![Percentage of Total Market Capitalisation [79]](https://www.researchgate.net/profile/Carmine-Ventre/publication/340172901/figure/fig9/AS:978408330764295@1610282132296/Percentage-of-Total-Market-Capitalisation-79_Q320.jpg)
Content uploaded by Carmine Ventre
Author content
All content in this area was uploaded by Carmine Ventre on Jan 10, 2021
Content may be subject to copyright.
Cryptocurrency Trading: A Comprehensive Survey
Fan Fang a,∗, Carmine Ventrea, Michail Basiosb, Hoiliong Kongb, Leslie Kanthanb,
David Martinez-Regob, Fan Wu band Lingbo Li b,∗
aKing’s College London, UK
bTuring Intelligence Technology Limited, UK
ARTICLE INFO
Keywords:
trading, cryptocurrency, machine learn-
ing, econometrics
ABSTRACT
In recent years, the tendency of the number of financial institutions including cryptocurrencies in
their portfolios has accelerated. Cryptocurrencies are the first pure digital assets to be included by
asset managers. Even though they share some commonalities with more traditional assets, they have a
separate nature of its own and their behaviour as an asset is still under the process of being understood.
It is therefore important to summarise existing research papers and results on cryptocurrency trading,
including available trading platforms, trading signals, trading strategy research and risk management.
This paper provides a comprehensive survey of cryptocurrency trading research, by covering 126
research papers on various aspects of cryptocurrency trading (e.g., cryptocurrency trading systems,
bubble and extreme condition, prediction of volatility and return, crypto-assets portfolio construction
and crypto-assets, technical trading and others). This paper also analyses datasets, research trends and
distribution among research objects (contents/properties) and technologies, concluding with some
promising opportunities that remain open in cryptocurrency trading.
1. Introduction
Cryptocurrencies have experienced broad market accep-
tance and fast development despite their recent conception.
Many hedge funds and asset managers have begun to in-
clude cryptocurrency-related assets into their portfolios and
trading strategies. The academic community has similarly
spent considerable efforts in researching cryptocurrency trad-
ing. This paper seeks to provide a comprehensive survey of
the research on cryptocurrency trading, by which we mean
any study aimed at facilitating and building strategies to
trade cryptocurrencies.
As an emerging market and research direction, cryp-
tocurrencies and cryptocurrency trading have seen consid-
erable progress and a notable upturn in interest and activ-
ity [103]. From Figure 1, we observe over 85% of papers
have appeared since 2018, demonstrating the emergence of
cryptocurrency trading as a new research area in financial
trading.
The literature is organised according tosix distinct as-
pects of cryptocurrency trading:
• Cryptocurrency trading software systems (i.e., real-
time trading systems, turtle trading systems, arbitrage
trading systems);
• Systematic trading including technical analysis, pairs
trading and other systematic trading methods;
fan.fang@kcl.ac.uk ( Fan Fang);
fan.fang@kcl.ac.uk ( Fan Fang);
carmine.ventre@kcl.ac.uk ( Carmine Ventre);
mike@turintech.ai ( Michail Basios); justin@turintech.ai
( Hoiliong Kong); leslie@turintech.ai ( Leslie Kanthan);
david@turintech.ai ( David Martinez-Rego);
fan@turintech.ai ( Fan Wu); lingbo@turintech.ai (
Lingbo Li)
www.cvr.cc,cvrŽsayahna.org ( Fan Fang)
ORCID(s): 0000-0002-5086-5614 ( Fan Fang);
0000-0002-3073-1352 ( Lingbo Li)
Figure 1: Cryptocurrency Trading Publications (cumulative)
during 2013-2019
• Emergent trading technologies including economet-
ric methods, machine learning technology and other
emergent trading methods;
• Portfolio and cryptocurrency assets including research
among cryptocurrency co-movements and crypto-asset
portfolio research;
• Market condition research including bubbles [106] or
crash analysis and extreme conditions;
• Other Miscellaneous cryptocurrency trading research.
In this survey we aim at compiling the most relevant re-
search in these areas and extract a set of descriptive indica-
tors that can give an idea of the level of maturity research in
this area has achieved.
We also summarise research distribution (among research
properties, categories and research technologies). The dis-
tribution among properties and categories identifies classifi-
cations of research objectives and contents. The distribution
among technologies identifies classifications of methodol-
First Author et al. Page 1 of 30
arXiv:2003.11352v3 [q-fin.TR] 7 Jan 2021
Cryptocurrency Trading: A Comprehensive Survey
ogy or technical methods in researching cryptocurrency trad-
ing. Specifically, we subdivide research distribution among
categories and technologies into statistical methods and ma-
chine learning technologies. Moreover, We identify datasets
and opportunities (potential research directions) that have
appeared in the cryptocurrency trading area. To ensure that
our survey is self-contained, we aim to provide sufficient
material to adequately guide financial trading researchers
who are interested in cryptocurrency trading.
There has been related work that discussed or partially
surveyed the literature related to cryptocurrency trading. Kyr-
iazis et al. [166] surveyed efficiency and profitable trading
opportunities in cryptocurrency markets. Ahamad et al. [4]
and Sharma et al. [221] gave a brief survey on cryptocur-
rencies. Ujan et al. [191] gave a brief survey of cryptocur-
rency systems. Ignasi et al. [186] performed a bibliometric
analysis of bitcoin literature. These relate work outcomes
focused on specific area in cryptocurrency, including cryp-
tocurrencies and cryptocurrency market introduction, cryp-
tocurrency systems / platforms, bitcoin literature review, etc.
To the best of our knowledge, no previous work has pro-
vided a comprehensive survey particularly focused on cryp-
tocurrency trading.
In summary, the paper makes the following contribu-
tions:
Definition. This paper defines cryptocurrency trading and
categorises it into: cryptocurrency markets, cryptocur-
rency trading models, and cryptocurrency trading strate-
gies. The core content of this survey is trading strate-
gies for cryptocurrencies while we cover all aspects
of it.
Multidisciplinary Survey. The paper provides a compre-
hensive survey of 126 cryptocurrency trading papers,
across different academic disciplines such as finance
and economics, artificial intelligence and computer
science. Some papers may cover multiple aspects and
will be surveyed for each category.
Analysis. The paper analyses the research distribution, datasets
and trends that characterise the cryptocurrency trad-
ing literature.
Horizons. The paper identifies challenges, promising re-
search directions in cryptocurrency trading, aimed to
promote and facilitate further research.
Figure 2depicts the paper structure, which is informed
by the review schema adopted. More details about this can
be found in Section 4.
2. Cryptocurrency Trading
This section provides an introduction to cryptocurrency
trading. We will discuss Blockchain, as the enabling tech-
nology, cryptocurrency markets and cryptocurrency trad-
ing strategies.
Figure 2: Tree structure of the contents in this paper
Figure 3: Workflow of Blockchain transaction
2.1. Blockchain
2.1.1. Blockchain Technology Introduction
Blockchain is a digital ledger of economic transactions
that can be used to record not just financial transactions, but
any object with an intrinsic value. [232]. In its simplest
form, a Blockchain is a series of immutable data records
with timestamps,which are managed by a cluster of ma-
chines that do not belong to any single entity. Each of these
data blocks is protected by cryptographic principle and bound
to each other in a chain (cf. Figure 3for the workflow).
Cryptocurrencies like Bitcoin are made on a peer-to-
peer network structure. Each peer has a complete history
of all transactions, thus recording the balance of each ac-
count. For example, a transaction is a file that says “A pays
X Bitcoins to B” that is signed by A using its private key.
This is basic public-key cryptography, but also the build-
ing block on which cryptocurrencies are based. After being
signed, the transaction is broadcast on the network. When a
peer discovers a new transaction, it checks to make sure that
the signature is valid (this amounts to use the signer’s public
key, denoted as the algorithm in Figure 3). If the verification
is valid then the block is added to the chain; all other blocks
added after it will “confirm” that transaction. For example,
if a transaction is contained in block 502 and the length of
the blockchain is 507 blocks, it means that the transaction
has 5 confirmations (507-502) [218].
2.1.2. From Blockchain to Cryptocurrencies
Confirmation is a critical concept in cryptocurrencies;
only miners can confirm transactions. Miners add blocks
to the Blockchain; they retrieve transactions in the previ-
ous block and combine it with the hash of the preceding
block to obtain its hash, and then store the derived hash into
First Author et al. Page 2 of 30
Cryptocurrency Trading: A Comprehensive Survey
the current block. Miners in Blockchain accept transactions,
mark them as legitimate and broadcast them across the net-
work. After the miner confirms the transaction, each node
must add it to its database. In layman terms, it has become
part of the Blockchain and miners undertake this work to
obtain cryptocurrency tokens, such as Bitcoin. In contrast
to Blockchain, cryptocurrencies are related to the use of to-
kens based on distributed ledger technology. Any transac-
tion involving purchase, sale, investment, etc. involves a
Blockchain native token or sub-token. Blockchain is a plat-
form that drives cryptocurrency and is a technology that acts
as a distributed ledger for the network. The network creates
a means of transaction and enables the transfer of value and
information. Cryptocurrencies are the tokens used in these
networks to send value and pay for these transactions. They
can be thought of as tools on the Blockchain, and in some
cases can also function as resources or utilities. In other
instances, they are used to digitise the value of assets. In
summary, Cryptocurrencies are part of an ecosystem-based
on Blockchain technology.
2.2. Introduction of cryptocurrency market
2.2.1. What is cryptocurrency?
Cryptocurrency is a decentralised medium of exchange
which uses cryptographic functions to conduct financial trans-
actions [90]. Cryptocurrencies leverage the Blockchain tech-
nology to gain decentralisation, transparency, and immutabil-
ity [187]. In the above, we have discussed how Blockchain
technology is implemented for cryptocurrencies.
In general, the security of cryptocurrencies is built on
cryptography, neither by people nor on trust [194]. For ex-
ample, Bitcoin uses a method called ”Elliptic Curve Cryp-
tography” to ensure that transactions involving Bitcoin are
secure [246]. Elliptic curve cryptography is a type of public-
key cryptography that relies on mathematics to ensure the
security of transactions. When someone attempts to circum-
vent the aforesaid encryption scheme by brute force, it takes
them one-tenth the age of the universe to find a value match
when trying 250 billion possibilities every second [118].
Regarding its use as a currency, cryptocurrency has the same
properties as money. It has a controlled supply. Most cryp-
tocurrencies limit the supply of tokens. E.g. for Bitcoin,
the supply will decrease over time and will reach its final
quantity sometime around 2,140. All cryptocurrencies con-
trol the supply of tokens through a timetable encoded in the
Blockchain.
One of the most important features of cryptocurrencies
is the exclusion of financial institution intermediaries [125].
The absence of a “middleman” lowers transaction costs for
traders. For comparison, if a bank’s database is hacked or
damaged, the bank will rely entirely on its backup to recover
any information that is lost or compromised. With cryp-
tocurrencies, even if part of the network is compromised,
the rest will continue to be able to verify transactions cor-
rectly. Cryptocurrencies also have the important feature of
not being controlled by any central authority [217]: the de-
centralised nature of the Blockchain ensures cryptocurren-
Figure 4: Total Market Capitalization and Volume of cryp-
tocurrency market, USD [238]
cies are theoretically immune to government control and in-
terference.
As of December 20, 2019, there exist 4,950 cryptocur-
rencies and 20,325 cryptocurrency markets; the market cap
is around 190 billion dollars [78]. Figure 4shows histor-
ical data on global market capitalisation and 24-hour trad-
ing volume [238]. The total market cap is calculated by
aggregating the dollar market cap of all cryptocurrencies.
From the figure, we can observe how cryptocurrencies expe-
rience exponential growth in 2017 and a large bubble burst
in early 2018. But in recent years, cryptocurrencies have
shown signs of stabilisation.
There are three mainstream cryptocurrencies: Bitcoin
(BTC), Ethereum (ETH), and Litecoin (LTC). Bitcoin was
created in 2009 and garnered massive popularity. On Oc-
tober 31, 2008, an individual or group of individuals oper-
ating under the pseudonym Satoshi Nakamoto released the
Bitcoin white paper and described it as: ”A pure peer-to-
peer version of electronic cash that can be sent online for
payment from one party to another without going through
a counterparty, ie. a financial institution.” [193] Launched
by Vitalik Buterin in 2015, Ethereum is a special Blockchain
with a special token called Ether (ETH symbol in exchanges).
A very important feature of Ethereum is the ability to create
new tokens on the Ethereum Blockchain. The Ethereum net-
work went live on July 30, 2015, and pre-mined 72 million
Ethereum. Litecoin is a peer-to-peer cryptocurrency created
by Charlie Lee. It was created according to the Bitcoin pro-
tocol, but it uses a different hashing algorithm. Litecoin
uses a memory-intensive proof-of-work algorithm, Scrypt.
Scrypt
Figure 5shows percentages of total cryptocurrency mar-
ket capitalisation; Bitcoin and Ethereum occupy the vast
majority of the total market capitalisation (data collected on
8 Jan 2020).
2.2.2. Cryptocurrency Exchanges
A cryptocurrency exchange or digital currency exchange
(DCE) is a business that allows customers to trade cryp-
tocurrencies. Cryptocurrency exchanges can be market mak-
ers, usually using the bid-ask spread as a commission for
services, or as a matching platform, by simply charging fees.
Table 1shows the top or classical cryptocurrency ex-
First Author et al. Page 3 of 30
Cryptocurrency Trading: A Comprehensive Survey
Table 1
Cryptocurrency exchanges Lists
Exchanges Category Supported currencies Fiat Currency Registration country Regulatory authority
CME Derivatives BTC and Ethereum [71] USD USA [73] CFTC [72]
CBOE Derivatives BTC [59] USD USA [58] CFTC [60]
BAKKT (NYSE) Derivatives BTC [15] USD USA [16] CFTC [15]
BitMex Derivatives 12 cryptocurrencies [31] USD Seychelles [32] -
Binance Spot 98 cryptocurrencies [27]EUR, NGN, RUB, TRY Malta [181] FATF [26]
Coinbase Spot 28 cryptocurrencies [76]EUR, GBP, USD USA [37] SEC [77]
Bitfinex Spot >100 cryptocurrencies [28]EUR, GBP, JPY, USD British Virgin Islands [29] NYAG [30]
Bitstamp Spot 5 cryptocurrencies [33]EUR, USD Luxembourg [34] CSSF [35]
Poloniex Spot 23 cryptocurrencies [213] USD USA [213] -
Figure 5: Percentage of Total Market Capitalisation [79]
changes according to the rank list, by volume, compiled
on “nomics” website [199]. Chicago Mercantile Exchange
(CME), Chicago Board Options Exchange (CBOE) as well
as BAKKT (backed by New York Stock Exchange) are reg-
ulated cryptocurrency exchanges. Fiat currency data also
comes from “nomics” website [199]. Regulatory authority
and supported currencies of listed exchanges are collected
from official websites or blogs.
2.3. Cryptocurrency Trading
2.3.1. Definition
Firstly we give a definition of cryptocurrency trading.
Definition 1. Cryptocurrency trading is the act of buying
and selling of cryptocurrencies with the intention of making
a profit.
The definition of cryptocurrency trading can be broken down
into three aspects: object, operation mode and trading strat-
egy. The object of cryptocurrency trading is the asset being
traded, which is “cryptocurrency”. The operation mode of
cryptocurrency trading depends on the means of transaction
in the cryptocurrency market, which can be classified into
“trading of cryptocurrency Contract for Differences (CFD)”
(The contract between the two parties, often referred to as
the “buyer” and “seller”, stipulates that the buyer will pay
the seller the difference between themselves when the po-
sition closes [11]) and “buying and selling cryptocurrencies
via an exchange”. A trading strategy in cryptocurrency trad-
ing, formulated by an investor, is an algorithm that defines
a set of predefined rules to buy and sell on cryptocurrency
markets.
2.3.2. Advantages of Trading Cryptocurrency
The benefits of cryptocurrency trading include:
Drastic fluctuations. The volatility of cryptocurrencies are
often likely to attract speculative interest and investors.
The rapid fluctuations of intraday prices can provide
traders with great money-earning opportunities, but it
also includes more risk.
24-hour market. The cryptocurrency market is available
24 hours a day, 7 days a week because it is a de-
centralised market. Unlike buying and selling stocks
and commodities, the cryptocurrency market is not
traded physically from a single location. Cryptocur-
rency transactions can take place between individuals,
in different venues across the world.
Near Anonymity. Buying goods and services using cryp-
tocurrencies is done online and does not require to
make one’s own identity public. With increasing con-
cerns over identity theft and privacy, cryptocurren-
cies can thus provide users with some advantages re-
garding privacy. Different exchanges have specific
Know-Your-Customer (KYC) measures used to iden-
tify users or customers [3]. The KYC undertook in
the exchanges allows financial institutions to reduce
the financial risk while maximising the wallet owner’s
anonymity.
Peer-to-peer transactions. One of the biggest benefits of
cryptocurrencies is that they do not involve financial
institution intermediaries. As mentioned above, this
can reduce transaction costs. Moreover, this feature
might appeal to users who distrust traditional systems.
Over-the-counter (OTC) cryptocurrency markets of-
fer, in this context, peer-to-peer transactions on the
Blockchain. The most famous cryptocurrency OTC
market is “LocalBitcoin [176]”.
Programmable “smart” capabilities. Some cryptocurren-
cies can bring other benefits to holders, including lim-
ited ownership and voting rights. Cryptocurrencies
may also include a partial ownership interest in phys-
ical assets such as artwork or real estate.
First Author et al. Page 4 of 30
Cryptocurrency Trading: A Comprehensive Survey
3. Cryptocurrency Trading Strategy
Cryptocurrency trading strategy is the main focus of this
survey. There are many trading strategies, which can be
broadly divided into two main categories: technical and fun-
damental. They are similar in the sense that they both rely
on quantifiable information that can be backtested against
historical data to verify their performance. In recent years,
a third kind of trading strategy, which we call quantitative,
has received increasing attention. Such a trading strategy is
similar to a technical trading strategy because it uses trad-
ing activity information on the exchange to make buying or
selling decisions. Quantitative traders build trading strate-
gies with quantitative data, which is mainly derived from
price, volume, technical indicators or ratios to take advan-
tage of inefficiencies in the market and are executed auto-
matically by trading software. Cryptocurrency market is dif-
ferent from traditional markets as there are more arbitrage
opportunities, higher fluctuation and transparency. Due to
these characteristics, most traders and analysts prefer using
quantitative trading strategies in cryptocurrency markets.
3.1. Cryptocurrency Trading Software System
Software trading systems allow international transactions,
process customer accounts and information, and accept and
execute transaction orders [50]. A cryptocurrency trading
system is a set of principles and procedures that are pre-
programmed to allow trade between cryptocurrencies and
between fiat currencies and cryptocurrencies. Cryptocur-
rency trading systems are built to overcome price manip-
ulation, cybercriminal activities and transaction delays [21].
When developing a cryptocurrency trading system, we must
consider the capital market, base asset, investment plan and
strategies [190]. Strategies are the most important part of an
effective cryptocurrency trading system and they will be in-
troduced below. There exist several cryptocurrency trading
systems that are available commercially, for example, Cap-
folio, 3Commas, CCXT, Freqtrade and Ctubio. From these
cryptocurrency trading systems, investors can obtain pro-
fessional trading strategy support, fairness and transparency
from the professional third-party consulting companies and
fast customer services.
3.2. Systematic Trading
Systematic Trading is a way to define trading goals,
risk controls and rules. In general, systematic trading in-
cludes high frequency trading and slower investment types
like systematic trend tracking. In this survey, we divide
systematic cryptocurrency trading into technical analysis,
pairs trading and others. Technical analysis in cryptocur-
rency trading is the act of using historical patterns of trans-
action data to assist a trader in assessing current and pro-
jecting future market conditions for the purpose of making
profitable trades. Price and volume charts summarise all
trading activity made by market participants in an exchange
and affect their decisions. Some experiments showed that
the use of specific technical trading rules allows generat-
ing excess returns, which is useful to cryptocurrency traders
and investors in making optimal trading and investment de-
cisions [116]. Pairs trading is a systematic trading strat-
egy that considers two similar assets with slightly different
spreads. If the spread widens, short the high stocks and buy
the low stocks. When the spread narrows again to a certain
equilibrium value, a profit is generated [94]. Papers shown
in this section involve the analysis and comparison of tech-
nical indicators, pairs and informed trading, amongst other
strategies.
3.3. Emergent Trading Technologies
Emergent trading strategies for cryptocurrency include
strategies that are based on econometrics and machine learn-
ing technologies.
3.3.1. Econometrics on Cryptocurrency
Econometric methods apply a combination of statistical
and economic theories to estimate economic variables and
predict their values [244]. Statistical models use mathe-
matical equations to encode information extracted from the
data [152]. In some cases, statistical modeling techniques
can quickly provide sufficiently accurate models [24]. Other
methods might be used, such as sentiment-based prediction
and long-and-short-term volatility classification based pre-
diction [64]. The prediction of volatility can be used to
judge the price fluctuation of cryptocurrencies, which is also
valuable for the pricing of cryptocurrency-related deriva-
tives [147].
When studying cryptocurrency trading using economet-
rics, researchers apply statistical models on time-series data
like generalised autoregressive conditional heteroskedastic-
ity (GARCH) and BEKK (named after Baba, Engle, Kraft
and Kroner, 1995 [96]) models to evaluate the fluctuation
of cryptocurrencies [55]. A linear statistical model is a
method to evaluate the linear relationship between prices
and an explanatory variable [196]. When there exists more
than one explanatory variable, we can model the linear re-
lationship between explanatory (independent) and response
(dependent) variables with multiple linear models. The com-
mon linear statistical model used in the time-series analysis
is the autoregressive moving average (ARMA) model [69].
3.3.2. Machine Learning Technology
Machine learning is an efficient tool for developing Bit-
coin and other cryptocurrency trading strategies [185] be-
cause it can infer data relationships that are often not di-
rectly observable by humans. From the most basic perspec-
tive, Machine Learning relies on the definition of two main
components: input features and objective function. The def-
inition of Input Features (data sources) is where knowledge
of fundamental and technical analysis comes into play. We
may divide the input into several groups of features, for ex-
ample, those based on Economic indicators (such as, gross
domestic product indicator, interest rates, etc.), Social indi-
cators (Google Trends, Twitter, etc.), Technical indicators
(price, volume, etc.) and other Seasonal indicators (time of
day, day of the week, etc.). The objective function defines
the fitness criteria one uses to judge if the Machine Learn-
First Author et al. Page 5 of 30
Cryptocurrency Trading: A Comprehensive Survey
Figure 6: Process of machine learning in predicting cryp-
tocurrency
ing model has learnt the task at hand. Typical predictive
models try to anticipate numeric (e.g., price) or categorical
(e.g., trend) unseen outcomes. The machine learning model
is trained by using historic input data (sometimes called
in-sample) to generalise patterns therein to unseen (out-of-
sample) data to (approximately) achieve the goal defined by
the objective function. Clearly, in the case of trading, the
goal is to infer trading signals from market indicators which
help to anticipate asset future returns.
Generalisation error is a pervasive concern in the appli-
cation of Machine Learning to real applications, and of ut-
most importance in Financial applications. We need to use
statistical approaches, such as cross validation, to validate
the model before we actually use it to make predictions.
In machine learning, this is typically called “validation”.
The process of using machine learning technology to pre-
dict cryptocurrency is shown in Figure 6.
Depending on the formulation of the main learning loop,
we can classify Machine Learning approaches into three
categories: Supervised learning, Unsupervised learning and
Reinforcement learning. Supervised learning is used to de-
rive a predictive function from labeled training data. La-
beled training data means that each training instance in-
cludes inputs and expected outputs. Usually, these expected
outputs are produced by a supervisor and represent the ex-
pected behaviour of the model. The most used labels in
trading are derived from in sample future returns of assets.
Unsupervised learning tries to infer structure from unla-
beled training data and it can be used during exploratory
data analysis to discover hidden patterns or to group data
according to any pre-defined similarity metrics. Reinforce-
ment learning utilises software agents trained to maximise
a utility function, which defines their objective; this is flex-
ible enough to allow agents to exchange short term returns
for future ones. In the financial sector, some trading chal-
lenges can be expressed as a game in which an agent aims
at maximising the return at the end of the period.
The use of machine learning in cryptocurrency trading
research encompasses the connection between data sources’
understanding and machine learning model research. Fur-
ther concrete examples are shown in a later section.
3.4. Portfolio Research
Portfolio theory advocates diversification of investments
to maximize returns for a given level of risk by allocating
assets strategically. The celebrated mean-variance optimisa-
tion is a prominent example of this approach [182]. Gener-
ally, crypto asset denotes a digital asset (i.e., cryptocurren-
cies and derivatives). There are some common ways to build
a diversified portfolio in crypto assets. The first method is
to diversify across markets, which is to mix a wide vari-
ety of investments within a portfolio of the cryptocurrency
market. The second method is to consider the industry sec-
tor, which is to avoid investing too much money in any one
category. Diversified investment of portfolio in the cryp-
tocurrency market includes portfolio across cryptocurren-
cies [175] and portfolio across the global market including
stocks and futures [140].
3.5. Market Condition Research
Market condition research appears especially important
for cryptocurrencies. A financial bubble is a significant in-
crease in the price of an asset without changes in its intrinsic
value [48]. Many experts pinpoint a cryptocurrency bubble
in 2017 when the prices of cryptocurrencies grew by 900%.
In 2018, Bitcoin faced a collapse in its value. This signif-
icant fluctuation inspired researchers to study bubbles and
extreme conditions in cryptocurrency trading.
4. Paper Collection and Review Schema
The section introduces the scope and approach of our
paper collection, a basic analysis, and the structure of our
survey.
4.1. Survey Scope
We adopt a bottom-up approach to the research in cryp-
tocurrency trading, starting from the systems up to risk man-
agement techniques. For the underlying trading system, the
focus is on the optimisation of trading platforms structure
and improvements of computer science technologies.
At a higher level, researchers focus on the design of
models to predict return or volatility in cryptocurrency mar-
kets. These techniques become useful to the generation of
trading signals. on the next level above predictive mod-
els, researchers discuss technical trading methods to trade
in real cryptocurrency markets. Bubbles and extreme condi-
tions are hot topics in cryptocurrency trading because, as
discussed above, these markets have shown to be highly
volatile (whilst volatility went down after crashes). Portfo-
lio and cryptocurrency asset management are effective meth-
ods to control risk. We group these two areas in risk man-
agement research. Other papers included in this survey in-
clude topics like pricing rules, dynamic market analysis,
regulatory implications, and so on. Table 2shows the gen-
eral scope of cryptocurrency trading included in this survey.
Since many trading strategies and methods in cryptocur-
rency trading are closely related to stock trading, some re-
searchers migrate or use the research results for the latter
to the former. When conducting this research, we only con-
sider those papers whose research focuses on cryptocurrency
markets or a comparison of trading in those and other finan-
cial markets.
Specifically, we apply the following criteria when col-
lecting papers related to cryptocurrency trading:
First Author et al. Page 6 of 30
Cryptocurrency Trading: A Comprehensive Survey
Table 2
Survey scope table
Trading (bottom up)
Trading System
Prediction (return)
Prediction (volatility)
Technical trading methods
Risk management Bubble and extreme condition
Porfolio and Cryptocurrency asset
Others
1. The paper introduces or discusses the general idea of
cryptocurrency trading or one of the related aspects of
cryptocurrency trading.
2. The paper proposes an approach, study or framework
that targets optimised efficiency or accuracy of cryp-
tocurrency trading.
3. The paper compares different approaches or perspec-
tives in trading cryptocurrency.
By “cryptocurrency trading” here, we mean one of the terms
listed in Table 2and discussed above.
Some researchers gave a brief survey of cryptocurrency [4,
221], cryptocurrency systems [191] and cryptocurrency trad-
ing opportunities [166]. These surveys are rather limited in
scope as compared to ours, which also includes a discussion
on the latest papers in the area; we want to remark that this
is a fast-moving research field.
4.2. Paper Collection Methodology
To collect the papers in different areas or platforms, we
used keyword searches on Google Scholar and arXiv, two
of the most popular scientific databases. We also choose
other public repositories like SSRN but we find that almost
all academic papers in these platforms can also be retrieved
via Google Scholar; consequently, in our statistical analysis,
we count those as Google Scholar hits. We choose arXiv as
another source since it allows this survey to be contempo-
rary with all the most recent findings in the area. The in-
terested reader is warned that these papers have not under-
gone formal peer review. The keywords used for searching
and collecting are listed below. [Crypto] means the cryp-
tocurrency market, which is our research interest because
methods might be different among different markets. We
conducted 6 searches across the two repositories just before
October 15, 2019.
- [Crypto] + Trading
- [Crypto] + Trading system
- [Crypto] + Prediction
- [Crypto] + Trading strategy
- [Crypto] + Risk Management
- [Crypto] + Portfolio
To ensure high coverage, we adopted the so-called snow-
balling [250] method on each paper found through these
keywords. We checked papers added from snowballing meth-
ods that satisfy the criteria introduced above until we reached
closure.
Table 3
Paper query results. #Hits, #Title, and #Body denote the
number of papers returned by the search, left after title filter-
ing, and left after body filtering, respectively.
Key Words #Hits #Title #Body
[Crypto] + Trading 555 32 29
[Crypto] + Trading System 4 3 2
[Crypto] + Prediction 26 14 13
[Crypto] + Trading Strategy 22 9 8
[Crypto] + Risk Management /
[Crypto] + Portfolio 120 14 14
Query - - 66
Snowball - - 60
Overall - - 126
4.3. Collection Results
Table 3shows the details of the results from our paper
collection. Keyword searches and snowballing resulted in
126 papers across the six research areas of interest in Sec-
tion 4.1.
Figure 7shows the distribution of papers published at
different research sites. Among all the papers, 45.24% pa-
pers are published in Finance and Economics venues such
as Journal of Financial Economics (JFE), Cambridge Centre
for Alternative Finance (CCAF), Finance Research Letters,
Centre for Economic Policy Research (CEPR) and Journal
of Risk and Financial Management (JRFM); 4.76% papers
are published in Science venues such as Public Library Of
Science one (PLOS one), Royal Society open science and
SAGE; 15.87% papers are published in Intelligent Engi-
neering and Data Mining venues such as Symposium Se-
ries on Computational Intelligence (SSCI), Intelligent Sys-
tems Conference (IntelliSys), Intelligent Data Engineering
and Automated Learning (IDEAL) and International Con-
ference on Data Mining (ICDM); 4.76% papers are pub-
lished in Physics / Physicians venues (mostly in Physics
venue) such as Physica A; 10.32% papers are published in
AI and complex system venues such as Complexity and In-
ternational Federation for Information Processing (IFIP); 17.46%
papers are published in Others venues which contains inde-
pendently published papers and dissertations; 1.59% papers
are published on arXiv. The distribution of different venues
shows that cryptocurrency trading is mostly published in Fi-
nance and Economics venues, but with a wide diversity oth-
erwise.
4.4. Survey Organisation
We discuss the contributions of the collected papers and
a statistical analysis of these papers in the remainder of the
paper, according to Table 4.
The papers in our collection are organised and presented
from six angles. We introduce the work about several dif-
ferent cryptocurrency trading software systems in Section
5. Section 6introduces systematic trading applied to cryp-
tocurrency trading. In Section 7, we introduce some emer-
gent trading technologies including econometrics on cryp-
tocurrencies, machine learning technologies and other emer-
First Author et al. Page 7 of 30
Cryptocurrency Trading: A Comprehensive Survey
Table 4
Review Schema
Classification Sec Topic
Cryptocurrency Trading Software System
5.1 Trading Infrastructure System
5.2 Real-time Cryptocurrency Trading System
5.3 Tur tle trading system in Cryptocurrency market
5.4 Arbitrage Trading Systems for Cryptocurrencies
5.5 Comparison of three cryptocurrency trading systems
Systematic Trading
6.1 Technical Analysis
6.2 Pairs Trading
6.3 Others
Emergent Trading Technologies
7.1 Econometrics on cryptocurrency
7.2 Machine learning technology
7.3 Others
Portfolio and Cryptocurrency Assets 8.1 Research among cryptocurrency pairs and related factors
8.2 Crypto-asset portfolio research
Market condition research 9.1 Bubbles and crash analysis
9.2 Extreme condition
Others 10 Others related to Cryptocurrency Trading
Summary Analysis of Literature Review
11.1 Timeline
11.2 Research distribution among properties
11.3 Research distribution among categories and technologies
11.4 Datasets used in cryptocurrency trading
Figure 7: Publication Venue Distribution
gent trading technologies in the cryptocurrency market. Sec-
tion 8introduces research on cryptocurrency pairs and re-
lated factors and crypto-asset portfolios research. In Sec-
tion 9we discuss cryptocurrency market condition research,
including bubbles, crash analysis, and extreme conditions.
Section 10 introduces other research included in cryptocur-
rency trading not covered above.
We would like to emphasize that the six headings above
focus on a particular aspect of cryptocurrency trading; we
give a complete organisation of the papers collected under
each heading. This implies that those papers covering more
than one aspect will be discussed in different sections, once
from each angle.
We analyse and compare the number of research papers
on different cryptocurrency trading properties and technolo-
gies in Section 11, where we also summarise the datasets
and the timeline of research in cryptocurrency trading.
We build upon this review to conclude in Section 12 with
some opportunities for future research.
5. Cryptocurrency Trading Software Systems
5.1. Trading Infrastructure Systems
Following the development of computer science and cryp-
tocurrency trading, many cryptocurrency trading systems/bots
have been developed. Table 5compares the cryptocurrency
trading systems existing in the market. The table is sorted
based on URL types (GitHub or Official website) and GitHub
stars (if appropriate).
Capfolio is a proprietary payable cryptocurrency trading
system which is a professional analysis platform and has an
advanced backtesting engine [51]. It supports five different
cryptocurrency exchanges.
3 Commas is a proprietary payable cryptocurrency trad-
ing system platform that can take profit and stop-loss orders
at the same time [1]. Twelve different cryptocurrency ex-
changes are compatible with this system.
CCXT is a cryptocurrency trading system with a unified
API out of the box and optional normalized data and sup-
ports many Bitcoin / Ether / Altcoin exchange markets and
merchant APIs. Any trader or developer can create a trading
strategy based on this data and access public transactions
through the APIs [61]. The CCXT library is used to con-
nect and trade with cryptocurrency exchanges and payment
processing services worldwide. It provides quick access to
market data for storage, analysis, visualisation, indicator de-
velopment, algorithmic trading, strategy backtesting, auto-
mated code generation and related software engineering. It
is designed for coders, skilled traders, data scientists and fi-
nancial analysts to build trading algorithms. Current CCXT
features include:
I. Support for many cryptocurrency exchanges;
II. Fully implemented public and private APIs;
III. Optional normalized data for cross-exchange analysis
and arbitrage;
First Author et al. Page 8 of 30
Cryptocurrency Trading: A Comprehensive Survey
IV. Out-of-the-box unified API, very easy to integrate.
Blackbird Bitcoin Arbitrage is a C++ trading system
that automatically executes long / short arbitrage between
Bitcoin exchanges. It can generate market-neutral strate-
gies that do not transfer funds between exchanges [36]. The
motivation behind Blackbird is to naturally profit from these
temporary price differences between different exchanges while
being market neutral. Unlike other Bitcoin arbitrage sys-
tems, Blackbird does not sell but actually short sells Bitcoin
on the short exchange. This feature offers two important
advantages. Firstly, the strategy is always market agnostic:
fluctuations (rising or falling) in the Bitcoin market will not
affect the strategy returns. This eliminates the huge risks of
this strategy. Secondly, this strategy does not require trans-
ferring funds (USD or BTC) between Bitcoin exchanges.
Buy and sell transactions are conducted in parallel on two
different exchanges. There is no need to deal with transmis-
sion delays.
StockSharp is an open-source trading platform for trad-
ing at any market of the world including 48 cryptocurrency
exchanges [227]. It has a free C# library and free trading
charting application. Manual or automatic trading (algo-
rithmic trading robot, regular or HFT) can be run on this
platform. StockSharp consists of five components that offer
different features:
I. S#.Designer - Free universal algorithm strategy app,
easy to create strategies;
II. S#.Data - free software that can automatically load and
store market data;
III. S#.Terminal - free trading chart application (trading
terminal);
IV. S#.Shell - ready-made graphics framework that can be
changed according to needs and has a fully open source
in C#;
V. S#.API - a free C# library for programmers using Vi-
sual Studio. Any trading strategies can be created in
S#.API.
Freqtrade is a free and open-source cryptocurrency trad-
ing robot system written in Python. It is designed to support
all major exchanges and is controlled by telegram. It con-
tains backtesting, mapping and money management tools,
and strategy optimization through machine learning [108].
Freqtrade has the following features:
I. Persistence: Persistence is achieved through SQLite
technology;
II. Strategy optimization through machine learning: Use
machine learning to optimize your trading strategy pa-
rameters with real trading data;
III. Marginal Position Size: Calculates winning rate, risk-
return ratio, optimal stop loss and adjusts position size,
and then trades positions for each specific market;
IV. Telegram management: use telegram to manage the
robot.
V. Dry run: Run the robot without spending money;
CryptoSignal is a professional technical analysis cryp-
tocurrency trading system [86]. Investors can track over 500
coins of Bittrex, Bitfinex, GDAX, Gemini and more. Auto-
mated technical analysis includes momentum, RSI, Ichimoku
Cloud, MACD, etc. The system gives alerts including Email,
Slack, Telegram, etc. CryptoSignal has two primary fea-
tures. First of all, it offers modular code for easy implemen-
tation of trading strategies; Secondly, it is easy to install with
Docker.
Ctubio is a C++ based low latency (high frequency)
cryptocurrency trading system [87]. This trading system can
place or cancel orders through supported cryptocurrency ex-
changes in less than a few milliseconds. Moreover, it pro-
vides a charting system that can visualise the trading ac-
count status including trades completed, target position for
fiat currency, etc.
Catalyst is an analysis and visualization of the cryp-
tocurrency trading system [57]. It makes trading strategies
easy to express and backtest them on historical data (daily
and minute resolution), providing analysis and insights into
the performance of specific strategies. Catalyst allows users
to share and organise data and build profitable, data-driven
investment strategies. Catalyst not only supports the trading
execution but also offers historical price data of all crypto
assets (from minute to daily resolution). Catalyst also has
backtesting and real-time trading capabilities, which enables
users to seamlessly transit between the two different trad-
ing modes. Lastly, Catalyst integrates statistics and machine
learning libraries (such as matplotlib, scipy, statsmodels and
sklearn) to support the development, analysis and visualiza-
tion of the latest trading systems.
Golang Crypto Trading Bot is a Go based cryptocur-
rency trading system [117]. Users can test the strategy in
sandbox environment simulation. If simulation mode is en-
abled, a fake balance for each coin must be specified for
each exchange.
5.2. Real-time Cryptocurrency Trading Systems
Amit et al. [21] developed a real-time Cryptocurrency
Trading System. A real-time cryptocurrency trading system
is composed of clients, servers and databases. Traders use
a web-application to login to the server to buy/sell crypto
assets. The server collects cryptocurrency market data by
creating a script that uses the Coinmarket API. Finally, the
database collects balances, trades and order book informa-
tion from the server. The authors tested the system with an
experiment that demonstrates user-friendly and secure expe-
riences for traders in the cryptocurrency exchange platform.
5.3. Turtle trading system in Cryptocurrency
market
The original Turtle Trading system is a trend following
trading system developed in the 1970s. The idea is to gen-
erate buy and sell signals on stock for short-term and long-
term breakouts and its cut-loss condition which is measured
by Average true range (ATR) [144]. The trading system will
adjust the size of assets based on their volatility. Essen-
tially, if a turtle accumulates a position in a highly volatile
First Author et al. Page 9 of 30
Cryptocurrency Trading: A Comprehensive Survey
Table 5
Comparison of existing cryptocurrency trading systems. #Exchange, Language, and
#Popularity denote the number of the exchanges that are supported by this software,
programming language used, and the popularity of the software (number of the stars in
Github).
Name Features #Exchange Language Open-Source URL #Popularity
Capfolio Professional analysis platform, 5 Not mentioned No Official website [51]
Advanced backtesting engine
3 Commas Simultaneous take profit and 12 Not mentioned No Official website [1]
stop loss orders
CCXT An out of the box unified API, 10 JavaScript / Python / PHP Yes GitHub [61] 13k
optional normalized data
BlackBird Strategy is market-neutral 8 C++ Yes GitHub [36] 4.7k
strategy not transfer funds between exchanges
StockSharp Free C# library, 48 C# Yes GitHub [227] 2.6k
free trading charting application
Freqtrade Strategy Optimization by machine learning, 2 Python Yes GitHub [108] 2.4k
Calculate edge position sizing
CryptoSignal Technical analysis trading system 4 Python Yes GitHub [86] 1.9k
Ctubio Low latency 1 C++ Yes GitHub [87] 1.7k
Catalyst Analysis and visualization of system 4 Python Yes GitHub [57]1.7k
seamless transition between live
and back-testing
GoLang Sandbox environment simulation 7 Go Yes GitHub [117] 277
market, it will be offset by a low volatility position. Ex-
tended Turtle Trading system is improved with smaller time
interval spans and introduces a new rule by using exponen-
tial moving average (EMA). Three EMA values are used to
trigger the “buy” signal: 30EMA (Fast), 60EMA (Slow),
100EMA (Long). The author of [144] performed backtest-
ing and comparing both trading systems (Original Turtle and
Extended Turtle) on 8 prominent cryptocurrencies. Through
the experiment, Original Turtle Trading System achieved an
18.59% average net profit margin (percentage of net profit
over total revenue) and 35.94% average profitability (per-
centage of winning trades over total numbers of trades) in
87 trades through nearly one year. Extended Turtle Trad-
ing System achieved 114.41% average net profit margin and
52.75% average profitability in 41 trades through the same
time interval. This research showed how Extended Turtle
Trading System compared can improve over Original Turtle
Trading System in trading cryptocurrencies.
5.4. Arbitrage Trading Systems for
Cryptocurrencies
Christian [205] introduced arbitrage trading systems for
cryptocurrencies. Arbitrage trading aims to spot the differ-
ences in price that can occur when there are discrepancies in
the levels of supply and demand across multiple exchanges.
As a result, a trader could realise a quick and low-risk profit
by buying from one exchange and selling at a higher price
on a different exchange. Arbitrage trading signals are caught
by automated trading software. The technical differences
between data sources impose a server process to be organ-
ised for each data source. Relational databases and SQL
are reliable solution due to the large amounts of relational
data. The author used the system to catch arbitrage oppor-
tunities on 25 May 2018 among 787 cryptocurrencies on 7
different exchanges. The research paper [205] listed the best
ten trading signals made by this system from 186 available
found signals. The results showed that the system caught
the trading signal of “BTG-BTC” to get a profit of up to
495.44% when arbitraging to buy in Cryptopia exchange
and sell in Binance exchange. Another three well-traded ar-
bitrage signals (profit expectation around 20% mentioned by
the author) were found on 25 May 2018. Arbitrage Trading
Software System introduced in that paper presented general
principles and implementation of arbitrage trading system
in the cryptocurrency market.
5.5. Comparison of three cryptocurrency trading
systems
Real-time trading systems use real-time functions to col-
lect data and generate trading algorithms. Turtle trading sys-
tem and arbitrage trading system have shown a sharp con-
trast in their profit and risk behaviour. Using Turtle trading
system in cryptocurrency markets got high returns with high
risk. Arbitrage trading system is inferior in terms of revenue
but also has a lower risk. One feature that turtle trading sys-
tem and arbitrage trading system have in common is they
performed well in capturing alpha.
6. Systematic Trading
6.1. Technical Analysis
Many researchers have focused on technical indicators
(patterns) analysis for trading on cryptocurrency markets.
Examples of studies with this approach include “Turtle Soup
pattern strategy” [233], “Nem (XEM) strategy” [236], “Amaz-
ing Gann Box strategy” [234], “Busted Double Top Pat-
tern strategy” [235], and “Bottom Rotation Trading strat-
egy” [237]. Table 6shows the comparison among these
five classical technical trading strategies using technical in-
dicators. “Turtle soup pattern strategy” [233] used a 2-day
breakout of price in predicting price trends of cryptocurren-
cies. This strategy is a kind of chart trading pattern. “Nem
(XEM) strategy” combined Rate of Change (ROC) indica-
tor and Relative Strength Index (RSI) in predicting price
trends [236]. “Amazing Gann Box” predicted exact points
of increase and decrease in Gann Box which are used to
First Author et al. Page 10 of 30
Cryptocurrency Trading: A Comprehensive Survey
Table 6
Comparison among five classical technical trading strategies
Technical trading strategy Core Methods Tecchnical tools/patterns
Tur tle Soup pattern [233] 2-daybreakout of price Chart trading patterns
Nem (XEM) [236] Price trends combined ROC & RSI Rate of Change indictor (ROC)
Relative strength index (RSI)
Amazing Gann Box [234]Predict exact points of rises and falls
in Gann Box (catch explosive trends)
Candlestick, boxcharts with
Fibonacci Retracement
Busted Double Top Pattern [235]Bearish reversal trading pattern that
generates a sell signal Price chart pattern
Bottom Rotation Trading [237]Pick the bottom before the reversal
happens Price chart pattern, box chart
catch explosive trends of cryptocurrency price [234]. Tech-
nical analysis tools such as candlestick and box charts with
Fibonacci Retracement based on golden ratio are used in this
technical analysis. Fibonacci Retracement uses horizontal
lines to indicate where possible support and resistance lev-
els are in the market. “Busted Double Top Pattern” used a
Bearish reversal trading pattern which generates a sell sig-
nal to predict price trends [235]. “Bottom Rotation Trading”
is a technical analysis method that picks the bottom before
the reversal happens. This strategy used a price chart pattern
and box chart as technical analysis tools.
Sungjoo et al. [122] investigated using genetic program-
ming (GP) to find attractive technical patterns in the cryp-
tocurrency market. Over 12 technical indicators including
Moving Average (MA) and Stochastic oscillator were used
in experiments; adjusted gain, match count, relative mar-
ket pressure and diversity measures have been used to quan-
tify the attractiveness of technical patterns. With extended
experiments, the GP system is shown to find successfully
attractive technical patterns, which are useful for portfolio
optimization. Hudson et al. [130] applied almost 15,000
to technical trading rules (classified into MA rules, filter
rules, support resistance rules, oscillator rules and channel
breakout rules). This comprehensive study found that tech-
nical trading rules provide investors with significant pre-
dictive power and profitability. Corbet et al. [82] analysed
various technical trading rules in the form of the moving
average-oscillator and trading range break-out strategies to
generate higher returns in cryptocurrency markets. By using
one-minute dollar-denominated Bitcoin close-price data, the
backtest showed variable-length moving average (VMA) rule
performs best considering it generates the most useful sig-
nals in high frequency trading.
6.2. Pairs Trading
Pairs trading is a trading strategy that attempts to ex-
ploit the mean-reversion between the prices of certain secu-
rities. Miroslav [105] investigated the applicability of stan-
dard pairs trading approaches on cryptocurrency data with
the benchmarks of Gatev et al. [115]. The pairs trading strat-
egy is constructed in two steps. Firstly, suitable pairs with
a stable long-run relationship are identified. Secondly, the
long-run equilibrium is calculated and pairs trading strategy
is defined by the spread based on the values. The research
also extended intra-day pairs trading using high frequency
data. Overall, the model was able to achieve a 3% monthly
profit in Miroslav’s experiments [105]. Broek [47] ap-
plied pairs trading based on cointegration in cryptocurrency
trading and 31 pairs were found to be significantly cointe-
grated (within sector and cross-sector). By selecting four
pairs and testing over a 60-day trading period, the pairs trad-
ing strategy got its profitability from arbitrage opportunities,
which rejected the Efficient-market hypothesis (EMH) for
the cryptocurrency market. Lintihac et al [174] proposed an
optimal dynamic pair trading strategy model for a portfolio
of assets. The experiment used stochastic control techniques
to calculate optimal portfolio weights and correlated the re-
sults with several other strategies commonly used by practi-
tioners including static dual-threshold strategies. Thomas et
al. [171] proposed a pairwise trading model incorporating
time-varying volatility with constant elasticity of variance
type. The experiment calculated the best pair strategy by
using a finite difference method and estimated parameters
by generalised moment method.
6.3. Others
Other systematic trading methods in cryptocurrency trad-
ing mainly include informed trading. Using USD / BTC ex-
change rate trading data, Feng et al. [104] found evidence
of informed trading in the Bitcoin market in those quantiles
of the order sizes of buyer-initiated (seller-initiated) orders
are abnormally high before large positive (negative) events,
compared to the quantiles of seller-initiated (buyer-initiated)
orders; this study adopts a new indicator inspired by the vol-
ume imbalance indicator [93]. The evidence of informed
trading in the Bitcoin market suggests that investors profit
on their private information when they get information be-
fore it is widely available.
7. Emergent Trading Technologies
7.1. Econometrics on cryptocurrency
Copula-quantile causality analysis and Granger-causality
analysis are methods to investigate causality in cryptocur-
rency trading analysis. Bouri et al. [41] applied a copula-
quantile causality approach on volatility in the cryptocur-
rency market. The approach of the experiment extended the
Copula-Granger-causality in distribution (CGCD) method
of Lee and Yang [170] in 2014. The experiment constructed
two tests of CGCD using copula functions. The paramet-
First Author et al. Page 11 of 30
Cryptocurrency Trading: A Comprehensive Survey
ric test employed six parametric copula functions to dis-
cover dependency density between variables. The perfor-
mance matrix of these functions varies with independent
copula density. Three distribution regions are the focus of
this research: left tail (1%, 5%, 10% quantile), central re-
gion (40%, 60% quantile and median) and right tail (90%,
95%, 99% quantile). The study provided significant evi-
dence of Granger causality from trading volume to the re-
turns of seven large cryptocurrencies on both left and right
tails. Elie et al. [42] examined the causal linkages among
the volatility of leading cryptocurrencies via the frequency-
domain test of Bodart and Candelon [38] and distinguished
between temporary and permanent causation. The results
showed that permanent shocks are more important in ex-
plaining Granger causality whereas transient shocks dom-
inate the causality of smaller cryptocurrencies in the long
term. Badenhorst [13] attempted to reveal whether spot
and derivative market volumes affect Bitcoin price volatility
with the Granger-causality method and ARCH (1,1). The
result shows spot trading volumes have a significant posi-
tive effect on price volatility while the relationship between
cryptocurrency volatility and the derivative market is uncer-
tain. Elie et al. [45] used a dynamic equicorrelation (DECO)
model and reported evidence that the average earnings equi-
librium correlation changes over time between the 12 lead-
ing cryptocurrencies. The results showed increased cryp-
tocurrency market consolidation despite significant price de-
clined in 2018. Furthermore, measurement of trading vol-
ume and uncertainty are key determinants of integration.
Several econometrics methods in time-series research,
such as GARCH and BEKK, have been used in the liter-
ature on cryptocurrency trading. Conrad et al. [81] used
the GARCH-MIDAS model to extract long and short-term
volatility components of the Bitcoin market. The technical
details of this model decomposed the conditional variance
into the low-frequency and high-frequency components. The
results identified that S&P 500 realized volatility has a nega-
tive and highly significant effect on long-term Bitcoin volatil-
ity and S&P 500 volatility risk premium has a significantly
positive effect on long-term Bitcoin volatility. Ardia et al. [8]
used the Markov Switching GARCH (MSGARCH) model
to test the existence of institutional changes in the GARCH
volatility dynamics of Bitcoin’s logarithmic returns. More-
over, a Bayesian method was used for estimating model pa-
rameters and calculating VaR prediction. The results showed
that MSGARCH models clearly outperform single-regime
GARCH for Value-at-Risk forecasting. Troster et al. [239]
performed general GARCH and GAS (Generalized Auto-
regressive Score) analysis to model and predict Bitcoin’s re-
turns and risks. The experiment found that the GAS model
with heavy-tailed distribution can provide the best out-of-
sample prediction and goodness-of-fit attributes for Bitcoin’s
return and risk modeling. The results also illustrated the
importance of modeling excess kurtosis for Bitcoin returns.
Charles et al. [65] studied four cryptocurrency markets in-
cluding Bitcoin, Dash, Litecoin and Ripple. Results showed
cryptocurrency returns are strongly characterised by the pres-
ence of jumps as well as structural breaks except the Dash
market. Four GARCH-type models (i.e., GARCH, APARCH,
IGARCH and FIGARCH) and three return types with struc-
tural breaks (original returns, jump-filtered returns, and jump-
filtered returns) are considered. The research indicated the
importance of jumps in cryptocurrency volatility and struc-
tural breakthroughs.
Some researchers focused on long memory methods for
volatility in cryptocurrency markets. Long memory meth-
ods focused on long-range dependence and significant long-
term correlations among fluctuations on markets. Chaim et
al. [63] estimated a multivariate stochastic volatility model
with discontinuous jumps in cryptocurrency markets. The
results showed that permanent volatility appears to be driven
by major market developments and popular interest levels.
Caporale et al. [52] examined persistence in the cryptocur-
rency market by Rescaled range (R/S) analysis and frac-
tional integration. The results of the study indicated that
the market is persistent (there is a positive correlation be-
tween its past and future values) and that its level changes
over time. Khuntin et al. [154] applied the adaptive market
hypothesis (AMH) in the predictability of Bitcoin evolving
returns. The consistent test of Dominguez and Lobato [89],
generalized spectral (GS) of Escanciano and Velasco [98]
are applied in capturing time-varying linear and nonlinear
dependence in bitcoin returns. The results verified Evolving
Efficiency in Bitcoin price changes and evidence of dynamic
efficiency in line with AMH’s claims.
Katsiampa et al. [150] applied three pair-wise bivariate
BEKK models to examine the conditional volatility dynam-
ics along with interlinkages and conditional correlations be-
tween three pairs of cryptocurrencies in 2018. More specifi-
cally, the BEKK-MGARCH methodology also captured cross-
market effects of shocks and volatility, which are also known
as shock transmission effects and volatility spillover effects.
The experiment found evidence of bi-directional shock trans-
mission effects between Bitcoin and both Ether and Lit-
coin. In particular, bi-directional shock spillover effects are
identified between three pairs (Bitcoin, Ether and Litcoin)
and time-varying conditional correlations exist with positive
correlations mostly prevailing. In 2019, Katsiampa [149]
further researched an asymmetric diagonal BEKK model
to examine conditional variances of five cryptocurrencies
that are significantly affected by both previous squared er-
rors and past conditional volatility. The experiment tested
the null hypothesis of the unit root against the stationar-
ity hypothesis. Once stationarity is ensured, ARCH LM
is tested for ARCH effects to examine the requirement of
volatility modeling in return series. Moreover, volatility
co-movements among cryptocurrency pairs are also tested
by the multivariate GARCH model. The results confirmed
the non-normality and heteroskedasticity of price returns in
cryptocurrency markets. The finding also identified the ef-
fects of cryptocurrencies’ volatility dynamics due to major
news. Hultman [131] set out to examine GARCH (1,1),
bivariate-BEKK (1,1) and a standard stochastic model to
forecast the volatility of Bitcoin. A rolling window approach
First Author et al. Page 12 of 30
Cryptocurrency Trading: A Comprehensive Survey
is used in these experiments. Mean absolute error (MAE),
Mean squared error (MSE) and Root-mean-square deviation
(RMSE) are three loss criteria adopted to evaluate the de-
gree of error between predicted and true values. The re-
sult shows the following rank of loss functions: GARCH
(1,1) > bivariate-BEKK (1,1) > Standard stochastic for all
the three different loss criteria; in other words, GARCH(1,1)
appeared best in predicting the volatility of Bitcoin. Wavelet
time-scale persistence analysis is also applied in the predic-
tion and research of volatility in cryptocurrency markets [202].
The results showed that information efficiency (efficiency)
and volatility persistence in the cryptocurrency market are
highly sensitive to time scales, measures of returns and volatil-
ity, and institutional changes. Adjepong et al. [202] con-
nected with similar research by Corbet et al. [85] and showed
that GARCH is quicker than BEKK to absorb new informa-
tion regarding the data.
7.2. Machine Learning Technology
As we have previously stated, Machine learning technol-
ogy constructs computer algorithms that automatically im-
prove themselves by finding patterns in existing data with-
out explicit instructions [128]. The rapid development of
machine learning in recent years has promoted its applica-
tion to cryptocurrency trading, especially in the prediction
of cryptocurrency returns.
7.2.1. Common Machine Learning Technology in this
survey
Several machine learning technologies are applied in cryp-
tocurrency trading. We distinguish these by the objective set
to the algorithm: classification, clustering, regression, rein-
forcement learning. We have separated a section specifically
on deep learning due to its intrinsic variation of techniques
and wide adoption.
Classification Algorithms. Classification in machine
learning has the objective of categorising incoming objects
into different categories as needed, where we can assign
labels to each category (e.g., up and down). Naive Bayes
(NB) [216], Support Vector Machine (SVM) [247], K-Nearest
Neighbours (KNN) [247], Decision Tree (DT) [109], Ran-
dom Forest (RF) [173] and Gradient Boosting (GB) [111]
algorithms habe been used in cryptocurrency trading based
on papers we collected. NB is a probabilistic classifier based
on Bayes’ theorem with strong (naive) conditional indepen-
dence assumptions between features [216]. SVM is a su-
pervised learning model that aims at achieving high mar-
gin classifiers connecting to learning bounds theory [256].
SVMs assign new examples to one category or another, mak-
ing it a non-probabilistic binary linear classifier [247], al-
though some corrections can make a probabilistic interpre-
tation of their output [153]. KNN is a memory-based or
lazy learning algorithm, where the function is only approx-
imated locally, and all calculations are being postponed to
inference time [247]. DT is a decision support tool algo-
rithm that uses a tree-like decision graph or model to seg-
ment input patterns into regions to then assign an associ-
ated label to each region [109]. RF is an ensemble learn-
ing method. The algorithm operates by constructing a large
number of decision trees during training and outputting the
average consensus as predicted class in the case of classi-
fication or mean prediction value in the case of regression
[173]. GB produces a prediction model in the form of an
ensemble of weak prediction models [111].
Clustering Algorithms. Clustering is a machine learn-
ing technique that involves grouping data points in a way
that each group shows some regularity [137]. K-Means is a
vector quantization used for clustering analysis in data min-
ing. K-means stores the 𝑘-centroids used to define the clus-
ters; a point is considered to be in a particular cluster if it is
closer to the cluster’s centroid than any other centroid [245].
K-Means is one of the most used clustering algorithms used
in cryptocurrency trading according to the papers we col-
lected.
Regression Algorithms. We have defined regression
as any statistical technique that aims at estimating a con-
tinuous value [164]. Linear Regression (LR) and Scatter-
plot Smoothing are common techniques used in solving re-
gression problems in cryptocurrency trading. LR is a lin-
ear method used to model the relationship between a scalar
response (or dependent variable) and one or more explana-
tory variables (or independent variables) [164]. Scatterplot
Smoothing is a technology to fit functions through scatter
plots to best represent relationships between variables [110].
Deep Learning Algorithms. Deep learning is a modern
take on artificial neural networks (ANNs) [257], made pos-
sible by the advances in computational power. An ANN is
a computational system inspired by the natural neural net-
works that make up the animal’s brain. The system “learns”
to perform tasks including the prediction by considering ex-
amples. Deep learning’s superior accuracy comes from high
computational complexity cost. Deep learning algorithms
are currently the basis for many modern artificial intelli-
gence applications [231]. Convolutional neural networks
(CNNs) [168], Recurrent neural networks (RNNs) [188],
Gated recurrent units (GRUs) [70], Multilayer perceptron
(MLP) and Long short-term memory (LSTM) [67] networks
are the most common deep learning technologies used in
cryptocurrency trading. A CNN is a specific type of neu-
ral network layer commonly used for supervised learning.
CNNs have found their best success in image processing
and natural language processing problems. An attempt to
use CNNs in cryptocurrency can be shown in [143]. An
RNN is a type of artificial neural network in which con-
nections between nodes form a directed graph with possi-
ble loops. This structure of RNNs makes them suitable for
processing time-series data [188] due to the introduction of
memory in the recurrent connections. They face neverthe-
less for the vanishing gradients problem [203] and so dif-
ferent variations have been recently proposed. LSTM [67]
is a particular RNN architecture widely used. LSTMs have
shown to be superior to nongated RNNs on financial time-
series problems because they have the ability to selectively
remember patterns for a long time. A GRU [70] is another
gated version of the standard RNN which has been used in
First Author et al. Page 13 of 30
Cryptocurrency Trading: A Comprehensive Survey
crypto trading [91]. Another deep learning technology used
in cryptocurrency trading is Seq2seq, which is a specific
implementation of the Encoder–Decoder architecture [251].
Seq2seq was first aimed at solving natural language process-
ing problems but has been also applied it in cryptocurrency
trend predictions in [226].
Reinforcement Learning Algorithms. Reinforcement
learning (RL) is an area of machine learning leveraging the
idea that software agents act in the environment to maximize
a cumulative reward [230]. Deep Q-Learning (DQN) [120]
and Deep Boltzmann Machine (DBM) [219] are common
technologies used in cryptocurrency trading using RL. Deep
Q learning uses neural networks to approximate Q-value
functions. A state is given as input, and Q values for all pos-
sible actions are generated as outputs [120]. DBM is a type
of binary paired Markov random field (undirected probabil-
ity graphical model) with multiple layers of hidden random
variables [219]. It is a network of randomly coupled random
binary units.
7.2.2. Research on Machine Learning Models
In the development of machine learning trading signals,
technical indicators have usually been used as input fea-
tures. Nakano et al. [193] explored Bitcoin intraday tech-
nical trading based on ANNs for return prediction. The ex-
periment obtained medium frequency price and volume data
(time interval of data is 15min) of Bitcoin from a cryptocur-
rency exchange. An ANN predicts the price trends (up and
down) in the next period from the input data. Data is pre-
processed to construct a training dataset that contains a ma-
trix of technical patterns including EMA, Emerging Markets
Small Cap (EMSD), relative strength index (RSI), etc. Their
numerical experiments contain different research aspects in-
cluding base ANN research, effects of different layers, ef-
fects of different activation functions, different outputs, dif-
ferent inputs and effects of additional technical indicators.
The results have shown that the use of various technical in-
dicators possibly prevents over-fitting in the classification
of non-stationary financial time-series data, which enhances
trading performance compared to the primitive technical trad-
ing strategy. (Buy-and-Hold is the benchmark strategy in
this experiment.)
Some classification and regression machine learning mod-
els are applied in cryptocurrency trading by predicting price
trends. Most researchers have focused on the comparison
of different classification and regression machine learning
methods. Sun et al. [229] used random forests (RFs) with
factors in Alpha01 [141] (capturing features from the his-
tory of the cryptocurrency market) to build a prediction model.
The experiment collected data from API in cryptocurrency
exchanges and selected 5-minute frequency data for back-
testing. The results showed that the performances are pro-
portional to the amount of data (more data, more accurate)
and the factors used in the RF model appear to have different
importance. For example, “Alpha024” and “Alpha032” fea-
tures appeared as the most important in the model adopted.
(The alpha features come from paper “101 Formulaic Al-
phas" [141].) Vo et al. [243] applied RFs in High-Frequency
cryptocurrency Trading (HFT) and compared it with deep
learning models. Minute-level data is collected when util-
ising a forward fill imputation method to replace the NULL
value (i.e., a missing value). Different periods and RF trees
are tested in the experiments. The authors also compared F-
1 precision and recall metrics between RF and Deep Learn-
ing (DL). The results showed that RF is effective despite
multicollinearity occurring in ML features, the lack of model
identification also potentially leading to model identifica-
tion issues; this research also attempted to create an HFT
strategy for Bitcoin using RF. Maryna et al. [260] inves-
tigated the profitability of an algorithmic trading strategy
based on training an SVM model to identify cryptocurren-
cies with high or low predicted returns. The results showed
that the performance of the SVM strategy was the fourth be-
ing better only than S&P B&H strategy, which simply buys-
and-hold the S&P index. (There are other 4 benchmark
strategies in this research.)The authors observed that SVM
needs a large number of parameters and so is very prone
to overfitting, which caused its bad performance. Barnwal
et al. [18] used generative and discriminative classifiers to
create a stacking model, particularly 3 generative and 6 dis-
criminative classifiers combined by a one-layer Neural Net-
work, to predict the direction of cryptocurrency price. A
discriminative classifier directly models the relationship be-
tween unknown and known data, while generative classifiers
model the prediction indirectly through the data generation
distribution [198]. Technical indicators including trend, mo-
mentum, volume and volatility, are collected as features of
the model. The authors discussed how different classifiers
and features affect the prediction. Attanasio et al. [10] com-
pared a variety of classification algorithms including SVM,
NB and RF in predicting next-day price trends of a given
cryptocurrency. The results showed that due to the hetero-
geneity and volatility of cryptocurrencies’ financial instru-
ments, forecasting models based on a series of forecasts ap-
peared better than a single classification technology in trad-
ing cryptocurrencies. Madan et al. [179] modeled the Bit-
coin price prediction problem as a binomial classification
task, experimenting with a custom algorithm that leverages
both random forests and generalized linear models. Daily
data, 10-minute data and 10-second data are used in the
experiments. The experiments showed that 10-minute data
gave a better sensitivity and specificity ratio than 10-second
data (10-second prediction achieved around 10% accuracy).
Considering predictive trading, 10-minute data helped show
clearer trends in the experiment compared to 10-second back-
testing. Similarly, Virk [242] compared RF, SVM, GB and
LR to predict the price of Bitcoin. The results showed that
SVM achieved the highest accuracy of 62.31% and preci-
sion value 0.77 among binomial classification machine learn-
ing algorithms.
Different deep learning models have been used in find-
ing patterns of price movements in cryptocurrency markets.
Zhengy et al. [258] implemented two machine learning mod-
els, fully-connected ANN and LSTM to predict cryptocur-
First Author et al. Page 14 of 30
Cryptocurrency Trading: A Comprehensive Survey
rency price dynamics. The results showed that ANN, in
general, outperforms LSTM although theoretically, LSTM
is more suitable than ANN in terms of modeling time series
dynamics; the performance measures considered are MAE
and RMSE in joint prediction (five cryptocurrencies daily
prices prediction). The findings show that the future state of
a time series for cryptocurrencies is highly dependent on its
historic evolution. Kwon et al. [165] used an LSTM model,
with a three-dimensional price tensor representing the past
price changes of cryptocurrencies as input. This model out-
performs the GB model in terms of F1-score. Specifically,
it has a performance improvement of about 7% over the GB
model in 10-minute price prediction. In particular, the ex-
periments showed that LSTM is more suitable when classi-
fying cryptocurrency data with high volatility. Alessandretti
et al. [5] tested Gradient boosting decision trees (includ-
ing single regression and XGBoost-augmented regression)
and the LSTM model on forecasting daily cryptocurrency
prices. They found methods based on gradient boosting de-
cision trees worked best when predictions were based on
short-term windows of 5/10 days while LSTM worked best
when predictions were based on 50 days of data. The rel-
ative importance of the features in both models are com-
pared and an optimised portfolio composition (based on ge-
ometric mean return and Sharpe ratio) is discussed in this
paper. Phaladisailoed et al. [207] chose regression mod-
els (Theil-Sen Regression and Huber Regression) and deep
learning-based models (LSTM and GRU) to compare the
performance of predicting the rise and fall of Bitcoin price.
In terms of two common measure metrics, MSE and R-
Square (R2), GRU shows the best accuracy. Fan et al. [100]
applied an autoencoder-augmented LSTM structure in pre-
dicting the mid-price of 8 cryptocurrency pairs. Level-2
limit order book live data is collected and the experiment
achieved 78% accuracy of price movements prediction in
high frequency trading (tick level). This research improved
and verified the view of Sirignano et al. [224] that univer-
sal models have better performance than currency-pair spe-
cific models for cryptocurrency markets. Moreover, “Walk-
through” (i.e., retrain the original deep learning model it-
self when it appears to no longer be valid) is proposed as
a method to optimise the training of a deep learning model
and shown to significantly improve the prediction accuracy.
Researchers have also focused on comparing classical
statistical models and machine/deep learning models. Rane
et al. [214] described classical time series prediction meth-
ods and machine learning algorithms used for predicting
Bitcoin price. Statistical models such as Autoregressive In-
tegrated Moving Average models (ARIMA), Binomial Gen-
eralized Linear Model and GARCH are compared with ma-
chine learning models such as SVM, LSTM and Non-linear
Auto-Regressive with Exogenous Input Model (NARX). The
observation and results showed that the NARX model is the
best model with nearly 52% predicting accuracy based on
10 seconds interval. Rebane et al. [215] compared tradi-
tional models like ARIMA with a modern popular model
like seq2seq in predicting cryptocurrency returns. The re-
sult showed that the seq2seq model exhibited demonstra-
ble improvement over the ARIMA model for Bitcoin-USD
prediction but the seq2seq model showed very poor perfor-
mance in extreme cases. The authors proposed performing
additional investigations, such as the use of LSTM instead
of GRU units to improve the performance. Similar models
were also compared by Stuerner et al. [228] who explored
the superiority of automated investment approach in trend
following and technical analysis in cryptocurrency trading.
Samuel et al. [206] explored the vector autoregressive model
(VAR model), a more complex RNN, and a hybrid of the two
in residual recurrent neural networks (R2N2) in predicting
cryptocurrency returns. The RNN with ten hidden layers is
optimised for the setting and the neural network augmented
by VAR allows the network to be shallower, quicker and
to have a better prediction than an RNN. RNN, VAR and
R2N2 models are compared. The results showed that the
VAR model has phenomenal test period performance and
thus props up the R2N2 model, while the RNN performs
poorly. This research is an attempt at optimisation of model
design and applying to the prediction on cryptocurrency re-
turns.
7.2.3. Sentiment Analysis
Sentiment analysis, a popular research topic in the age of
social media, has also been adopted to improve predictions
for cryptocurrency trading. This data source typically has to
be combined with Machine Learning for the generation of
trading signals.
Lamon et al. [167] used daily news and social media
data labeled on actual price changes, rather than on positive
and negative sentiment. By this approach, the prediction
on price is replaced with positive and negative sentiment.
The experiment acquired cryptocurrency-related news arti-
cle headlines from the website like “cryptocoinsnews” and
twitter API. Weights are taken in positive and negative words
in the cryptocurrency market. Authors compared Logistic
Regression (LR), Linear Support Vector Machine (LSVM)
and NB as classifiers and concluded that LR is the best
classifier in daily price prediction with 43.9% of price in-
creases correctly predicted and 61.9% of price decreases
correctly forecasted. Smuts [225] conducted a similar bi-
nary sentiment-based price prediction method with an LSTM
model using Google Trends and Telegram sentiment. In
detail, the sentiment was extracted from Telegram by us-
ing a novel measure called VADER [132]. The backtest-
ing reached 76% accuracy on the test set during the first
half of 2018 in predicting hourly prices. Nasir et al. [195]
researched the relationship between cryptocurrency returns
and search engines. The experiment employed a rich set
of established empirical approaches including VAR frame-
work, copulas approach and non-parametric drawings of time
series. The results found that Google searches exert signif-
icant influence on Bitcoin returns, especially in the short-
term intervals. Kristoufek [162] discussed positive and neg-
ative feedback on Google trends or daily views on Wikipedia.
The author mentioned different methods including Cointe-
First Author et al. Page 15 of 30
Cryptocurrency Trading: A Comprehensive Survey
gration, Vector autoregression and Vector error-correction
model to find causal relationships between prices and searched
terms in the cryptocurrency market. The results indicated
that search trends and cryptocurrency prices are connected.
There is also a clear asymmetry between the effects of in-
creased interest in currencies above or below their trend val-
ues from the experiment. Young et al. [156] analysed user
comments and replies in online communities and their con-
nection with cryptocurrency volatility. After crawling com-
ments and replies in online communities, authors tagged the
extent of positive and negative topics. Then the relation-
ship between price and the number of transactions of cryp-
tocurrency is tested according to comments and replies to
selected data. At last, a prediction model using machine
learning based on selected data is created to predict fluctu-
ations in the cryptocurrency market. The results show the
amount of accumulated data and animated community ac-
tivities exerted a direct effect on fluctuation in the price and
volume of a cryptocurrency.
Phillips et al. [212] applied dynamic topic modeling and
Hawkes model to decipher relationships between topics and
cryptocurrency price movements. The authors used Latent
Dirichlet allocation (LDA) model for topic modeling, which
assumes each document contains multiple topics to different
extents. The experiment showed that particular topics tend
to precede certain types of price movements in the cryp-
tocurrency market and the authors proposed the relation-
ships could be built into real-time cryptocurrency trading.
Li et al. [172] analysed Twitter sentiment and trading vol-
ume and an Extreme Gradient Boosting Regression Tree
Model in the prediction of ZClassic (ZCL) cryptocurrency
market. Sentiment analysis using natural language process-
ing from the Python package “Textblob” assigns impactful
words a polarity value. Values of weighted and unweighted
sentiment indices are calculated on an hourly basis by sum-
ming weights of coinciding tweets, which makes us com-
pare this index to ZCL price data. The model achieved
a Pearson correlation of 0.806 when applied to test data,
yielding a statistical significance at the 𝑝 < 0.0001 level.
Flori [107] relied on a Bayesian framework that combines
market-neutral information with subjective beliefs to con-
struct diversified investment strategies in the Bitcoin mar-
ket. The result shows that news and media attention seem to
contribute to influence the demand for Bitcoin and enlarge
the perimeter of the potential investors, probably stimulating
price euphoria and upwards-downwards market dynamics.
The authors’ research highlighted the importance of news
in guiding portfolio re-balancing. Elie et al. [39] compared
the ability of newspaper-based metrics and internet search-
based uncertainty metrics in predicting bitcoin returns. The
predictive power of Internet-based economic uncertainty-
related query indices is statistically stronger than that of
newspapers in predicting bitcoin returns.
Similarly, Colianni et al. [80], Garcia et al. [113], Za-
muda et al. [254] et al. used sentiment analysis technol-
ogy applying it in the cryptocurrency trading area and had
similar results. Colianni et al. [80] cleaned data and ap-
plied supervised machine learning algorithms such as logis-
tic regression, Naive Bayes and support vector machines,
etc. on Twitter Sentiment Analysis for cryptocurrency trad-
ing. Garcia et al. [113] applied multidimensional analy-
sis and impulse analysis in social signals of sentiment ef-
fects and algorithmic trading of Bitcoin. The results veri-
fied the long-standing assumption that transaction-based so-
cial media sentiment has the potential to generate a posi-
tive return on investment. Zamuda et al. [254] adopted new
sentiment analysis indicators and used multi-target portfo-
lio selection to avoid risks in cryptocurrency trading. The
perspective is rationalized based on the elastic demand for
computing resources of the cloud infrastructure. A gen-
eral model evaluating the influence between user’s network
Action-Reaction-Influence-Model (ARIM) is mentioned in
this research. Bartolucci et al. [19] researched cryptocur-
rency prices with the “Butterfly effect”, which means “is-
sues” of the open-source project provides insights to im-
prove prediction of cryptocurrency prices. Sentiment, po-
liteness, emotions analysis of GitHub comments are applied
in Ethereum and Bitcoin markets. The results showed that
these metrics have predictive power on cryptocurrency prices.
7.2.4. Reinforcement Learning
Deep reinforcement algorithms bypass prediction and
go straight to market management actions to achieve high
cumulated profit [126]. Bu et al. [49] proposed a combina-
tion of double Q-network and unsupervised pre-training us-
ing DBM to generate and enhance the optimal Q-function in
cryptocurrency trading. The trading model contains agents
in series in the form of two neural networks, unsupervised
learning modules and environments. The input market state
connects an encoding network which includes spectral fea-
ture extraction (convolution-pooling module) and temporal
feature extraction (LSTM module). A double-Q network
follows the encoding network and actions are generated from
this network. Compared to existing deep learning models
(LSTM, CNN, MLP, etc.), this model achieved the high-
est profit even facing an extreme market situation (recorded
24% of the profit while cryptocurrency market price drops
by -64%). Juchli [138] applied two implementations of rein-
forcement learning agents, a Q-Learning agent, which serves
as the learner when no market variables are provided, and
a DQN agent which was developed to handle the features
previously mentioned. The DQN agent was backtested un-
der the application of two different neural network architec-
tures. The results showed that the DQN-CNN agent (convo-
lutional neural network) is superior to the DQN-MLP agent
(multilayer perceptron) in backtesting prediction. Lucarelli
et al. [177] focused on improving automated cryptocurrency
trading with a deep reinforcement learning approach. Dou-
ble and Dueling double deep Q-learning networks are com-
pared for 4 years. By setting rewards functions as Sharpe
ratio and profit, the double Q-learning method demonstrated
to be the most profitable approach in trading cryptocurrency.
First Author et al. Page 16 of 30
Cryptocurrency Trading: A Comprehensive Survey
7.3. Others
Atsalakis et al. [9] proposes a computational intelligence
technique that uses a hybrid Neuro-Fuzzy controller, namely
PATSOS, to forecast the direction in the change of the daily
price of Bitcoin. The proposed methodology outperforms
two other computational intelligence models, the first be-
ing developed with a simpler neuro-fuzzy approach, and the
second being developed with artificial neural networks. Ac-
cording to the signals of the proposed model, the investment
return obtained through trading simulation is 71.21% higher
than the investment return obtained through a simple buy
and hold strategy. This application is proposed for the first
time in the forecasting of Bitcoin price movements. Topo-
logical data analysis is applied to forecasting price trends of
cryptocurrency markets in [155]. The approach is to har-
ness topological features of attractors of dynamical systems
for arbitrary temporal data. The results showed that the
method can effectively separate important topological pat-
terns and sampling noise (like bid–ask bounces, discrete-
ness of price changes, differences in trade sizes or infor-
mational content of price changes, etc.) by providing theo-
retical results. Kurbucz [163] designed a complex method
consisting of single-hidden layer feedforward neural net-
works (SLFNs) to (i) determine the predictive power of the
most frequent edges of the transaction network (a public
ledger that records all Bitcoin transactions) on the future
price of Bitcoin; and, (ii) to provide an efficient technique
for applying this untapped dataset in day trading. The re-
search found a significantly high accuracy (60.05%) for the
price movement classifications base on information that can
be obtained using a small subset of edges (approximately
0.45% of all unique edges). It is worth noting that, Kondor
et al. [157,159] firstly published some papers giving analy-
sis on transaction networks on cryptocurrency markets and
applied related research in identifying Bitcoin users [139].
Abay et al. [2] attempted to understand the network dynam-
ics behind the Blockchain graphs using topological features.
The results showed that standard graph features such as the
degree distribution of transaction graphs may not be suffi-
cient to capture network dynamics and their potential im-
pact on Bitcoin price fluctuations. Maurice et al [202] ap-
plied wavelet time-scale persistence in analysing returns and
volatility in cryptocurrency markets. The experiment exam-
ined the long-memory and market efficiency characteristics
in cryptocurrency markets using daily data for more than
two years. The authors employed a log-periodogram re-
gression method in researching stationarity in the cryptocur-
rency market and used ARFIMA-FIGARCH class of mod-
els in examining long-memory behaviour of cryptocurren-
cies across time and scale. In general, experiments indicated
that heterogeneous memory behaviour existed in eight cryp-
tocurrency markets using daily data over the full-time period
and across scales (August 25, 2015 to March 13, 2018).
8. Portfolio and Cryptocurrency Assets
8.1. Research among cryptocurrency pairs and
related factors
Ji et al. [135] examined connectedness via return and
volatility spillovers across six large cryptocurrencies (col-
lected from coinmarketcap lists from August 7 2015 to Febru-
ary 22 2018) and found Litecoin and Bitcoin to have the
most effect on other cryptocurrencies. The authors followed
methods of Diebold et al. [88] and built positive/negative re-
turns and volatility connectedness networks. Furthermore,
the regression model is used to identify drivers of various
cryptocurrency integration levels. Further analysis revealed
that the relationship between each cryptocurrency in terms
of return and volatility is not necessarily due to its mar-
ket size. Adjepong et al. [201] explored market coherence
and volatility causal linkages of seven leading cryptocurren-
cies. Wavelet-based methods are used to examine market
connectedness. Parametric and nonparametric tests are em-
ployed to investigate directions of volatility spillovers of the
assets. Experiments revealed from diversification benefits to
linkages of connectedness and volatility in cryptocurrency
markets. Elie et al. [43] found the presence of jumps was
detected in a series of 12 cryptocurrency returns, and signif-
icant jumping activity was found in all cases. More results
underscore the importance of the jump in trading volume for
the formation of cryptocurrency leapfrogging.
Some researchers explored the relationship between cryp-
tocurrency and different factors, including futures, gold, etc.
Hale et al. [123] suggested that Bitcoin prices rise and fall
rapidly after CME issues futures consistent with pricing dy-
namics. Specifically, the authors pointed out that the rapid
rise and subsequent decline in prices after the introduction
of futures is consistent with trading behaviour in the cryp-
tocurrency market. Werner et al. [161] focused on the asym-
metric interrelationships between major currencies and cryp-
tocurrencies. The results of multiple fractal asymmetric de-
trending cross-correlation analysis show evidence of signif-
icant persistence and asymmetric multiplicity in the cross-
correlation between most cryptocurrency pairs and ETF pairs.
Bai et al. [14] studied a trading algorithm for foreign ex-
change on a cryptocurrency Market using the Automated
Triangular Arbitrage method. Implementing a pricing strat-
egy, implementing trading algorithms and developing a given
trading simulation are three problems solved by this research.
Kang et al. [146] examined the hedging and diversification
properties of gold futures versus Bitcoin prices by using
dynamic conditional correlations (DCCs) and wavelet co-
herence. DCC-GARCH model [95] is used to estimate the
time-varying correlation between Bitcoin and gold futures
by modeling the variance and the co-variance but also this
two flexibility. Wavelet coherence method focused more on
co-movement between Bitcoin and gold futures. From ex-
periments, the wavelet coherence results indicated volatility
persistence, causality and phase difference between Bitcoin
and gold. Dyhrberg et al [92] applied the GARCH model
and the exponential GARCH model in analysing similarities
First Author et al. Page 17 of 30
Cryptocurrency Trading: A Comprehensive Survey
between Bitcoin, gold and the US dollar. The experiments
showed that Bitcoin, gold and the US dollar have similari-
ties with the variables of the GARCH model, have similar
hedging capabilities and react symmetrically to good and
bad news. The authors observed that Bitcoin can combine
some advantages of commodities and currencies in finan-
cial markets to be a tool for portfolio management. Baur et
al. [20] extended the research of Dyhrberg et al.; the same
data and sample periods are tested [92] with GARCH and
EGARCH-(1,1) models but the experiments reached differ-
ent conclusions. Baur et al. found that Bitcoin has unique
risk-return characteristics compared with other assets. They
noticed that Bitcoin excess returns and volatility resemble a
rather highly speculative asset with respect to gold or the
US dollar. Bouri et al. [40] studied the relationship be-
tween Bitcoin and energy commodities by applying DCCs
and GARCH (1,1) models. In particular, the results showed
that Bitcoin is a strong hedge and safe haven for energy
commodities. Kakushadze [142] proposed factor models for
the cross-section of daily cryptoasset returns and provided
source code for data downloads, computing risk factors and
backtesting for all cryptocurrencies and a host of various
other digital assets. The results showed that cross-sectional
statistical arbitrage trading may be possible for cryptoas-
sets subject to efficient executions and shorting. Beneki et
al. [25] tested hedging abilities between Bitcoin and Ethereum
by a multivariate BEKK-GARCH methodology and impulse
response analysis within VAR model. The results indicated
a volatility transaction from Ethereum to Bitcoin, which im-
plied possible profitable trading strategies on the cryptocur-
rency derivatives market. Guglielmo et al. [54] examined
the week effect in cryptocurrency markets and explored the
feasibility of this indicator in trading practice. Student 𝑡-test,
ANOVA, Kruskal–Wallis and Mann–Whitney tests were car-
ried out for cryptocurrency data in order to compare time
periods that may be characterised by anomalies with other
time periods. When an anomaly is detected, an algorithm
was established to exploit profit opportunities (MetaTrader
terminal in MQL4 is mentioned in this research). The re-
sults showed evidence of anomaly (abnormal positive re-
turns on Mondays) in the Bitcoin market by backtesting in
2013-2016.
8.2. Crypto-asset Portfolio Research
Some researchers applied portfolio theory for crypto as-
sets. Corbet et al. [83] gave a systematic analysis of cryp-
tocurrencies as financial assets. Brauneis et al. [46] ap-
plied the Markowitz mean-variance framework in order to
assess the risk-return benefits of cryptocurrency portfolios.
In an out-of-sample analysis accounting for transaction cost,
they found that combining cryptocurrencies enriches the set
of ‘low’-risk cryptocurrency investment opportunities. In
terms of the Sharpe ratio and certainty equivalent returns,
the 1∕𝑁-portfolio (i.e., “naive” strategies, such as equally
dividing amongst asset classes) outperformed single cryp-
tocurrencies and more than 75% in terms of the Sharpe ratio
and certainty equivalent returns of mean-variance optimal
portfolios. Castro et al. [56] developed a portfolio optimi-
sation model based on the Omega measure which is more
comprehensive than the Markowitz model and applied this
to four crypto-asset investment portfolios by means of a nu-
merical application. Experiments showed crypto-assets im-
proves the return of the portfolios, but on the other hand,
also increase the risk exposure.
Bedi et al. [22] examined diversification capabilities of
Bitcoin for a global portfolio spread across six asset classes
from the standpoint of investors dealing in five major fiat
currencies, namely US Dollar, Great Britain Pound, Euro,
Japanese Yen and Chinese Yuan. They employed modified
Conditional Value-at-Risk and standard deviation as mea-
sures of risk to perform portfolio optimisations across three
asset allocation strategies and provided insights into the sharp
disparity in Bitcoin trading volumes across national curren-
cies from a portfolio theory perspective. Similar research
has been done by Antipova et al. [7], which explored the
possibility of establishing and optimizing a global portfolio
by diversifying investments using one or more cryptocur-
rencies, and assessing returns to investors in terms of risks
and returns. Fantazzini et al. [102] proposed a set of models
that can be used to estimate the market risk for a portfo-
lio of crypto-currencies, and simultaneously estimate their
credit risk using the Zero Price Probability (ZPP) model.
The results revealed the superiority of the t-copula/skewed-t
GARCH model for market risk, and the ZPP-based models
for credit risk. Qiang et al. [134] examined the common
dynamics of bitcoin exchanges. Using a connectivity met-
ric based on the actual daily volatility of the bitcoin price,
they found that Coinbase is undoubtedly the market leader,
while Binance performance is surprisingly weak. The re-
sults also suggested that safer asset extraction is more im-
portant for volatility linkages between Bitcoin exchanges
relative to trading volumes.
Trucios et al. [240] proposed a methodology based on
vine copulas and robust volatility models to estimate the
Value-at-Risk (VaR) and Expected Shortfall (ES) of cryp-
tocurrency portfolios. The proposed algorithm displayed
good performance in estimating both VaR and ES. Hrytsiuk
et al. [129] showed that the cryptocurrency returns can be
described by the Cauchy distribution and obtained the an-
alytical expressions for VaR risk measures and performed
calculations accordingly. As a result of the optimisation,
the sets of optimal cryptocurrency portfolios were built in
their experiments.
Jiang et al. [136] proposed a two-hidden-layer CNN that
takes the historical price of a group of cryptocurrency assets
as an input and outputs the weight of the group of cryp-
tocurrency assets. This research focused on portfolio re-
search in cryptocurrency assets using emerging technolo-
gies like CNN. Training is conducted in an intensive man-
ner to maximise cumulative returns, which is considered a
reward function of the CNN network. The performance of
the CNN strategy is compared with the three benchmarks
and the other three portfolio management algorithms (buy
and hold strategy, Uniform Constant Rebalanced Portfolio
First Author et al. Page 18 of 30
Cryptocurrency Trading: A Comprehensive Survey
and Universal Portfolio with Online Newton Step and Pas-
sive Aggressive Mean Reversion); the results are positive
in that the model is only second to the Passive Aggressive
Mean Reversion algorithm (PAMR). Estalayo et al. [99] re-
ported initial findings around the combination of DL mod-
els and Multi-Objective Evolutionary Algorithms (MOEAs)
for allocating cryptocurrency portfolios. Technical rationale
and details were given on the design of a stacked DL recur-
rent neural network, and how its predictive power can be ex-
ploited for yielding accurate ex-ante estimates of the return
and risk of the portfolio. Results obtained for a set of exper-
iments carried out with real cryptocurrency data have veri-
fied the superior performance of their designed deep learn-
ing model with respect to other regression techniques.
9. Market Condition Research
9.1. Bubbles and Crash Analysis
Phillips and Yu proposed a methodology to test for the
presence of cryptocurrency bubble [68], which is extended
by Shaen et al. [84]. The method is based on supremum
Augmented Dickey–Fuller (SADF) to test for the bubble
through the inclusion of a sequence of forwarding recur-
sive right-tailed ADF unit root tests. An extended method-
ology generalised SADF (GSAFD), is also tested for bub-
bles within cryptocurrency data. The research concluded
that there is no clear evidence of a persistent bubble in cryp-
tocurrency markets including Bitcoin or Ethereum. Bouri
et al. [44] date-stamped price explosiveness in seven large
cryptocurrencies and revealed evidence of multiple periods
of explosivity in all cases. GSADF is used to identify mul-
tiple explosiveness periods and logistic regression is em-
ployed to uncover evidence of co-explosivity across cryp-
tocurrencies. The results showed that the likelihood of ex-
plosive periods in one cryptocurrency generally depends on
the presence of explosivity in other cryptocurrencies and
points toward a contemporaneous co-explosivity that does
not necessarily depend on the size of each cryptocurrency.
Extended research by Phillips et al. [208,209] (who ap-
plied a recursive augmented Dickey-Fuller algorithm, which
is called PSY test) and Landsnes et al. [97] studied pos-
sible predictors of bubble periods of certain cryptocurren-
cies. The evaluation includes multiple bubble periods in all
cryptocurrencies. The result shows that higher volatility and
trading volume is positively associated with the presence of
bubbles across cryptocurrencies. In terms of bubble predic-
tion, the authors found the probit model to perform better
than the linear models.
Phillips et al. [210] used Hidden Markov Model (HMM)
and Superiority and Inferiority Ranking (SIR) method to
identify bubble-like behaviour in cryptocurrency time se-
ries. Considering HMM and SIR method, an epidemic de-
tection mechanism is used in social media to predict cryp-
tocurrency price bubbles, which classify bubbles through
epidemic and non-epidemic labels. Experiments have demon-
strated a strong relationship between Reddit usage and cryp-
tocurrency prices. This work also provides some empirical
evidence that bubbles mirror the social epidemic-like spread
of an investment idea. Guglielmo et al. [53] examined the
price overreactions in the case of cryptocurrency trading.
Some parametric and non-parametric tests confirmed the pres-
ence of price patterns after overreactions, which identified
that the next-day price changes in both directions are bigger
than after “normal” days. The results also showed that the
overreaction detected in the cryptocurrency market would
not give available profit opportunities (possibly due to trans-
action costs) that cannot be considered as evidence of the
EMH. Chaim et al. [62] analysed the high unconditional
volatility of cryptocurrency from a standard log-normal stochas-
tic volatility model to discontinuous jumps of volatility and
returns. The experiment indicated the importance of in-
corporating permanent jumps to volatility in cryptocurrency
markets.
9.2. Extreme condition
Differently from traditional fiat currencies, cryptocur-
rencies are risky and exhibit heavier tail behaviour. Paraskevi
et al. [151] found extreme dependence between returns and
trading volumes. Evidence of asymmetric return-volume re-
lationship in the cryptocurrency market was also found by
the experiment, as a result of discrepancies in the correlation
between positive and negative return exceedances across all
the cryptocurrencies.
There has been a price