020, 81, 106527.
54. McNally, S. Predicting the price of Bitcoin using Machine Learning. Ph.D. Thesis, National College of Ireland, Dublin, Ireland, 2016.
55. Kim, Y.B.; Kim, J.G.; Kim, W.; Im, J.H.; Kim, T.H.; Kang, S.J.; Kim, C.H. Predicting fluctuations in cryptocurrency transactions
based on user comments and replies. PLoS ONE 2016, 11, e0161197. [CrossRef]
56. Alessandretti, L.; ElBahrawy, A.; Aiello, L.M.; Baronchelli, A. Anticipating cryptocurrency prices using machine learning.
Complexity 2018, 2018, 8983590. [CrossRef]
57. Peng, Y.; Albuquerque, P.H.M.; de Sá, J.M.C.; Padula, A.J.A.; Montenegro, M.R. The best of two worlds: Forecasting high
frequency volatility for cryptocurrencies and traditional currencies with Support Vector Regression. Expert Syst. Appl. 2018,
97, 177–192. [CrossRef]
58. Lamon, C.; Nielsen, E.; Redondo, E. Cryptocurrency price prediction using news and social media sentiment. SMU Data Sci. Rev.
2017, 1, 1–22.
59. Sun, X.; Liu, M.; Sima, Z. A novel cryptocurrency price trend forecasting model based on LightGBM. Financ. Res. Lett. 2020,
32, 101084. [CrossRef]
Algorithms 2022, 15, 428 25 of 26

60. Lahmiri, S.; Bekiros, S. Intelligent forecasting with machine learning trading systems in chaotic intraday Bitcoin market. Chaos
Solitons Fractals 2020, 133, 109641. [CrossRef]
61. Kotu, V.; Deshpande, B. Chapter 2: Data Mining Process. In Predictive Analytics and Data Mining; Elsevier: Amsterdam,
The Netherlands, 2015; p. 26.
62. Sun, Y.; Fang, M.; Wang, X. A novel stock recommendation system using Guba sentiment analysis. Pers. Ubiquitous Comput. 2018,
22, 575–587. [CrossRef]
63. Long, W.; Lu, Z.; Cui, L. Deep learning-based feature engineering for stock price movement prediction. Knowl.-Based Syst. 2019,
164, 163–173. [CrossRef]
64. Bekiros, S.D. Fuzzy adaptive decision-making for boundedly rational traders in speculative stock markets. Eur. J. Oper. Res. 2010,
202, 285–293. [CrossRef]
65. Kim, Y.; Ahn, W.; Oh, K.J.; Enke, D. An intelligent hybrid trading system for discovering trading rules for the futures market
using rough sets and genetic algorithms. Appl. Soft Comput. 2017, 55, 127–140. [CrossRef]
66. Yang, H.; Liu, X.Y.; Zhong, S.; Walid, A. Deep reinforcement learning for automated stock trading: An ensemble strategy.
In Proceedings of the First ACM International Conference on AI in Finance (ICAIF ’20), New York, NY, USA, 15–16 October 2020.
67. Arulkumaran, K.; Deisenroth, M.P.; Brundage, M.; Bharath, A.A. Deep reinforcement learning: A brief survey. IEEE Signal
Process. Mag. 2017, 34, 26–38. [CrossRef]
68. Li, Y. Deep reinforcement learning: An overview. arXiv 2017, arXiv:1701.07274.
69. Sutton, R.S.; Barto, A.G. Introduction to Reinforcement Learning; MIT Press: Cambridge, MA, USA, 1998; Volume 135.
70. Rosenstein, M.T.; Barto, A.G.; Si, J.; Barto, A.; Powell, W.; Wunsch, D. Supervised actor-critic reinforcement learning. Learning and
Approximate Dynamic Programming: Scaling Up to the Real World; IEEE: Piscataway, NJ, USA, 2004; pp. 359–380.
71. Chinnamgari, S.K. R Machine Learning Projects: Implement Supervised, Unsupervised, and Reinforcement Learning Techniques Using R
3.5; Packt Publishing Ltd.: Birmingham, UK, 2019.
72. Botvinick, M.; Ritter, S.; Wang, J.X.; Kurth-Nelson, Z.; Blundell, C.; Hassabis, D. Reinforcement learning, fast and slow. Trends
Cogn. Sci. 2019, 23, 408–422. [CrossRef]
73. Mnih, V.; Kavukcuoglu, K.; Silver, D.; Rusu, A.A.; Veness, J.; Bellemare, M.G.; Graves, A.; Riedmiller, M.; Fidjeland, A.K.;
Ostrovski, G.; et al. Human-level control through deep reinforcement learning. Nature 2015, 518, 529–533. [CrossRef] [PubMed]
74. Silver, D.; Huang, A.; Maddison, C.J.; Guez, A.; Sifre, L.; Van Den Driessche, G.; Schrittwieser, J.; Antonoglou, I.; Panneershelvam,
V.; Lanctot, M.; et al. Mastering the game of Go with deep neural networks and tree search. Nature 2016, 529, 484–489. [CrossRef]
[PubMed]
75. Sutton, R.S.; Barto, A.G. Reinforcement Learning: An Introduction; MIT Press: Cambridge, MA, USA, 2018.
76. Zhang, H.; Yu, T. Taxonomy of Reinforcement Learning Algorithms. In Deep Reinforcement Learning; Springer: Berlin/Heidelberg,
Germany, 2020; pp. 125–133.
77. Feinberg, V.; Wan, A.; Stoica, I.; Jordan, M.I.; Gonzalez, J.E.; Levine, S. Model-based value estimation for efficient model-free
reinforcement learning. arXiv 2018, arXiv:1803.00101.
78. Kaiser, L.; Babaeizadeh, M.; Milos, P.; Osinski, B.; Campbell, R.H.; Czechowski, K.; Erhan, D.; Finn, C.; Kozakowski, P.; Levine,
S.; et al. Model-based reinforcement learning for atari. arXiv 2019, arXiv:1903.00374.
79. Nachum, O.; Norouzi, M.; Xu, K.; Schuurmans, D. Bridging the gap between value and policy based reinforcement learning.
arXiv 2017, arXiv:1702.08892.
80. Kumar, A.; Zhou, A.; Tucker, G.; Levine, S. Conservative q-learning for offline reinforcement learning. arXiv 2020,
arXiv:2006.04779.
81. Masson, W.; Ranchod, P.; Konidaris, G. Reinforcement learning with parameterized actions. In Proceedings of the Thirtieth AAAI
Conference on Artificial Intelligence, Phoenix, AZ, USA, 12–17 February 2016.
82. Henderson, P.; Islam, R.; Bachman, P.; Pineau, J.; Precup, D.; Meger, D. Deep reinforcement learning that matters. In Proceedings
of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018; Volume 32.
83. Meng, W.; Zheng, Q.; Shi, Y.; Pan, G. An Off-Policy Trust Region Policy Optimization Method with Monotonic Improvement
Guarantee for Deep Reinforcement Learning. IEEE Trans. Neural Netw. Learn. Syst. 2021, 33, 2223–2235. [CrossRef]
84. Nachum, O.; Norouzi, M.; Xu, K.; Schuurmans, D. Trust-pcl: An off-policy trust region method for continuous control. arXiv
2017, arXiv:1707.01891.
85. Jaderberg, M.; Mnih, V.; Czarnecki, W.M.; Schaul, T.; Leibo, J.Z.; Silver, D.; Kavukcuoglu, K. Reinforcement learning with
unsupervised auxiliary tasks. arXiv 2016, arXiv:1611.05397.
86. Mirowski, P.; Pascanu, R.; Viola, F.; Soyer, H.; Ballard, A.J.; Banino, A.; Denil, M.; Goroshin, R.; Sifre, L.; Kavukcuoglu, K.; et al.
Learning to navigate in complex environments. arXiv 2016, arXiv:1611.03673.
87. Liang, C.; Berant, J.; Le, Q.; Forbus, K.D.; Lao, N. Neural symbolic machines: Learning semantic parsers on freebase with weak
supervision. arXiv 2016, arXiv:1611.00020.
88. Cao, Q.; Lin, L.; Shi, Y.; Liang, X.; Li, G. Attention-aware face hallucination via deep reinforcement learning. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA, 21–26 July 2017; pp. 690–698.
89. Koker, T.E.; Koutmos, D. Cryptocurrency Trading Using Machine Learning. J. Risk Financ. Manag. 2020, 13, 178. [CrossRef]
90. Makarov, I.; Schoar, A. Trading and arbitrage in cryptocurrency markets. J. Financ. Econ. 2020, 135, 293–319. [CrossRef]
Algorithms 2022, 15, 428 26 of 26

91. Huck, N. Large data sets and machine learning: Applications to statistical arbitrage. Eur. J. Oper. Res. 2019, 278, 330–342.
[CrossRef]
92. Sadighian, J. Deep Reinforcement Learning in Cryptocurrency Market Making. arXiv 2019, arXiv:1911.08647.
93. Sattarov, O.; Muminov, A.; Lee, C.W.; Kang, H.K.; Oh, R.; Ahn, J.; Oh, H.J.; Jeon, H.S. Recommending Cryptocurrency Trading
Points with Deep Reinforcement Learning Approach. Appl. Sci. 2020, 10, 1506. [CrossRef]
94. Lee, K.; Ulkuatam, S.; Beling, P.; Scherer, W. Generating synthetic Bitcoin transactions and predicting market price movement via
inverse reinforcement learning and agent-based modeling. J. Artif. Soc. Soc. Simul. 2018, 21, 5. [CrossRef]
95. Schnaubelt, M. Deep reinforcement learning for the optimal placement of cryptocurrency limit orders. Eur. J. Oper. Res. 2022, 296,
993–1006. [CrossRef]
96. Lucarelli, G.; Borrotti, M. A deep Q-learning portfolio management framework for the cryptocurrency market. Neural Comput.
Appl. 2020, 32, 17229–17244. [CrossRef]
97. Ye, Y.; Pei, H.; Wang, B.; Chen, P.Y.; Zhu, Y.; Xiao, J.; Li, B. Reinforcement-learning based portfolio management with augmented
asset movement prediction states. In Proceedings of the AAAI Conference on Artificial Intelligence, New York, NY, USA, 7–12
February 2020; Volume 34, pp. 1112–1119.
98. Jiang, Z.; Liang, J. Cryptocurrency portfolio management with deep reinforcement learning. In Proceedings of the 2017 Intelligent
Systems Conference (IntelliSys), London, UK, 7–8 September 2017; pp. 905–913.
99. Hegazy, K.; Mumford, S. Comparitive Automated Bitcoin Trading Strategies. Available online: http://cs229.stanford.edu/proj2
016/report/MumfordHegazy-ComparitiveAutomatedBitcoinTradingStrategies-report.pdf (accessed on 1 November 2022).
100. Betancourt, C.; Chen, W.H. Deep reinforcement learning for portfolio management of markets with a dynamic number of assets.
Expert Syst. Appl. 2021, 164, 114002. [CrossRef]
101. Aamir, M.; Zaidi, S.M.A. DDoS attack detection with feature engineering and machine learning: The framework and performance
evaluation. Int. J. Inf. Secur. 2019, 18, 761–785. [CrossRef]
102. Weng, L.; Sun, X.; Xia, M.; Liu, J.; Xu, Y. Portfolio trading system of digital currencies: A deep reinforcement learning with
multidimensional attention gating mechanism. Neurocomputing 2020, 402, 171–182. [CrossRef]
103. Marcot, B.G.; Penman, T.D. Advances in Bayesian network modelling: Integration of modelling technologies. Environ. Model.
Softw. 2019, 111, 386–393. [CrossRef]
104. Corbet, S.; Hou, Y.G.; Hu, Y.; Larkin, C.; Oxley, L. Any port in a storm: Cryptocurrency safe-havens during the COVID-19
pandemic. Econ. Lett. 2020, 194, 109377. [CrossRef] [PubMed]
105. Vidal-Tomás, D. Transitions in the cryptocurrency market during the COVID-19 pandemic: A network analysis. Financ. Res. Lett.
2021, 43, 101981. [CrossRef] [PubMed]
106. Surden, H. Machine learning and law: An overview. Research Handbook on Big Data Law; Edward Elgar Publishing: Cheltenham,
UK, 2021.
107. Nguyen, T.V.H.; Nguyen, B.T.; Nguyen, T.C.; Nguyen, Q.Q. Bitcoin return: Impacts from the introduction of new altcoins. Res.
Int. Bus. Financ. 2019, 48, 420–425. [CrossRef]
108. Chakraborty, G.; Chandrashekhar, G.; Balasubramanian, G. Measurement of extreme market risk: Insights from a comprehensive
literature review. Cogent Econ. Financ. 2021, 9, 1920150. [CrossRef]
109. Hong, Y.; Liu, Y.; Wang, S. Granger causality in risk and detection of extreme risk spillover between financial markets. J. Econom.
2009, 150, 271–287. [CrossRef]
110. Ahmed, M.; Choudhury, N.; Uddin, S. Anomaly detection on big data in financial markets. In Proceedings of the 2017 IEEE/ACM
International Conference on Advances in Social Networks Analysis and Mining (ASONAM), Sydney, Australia, 31 July–3 August
2017; pp. 998–1001.